IFAC PROCEEDINGS SERIES

Editor-in-Chief
J A N O S GERTLER, Department of Computer and Electrical Engineering,
George Mason University, Fairfax, Virginia, U S A

GERTLER 8c KEVICZKY (General Editors): A Bridge Between Control Science and T e c h n o l o g y

(Ninth Triennial World Congress, in 6 volumes)
Analysis and Synthesis o f Control Systems (1985, No. 1)
Identification, Adaptive and Stochastic Control (1985, No. 2)
Large-scale Systems, Decision-making, Mathematics o f Control (1985, No. 3)
Process Industries, Power Systems (1985, No. 4)
Manufacturing, Man—Machine Systems, Computers, Components, Traffic Control, Space Applications
(1985, No. 5)
Biomedical Applications, Water Resources, Environment, Energy Systems, Development, Social Effects,
SWIIS, Education (1985, No. 6)
BARKER 8c Y O U N G : Identification and System Parameter Estimation (1985) (1985, No. 7)

N O R R I E 8c T U R N E R : Automation for Mineral Resource D e v e l o p m e n t (1986, No. 1)

C H R E T I E N : Automatic Control in Space (1986, No. 2)
DA C U N H A : Planning and Operation o f Electric Energy Systems (1986, No. 3)
V A L A D A R E S T A V A R E S 8c E V A R I S T O D A SILVA: Systems Analysis Applied to Water and Related Land
Resources (1986, No. 4)
LARSEN 8c H A N S E N : Computer A i d e d Design in Control and Engineering Systems (1986, No. 5)
PAUL: Digital Computer Applications to Process Control (1986, No. 6)
Y A N G J I A C H I : Control Science and T e c h n o l o g y for D e v e l o p m e n t (1986, No. 7)
M A N C I N I , J O H A N N S E N 8c M A R T E N S S O N : Analysis, Design and Evaluation of M a n - M a c h i n e Systems
(1986, No. 8)
B A S A N E Z , F E R R A T E 8c SARIDIS: Robot Control "Syroco '85" (1986, No. 9)
J O H N S O N : Modelling and Control o f Biotechnological Processes (1986, No. 10)

T A L ' : Information Control Problems in Manufacturing T e c h n o l o g y (1987, No. 1)

S I N H A 8c T E L K S N Y S : Stochastic Control (1987, No. 2)
R A U C H : Control o f Distributed Parameter Systems (1987, No. 3)
F L O R I A N 8c H A A S E : Software for Computer Control (1987, No. 4)
M A R T O S , P A U 8c Z I E R M A N N : Dynamic Modelling and Control o f National Economies (1987, No. 5)
GENSER, E T S C H M A I E R , H A S E G A W A 8c S T R O B E L : Control in Transportation Systems (1987, No. 6)
A D A L I & T U N A L L Microcomputer Application in Process Control (1987, No. 7)
W A N G P I N G Y A N G : Power Systems and Power Plant Control (1987, No. 8)
B A L C H E N : Automation and Data Processing in Aquaculture (1987, No. 9)
Y O S H I T A N I : Automation in Mining, Mineral and Metal Processing (1987, No. 10)
GEERING 8c M A N S O U R : Large Scale Systems: T h e o r y and Applications (1987, No. 11)
ROOS: Economics and Artificial Intelligence (1987, No. 12)
T R O C H , KOPACEK 8c BREITENECKER: Simulation o f Control Systems (1987, No. 13)
KAYA 8c WILLIAMS: Instrumentation and Automation in the Paper, Rubber, Plastics and Polymerization
Industries (1987, No. 14)

H R U Z 8c CICEL: Automatic Measurement and Control in Woodworking Industry (1988, No. 1)

A L B E R T O S 8c DE LA P U E N T E : C o m p o n e n t s , Instruments and Techniques for Low Cost Automation and
Applications (1988, No.2)
KOPACEK, T R O C H 8c DESOYER: T h e o r y of Robots (1988, No. 3)
McGREAVY: Dynamics and Control o f Chemical Reactors and Distillation Columns (1988, No. 4)
R A M A M O O R T Y : Automation and Instrumentation for Power Plants (1988, No. 5)
I S E R M A N N : Automatic Control (Tenth Triennial World Congress) (1988, Nos. 6-15)

NOTICE TO READERS
If your library is not already a standing/continuation order customer or subscriber to this series, may we recommend that you place a standing/
continuation or subscription order to receive immediately upon publication all new volumes. Should youfindthat these volumes no longer serve your
needs your order can be cancelled at any time without notice.

Copies of all previously published volumes are available. A fully descriptive catalogue will be gladly sent on request.

ROBERT MAXWELL
Publisher
DYNAMICS AND CONTROL OF
CHEMICAL REACTORS AND
DISTILLATION COLUMNS
Selected Papers from the IFAC Symposium,
Bournemouth, UK, 8—10 December 1986

Edited by

C. M c G R E A V Y
University of Leeds, Yorhhire, UK

Published for the

INTERNATIONAL FEDERATION OF AUTOMATIC CONTROL

by

PERGAMON PRESS

OXFORD · NEW YORK · BEIJING · FRANKFURT

SÂO PAULO · SYDNEY · TOKYO TORONTO
U.K. Pergamon Press, Headington Hill Hall, Oxford O X 3 OBW, England

U.S.A. Pergamon Press, Maxwell House, Fairview Park, Elmsford, New York 10523, U.S.A.

PEOPLE'S REPUBLIC Pergamon Press, Room 4037, Qianmen Hotel, Beijing, People's Republic of China
OF CHINA

F E D E R A L REPUBLIC Pergamon Press, Hammerweg 6, D-6242 Kronberg, Federal Republic of Germany

OF G E R M A N Y

BRAZIL Pergamon Editora, Rua Eça de Queiros, 346, CEP 04011, Paraiso, Sâo Paulo, Brazil

AUSTRALIA Pergamon Press Australia, P.O. Box 544, Potts Point, N.S.W. 2011, Australia

JAPAN Pergamon Press, 8th Floor, Matsuoka Central Building, 1-7-1 Nishishinjuku, Shinjuku-ku,
Tokyo 160, Japan

CANADA Pergamon Press Canada, Suite No. 271, 253 College Street, Toronto, Ontario, Canada M 5 T 1R5

Copyright © 1988 IFAC

All Rights Reserved. No part of this publication may be reproduced, stored in a retrieval system or transmitted in
any form or by any means: electronic, electrostatic, magnetic tape, mechanical, photocopying, recording or other-
wise, without permission in writing from the copyright holders.

First edition 1988

Library o f C o n g r e s s C a t a l o g i n g i n P u b l i c a t i o n Data
Dynamics and control of chemical reactors and
distillation columns.
(IFAC proceedings series; 4)
"IFAC Symposium on Dynamics and Control of Chemical
Reactors and Distillation Columns, sponsored by
International Federation of Automatic Control (IFAC),
Technical Committee on Application; co-sponsored by
European Federation of Chemical Engineering [and]
Institute of Chemical Engineers, U K " — P.
Includes indexes.
1. Chemical reactors — Congresses. 2. Distillation
appartus — Congresses. I. McGreavy, C. II. IFAC
Symposium on Dynamics and Control of Chemical Reactors
and Distillation Columns (1986: Bournemouth, Dorset)
III. International Federation of Automatic Control.
Technical Committee on Applications. IV. European
Federation of Chemical Engineering. V. Institution of
Chemical Engineers (Great Britain) VI. Title.
VII. Series: IFAC proceedings series; no. 4.
TP157.D96 1988 660.2'83 88-4047

B r i t i s h Library C a t a l o g u i n g i n P u b l i c a t i o n Data
Dynamics and control of chemical reactors
and distillation columns: selected papers
from the IFAC symposium, Bournemouth, UK,
8-10 December 1986.
1. Chemical engineering equipment: Reactors
2. Distillation
I. McGreavy, C. II, International
Federation of Automatic Control III. Series
660.2'83
ISBN 0-08-034917-X

These proceedings were reproduced by means of the photo-offset process using the manuscripts supplied by the
authors of the different papers. The manuscripts have been typed using different typewriters and typefaces. The
lay-out, figures and tables of some papers did not agree completely with the standard requirements: consequently
the reproduction does not display complete uniformity. To ensure rapid publication this discrepancy could not be
changed: nor could the English be checked completely. Therefore, the readers are asked to excuse any deficiencies
of this publication which may be due to the above mentioned reasons.

The Editor

Printed in Great Britain by A. Wheaton & Co. Ltd., Exeter

IFAC SYMPOSIUM O N DYNAMICS A N D C O N T R O L OF
CHEMICAL REACTORS A N D DISTILLATION COLUMNS

Sponsored by
International Federation of Automatic Control (IFAC)
Technical Committee on Application

Co-sponsored by
European Federation of Chemical Engineering
Institute of Chemical Engineers, UK

Organized by
Institute of Measurement and Control, UK

International Programme Committee

T. F. Edgar, USA W. H. M. Ray, USA
E. D. Gilles, FRG C. F. H. van Rijn, T h e Netherlands
P. J. de Jong, T h e Netherlands C.J. Ryskamp, U S A
S. B. Jorgensen, Denmark J. H. Seinfeld, USA
M. Kummel, Denmark (Chairman) T. Takamatsu, Japan
J. F. McGregor, Canada K. Walter, Finland
T. McAvoy, U S A C. While, UK
C. McGreavy, UK R. K. Wood, Canada
M. Morari, U S A

National Organizing Committee

J. S. Anderson
L. Kershenbaum
J. Love
C. McGreavy (Chairman)
C. While
P. L. Yue
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, U K 1986

DISTILLATION CONTROL SYSTEM

STRUCTURES
Κ. V. Waller
Process Control Laboratory, Department of Chemical Engineering, Abo Akademi,
20500 ABO, Finland

Abstract. Much of the recent research on distillation control has been treating the
question of control system configuration. Dual-composition control by controlling tem-
perature differences and sums in a column, various ways of reducing interaction by use
of flow ratios as manipulators, and use of decoupling strategies are examples on dis-
cussed structures.
In a common framework the different suggestions can be viewed as control systems of the
same basic multivariable structure. The approaches then mainly differ in their ways of
designing the various parts of the control system. It seems that what has been labeled
as a structural problem could equally well be considered an algorithmic problem within
the general multivariable structure.

Keywords. Distillation control; process control; control system analysis; decoupling;

multivariable control systems; control systems synthesis.

INTRODUCTION teraction effects between the control loops, simi-

It has often been stated that the most important
problem in chemical process control is not the de-
velopment of more sophisticated algorithms, but
rather the establishment of a structural framework
for selecting the manipulated and measured varia-
bles and linking them appropriately (Lau, Alvarez
and Jensen, 1985).
The theoretically optimal structure where all out-
puts are utilized for computation of the control
action, i.e. all outputs are connected to all in-
puts, has not so far gained much acceptance in the
process industries. Instead the control system de-
sign has aimed at single input single output
(SIS0) control loops, and much work has been de-
voted to finding suitable combinations of inputs
and outputs in order to decrease or eliminate the
interaction between the loops.
Generally, this area of variable transformations
is a field of significant potential in process con-
trol (Waller, 1980, 1982) where applications may
be to
ο eliminate or decrease interaction between con-
trol loops
ο eliminate or decrease nonlinearities
ο decrease dimensionality of high-dimensional prob-
lems
Some successful applications of variable transfor-
mations have already been reported. One example
outside the distillation field is the modeling and
control of pH (Gustafsson and Waller, 1983) (Gus-
tafsson, 1984). Other applications are reviewed by
Waller and Màkilâ (1981).
In distillation control various transformations of
variables have been shown much interest during the
last decades. So far the main interest has been
devoted to dual composition control, i.e. control-
ling both the product compositions from a continu-
ously operating two-product distillation column.
Since the aim has been to reduce detrimental in-
lar approaches can be expected to be used in the
future for control of more tightly (heat) inte-
grated trains of columns. Another field, not much
explored as yet, is control of the compositions
from multicomponent distillation columns with sig-
nificant sidestreams.
Structural aspects are emphasized in the follow-
ing review of various approaches that have been
suggested for dual-composition control of distil-
lation.
CHOICE OF CONTROLLED VARIABLES
It has been known for a long time that the inter-
action between the two composition control loops
in a dual-composition control system for distilla-
tion, as illustrated in Fig. 1, may degrade con-
trol quality. As a remedy Rosenbrock (1962) al-
ready in 1962 suggested the control system in Fig.
2, which according to him often can be shown to be
noninteracting. The rationale behind the control
scheme of Fig. 2 is expressed by Rosenbrock in the
following way. "The 'top loop' now manipulates the
flow of top product to control the sum of the two
measured compositions. The 'bottom loop' manipu-
lates the boil-up and (with fixed flow of top prod-
uct) the reflux to control the difference between
the two compositions. Thus the two loops respec-
tively determine the vertical position of the com-
position curve, and its slope, by manipulating two
appropriate variables (cf. Fig. 3)."
A few years later Davison (1967) extended Rosen-
brock's treatment and suggested, for a column
studied, the scheme in Fig. 4 for the case where
pressure is constant. The control scheme suggested
is obtained by Rosenbrock's modal analysis proce-
dure. The criterion is that "the dominant time con-
stant of the controlled plant is minimized".
The multiloop SIS0 (single-input single-output)
scheme in Fig. 1 can be drawn as in Fig. 5. The
interaction between the two loops is caused by the
elements G and G .
01 1 Q

I
2 Κ. V. Waller
The interaction can be eliminated or reduced by in-
serting compensators, much like feedforward con-
trollers, between the two primary feedback loops.
The approach is called (external) decoupling in
the distillation control literature and has been
shown much interest since Luyben's (1970) early pa-
per.
A decoupling scheme is shown in Fig. 6. To decou-
ple the two feedback loops, the decouplers have to
be chosen in such a way that they counteract the
interaction caused by the two process elements G-j2
and G 2 1 · The decouplers in Fig. 6 then become
-G21/G22 and -G12/G1V
The decoupling scheme of Fig. 6 has the structure
of a multivariable control system where each of
the two outputs are connected to each input. In
standard multivariable control theory the system
is treated as a whole when the controllers are de-
signed. This has the implication that interaction
in the system that might be beneficial for disturb-
ance rejection is taken advantage of in the design.
In the decoupling approach the basic idea is to
eliminate all interaction, be it detrimental or
not, and make the system behave as two isolated
SISO-loops. It has been stated that decoupling con-
trol is the opposite to multivariable control.
Rosenbrock's and Davison's suggestions in Figs 2
and 4 can be compared with the decoupling scheme
in Fig. 6. Rosenbrock's scheme can be drawn as in
Fig. 7. Fig. 7 can be redrawn as in Fig. 8 to show
the structural equivalence between Rosenbrock's
scheme and the decoupling scheme of Fig. 6.
Davison's scheme of Fig. 4 can analogously be re-
drawn as in Fig. 9.
A comparison between Rosenbrock's and Davison's
schemes as plotted in Figs 8 and 9 and the decoup-
ling scheme of Fig. 6 shows that the elements in
Rosenbrock's and Davison's schemes that correspond
to the decouplers in Fig. 6 are not directly based
on the properties of the process (only indirectly
through the controllers and their tuning). It
therefore seems that they cannot generally be ex-
pected to decouple the loops and make the system
noninteracting. Actually this was not the basic
aim of Davison's system, and Figs 8 and 9 are here
used only to show the structural similarity be-
tween the schemes and the decoupling scheme of Fig.
6 and to emphasize that all three schemes can be
structurally viewed as general multivariable con-
trol systems where all (in this case two) outputs
are connected to all inputs, as in Fig. 7.
Also Ryskamp (1982) discusses schemes where one
temperature above the feed plate and one below the
feed are used as the two measurements from which
product composition is inferred. Ryskamp suggests
a scheme where heat input is set by the sum of the
temperatures and the difference between them is
used to set reflux.
Structurally Ryskamp's scheme resembles Rosen-
brock's and Davison's schemes. It differs, however,
in two respects. Firstly, Rosenbrock and Davison
use distillate flow D as a manipulator in addition
to boil up V, whereas Ryskamp use reflux flow L in
addition to V. Secondly, the temperature difference
is used to set boil up in Rosenbrock's and Davison's
schemes,1 whereas the sum of temperatures is used in
Ryskamp s scheme.
Recently Bequette and Edgar (1986) used so called
singular value analysis to design a control scheme
for a simulated column. A structurally non-inter-
acting system was obtained when distillate flow D
was paired with the sum of two temperatures, and
reboiler heat duty was paired with the difference
in tray temperatures, i.e. the system had the same
structure as has Rosenbrock's scheme in Fig. 2.
Bequette and Edgar state that the pairing for
their column should not be the same as suggested
by Ryskamp (1982).
Different design methods to obtain non-interacting
control were studied by Bequette and Edgar and
they all resulted in essentially the same scheme,
i.e. one using the sum and the difference between
two temperature measurements as inputs to the con-
trollers. The methods studied were (1) column pro-
file control, (2) implicit decoupling, (3) modal
control, (4) output decoupling, and (5) extensive
variable control (Georgakis, 1986). It should be
noted, however, that the study aimed at a control
system for a column where the manipulators were
chosen to be D and V.
A multivariable structure where both outputs are
connected to both inputs is obtained also by other
combinations of measured variables from the two
ends of the column. In an attempt to get linear-
ization in addition to non-interaction, Weber and
Gaitonde (1982) use the conventional manipulators
reflux flow L and vapor boil up V to control cer-
tain combinations of top and bottom compositions.
The variables controlled are called
fractionation = χ
β+ K(1-x )
D
which is controlled by reflux L, and
χ χ
cutpoint = (1- [))/ β
which is controlled by boil up V.
x
Weber and Gaitonde calculate the constant Κ from
1 X
the normal operating point as B ^ " D O ^ *
0
McAvoy (1983) discusses Weber's and Gai tonde's
scheme and states that a "better", i.e. less in-
teracting, choice for Κ would be Xp/(1-Xp).
McAvoy (1983) suggests a scheme where a variable
ξ defined as
? = +X X
B T^F D
is controlled by distillate flow D.
McAvoy state that use of ξ as a controlled varia-
ble results in an almost one-way steady-state de-
coupled system, i.e. changes in V do not signifi-
cantly affect ξ . The relationship between the con-
trolled and the manipulative variable is further
stated to be linear (McDonald and McAvoy, 1983).
Instead of ξ as a controlled variable, resulting
in an almost one-way steady-state decoupled sys-
tem, a similar variable could be used giving per-
fect one-way steady-state decoupling when the ma-
nipulators are D and V. This variable is (Haggblom
and Waller, 1986)
ξ= x
B+ K'x
D
where the constant K' can be calculated from
steady-state compositions or flows as
X x D
F o - B ox _ o
Β
Do " F o
In McAvoy's scheme (McAvoy, 1983, McDonald and
McAvoy, 1983) the second controlled variable is
separation factor S, which is controlled by boil up
V. The separation factor is defined as
 Distillation Control System Structures
B A list of ratio control schemes suggested in the
χ 1-x
D literature is given in the book by Rademaker,
S = - -
1-x x
D B
This loop is not decoupled from the other one,
changes in D affect the separation factor S.
Separation factor control is also discussed by
Shinskey (1984). It could be added that Boyd (1975)
and Ya and Luyben (1984) also have discussed con-
trol schemes where temperature differences, in
their schemes differences between differences, are
used as controlled variables.
It is interesting to note the similarities and dif-
ferences between the schemes discussed above. For
symmetrical separations, i.e. D = B , 1-xrj = Χβο>
0 Q 0
and x p = 0 . 5 , the controlled variables and mani-
0
pulators become the ones shown in Table 1.
CHOICE OF CONTROL VARIABLES
In the previous section were reviewed various sug-
gestions for transformation of controlled varia-
bles mainly to make the two control loops noninter-
acting for dual-composition control of distilla-
tion.
At the other end of the control system there are
the control variables or manipulators used to con-
trol the process. They, too, can be chosen in var-
ious ways, and this is also a subject much discus-
sed in the recent distillation control literature.
For feedback composition control of a distillation
column with two products there are the following
four primary control variables: distillate flow D,
reflux L, bottoms flow B, and vapor boil up V (the
last one indirectly manipulated through heat input
to the reboiler).
The manipulators in the so called conventional or
energy balance scheme are L and V. They are the
manipulators in Fig. 1, whereas D and V are the
manipulators in Rosenbrock"s scheme in Fig. 2 and
Davison's scheme in Fig. 4. Schemes where either D
or Β are manipulated to control composition are
usually referred to as material balance control
schemes.
McDonald and McAvoy (1983) use the "material bal-
ance variable, D or B", to control a linear combi-
nation of product compositions (see previous sec-
tion). The "energy balance variable", V or L, is
used to control the separation factor.
Between the two base cases of energy balance con-
trol and material balance control there are a num-
ber of combinations. Many suggestions to use vari-
ous ratios between flows as manipulators can be
found in the literature. An early example is given
by Rijnsdorp (1965), who suggests the ratio of re-
flux flow and top vapor flow as a manipulator for
the top loop (L/(L+D) «-x ) (here the arrow point-
D
ing backwards denotes feedback). Stainthorp and
Jackson (1974) experimentally studied a scheme in
which the top loop manipulator was L/D. For a num-
ber of simulated columns McAvoy (1977) studied the
steady-state interaction also for Rijnsdorp's sug-
gestion, both in combination with (V-*-x ) and - as
a combination of Rijnsdorp's scheme and a material
B to each input. However, it seems reasonable that
balance scheme - with ( B - e x J . McAvoy found the
smallest amount of steady-sxate interaction for
the last scheme, (L/(L+D) X Q ) ( B ^ - X ) . He also
b
extended the idea in search of such functional com-
binations of manipulated variables as would make
the loops noninteracting, the final result was de-
generacy (Jafarey and McAvoy, 1978).
3
Rijnsdorp, and Maarleveld (1975). The book discus-
ses various ratio-control schemes, among them also
two-ratio schemes, in which both the manipulators
are flow ratios.
One of the schemes recently most discussed where a
flow ratio is used as manipulator is a scheme sug-
gested by Ryskamp (1980). The manipulators for
composition control in the scheme are D/(L+D) and
V.
The rationale for the scheme is expressed by Rys-
kamp (1982) as follows. The scheme "holds reflux
ratio constant if the top AC output is constant.
An increase of heat input from the bottom AC does
not make top product as impure as would occur with
reflux constant (conventional control) nor as over-
pure as would occur with distillate flow constant
(material balance control)". Thus, this property
of the scheme results in a certain decoupling ef-
fect and the scheme is often said to result in
"implicit decoupling", in contrast to "explicit
decoupling" accomplished by external decoupling
elements, discussed in the previous section and
later in this paper treated more in detail.
1
A modification of Ryskamp s scheme to a scheme
where the manipulators are D/(L+D) and V/B has
been suggested by Takamatsu, I. Hashimoto and Y.
Hashimoto (1982) (1984) and by Shinskey (1984).
1
An implementation of Ryskamp s scheme can be il-
lustrated as done in Fig. 10 for the top of the
column.
An important difference between the manipulators
in the scheme of Fig. 10 and the schemes previ-
ously discussed with D, L, V, and Β as manipula-
tors should be noted. In the schemes where two of
the basic manipulators D, L, V, and Β are used to
control composition, the other two are used for
level control (as illustrated e.g. by Fig. 4 ) .
These level control loops have not been considered
in the discussion above, in order to simplify the
treatment. However, strictly speaking, the control
problem discussed is concerned with control of
four outputs using four manipulators, i.e. a 4x4
problem.
In Ryskamp's (1980) scheme in Fig. 10 both the m a -
nipulators L and D at the top end of the column
are simultaneously used for control of both compo-
sition and level, i.e. each output is connected to
each input, as illustrated in Fig. 11, which shows
the structure of Ryskamp's scheme in block diagram
form.
If, in analogy with the case discussed above, the
bottoms composition x and the reboiler level are
B
controlled by a flow ratio, like V/B, this means
that both these outputs are connected to both the
inputs V and B.
The various schemes using various flow ratios as
manipulators have as a rule been obtained by heu-
ristic reasoning. Obviously they have been found
advantageous in industrial practice. Theoretically
this is not unexpected: they have the structure of
an optimal system where each output is connected
still better control can be anticipated if other
algorithms than pure division between flows were
used.
Furthermore, the four outputs and four inputs have
been split into two groups. Still better control
can be expected if all four outputs are connected
to all four inputs, i.e. if information about what
happens at one end of the column is transmitted
4 Κ. V. Waller
also through the control system to the other end of
the column.
EXTERNAL DECOUPLING
Inserting compensators between interacting loops,
as shown in Fig. 6, is usually called (external)
decoupling in the distillation control literature.
This subject has attracted much interest in the
literature ever since Luyben's paper in 1970
(Luyben, 1970), and it was discussed in connection
with Rosenbrocks's and Davison's schemes (see Figs
6, 8, and 9 ) . Below some complementary aspects are
given.
The decoupling scheme as drawn in Fig. 6 is actual-
ly one of several possible decoupling schemes.
Which one to use is usually determined by realiza-
bility aspects (Waller, 1974).
Two-way decoupling (where there are two decouplers
between the two primary loops, as in Fig. 6) and
one-way decoupling (only one decoupler) are com-
pared by Fagervik, Waller, and Hammarstrbm (1983).
Results obtained in the study also indicate that,
in general, the best disturbance rejection is not
obtained by perfect decoupling, but that the best
response is obtained by some tuning of both decou-
plers and feedback controllers.
We are then left with a parametric optimization
problem of a considerable size. Even if the types
of the feedback controllers and the form of the de-
couplers are decided upon, the problem of simulta-
neous tuning of "decouplers" and feedback control-
lers is still considerable. It is further desira-
ble that various decoupling structures (such as
one-way versus two-way decoupling in a 2x2 system)
can be tried, as well as various forms for the de-
couplers.
Use of the Inverse Nyquist Array (INA) technique
for simultaneous feedback controller and "decoup-
ler" tuning is demonstrated by Waller, Wikman, and
Gustafsson (1985). The paper also shows that mini-
mizing interaction at the critical frequency often,
but not always, is a good criterion for design of
the decouplers. It is also shown that it may be
enough to use pure gains as decouplers; adding dy-
namics need not significantly improve control
quality.
It should further be emphasized that although the
control quality, as measured e.g. by error inte-
grals after step disturbances, is not very differ-
ent for different decoupling schemes, the robust-
ness of the scheme may significantly speak in fa-
vor of one of the schemes, as is quantitatively il-
lustrated by Fagervik, Waller, and Hammarstrbm
(1983).
ASPECTS ON HOLISTIC MULTIVARIABLE CONTROL
In the previous treatment it was found that many
of the schemes for dual-composition control dis-
cussed in the literature and there treated as mul-
tiloop SISO systems can be viewed as ΜΙΜΟ (multi
input multi output) systems.
An alternative would then be to start from the
structure of the system (be it 2x2, 3x3, or 4x4)
and use a true multivariable design method and cal-
culate the controller by some optimization proce-
dure, instead of determining part of the control-
ler by more or less heuristic choices. One such
method is the so-called linear quadratic (LQ) de-
sign. A number of simulation studies of LQ-design
have been reported, reviews can be found in Edgar
and Schwanke (1977) and Waller (1982).
The LQ-design results in a scheme where all the
state variables are used for feedback. The size of
the feedback coefficients may, however, differ con-
siderably. By neglecting small feedback coeffi-
cients, Oakley and Edgar (1976) concluded for a
specific example analyzed that the resulting
scheme could be approximated with good accuracy by
a scheme having the same structure as a one-way
decoupled scheme, where changes in reboiler heat
duty are fed to the top loop, not only through the
column but also through the decoupler. The same
conclusion is obtained in a study by Tung and
Edgar (1978).
In chemical process control, the number of sensors
is usually much lower than the number of state
variables used to describe the process. In that
case an optimal multivariable (LQ) controller con-
tains an observer or state estimator, by which the
unmeasured states are calculated before they are
used for feedback.
An LQ-design can very well start from a system
model consisting of simple transfer functions con-
taining dead times. An illustration of the various
steps of the design starting from experimentally
obtained transfer functions is given in Hammar-
strbm, Waller, and Fagervik (1982). The paper is
focused on how errors in the process models affect
the control properties of the multivariable con-
trol system obtained. Connections between model
structure, the performance index, the control
quality, and the sensitivity are illustrated.
One further aspect on state estimation may be men-
tioned here, since it has to do with the structure
of the state estimator. When there are disturb-
ances with non-zero mean, such as step disturb-
ances, estimation e.g. by a Kalman-fliter results
in steady-state estimation offset if a model of
the deterministic disturbances is not included in
the estimator. Also the resulting control quality
may be drastically reduced, compared to the case
when a model of the disturbances is included in
the estimator. In an distillation example treated
by Hammarstrbm and Waller (1974), the control was
improved by several orders of magnitude simply by
including estimation of an occuring step disturb-
ance in feed composition. It should be noted that
this improvement was accomplished not by any added
sensors or disturbance measurements, but by
changing the structure of the estimator so that
the control system was informed about the fact
that there might be disturbances of step type.
Estimation of more deterministic disturbances
(such as steps both in feed composition and flow)
than actually occured (in feed composition) did
not degrade control quality. However, not more
disturbances can be estimated than there are meas-
urements on the process (Hammarstrbm, 1980).
Further illustration of LQ-control in distillation
can be found in Hammarstrbm (1980).
NUMBER AND LOCATION OF SENSORS
The number and location of sensors are important
issues directly related to the control system
structures.
Waller, Gustafsson, and Hammarstrbm (1974) illus-
trated, through simulation, how drastic an effect
the location of a third sensor (two fixed in the
product lines) may have on distillation control
quality. Also when state estimation (Kalman-fli-
ter) was used (Hammarstrbm and Waller, 1974) the
location of the third sensor strongly affected
control quality in LQ-design. This could easily be
understood if the measurements were inaccurate,
but in the study mentioned perfect measurements
were assumed.
 Distillation Control System Structures
A frequent approach in the literature for sensor
location is to use some kind of observability or
sensitivity consideration. In recent literature
so-called singular value analysis has been popular.
Already in 1974, Hammarstrbm and Waller (1974)
studied the relation between control quality and
observability properties of a distillation system.
No correlation was found between the observability
index used and control quality. Observability con-
siderations are probably not enough to solve the
sensor location problem. This seems also to be the
opinion of Mellefont and Sargent (1978) who sug-
gest an implicit enumeration algorithm for the se-
lection of measurements to be used in optimal feed-
back control of linear stochastic systems.Applica-
tion of the algorithm to a distillation system
gave results which were reported to be "consider-
ably different from what would be expected for
pure estimation" (Mellefont and Sargent, 1978).
The location and number of sensors are also direct-
ly related to the optimal size and complexity of
the process model used for controller design.
Dahlqvist (1980) investigated, both experimentally
and through simulation on an 11-plate pilot column,
LQ-controllers based on various model sizes. The
measured states were obtained thorugh state esti-
mation. Using two sensors, a second order model
gave better results than did a sixth order model.
With three sensors, a sixth order model gave
better control than a 13th order model did, and
also better control than a fourth order model gave.
The relation between sensor number and location
and process model size and complexity in multivar-
iable control is likely to be a function of the
quality of the measurements, such as accuracy and
dynamics. These relations are also functions of
the properties (both static and dynamic) of the
process to be controlled, as well as of the prod-
uct specifications.
Another closely related problem, not much studied
so far, is how to combine slow and direct (desir-
able) measurements (like those of a gas chromato-
graph) with fast and indirect ones (like tempera-
tures).
DISCUSSION
Dual-composition control of distillation has been
treated emphasizing the structure of various con-
trol schemes. Many suggested SISO schemes have
been shown to be structurally equal to a general
multivariable scheme. For the problem studied, it
then seems that the structural question could
equally well be viewed as an algorithmic question.
So far only feedback control strategies have been
treated. For distillation feedforward control is
often used to cope with changes in feed flow rate.
If feed composition can be measured, feedforward
can be used also to compensate disturbances in
feed composition. In this way the two feedback
loops normally used in dual composition control
could be replaced by one feedforward loop and one
feedback loop, as suggested by Jafarey and McAvoy
(1980).
An issue that seems to be, at least partly, open
when controlling various (linear) combinations of
concentrations (or temperatures) instead of the
product compositions themselves is the following.
Does good control of a combination of concentra-
tions (or temperatures) also implay good control
of the product concentrations which usually are
the important variables to control? What about ro-
bustness? How should e.g. integral action in the
control system be implemented to bring the product
5
compositions to desired values? This last question
was recently discussed by Georgakis (1986).
Among interesting problems for future research in
distillation control can be mentioned the question
of sensor number and location (observability con-
siderations do not give the whole picture) and
their relations to the complexity of the process
models used for design. Here the combination of
various types of sensors is an interesting and
important subproblem.
NOTATION
Β bottoms flow
D distillate flow
F feed flow
L reflux flow
0 as index denotes steady state
Τ temperature
V boil up
X composition
REFERENCES
Bequette, B.W., and T.F. Edgar (1986). The equiv-
alence of non-interacting control system de-
sign methods in distillation. Proc. ACC,
Seattle, WA, USA, 31-38.
Boyd, D.M. (1975). Fractionation column control.
Chem. Eng. Progr. 71 No 6, 55-60.
Dahlqvist, S J \ . (1980). Control of a distillation
column - application of sampled data control
to a large pilot plant. Tech. Dr thesis,
Royal Institute of Technology, Stockholm,
Sweden.
Davison, E.J. (1967). Control of a distillation
column with pressure variation. Trans. Instn
Chem. Engrs 45, T229-T250.
Edgar, T.F., and C O . Schwanke (1977). A review
of the application of modern control theory
to distillation columns. Proc. JACC,
1370-1376.
Fagervik, K.C., K.V. Waller, and L.G. Hammarstrbm
(1983). Two-way or one-way decoupling in dis-
tillation? Chem.Eng.Comm. 21, 235-249.
Georgakis, C. (1986). On the use of extensive
variables in process dynamics and control.
Chem.Eng.Sci. 41, 1471-1484.
Gustafsson, T.K. (1984). A study of modeling and
control of pH in fast acid-base reaction proc-
esses based on chemical reaction invariance.
Tech. Dr thesis, Abo Akademi, Finland.
Gustafsson, T.K., and K.V. Waller (1983). Dynamic
modeling and reaction invariant control of pH.
Chem.Eng.Sci. 38, 389-398.
Haggblom, K.E., and K.V. Waller (1986). In prepa-
ration. Abo Akademi, Finland.
Hammarstrbm, L.G. (1980). An investigation of
multivariable control strategies for high-
order processes with few sensors. Report 80-7,
Process Control Lab., Abo Akademi, Finland.
Hammarstrbm, L.G., and K.V. Waller (1974). State
estimation in multivariable ΡΙ-control. Report
74-10, Process Control Lab., Abo Akademi,
Finland.
Hammarstrbm, L.G., K.V. Waller, and K.C. Fagervik
(1982). On modeling accuracy for multivariable
distillation control. Chem.Eng.Comm. 19,
77-90.
Jafarey, Α., and T.J. McAvoy (1978). Degeneracy of
decoupling in distillation columns. I&EC Proc.
Pes.Dev. 17, 485-490.
Jafarey, Α., and T.J. McAvoy (1980). Steady state
feedforward control algorithms for reducing
energy costs in distillation. ISA Trans. 19,
No. 4, 89-100.
6 Κ. V. Waller
Lau, H., J . Alvarez, and K.F. Jensen (1985). Syn-
thesis of control structures by singular value
analysis: dynamic measures of sensitivity and
interaction. AIChE Journal 31, 427-439.
Luyben, W.L. (1970). Distillation decoupling. AIChE
Journal 16, 198-203.
McAvoy, T.J. (1977). Interacting control systems:
steady state treatment of dual composition con-
trol in distillation columns. ISA trans. 16,
No. 4, 83-90.
McAvoy, T.J. (1983). Interacti on an alys i s. ISA
McDonald, K.A., and T.J. McAvoy (1983). Decoupling
dual composition controllers. 1. Steady state
results. Proc. ACC, San Francisco, USA,
176-184.
Mellefont, D.J., and R.W.H. Sargent (1978). Selec-
tion of measurements for optimal feedback con-
trol . I&EC Proc.Pes.Dev. 17, 549-552.
Oakley. D.R., and T.F. Edgar (1976). Optimal feed-
back control of a binary distillation column.
Proc. JACC, 551-557.
Rademaker, 0., J.E. Rijnsdorp, and A. Maarleveld
(1975). Dynamics and control of continuous dis-
tillation units. Elsevier, Amsterdam_0xford-
New York.
Rijnsdorp, J.E. (1965). Interaction in two-variab-
le control systems for distillation columns.
Automatica 3, 15-51.
Rosenbrock, H.H. (1962). The control of distilla-
tion columns. Trans.Instn.Chem.Engrs 40,
35-53.
Ryskamp, C.J. (1980). New Strategy improves dual
composition column control. Hydrocarbon proc-
essing 59, No. 6, 51-59.
Ryskamp, C.J. (1982). Explicit versus implicit de-
coupling in distillation control. In Seborg
D.E. and T.F. Edgar (Eds), Chemical process
control 2, Engineering Foundation/AIChE 361-
375.
Shinskey, F.G. (1984). Distillation control. 2nd
ed, McGraw-Hill.
Stainthorp, F.P., and C.B. Jackson (1974). Control
of a fractionating column subject to product
rate demand changes. Proc. Ill IFAC symp. on
multiv. techn.syst., Paper S 39.
Takamatsu, Τ,. I. Hashimoto, and Y. Hashimoto
(1982). Multivariable control system design of
distillation columns system. Proc. PSE 82,
Kyoto, Japan, Tech.Session 243-252.
Takamatsu, T., I. Hashimoto, and Y. Hashimoto
(1984). Dynamic decoupler sensitivity analysis
and its application in distillation control.
Preprints 9th IFAC World Congress, Budapest,
Hungary, Vol. Ill, 98-103.
Tung, L.S., and T.F. Edgar (1978). Analysis of
control-output interaction in dynamic systems.
71th annual meeting AIChE, Miami.
Waller, K.V. (1974). Decoupling in distillation.
AIChE Journal 20, 592-594.
Waller. K.V. (1980). Current problems in process
control. Paper and Timber 62, 128-132.
Waller, K.V. (1982). University research on dual
composition control of distillation: a review.
In Seborg, D.E. and T.F. Edgar (Eds) Chemical
process control 2, Engineering Foundation/
AIChE 395-412.
Waller, K.V., and P.M. Makilâ (1981). Chemical
reaction invariants and variants and their use
in reactor modeling, simulation, and control.
I&EC Proc. Pes.Dev. 20, 1-11.
Waller, K.V., S.E. Gustafsson, and L.G. Hammar-
strbm (1974). Incomplete state-feedback inte-
gral control of distillation columns. Proc.
Kern. Tek 3, Copenhagen, Denmark, Paper C.8.
Waller, K.V., K.E. Wikman, and S.E. Gustafsson
(1985). Decoupler design and control system
tuning by INA for distillation composition con-
trol. Chem.Eng.Comm. 23, 149-174.
Weber, R., and N.Y. Gaitonde (1982). Non-inter-
active distillation tower control. Proc. ACC,
Arlington, VA, 87-90.
Ya, C.C., and W.L. Luyben (1984). Use of multiple
temperatures for the control of multicomponent
distillation columns. I&EC Proc.Pes.Dev. 23,
590-597.
 Distillation Control System Structures 7

TABLE 1 Controlled variables and manipulators in

discussed schemes for symmetric separations

Controlled variable(s) Manipulator(s)

X +x
Rosenbrock (1962) D
1 2
V
T + T
Ryskamp (1982) V
1 2
L
X X
Weber and Gaitonde (1982) L
D " +B
McAvoy (1983) x x (= ξ ) D
T D +B T
Bequette and Edgar (1986) D
T1 T2
V
1 " 2
x
Haggblom and Waller (1986) D+ χ
Β ( = ξ ' )

D x,

En I

P-
I
I
I
I

ne Β χ.

Fig. 1. Dual composition Fig. 2. An alternative scheme for

control of distillation dual composition control
column. (Rosenbrock, 1962).
8 Κ. V. Waller

Fig. 2. An alternative scheme for

Fig. 2. An alternative scheme for
Fig. 2. An alternativ
Fig. 2. An alterna
Fig. 2. An alt

'ci '11

Fig. 5. Scheme of Fig. 1 in block

diagram form.

'12

?
-K> c2 >22

Process
 Distillation Control System Structures 9

'ci '11

- G 21 -G '12
3
22 G 11 Fig. 6. Decoupling scheme for
3 2x2 system.
12

Control system Process

1
22
4

-6
Fig. 7. Rosenbrock's scheme.

'12

Control system
4 3

Process
22

1
21

Gi Fig. 8. Redrawn version of Fig. 7,

Rosenbrock's scheme.
'12

Control system Process

3
22 -6
10 Κ. V. Waller

-0.6K . Fig. 9. Davison's scheme.

'12

0.6K ,

Control system
7

Process
22
4

D / (L+D)
AC)

Fig. 10. Implementation of control

scheme suggested by Ryskamp
(1980).

^D. set
Controller 1 χ

ω
ω
u
ο

'OA
level, set i l c level
Controller 2 CL

Fig. 11. Redrawn version of scheme

in Fig. 10.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, U K 1986

FIXED BED REACTOR DYNAMICS AND

CONTROL - A REVIEW
S. B. j0rgensen
Instituttet for kemiteknik, Technical University of Denmark, DK 2800,
Lyngby, Denmark

ABSTRACT : The industrial diversity of fixed bed reactors offers a challenging

and relevant set of control problems. These intricate problems arise due to the
rather complex dynamics of fixed bed reactors and to the complexity of actual
reactor configurations. Many of these control problems are nonlinear and multi-
variable. During the last decade fixed bed reactor control strategies have been
proposed and investigated experimentally. This paper reviews research on these
complex control problems with an emphasis upon solutions which have been demon-
strated to work in the laboratory and hold promise for industrial application.

Keywords : Packed bed reactor, Chemical reactor, Dynamics, Model Representation,

Optimal control, Adaptive control, Optimizing control.

1. INTRODUCTION

Reactor Control in general has been reviewed in a
number of cases over the last two decades, e.g.:
Denn (1969), Padmanaban and Lapidus (1977), and
McGreavy (1983). Recently Schnelle and Richards
(1986) presented an interesting account of reactor
control as seen from an industrial point of view.
The specific area of the present review was co-
vered by Ray (1972), therefore very little litté-
rature prior to that date is included in the
present review. The amount of work produced since
the early review of Ray (1972) is quite large,
consequently the scope in the present review is
limited to cover mainly experimental investiga-
tions of fixed-bed reactor dynamics and control.
Even with this constraint important contributions
may have been missed unintensionally. In particu-
lar the sovjet littérature is not given adequate
coverage.
Packed beds of catalyst particles is the most
widely used reactor type for gas phase reactants
in the chemical industry. Inherent in the design
of fixed-bed reactors rest some of the more diffi-
cult control problems found in industry. These
problems are due to the process being distributed,
nonlinear, having nonminimum phase characteristics
and deadtime. In addition other complicating phe-
nomena such as extreme parametric sensitivity of
the steady state profiles may occur in some reac-
tors .
The incentive for using on-line control is ob-
vious on a fixed-bed reactor with slowly varying
catalyst activity. Here constant conversion can be
achieved by either increasing inlet temperature,
reducing feed flow rate or increasing recycle
flow in all cases until the design limit is
reached and provided the reactor remains stable.
Control of fixed-bed reactors is also often re-
quired because of sensitivity towards disturban-
ces, or more seldom because of lack of stability
of the reactor. For the latter cases careful reac-
tor design is very important in order to reduce or
minimize the complexity of the final control pro-
blem. In order to distinguish between different
control problems it may be helpful to partition
the disturbances acting on any process. These may
be categorized either as stationary disturbances,
with poles strictly in the left half plane, or as
instationary disturbances, with poles on the ima-
ginary axis. In practice the poles of the models
of disturbances which are several times slower
than the dominant reactor timeconstants may be
assumed to be zero. The advantage of this parti-
tioning is that stationary disturbances are rela-
tively fast varying with a predicted value which
is essentially zero after a relatively short time.
Therefore their influence is suppressed using
regulatory control only. The instationary distur-
bances contain persistent and/or periodic compo-
nents, which most often are present in fixed-bed
reactors, and create offset and/or periodically
varying states and outputs. In order to handle the
persistent disturbances it is necessary to have
some type of offset elimination in the regulatory
control. This action may lead to unprofitable
operating conditions, therefore it can be economi-
cally feasible to include setpoint evaluation
either off-line or on-line to maximize a profit
function. The varying setpoints are implemented
using a setpoint tracking control design.
It follows from the preceding paragraph that the
basic fixed-bed reactor control objective is that
of operability, i.e. to keep the plant in an
operable stable region, in spite of the influence
of disturbances. The origin of this basic objec-
tive may be operational requirements, safety con-
siderations, environnemental regulations etc. The
basic objective may be economically most feasible
in some situations even though the trend in indu-
stry seems to be to automate further. Incentives
for additional control objectives are derived from
economical considerations and lead to modification
of the operating conditions in order to operate in
a more or the most profitable region. Examples of
additional control objectives are to achieve the
desired production rate, quality or yield. Another
additional control objective is to maximize a
profit function which in some situations can be
economical. Schnelle and Richards (1986) point out
that there can be a significant incentive to in-
crease productivity of sold out products and to
improve the quality in a competitive market. For
bulk products there is also a significant incen-
tive to improve raw material yield.
The above control objectives may be fulfilled
using techniques from various levels in the actual
control hierachy. These levels are for the present

11
12 S. B. J0rgensen
purpose defined as (MacGre gor et al. 1983):
Manual: All but the most fundamental regulatory
loops are in manual control. The operator sets
actuators and flow setpoints. This level may
have an advisory computor system.
Single Loop: The basic measurements are under
single loop control. Such loops may require
advancxed techniques as adaptive or nonlinear
control.
Multi Loop: Coordinated yet simple interaction
occur between two or more loops, such as cascade,
feedforward, interaction override or compensation.
Supervisory: Control coordination occur between
data from several sources. This coordination in-
cludes handling filtered data, e.g. from process
analyzers and executing multivariable control,
e.g. model based.
Optimizing: Plant steady state or dynamic behavior
is evaluated to optimize a profit criterion.
Solution of the basic control objective can on
some fixed-bed reactors be possible using a single
loop, which however may be non-minimum phase, and
therefore model based control is advantageous.
However often additional requirements are imposed
such as keeping the reaction zone position and/or
the maximun temperature. In these cases it may be
necessary or advantageous to use more than one
input variable. Therefore fixed-bed reactor con-
trol problems tend to be multivariable, and to be
approached using modelbased control designs. Con-
sequently most of the control designs to be dis-
cussed subsequently have been carried out at the
supervisory level.
The present paper aims at reviewing the control
design methods which have been applied on fixed-
bed reactors to fulfill one or more of the control
objectives, at pinpointing limitations of the
techniques; and thereby indicating possible direc-
tions for future research.
A procedure for fulfillment of a particular set oj
control objectives using a model based approach
may be achieved through the following steps:
i) Investigation of process statics and dynamics
ii) Selection of model representation.
iii) Selection of identification and control
strategy.
iv) Selection of number and location of sensors.
v) Process identification either off- or on-line
vi) Control design for stabilization, disturbance
rejection and set-point tracking.
vii) Optimizing control design.
The following sections contain a discussion of
most of these steps as they have been dealt with
in the available investigations. First some typi-
cal fixed-bed reactor configurations will be clas-
sified in section 2. Section 3 starts with discus-
sing fixed-bed rector models and summarizing ob-
served static behaviors. Then follows a presenta-
tion of the basis for the following treatment
which is the underlying process dynamics with
special focus upon experimental investigations and
demonstrated abilities of the applied mathematical
models. Among these abilities has the locally
linear behavior been extensively exploited. Sec-
tion 4 covers process identification and starts
with a presentation of locally linear model repre-
sentations, then model order reduction procedures
are touched upon since these techniques may be
important in many cases. Thereafter techniques
applied for parameter estimation,which may use
either off- or on-line data acquisition, are re-
viewed briefly. Section 5 contains a review of
application of control designs for fulfilling the
primary and some additional control objectives
i.e. designs for stabilization, disturbance rejec-
tion and set-point tracking. This section pre-
sents first linear quadratic control design and
then reviews selection of the number and location
of sensors. The largest part of section 5 contains
a review of the applied control designs mainly
grouped according to the applied model representa-
tion. Section 6 contains a brief review of optimi-
zing control. Finally the main conclusions are
summarized.
2. FIXED-BED REACTOR CONFIGURATIONS
Catalytic gas-phase chemical reactions play an
important role in chemical industry, and has also
found important applications in pollution abate-
ment of exhaust gases. Contacting of the reactant
gas with the catalyst is mainly effected in reac-
tors where the catalyst is fixed and to a much
smaller extent in fluid beds. The various types of
beds with fixed catalyst employed in present day
reactors cover a wide spectrum. A large fraction
of these reactors may be described by one or a
combination of some generic types:
a) A single tubular bed with axial reactant flow.
b) A single annular bed with radial reactant flow.
c) A channel with wall deposited catalyst.
The type a bed is the traditional catalytic bed.
Type b beds have developed to make it possible to
use large amounts of catalyst with small diameter
particles without getting excessive pressure
drops. Type c, which clearly is not a packed bed,
have been developed for handling exhaust or emis-
sion gases, with still lower pressure drops. The
channel geometry may vary as well as the support
construction material. Type c is used with many
parallel channels, whereas type a and b beds may
be applied as single beds. These single beds may
approach adiabaticity if well insulated and/or of
large diameter (type a) or length (type b ) . How-
ever in many cases external cooling or heating is
desirable or necessary to reduce parametric sensi-
tivity. In these cases two types of configurations
may be used or possibly combined:
d) Multiple parallel beds (type a) which are ex-
ternally cooled (heated) in co- or countercur-
rent flow.
e) Multiple consecutive beds (type a or b) with
interstage cooling (heating), e.g. using a
quench stream.
The feed stream in exothermic processes is often
used for cooling purposes through heat exchange in
type d and between type e beds or as a quench
stream between type e beds. In these cases auto-
thermal operation may be possible. In some proces-
ses (partial) recycle of effluent gas is used,
e.g. to obtain full conversion of an expensive
reactant.
The beds whose dynamics and control have been
investigated in the laboratory have mainly been of
type a and d; and in a few cases in configura-
tions with autothermal opeartion or with reactant
recycle. Some of the experimental investigations
are summarized in tables 2 and 3, where the inve-
stigations are grouped according to the bed type
or configuration. The term cooling is used in the
tables mainly to indicate that the cooling tempe-
rature possibly could be manipulated. It is clear
from these tables that many of the reactions are
only model reactions, which are used to elucidate
the dynamic behavior of fixed-bed reactors and to
 Fixed Bed Reactor Dynamics and Control 13

investigate the possibilities for controlling

fixed-bed reactors. Clearly many factors influence
the general validity of the obtained results in
laboratory reactors. The influence of the thermal
wall capacity in single beds can be significant if
no special precautions are taken such as using a
large diameter or a thin wall. Careful considera-
tion should also be given to the placement of
sensors in order to avoid distortion of the reac-
C — 1 : >£ p
rmiw.i
tant flow or the bed geometry. The experimental t
studies in general have been carried out at much
lower Reynolds numbers than those of industrial n qot i: r1 Dcy ngaum yair jt _
t int tne rdnimca tg(iD
roaY
) QrfUoIacDodo l e
reactors. t yep af i xde dr-eb raecwht o
iet h i r r eev ecrosnn i vbel r i i i o
A>- p r o d. u c t ( s )
3. PACKED BED REACTOR MODELS AND BEHAVIOR
conversion of reactant A into products. The axial
Modelling of packed beds have been covered in a extension of the type a bed is indicated by the
number of surveys a.o. Froment (1974), Schmitz horizontal length and the relative quantity capa-
(1976) , Hofmann (1979), Hlavacek and Votruba cities by the width of the symbols. The quantity
(1977) and Hlavacek and van Rompay (1981). There- fluxes are shown with wide arrows and the interac-
fore a complete review will not be attempted in tions are indicated by thin arrows. The dominating
this section, only some specific points relevant thermal capacitance over the mass capacitances is
for fixed-bed reactor dynamics and control will be immediately obvious. A set of modelling equations
touched upon. are given below in nondimensional form, using the
continuous axial dispersion model:
TABLE 1: Fixe
d Bed Reacto
r Model
s
2
# spatia
l dim.
A. Pseudo
-
one phase
Β. Two phas
e
8C 1 d C
AC
F R(C,T) (1)
one 1. Plu
g Flo
w
1. Externa
l resistanc
e onl
y at Pe dz dz
one 2. Axia
l dispersio
n
2. Externa
l and interna
l res
two 3. Axia
l and radia
l Externa
l heat and 2
dispersio
n interna
l mas
s resist
* 3T I A Τ
AT
(1+H) — β R(C,T) + M(T -T) (2)
at Pe, a z az
For the generic type a fixed bed a number of η
mathematical models have been proposed. A spectrum
of models is shown in Table 1 (Froment 1974). The R(C,T) - Dm exp( (l-l/T))C (3)
fluid phases may be modelled using either a cell 7
or a continuum model. The models and their assump- with the Danckwerts (1953) boundary conditions,
tions will not be discussed in detail here, except which have been used most often in the transient
to state that both model types appear to be rea- case (Hlavacek and van Rompay 1981):
sonable for design purposes, and in most cases
also for modelling the dynamic behavior. When it ac(t ,O) ac(t,im
)
comes to the finer details then the standard dis- F Pe C ( t , l ) - 0 ; 0 (4,5)
persion model seems to be deficient for modelling az az
flow of interstitial fluid (Sundaresan et al.
(1980)). A few comments concerning determination AT(T,O) 3T(t,l)
of model parameters will be given. In the two - Pe,T(t,l) - 0 ; - 0 (6,7)
phase models quantity (i.e. mass and energy) ba- η
az az
lances are formulated for both the gas and the
fluid phases. In the pseudo-one phase models it is
assumed that interphase potential differences are
sufficiently small to make it possible to account In the case of significant wall heat capacity it
for the presence of the two phases by applying is necessary to include an additional energy ba-
suitably modified effective parameters in a pseu- lance for the wall and similarly for the coolant
do-one phase model. Determination of such effec- in the case of a nonadiabatic reactor. In gas
tive transport parameters have been the subject of phase reactors is the ratio between the solid and
considerable research especially in the case of the fluid phase heat capacities Η » 1. This leads
heat transport. Recently Dixon and Cresswell often to the assumption of quasistationarity of
(1986) presented a theory for prediction of effec- the mass balance eq. (1). Usually the axial Peclet
tive heat transport parameters in transient model- number in the energy balance is smaller than that
ling. They shown that effective axial Peclet num- in the mass balance. The above model equations may
bers in the pseudo-one phase dispersion model has be solved using one of several types of techniqu-
a finite value even at infinite Reynolds number. es: Crank-Nicolsen (Eigenberger and Butt 1974),
This shows that it is essential to account for orthogonal collocation (Michelsen et al. 1973),
effective axial dispersion when using the pseudo- orthogonal collocation on finite elements (Garey
one phase model for transient phenomena. Thus and Finlayson 1975) or transformation to integral
model A-l ,which is widely used for the steady equation (Lubeck 1974). However for developing
state design of fixed-bed reactors, is not advi- continuum based dynamic model representations
seable for simulating transients. Dixon and Cres- orthogonal collocation has been the dominant me-
swell (1986) also show that the ratio of fluid thod.
phase and solid phase heat capacities may be used
as an effective parameter in order to make the The steady state and dynamic behavior of fixed bed
pseudo-one phase model both reproduce the steady reactors can be very complex even when exotic
state and reasonably fit the transient response of reaction rate expressions are excluded. It will
the two phase model. not be attempted to review the very extensive
littérature covering the entire subject area.
However a few points relevant for control of fix-
Much of the control work has been carried out ed-bed reactors will be made by first considering
using the pseudo-one phase axial (and radial) multiplicity, sensitivity and stability, and then
dispersion model. A pictorial representation of considering fixed-bed reactor dynamics. Understan-
the dynamic quantity interactions present in such ding of the latter subject is clearly essential
a model is given in Figure 1 for the irreversible for model based control design.
14 S. B. J0rgensen
Multiplicity, sensitivity and stability: Fixed-bed reactor dynamics

The occurrence of more than one steady state solu-
tion for the same operating conditions and nonpa-
thological rate expressions, can be demonstrated
theoretically for all models, except A-l, in Table
I, . This possibility is due to interaction of
axial dispersion and reaction. This subject has
been treated in the above reviews in particular by
Hlavacheck and van Rompay (1981) and Jensen and
Ray (1982). They both review experimental findings
of multiple stady states. In the cases where mul-
tiple steady states may occur, it is possible to
have high sensitivity of the reactor conditions
towards changes in inlet conditions or in parame-
ter values. Cases of high parametric sensitivity
occur near bifurcation of steady states i.e. where
multiplicity of steady states shows up. Puszynski
et al.(1981) show using an A-3 model that mul-
tiple steady states may occur in nonadiabatic
packed beds for strongly exothermic reactions even
when the Peclet number is large, as in most indu-
strial reactors. For the pseudo-one phase axial
dispersion model with a first order irreversible
reaction the sufficient conditions for uniqueness
show that the solution will be unique for suffi-
ciently high values of the unequal Peclet numbers,
large heat transfer coefficient or small values of
the Damkohler number. It has been shown computa-
tionally that the region of multiplicity is enlar-
ged and shifted towards lower Damkohler numbers by
increasing adiabatic temperature rise or the acti-
vation energy. An increase in the reaction order
reduces the region of multiplicity. In the adiaba-
tic reactor a multiplicity of three is possible,
whereas more steady states can be found in the
nonadiabatic fixed-bed reactor. Jensen and
Ray(1982) investigate theoretically the bifurca-
tion behavior especially for the pseudo-one phase
axial dispersion model. Their results show a
wealth of bifurcations at low Peclet numbers.
However at Peclet numbers relevant for industrial
packed beds it is concluded that bifurcations to
periodic solutions should not occur.
Experimental investigations in fixed bed reactor
dynamics are summarized in Table 2. There are two
sets of investigations; those concerned with the
dynamics around steady states (entries 2,3,6-8,10,
and 13-16) and those concerned with the dynamic
behavior when transition between steady states
occur (entries 1,4,5,9, and 12). The latter set of
investigators have focussed on CO and ehtane oxi-
dation as model reactions. The main results of the
latter group of investigators are that transition
between multiple steady states occur as propaga-
ting fronts, where the front shape changes very
little and the speed of propagation is very low
compared to the thermal wave speed. This behavior
can in general be simulated reasonably using A-2
models. In some instances it appears nescessary to
account for radial velocity profiles (Vortmeyer
and Winther 1982). Puszynsky et al. 1982 observed
oscillatory behavior, which could not be simulated
using an A-2 model.
The main results of the first group of investiga-
tions concerning the local behavior around a stea-
dy state are that the observed local phenomena can
in general be modelled quite well using pseu-
do-one phase models with dispersion or two phase
models. An exception is the study by van Doesburg
and de Jong (1976), where however the wall heat
capacity had a major influence. The basic dynamic
features of a type a bed can bee qualitatively
outlined using the DYQUID in figure 1. The signi-
ficant difference between the mass and thermal
residence times has a decisive influence upon the
dynamic behavior. Inverse response, or 'wrong
way' behavior (Crider and Foss 1966), have been
observed upon inlet temperature and flow rate
disturbances. The presence of a concentration node
point (Crider and Foss 1966, 1968) has also been
observed in several cases. Locally linear models
give a reasonable description of the local dyna-
mics and are therefore obvious candidates for use
in model based control designs.

TABLE 2: Experimental inveatigationa of ga. phaae fbed-bed reactor dynamics,

Bed
Multiple steady states can occur in autothermal type: Reaction:
Phenomenon or di sturhnnce
type: inlet: mocho!:
reactors, where energy is recycled and thus the
Single bed
effect of backmixing is significantly increased. 1. CO Padberg and Wicke (1967) propagt. fronts
For reciculation of mass Root and Schmitz oxidat. Fleguth and Wicke (1971) propagt. fronts
Wicke and Onken (1986)
(1969,1970) found multiple steady states in a 2. Ethanol
noise origin

tubular reactor. In a fixed-bed reactor with reac- Dehydrat. Lehr, Yurchak and Kabel (1968) step B-1
tant recycle bifurcation to periodic solutions is 3. H
2
possible Jergensen et al.(1986). Most of the work Oxidat. Hoiberg, Lyche and Foss (1971) sine T,C B-1,2
referenced above has been concerned with exother- 4. Ethane Vortmeyer and Jahnel (1972) propagt. fronts
mic reactions. Gilles et al.(1978) showed by using oxldat. Simon and Vortmeyer (1978) propagt. fronts
a B-l model for selfinhibiting reaction kinetics 5. CO Hlavacek and Votruba (1974) propagt . fronts
that it is possible to obtain bifurcation to peri- oxldat.,

odic solutions for endothermic reactions. Catalyst 6. H

2 Hansen and Jergensen(1974, 76a,b) pulse T,C A-2&3
Oildat. Jergensen and Hansen (1976) multlfreq.
deactivation may give rise to thermal effects in T,C A-2
Serensen (1976) start-up A-3
excess of the adiabatic temperature rise (Hlavacek shut-down
7. ~etha­
and van Rompay (1981)). Although this subject is natIon vanDoesburg and de Jong(1976a,b) step T,C A-I
very important in catalyst design, the transient
8. CO Sharma and Hughes (l979a,b)
effects of catalyst deactivation has been the Oxidat.
step, pulse T,C B-2

subject of relatively few experimental studies.

COOled single bed:
Christiansen and Andersen (1980) simulate the 9. CO Hlavachek and Votruba (1974) propgt. frts
transient sulpher profiles in sulpher poisoning of oxldat. Puszynski and Hlavachek (1980) steps,p.f.
Hlavachek, Puszynski and propgt. frts A-2
steam reformers, adn find qualitatively reasonable van Rompay (1982)
agreement with scarce data and with industrial Puszynskl and Hlavachek (1984) propgt . frts
experience. 10. Butane Jutan, Tremblay, MacGregor
hydroge- and Wright (1977a,b) PRBS A-2?
nolysis
The variety of phenomena described above shows H. Acetylene Lee and Agnew (1977a,b) step B-1
that many different types of exotic behavior can hydrochln.
be expected for a fixed-bed reactor. The design 12. CO Vortlleyer and Winther (1982) propgt. frts A-2 u(r)
goal is clearly to choose the least sensitive oXidat.
reactor configuration which provides the desired 13. O-xylene Kershenbaum and step', B-2
conversion per unit catalyst, thus giving a less oxidat. Lopez- Isunza(1982) start-up
complex control problem. One interesting approach 14. Toluene Balker and Bergougnan (1985a, b) pulse - trlang. e, F, A2-3
to a systematic process and control design is hydrognt. sine T
a
given by Chylla et al. (1986) for an autothermal Autotheraal
reactor. 15. Water-gas Wong, Bonvin, MeHichamp and pulse A-2
shift Rinker (1983)

Recycle
Jergensen, Bortolotto, step T,e, A·2
16. :iidat.
and Kaas· Petersen F
 Fixed Bed Reactor Dynamics and Control
4. PROCESS IDENTIFICATION
Process identification is concerned with obtaining
process models and the necessary estimates from
experimental data, in the present context with the
purpose of control synthesis. Below follows a
presentation of locally linear model representa-
tions and of model reduction techniques whereby
the model structures may be obtained. Thereafter
estimation of model parameters is treated.
Locally linear model representations
The linear models to be used for control design
represents generally either internal states i.e.
are in state space form or process inputs and
outputs only, i.e. are in input-output form. One
state space and two input output forms will be
presented below.
A state space representation may be obtained from
the partial differential equations (1-7) by appli-
cation of the quasistationarity assumption to the
gas phase component balance, through linearization
around a steady state profile and by use of ortho-
gonal collocation to discretize the axial coordi-
nate, as shown for the state space model by Mi-
chelsen et al (1973). The resulting model may be
written:
dx
Α χ + Β u + Β e (8)
dt
Where χ is the n-dimensional vector of state vari-
ables, u the m-dimensional vector of controls,
and e a zero mean white noise vector process with
intensity matrix Q. The measurements may be col-
lected in a ρ-dimensional y-vector, where ν is a
zero mean white noise vector process with inten-
sity matrix :
y = Cx + Du + v
The use of LaGrange interpolation (Villadsen and
Michelsen 1978), makes it possible to interpolate
from the collocation points to arbitrary axial
measurement positions with high accuracy. The
second term on the right in eq.(9), which repre-
sents direct transmission from the inlet to the
measurements is due to the quasistationarity as-
sumption and the global LaGrange interpolation.
The above two equations are shown in continuos
time form, whereas practical application usually
occurs at discrete points in time, due to the
discrete time action of the computer. The measure-
ments are available at the sampling instants:
y(t) - C x(t) + D u(t-l) + v ( t )
d reflects only the controllable and the observable
where time is normalised with the sampling inter-
val (=1) for convenience, and the discrete time
noise covariance is V^. The corresponding dis-
crete time state space model is obtained by inte-
grating eq.(8) over the sampling interval (Âstrom
1970, Sorensen et al 1980), giving:
x(t+l) - A
dx(t) + B
du(t) + e (t+1)
with A
d - A d( l ) _; A1 (t) - exp(At),
B - ( A - I ) A B , e (t+l) -JjA(l-t')B e(t')dt'
Qd= d ,B
t B Ad( t d t
d ioV ) nQ n d '> '
Q high pass filtering the inputs and outputs,e.g. by
d is the discrete time noise covariance. The
advantage of using the discrete time state space
model in eq.(ll) is that it is based upon the
continuous time noise description. Thus any subse-
quent design is potentially less dependent upon
sampling time than designs starting with a dis-
crete time noise with a constant covariance. The
B^ matrix is included to allow some freedom in the
disturbance modelling, i.e. the disturbances may
15
be modelled as filtered white noise. In the pre-
sentation above the disturbances have been assumed
to be stochastic. In some cases are the distur-
bances predominantly deterministic. In that case
the above model representation also can be used,
but then e(t) represents a sequence of impulses
that occur irregularly but relatively seldom.
One input-output model representation, which by
tradition is popular, is the frequency response
representation; or in the discrete time case the
pulse transfer matrices, which may be written:
1
y(t) - Giq' ) u(t) + v(t) (12)
where stationary noise may be modelled as:
1
v(t) - IKq" ) e(t) (13a)
1 1
where Giq" ) and HCq" ) are rational transfer
function matrices, which may be expressed as
polynomials of the general form:
_ 1 1 2
X(q ) - X + Xjq" + X q " + ... (14)
q 2
where G - 0 and H - I.
ο ο
When the transfer function representation is ap-
plied the linear transfer functions may be deter-
mined directly from input-output data using multi-
variate time series analysis techniques, e.g.
spectral analysis as described in Box and Jenkins
(1970).
Another input-output model representation is the
ARMAX model:
1 1 l
Piq" ) y(t) = CKq" ) u(t) + R(q~ ) e(t) (15)
where u and y are input and output vectors as
above, and e is discrete time white noise. P, Q
(9) an<jl R are matrix polynomials in the delay operator
q , of the general form:
1 1 2 Π
Xiq' ) - X + X^' + X q" + . . . + Χ ς* Χ (16)
q 2 χ
where Ρ - R - I whereas Q = 0 . The left hand
side term in eq. (15) is an autoregression of the
outputs, the noise term on the rigth hand side
represents a moving average and the input is exo-
genous .
The three model representations given above can be
shown to be input-output equivalent for stable
plants. For unstable plants predictors based upon
the three representations can be shown to be
(10) equivalent. The transfer function representation
part of the process. Whereas the state space and
the ARMAX models may include unobservable and
uncontrollable modes. The state space and the
ARMAX models may be transformed into nonunique
cannonical forms, i.e. one which is observable,
called left matrix fraction description in the
ARMAX case, and one which is controllable, called
rigth matrix fraction description in the ARMAX
(11)
case.
and
The noise in the above representations has been
assumed stationary. Most often the noise in chemi-
n cal processes contains a nonstationary component,
which may be observed as drift in the inputs and
outputs. This nonstationarity may be removed by
using the incremental value y(t) - y(t-l). This
filtering may however introduce a zero on the unit
circle in the noise polynomial R(q" ) in the ARMAX
case. An alternative procedure is to model the
noise as nonstationary, e.g. by differencing the
noise. In this case eq.(13a) may be written:
1 _ 1
( 1 - q " ) ! v(t) - H ( q ) e(t) (13b)
16 S. B. J0rgensen
The resulting model representation contains an
integrated mowing average (IMA) of the noise.
Model order reduction
Various types of model order reduction techniques
have been applied on fixed-bed reactors. Some of
these techniques will be briefly discussed.
The number of significant eigenmodes in the axial
dispersion model eq.(l-7), is very high for most
practical values of the axial Peclet number. It
turns out however by application of orthogonal
collocation that a limited number of collocation
points gives a satisfactory description of the bed
dynamics. Thus application of orthogonal colloca-
tion renders by itself a very useful model order
reduction technique when a single bed is modelled.
In several cases, e.g. when wall or environmental
dynamics have to be modelled further model order
reduction may be desirable. Silva et al.(1979) ap-
plied Davison (1966)'s method to reduce a 14'th
to a 7'th order model. Bonvin and Mellichamp
(1982b) compare various model order reduction tech-
niques, and in Bonvin et al.(1983c) they apply
Litz (1979a,b)'s method to reduce their 24'th to a
5'th order model. Bonvin and Mellichamp (1982b)
demonstrate that information may be obtained
about the structural dominance characteristics.
This information may be applied advantageously to
compare structural aspects of different reactor
designs as used by Chylla et al (1986). Rivera and
Morari (1985) have developed a frequency domain
model reduction procedure that yield near optimal
models with respect to user designed closed loop
performance.
A different approach to model order reduction is
taken by Hallager and Jergensen (1981,1983) and
Jergensen et al.(1985). The convective feature of
the fundamental thermal delay in the fixed bed and
process knowledge in a DYQUID is used to directly
setup a discrete time model structure of an ARMAX
model. The model structure can include the possi-
bility of variable flow rate, i.e. of variable
process delay. The advantage of this procedure is
that the resulting model structure is parsimo-
nious , which speeds up convergence of parameter
estimation and that the model structure is valid
over a relatively wide range of operating condi-
tions, when the model parameters are estimated on-
line.
Epple (1986) uses the interpretation of the
reaction zone as a moving wavefront to propose a
nonlinear model reduction technique, where the
reactor response to inlet or cooling temperature
changes at an arbitrary location may be described
by a sixth order model.
Parameter estimation
The early model based fixed bed reactor control
studies used mechanistic models wherein the para-
meters either were estimated using offline data
from the reactor (Jutan et al.1977b), or were
based mainly, but not entirely, upon prior infor-
mation (Serensen 1977), or were based upon a com-
bination thereof (Wong et al. 1983), and (Silva et
al. 1979). In the latter case a few parameters
were updated prior to the experimental run . Jutan
et al.(1977b) collected their data under univa-
riate closed loop operation, presumably due to
nonstationary character of the stochastic distur-
bances, using an external dither added to the
feedback signal. The remaining investigators used
open loop experiments as a basis for parameter
estimation. It is noteworthy that in no case were
the model parameters derived entirely from inde-
pendent experiments. It is a conclusion in several
of these early works (Jutan et al. 1977, Wallmann
et al. 1979) that the effort necessary for the
model development seems to preclude industrial use
of mechanistic models.
MacGregor and Wong (1980) use an empirical trans-
fer function approach, where the input-output
characteristics are described with an autoregres-
sive integrated mowing average (ARIMA) model, i.e.
eq.(12 and 13b). They then determine a minimal
state space realization. Kozub et al.(1986) also
use an ARIMA model. In both cases the parameters
are determined offline using a multivariate maxi-
mum likelihood method.
Harris et al.(1980) use black box ARMAX models and
determine the parameters on-line, with recursive
least squares. A similar procedure is used by
McDermott et al.(1986a,b), where the approximate
model orders are detemined using simulation stu-
dies. A different approach to using ARMAX models
were taken by Hallager et al. (1981,1983 and 1984)
and Jergensen et al. (1985). These investigators
use a procès knowledge based approach to specify
only the unknown parameters in ARMAX model struc-
tures. This usage of parsimonious model structures
makes it possible to obtain relatively fast con-
vergence to new parameter values during on-line
operation, when necessary.
The fixed-bed reactor identification studies de-
monstrate a clear tendency in that the early stu-
dies use mainly mechanistic based, state space
models, which were fairly manpower demanding to
develop and require nonlinear regression methods
for obtaining parameter estimates. Whereas later
work has concentrated around developing and iden-
tifying models at a significantly lower manpower
cost. These models were either transfer function
based models, where the parameters are obtained
off-line using spectral analysis, or ARMAX models
which are linear in the parameters which therefore
can be estimated on-line.
5. PROCESS CONTROL DESIGN
This section starts with a brief summary of linear
qudratic control design including state recon-
struction. Thereafter follows a brief discussion
of sensor selection and location for control pur-
poses . Then follows a summary of the investigated
control designs. This major part is grouped accor-
ding to the applied model representations, star-
ting with the mechanistic model based control
designs, followed by the transfer model based and
finally the adaptive control designs which mainly
are ARMAX model based.
Most of the fixed bed reactor control experiments
listed in Table 3 have been using linear qudratic
(LQ) control, or linear quadratic Gaussian (LQG)
control when the disturbances are assumed to be
Gaussian stochastic. In this socalled optimal
control approach the ideal continuos time purpose
is to minimize (for stochastic disturbances: the
expectation of) the objective function:
Nv T T T T
j - f ( n F n+2u F n+u F u)dt+n (N)F n ( N ) (17)
C J o ne une uc oc
where Ν is the time horizon of the optimal con-
trol, and η can be the tracking error y - y ,
the process output y or another linear combina-
tion of states and inputs :
n(t) = C x(t) + D u(t) (18)
η η
using the state space model and the knowledge that
the controls are constant over a sampling period
eq.(17) may be written:
N-1A
χρΤA T A
T T
J„ =7 ( x+2u F x+u F u)+n (N)F n(N) (19)
C 4V»Q χ ux u oc
where F , F , and F are integral functions of
the weighting matrices in eq. (17). These inte-
grals are evaluated over the sampling interval,
and depends also upon the process dynamics.
 Fixed Bed Reactor Dynamics and Control
Most often an objective function of the following
form is applied:
J N-l( n nT +F 2 nu +T uuF) T +F
D -^ nd und ud ^ W F ^ ) (20)
This function may also be brougth into the form of
eq.(19), but the weigthing matrices are then sim-
ple functions of the weigthing matrices in eq.
(20) and the process dynamics. Differences between
the two approaches can thus be expected when there
are significant variation in the process state
over the sampling interval. Such differences may
be reduced by reducing the sampling interval.
The minimization of either form of eq. (19) re-
sults in a control which is a linear function of
the states or the state estimates. The control
gain matrix is in general time varying, but often
the stationary solution obtained for Ν approaching
infinity is used. The calculation of the feedback
reveals that it is desirable to have access to the
full state vector, which is rarely possible in
fixed bed reactors. Instead state reconstruction
can be accomplished. In the case of Gaussian dis-
turbances it is possible to determine the optimal
estimator, which minimizes the variance of the one
step prediction error. This procedure gives the
Kalman gain matrix Κ to be used in the Kalman
filter:
x ( t + y t ) - A x(tyt-l)+B u(t)+K(y(t)-y(tyt-l)) (21)
d d
where y(t^t-l) - C x ^ t - l ) + Du(t-l)
The filter gain matrix Κ is in general time va-
rying, but most often the stationary filter gain
have been used.
The theoretical basis for using Linear Quadratic
control is the validity of the separation theorem
for linear process models. This implies that the
optimal control strategy can be split into two
parts: one which estimates the states from the
measured outputs and one which computes the linear
feedback from either the estimated or the measured
states. The obtained closed loop is guaranted to
be stable, provided the process is linear. However
fixed-bed reactors are not linear, therefore an
important issue is how to tune the linear quadra-
tic control, i.e. to select the control weigthing
matrices in eq. 17 or 20, in order to obtain a
control design which is robust against parameter
variations and against possible variations in
disturbance characteristics. Although the specific
form of the objective functions in eq.'s (18 and
20) does not have a physical background the LQ(G)
design approach have been widely applied in fixed-
bed reactor control studies. One reason for this
popularity may be the relatively direct interpre-
tation of penalizing the variance of the chosen
variables η and inlets in the objective function
through the choise of weights. Thus providing a
direct link to the control objective.
Sensor location
In control system design both selection of the
sensor types and their position are design vari-
ables. In practice however there may not be reli-
able on-line sensors available for measurement of
product compositions. Similarly for load distur-
bances, whether they are of internal origin e.g.
catalyst activity variations or of external ori-
gin, they may not be feasibly measured e.g. due to
inhospitality of the environnement. Therefore
variables of secondary interest with regard to the
purpose of control may often be measured and used
in a state reconstructor to estimate the vari-
able (-s) of primary interest for the control ob-
jective. In this section mainly the principal
aspects of sensor location for control will be
discussed. A few particular examples will be men-
tioned and other are covered in the discussion of
control designs.
17
The sensor selection and location problem differ
in principle according to the type of feedback
used. If output feedback is used then the sensor
types are dictated by the control objective and
their location by the location of the control
zone. If sensors are not available or only can
provide measurements less frequent than desired
then an estimate may be evaluated based upon se-
condary measurements. The latter application was
demonstrated by Tremblay and Wrigth (1977). In
their case the predictor parameters were updated
on-line when the less frequent concentration mea-
surements became available.
When a detailed process model is available in the
form of a state space model then it is possible to
use a state reconstructor to reduce the effects of
process and measurement noise, and to provide
estimates of unmeasured states for feedback. Here
the sensors should be positioned such that the
model is observable, which usually is not a pro-
blem in fixed bed reactors due to the predominant
convective behavior, and such that optimal esti-
mates are obtained for the control objective. This
problem have been investigate by several groups
e.g. Mellefront and Sargent (1978) and Harris et
al.(1980). When state reconstruction is used then
in addition unmeasurable disturbances may be esti-
mated. The minimum number of sensors necessary to
observe external disturbances is equal to the
number of linearly independent external disturban-
ces (Morari and Stephanopoulos 1980, and Jcrgensen
et al. 1984). In the latter study it is recommen-
ded to locate temperature sensors just upstream of
the control zone for estimation of distributed
acting disturbances. If the disturbance is a boun-
dary forcing to a delay ( e.g. inlet temperature )
then it is important to obtain a measurement as
close to the disturbance entry point as possible.
Generally it is advisable to use redundant infor-
mation for both state and disturbance estimation.
It was however demonstrated by Joseph and Brosilow
(1978a,b) that a high number of sensors increases
the sensitivity towards modelling error. Alvarez
et al.(1981) and Romagnoli et al.(1981) propose to
use a variable location of the sensors according
to the present operating conditions, and suggest a
fast algorithm for this purpose. Kuruoglu et al.
(1981) develope a steady state distributed filter
for estimation of concentration and temperature
profiles, and a dynamic filter for estimation of
slow catalytic activity changes. The combined
filters are simulated using the Clough and Ramirez
(1976) fixed-bed reactor model for styrene produc-
tion by dehydrogenation of ethylbenzene. They
demonstrate for the profile estimation that inter-
nal measurements allow for subdivision of the
spatial domain and thus yield improved profile
estimates.
Mechanistic model based control designs
Vakil et al.(1973) used the continuous time LQG
approach on a simulated adiabatic fixed bed reac-
tor, with the parameters of the Sinai and Foss
(1970) liquid phase packed bed reactor. The state
space model is derived using orthogonal colloca-
tion in space. The control problem was to keep the
reactor close to the steady state operating condi-
tions using manipulation of the temperature or
concentration of a secondary reactant stream in-
jected partway along the bed to compensate for
upstream disturbances in the feed concentration
and temperature. A temperature measurement at the
injection point is used as the only output. Thus
the sensor and the control input locations were
coupled. This work inspired a number of subsequent
studies in several ways. Cinar (1984) studied the
location of the sensor and control input in a
similar two bed gas phase fixed-bed reactor, using
the parameters of the Hoiberg et al. (1971) gas
phase packed bed reactor.
18 S. B. J0rgensen

Jutan et al.(1977c) applied discrete time LQG Clement et al.(1980) show experimentally and theo-
control to a cooled type a fixed bed reactor with retically that usage of the discrete time control
three independent reaction components. An objec- may result in oscillations in the discrete time
tive function similar to eq.(20) was used with η axial temperature variance profiles at relatively
being the measured temperatures. The control ob- large sampling time. These variance oscillations
jective was to keep the reactor close to the reduce the utility of dicrete time variance pro-
steady state operating conditions when upset by files for design. Instead continuous time variance
stochastic disturbances and an induced determini- profiles representing a time average over the
stic load disturbance in the coolant temperature. sampling interval should be used (Ahrensberg et
The manipulated variables were the Hydrogen and al. (1983)). The latter investigators also eva-
Butane flow rates. The LQG control performed well luate the effect of using the discrete and the
and it was demonstrated (Jutan et al.1977a,b) that continuous time objective functions eq.20 and 18.
it was possible to infer exit concentrations rea- The results at the investigated conditions show
sonably well from the measured flowrates and that there are only minor differences between the
(nine) temperatures . The control design appeared two objective functions, at the relatively mild
to be robust towards moderate changes in operating operating conditions.
conditions and especially in catalyst activity. In
a subsequent study Jutan et al.(1984) compared the Silva et al.(1979) and Wallmann et al.(1979)
above controller with white process and measure- investigate experimentally LQG control of a two-
ment noise to another LQG controller where a third bed catalytic reactor system using quench flow
order identified multivariable time series model rate and temperature as possible manipulated
of the process disturbances and negligible mea- variables. The individual beds were similar to
surement noise is used. The first controller has that studied by Hoiberg and Foss (1971).In Silva
the best performance to a load disturbance in et al.(1979) the control objective is to maintain
coolant temperature, but apparently the closed the effluent concentration when the process is
loop behavior is relatively insensitive to the subjected to zero mean disturbances at the inlet
noise model. of the first bed. It was found that a simple first
order shaping filter was sufficient to model the
TABLE 3
: Experimenta
l Investigation
s of fixed-be
d reacto
r estimatio
n and co disturbances entering the second bed. A fairly
Identificatio
n Contro
l simple solution results: The mixing temperature at
Bed model estimât.
type: p«- ate- me- the quench inlet is controlled locally, whereas
repr: ran: te: thod
: class: the quench flow rate is used for univariate con-
Single bed
:
1. H Serense
n (1977
) · off
•. KF LQ
C ΜΙΜ
Ο ί trol of the effluent concentration from the second
Oxldat
. Buchholt,Clement.Jergenaen(1979
) ARMAX RLS CK
V SIM
O c
Clement,Jergenaen.Serenaen(1980
) e
a off . KF LQ
G ΜΙΜ
Ο ί
bed, based upon an estimate of this concentration
Clement and Jergense
n (1981
) aa off . IK
A ΜΙΜ
Ο c from measuring four temperatures in the second
Hallager et al.(1981,1983,1984
) ARMA X RELS QL ΜΙΜ
Οc
Coldachmldt,Hallage
r bed (Wallman and Foss 1981). In the adjoining
and Jergense
n (1985
, 1986
) ARMA
X RELS QL ΜΙΜ
Ο < paper Wallman et al.(1979) consider maintaining
Cooled singl
e bed: average quench flow rate and effluent concentra-
2. Butane Jutan, KacCregor
and Wrigh t (1977a,b.c
) s a off
. KF LQC ΜΙΜ Ο1 tion or temprature upon sustained disturbances in
noly-
Tremblay and Wrlgth (1977
) a a off
. MRA
C ΜΙΜΟ< feed concentration. In this study proportional and
MacCregor and Wong (1980
) TF.a
a off. KF LQC ΜΙΜ Οi
ala Harrls.MacCregor. Wrigth(1980
) ARMA
X RLS VM SIS
O< integral LQG control is achieved using the objec-
Kozub and MacCregor (1986
) F T off
. LQ.IM
C ΜΙΜΟ <
tive function J in eq. 20 with η including incre-
3. Butan
e Lee and Lee (1985
) ARMAX,H
S RLS,I
V V
M SIM
O ο
d
mental states and the selected states to be inte-
oxldat
.
grally controlled. In this approach incremental
C o n s e q u teW beds
: controls are used in eq.(20). Both the effluent
4. H, Silva
, Wallman and Foe
s (1979
) as off
. KF LQC ΜΙΜ
Ο fi
Oxldat . Wallman , Silv
a and Foaa (1979
) sa off
. KF LQG ΜΙΜ
Ο s concentration and temperature controller operated
Wallman and Foes (1981
) as off
. DE LQC SIS
O fi
satisfactorily when four temperatures were mea-
Autothermal: sured in the second bed. Using only one tempera-
5. Water- Wong, Bonvin
, Mellicham
p
gas an
d Rlnke
r (1983
) ture measurement gave considerable control errors.
shift McDermott, Mellicham
p The advantage of slow integral action on the
and Rlnke
r (1986a.b
) PP SIS
O ο +
ΜΙΜΟ ο + quench flow is avoidance of control input satura-
Recycle
:
tion. Quench flow is effectively used for rapid
6. H
?
Oxldat
.
Bortolott
o and Jergense
n (1985
) ARMA
X REL
S
control actions, since it forces the bed distribu-
Serensen (1977) applied approximate continous time tedly, whereas the slow control action is taken
LQ control to the adiabatic reactor of Hansen and care of by the temperature at the quench point.
Jergensen (1976a). The control problem was to
maintain the steady state operating conditions in
the face of deterministic upstream disturbances in This essentially completes the presently available
temperature, concentration and mass flow rate studies of application of LQ(G) methods based upon
using inlet temperature and concentration as mani- a priory developed mathematical models with state
pulated variables. The upstream disturbances were reconstruction. Three major conclusions emerge:
modelled using first oder filters, and thus esti- i) The LQ(G) approach works on fixed-bed
mated from the internal temperature and exit con- reactors.
centration measurement. The control performed ii) Redundant measurements and feedback
well, but it was not possible to estimate simulta- should be emphasized due to limited
neous upstream disturbances in concentration and model accuracy,
mass flow rate. This estimation difficulty was due iii) The control design seems robust towards
to a nearly linear dependence of the two distur- inaccuracies in process noise models.
bances at the particular flowrate used. In a sub- Two impediments for industrial application seem
sequent study Jergensen and Clement (1977) demon- obvious :
strated that it is possible to estimate all three iv) Model development may be a significant
disturbances at a lower flowrate using the minimum burden.
of three carefully located sensors, and to main- v) The robustness towards errors in model
tain the steep part of the temperature profile at structure is limited.
a desired position using LQ control. The design problem of choosing the controller
weights does not seem to be a significant problem
A comparison of continuous and discrete time Kal- in that diagonal matrices seem to be sufficient.
man Filtering and optimal control is performed by The choise of weigths do however require experi-
Serensen et al.(1980). It is shown that the dis- ence .
crete state space model eq.(ll) is convenient for
filter design calculations due to the relatively A couple of frequency domain design techniques
low sensitivity of the filter design towards chan- have been investigated. Foss et al. (1980) used the
ges in sampling time. In the subsequent paper characteristic locus analysis method to design a
 Fixed Bed Reactor Dynamics and Control
three-input multivariable control system for the
two-bed reactor system described above with the
objective of regulating the product concentration
and temperature. Dynamic compensators are deve-
loped such that interaction among variables is
suppressed and the effects of concentration di-
sturbances are controlled. The control is effi-
cient, but uses about eigth times the control
effort needed for concentration control alone due
to a near singular system matrix. It is concluded
by the authors that proper design of dynamic com-
pensators relies heavily on the designers experi-
ence. Clement and Jorgensen (1981), Table 3.1
investigated the direct Nyquist array (DNA) design
with constant pre and post compensator matrices.
The control objective was to maintain the steady
state profile, when the reactor was upset by up-
stream disturbances. The controls utilized were
inlet concentration and temperature. However a
relatively poor behavior was obtained for the
controlled reactor. This seemed to be connected to
the thermal delay, therefore a delay compensation
was investigated, however without much success
(Terndrup et al.(1983)).
Economu and Morari (1986) develop and investigate
through simulation a multiloop IMC design ap-
proach. The method treats interaction terms as
additive pertubations and uses an interaction
measure, which is closely related to the Gershgo-
rin bands of DNA, to determine either a reasonable
configuration or to decide for a full multivari-
able IMC design (Garcia and Morari 1985a and b ) .
The multiloop method gives a rather good design
for the Foss et al. (1980) problem. Whereas a
multivariable design seems most promising for a
methanation reactor where outlet and hotspot tem-
perature is controlled using inlet flow rate and
temperature as manipulated variables.
The frequency domain design techniques employed
thus far rely on a trial and error procedure for
determining simple compensators. In the multiloop
IMC design approach only the univariate control-
lers are detuned using simple filters to account
for interactions and the performance can be di-
rectly evaluated.
Modal design has also been investigated in a coup-
le of instances. Georgakis et al.(1977b,c) inve-
stigate the stabilization of a cooled tubular
reactor at an unstable steady state, using coolant
flow as manipulated variable and measuring tempe-
ratures and concentrations at the collocation
points. Modal design is demonstrated computa-
tionally to give a fairly large domain of attrac-
tion when the two largest eigenvalues are relo-
cated. It is also shown that an observer can be
used to estimate the control variable directly
from temperature measurements only. Wong et
al.(1983), Table 3.5 use modal design with a re-
duced order model to stabilize the autothermal
reactor at an unstable state manipulating feed
stream heat input. The number of states in the
reduced order model were measured directly, The
control design performed well in disturbance re-
jection both with a fifth and a third order re-
duced model.
The problem of robustness of control design is
treated by Mandler et al. (1986) in a computatio-
nal study of univariate control of a cooled metha-
nation reactor operating close to ignition. The
inlet temperature is used to control the product
temperature. The control design is based upon a
reduced order model. Internal model control (Gar-
cia and Morari 1982a) is used for control design
coupled with Doyle's (1982) structured singular
value analysis, whereby a tradeoff between robust
stability and performance is accomplished.
19
Transfer function based design
MacGregor and Wong (1980),table 3.2 use an off-line
identified transfer function model developed from
production rates which are inferred using a linear
regression on current temperatures and flowrates.
The regression coefficients are updated every time
a gas chromatograph measurement is available
(Wright et al.1977). The transfer function model
with nonstationary noise is transformed into a
minimal order state space realization. The resul-
ting fourth order model is used in a LQG design to
control the production rates of propane and bu-
tane. The objective function is eq.(20) with η
containing the two outputs and using incremental
controls of hydrogen and butane to cope with non-
stationary noise. The resulting control perform
well in rejection of a coolant temperature load
disturbance. Drift is observed in the manipulated
variables possibly due to catalyst activity chan-
ges.
In their recent study Kozub et al. (1986), table
3.2, use an off-line identified empirical trans-
fer function representation similar to that of
MacGregor and Wong (1980). Kozub et al. (1986)
transform their model into a right matrix fraction
description, which is convenient for control de-
sign. They investigate two different control de-
signs: LQ output feedback control using eq.(20)
with η being the tracking errors and incremental
controls; and deterministic internal model control
(IMC). IMC may be tuned to achieve satisfactory
performance and robustness towards model plant
mismatch using diagonal filter matrices on the
feedback and setpoint signals. The manipulated
variables were setpoints to the hydrogen flowrate
and to a hotspot temperature controller. The hot-
spot temperature is controlled by butane flow
rate.The results indicate that both designs were
restricted by model plant mismatch, due to nonli-
nearities and catalyst activity changes. Trial and
error adjustment of the tuning matrices led to
robust designs with acceptable regulation and
setpoint change performance over a fairly wide
operating region.
Adaptive control design
A number of studies have been concerned with inve-
stigating various adaptive control algorithms for
fixed bed reactor control. The main purpose has
been to reduce the control sensitivity towards
model plant mismatch, using input-output data to
provide information for on-line tuning of the
control. Two main approaches to adaptive control
have been used. In one case are the controller
parameters estimated directly using an implicit
model, in the other case is model parameters ex-
plicitly estimated, and then used to calculate the
control. This general area has recently been re-
viewed by Seborg et al.(1986).
Much of the early theoretical development was
carried out by the group in Lund, Sweden, where
the univariate selftuning regulator was described
by Âstrom and Wittenmark (1973) and Àstrom et al
(1977). They use implicit minimum variance (MV)
control, i.e. eq.(20) with F =0. Borisson (1979)
derived an implicit MV controYler for the multi-
variable case. Clarke and Gawthrop (1979) derived
an implicit controller which minimize the univa-
riate objective function in eq.(20) with a one
step time horizon. This controller is often called
a generalised minimum variance controller (GMV). A
multivariable version of the implicit GMV control-
ler was developed by Koivo (1980) for an equal
number of inputs and outputs (square plants) and
by Grimble (1981) for nonsquare plants. An impli-
cit univariate LQG controller has been developed
by Zhao Ying and Âstrom (1981) and Grimble (1984).
Univariate pole-placement control design algo-
20 S. B. J0rgensen
rithms have been proposed by Wellstead et
al.(1979) and Àstrom and Wittenmark (1980). Prager
and Wellstead (1980) proposed a multivariate ver-
sion. These algorithms are mainly of the explicit
model type.
The first attempts at implementing adaptive con-
trol design used model reference adaptive control
(MRAC). Here the objective is to make the output
of an unknown plant assymptotically approach that
of a specified reference model. The initial at-
tempts lacked theoretical background. The MRAC
approach can be shown to be similar to the selftu-
ning regulator in a number of cases. Tremblay and
Wright (1977), Table 3.2, performed an early inve-
stigation of MRAC. They used a state space model
for the extents of the reaction augmented with the
production rate of ethane and propane. The slecti-
vity of propane was to be controlled using the
flow rates of hydrogen and butane as manipulated
variables. The model reference algorithm was based
upon Liapunov's direct method. The results were
not satisfactory. However in the course of this
study a linear regression model was developed for
the effluent concentrations as a function of three
functionals of the reactor temperature and the
inlet flow rates. This regression provided predic-
tions for the concentrations at every sampling
instant. The regression coefficients were updated
using recursive least squares (RLS) estimation
when actual concentration measurements were avai-
lable at every 12'th sample (Wright et al.1977).
Harris et al.(1980), Table 3.2, apply univariate
self-tuning control in various forms to control
the hotspot temperature, independent of position,
manipulating butane flow rate. A GMV controller
with incremental control signal provided conside-
rably better control than a conventional PI con-
troller. In a recent study Underwater et al.(1986)
have used a GMV controller to tune the same loop
using a logarithmic transformation of the hotspot
temperature. The input weight in eq.(20) was used
to provide a trade off between acceptable hotspot
stabilization and smooth variation of the butane
flow.
McDermott et al.(1986a), Table 3.5 apply a uni-
variate pole-placement controller with the abili-
ty to auto-tune a pole location on-line. They
control a temperature just upstream of the hotspot
manipulating the heat input to the entering reac-
tant gas. Good disturbance rejection behavior is
demonstrated both at an open loop unstable and at
an upper stable steady state and similarly for
setpoint changes between the two operating condi-
tions. The good setpoint behavior is obtained by
on-line adjustment - auto-tuning - of a desired
pole-position by minimizing the error variance η
- y - y (McDermott and Mellichamp (1984)). In
the adjoining paper (McDermott et al.(1986b)) a
multiloop version of the pole-placement controller
is applied to maintain the temperature profile
with two loops : in one the temperature near the
hotspot is controlled by the top heater (i.e. at
the turning point of the gas), in the other loop
a temperature near the exit is controlled by mani-
pulating either - in simulations - the heat input
to the entering reactant gas or - in an experi-
mental test - the gas flow rate . A dynamic de-
coupler is used to minimize loop interactions. The
model structures, i.e. the polynomial orders η ,n
and n^, were chosen based upon simulation explri-
ence. The above configuration behaved well in
simulations and also in an experimental test
around the upper stable state.
Buchholt et al.(1979), Table 3.1, investigate
disturbance rejection of the adiabatic fixed-bed
reactor using inlet temperature as manipulated
variable with five equidistant axial temperature
measurements. The control design was GMV and RLS
was used for parameter estimation. The control
results were satisfactory. In subsequent works has
the control design been improved to an an assymp-
totic LQ design, which has been tested on the
fixed-bed reactor under more and more demanding
conditions. The control design utilizes the objec-
tive function in eq.(20) with
T T T
n(t) = (y (t-l),..,y (t-n l),u (t-2), .
T T p T
.,u (t-n -l),z (t-l)) (22)
q
where z(t) contains integral outputs, if included.
The control horizon is set to infinity and the
Riccati equation is iterated by direct substitu-
tion once at each sampling instant. The feedback
gains will converge to the optimal values when
the parameters are constant. The feedback gains
will vary relatively slowly when the model parame-
ter values change. The weigths in the control
performance criterion are chosen using available
knowledge about the limitations of the model
structure in order to obtain robustness towards
unmodelled dynamics. Hallager and Jergensen (1981)
investgate this design experimentally using incre-
mental inputs in eq.(20) to achieve offset elimi-
nation in a least squares sense, since there are
maximally two inputs: inlet temperature and con-
centration and five outputs. Effective rejection
of upstream pulse disturbances in temperature,
concentration and flowrate is demonstrated when
the reactor is operated at low inlet concentra-
tion. In Hallager and Jergensen (1983) is the
model structuring expanded to be able to handle
disturbances in flow rate, i.e. the thermal delay;
and to include integral action on specific out-
puts. The ability to reject an upstream tempera-
ture disturbance is demonstrated to be satisfacto-
ry and to be nearly independent of sampling time,
provided the model structure is modified accor-
dingly, when the distributed inlets also are used
as manipulated variables. Hallager et al.(1984)
investigate various identification aspects using
RELS with variable forgetting factor and show in
particular that the adaptive controller is able to
reject disturbances in upstream temperature and
flowrate also at medium high oxygen concentration.
The parameter estimation, which is RELS with va-
riable forgetting, is relatively rapidly able to
estimate varying parameters after a flow rate
change. Goldschmidt et al.(1986), table 3.1, inve-
stigate the servo behaviour obtained through a
first order filtered change of the integral refe-
rence states. During a sequence of setpoint chan-
ges the reactor is brought from almost no reaction
up to a very active reaction, with a high thermal
gain. The adaptive control behaves very well and
is able to follow the parameter changes even
though the process in addition is slowly time-
varying due to catalyst activity changes.
Bortolotto and Jergensen (1985), investigate di-
sturbance rejection and set point tracking of a
reactor with recirculation of unreacted reactants,
using the same control design as above. The mani-
pulated variables are inlet temperature to the
reactor and limiting reactant mass flow rate to
the recirculation loop. The outputs are five
temperature measurements and the exit reactant
concentration. In this case the control weights in
the objective function were selected carefully in
order to account for model structure limitations.
Disturbance rejection is effective and a setpoint
change which brings the reactor close to a Hopf
bifurcation point is also very well handled.
Both the transfer function based and the ARMAX
based adaptive control designs lead relatively
quickly to on-line implementation. Especially the
process knowledge based model structuring proce-
dure seems promising in this regard. Most of the
early control design approaches include a step,
which often can be of trial and error type, where
the designers artistic abilities often play an
 Fixed Bed Reactor Dynamics and Control
important role for a succesful result. The on-
going development of how to apply control design
methods such that they are robust against known
inaccuracies in model and/or disturbance charac-
teristics holds promise for reduction if not eli-
mination of artistic elements in this design step.
6. OPTIMIZING CONTROL
Long term optimal operating conditions for fixed-
bed reactors may be difficult to determine due to
timevarying process characteristics such as cata-
lyst deactivation and the associated modelling
difficulties, which persist even though sinificant
progress have been achieved in this area. There-
fore it is not desirable to apply open-loop opti-
mizing control polices to fixed-bed reactors. Due
to the time-varying physics this is an obvious
area for application of on-line identification
techniques.
Hong and Lee (1985) propose on line computation of
a measure of catalyst activity using measurement
of inlet temperature, exit concentration, and for
a cooled bed also of coolant temperatures. They
apply their measure of catalyst activity for eva-
luation of the desired quasi steady inlet tempera-
ture to maintain the product concentration within
a specified range. Thus their control objective
could presumably be met with techniques from the
previous section which also can address the prima-
ry control objective. Lee and Lee (1985), table
3.3, study optimization by simulation and expe-
rimentally on a non-adiabatic fixed-bed reactor
for the production of maleic anhydride using coo-
lant temperature and flow rate as manipulated
variables, with a temperature near the hot spot
and the inferred product concentration as out-
puts. The temperature near the hot-spot was con-
trolled in an adaptive loop by the coolant tempe-
rature with a feedforward from flow rate. Assuming
this control to be perfect a second order Hammer-
stein model was assumed for maleic anhydride con-
centration as a function of hot-spot setpoint and
flow rate. This model is a quadratic function of
the input variables, but linear in the parameters.
These parameters are determined using recursive
instrumental variables. The steady state relation-
ship from this model is used in a profit function:
where the first right hand side term is the net
profit per unit time and the second a penalty term
on high bed temperatures. The maximization of J is
locally constrained to the region where the iden-
tified model is valid. The search direction is
determined, in the experiments at every 100'th
sample, using the random search method . The two
inputs are perturbed by PRBS sequences during the
run, the amplitude of these signals are reduced as
the optimum is approached. The algorithm is demon-
strated to perform well in two experimental opti-
mization runs.
7. CONCLUSIONS
The experimental investigations of fixed-bed reac-
tor dynamics and control have mainly been carried
out with type a beds in various laboratory reactor
configurations, where the Reynolds number often is
significantly lower than in industrial practise.
The dynamic behavior of fixed-bed reactors is
qualitatively well understood. The dynamic beha-
vior can be quantitatively modelled using conser-
vation principles with reasonable accuracy when
the reaction kinetics are known. However a number
of theoretical modelling issues remain unsolved.
Theoretical results indicate that axial thermal
dispersion also is important for the dynamics of
industrial fixed-beds. It has been demonstrated
21
that linear(-ised) mathematical models can des-
cribe the dynamic behavior in a reasonable neigh-
borhood arouund an operating point. In cases with
very high sensitivity it may be necessary or ad-
vantageous to use nonlinear dynamic models.
Three somewhat different types of linear model
representations have been applied in identifica-
tion studies on fixed-bed reactors. The most pro-
mising approach for practical application seems to
be ARMAX models with parsimonious model structures
where the parameter to be estimated are selected
based upon knowledge of qualitative process dyna-
mics . The advantage of this approach is that the
resulting model structures are valid over a wide
range of operating conditions and that the rela-
tively low number of parameters may be determined
relatively fast on-line.
All three model types and various identification
methods have been successfully applied in control
studies. The sucessful control design methods have
predominantly been based upon LQ(G) designs where
both the primary but also additional control ob-
jectives have been investigated. Modal design have
mainly been used for satisfying the primary design
objective. The mechanistic modelbased control
design approaches have in some cases reported low
robustness towards varying parameters or modelling
errors. The recent design procedures which addres-
ses the problem of providing robust performance
and stabilization represent a promising develop-
ment towards reducing the amount of trial and
error necessary for multivariable control design.
It is noteworthy that sequences of stepchanges
covering a relatively large operating region have
mainly been reported for robust control and for
adaptive control methods. Thus these two methods
seem most promising for practical application.
Optimizing control has been demonstrated to work
well in one case, where adaptive control was used
to satisfy the primary control objective. This
result is also a most promising development.
A few attempts have been made towards developing
a methodology for carrying out reactor design
while accounting for poosible control problems
this is a very important area to pursue in order
to obtain resilient process designs which enable
practical solution of additional control problems
including optimizimg conntrol.
In summary the area of fixed-bed reactor control
have matured significantly and practical applica-
tions of advanced modelbased control techniques
seem most likely. However several aspects deserve
additional theoretical and experimental studies.
NOTATION
A,Β Matrices in state space model
C Concentration
Dm Damkohler number
F ^ Mass flow rate or weighting matrix
G(q ) Transfer function matrix polynomial.
H ^ Solid to gas thermal heat capacity ratio
H(q ) Transfer function matrix polynomial.
M Thermal residence time to heat transfer
time constant ratio,
η ^ vector variables in control objective.
P(q ) Autoregression matrix polynomial.
Pe^ Peclet number (axial)
q ^ Backward shift operator.
Q(q ^) Input matrix polynomial.
R(q ) Noise matrix polynomial.
R(C,T) Reaction rate
Τ Temperature
u vector of inputs (controls),
ν measurement noise vector,
χ state vector,
y output vector.
ζ axial distance, integral output.
22 S. B. J0rgensen
subscripts
d discrete time
h heat
m mass
η noise or output function
u input
Greek letters
β adiabatic temperature rise
7 activation energy
REFERENCES
Ahrensberg,A., Goldschmidt,L.; Clement,Κ.; Seberg,
H.; Jergensen,S.B. 1983 :"A Note on Discrete
Time Optimal Control of Continous Time Systems"
Proc.IASTED Symp. Applied Control and Identifi-
cation ρ. 2.6 - 2.10, Copenhagen
Alvarez,J.; Romagnoli,J., Stephanoupolos,G. 1981 :
"Variable Measurement Structures for the Control
of a Tubular Reactor" Chem.Eng.Science 36
1695-1712
Âstrom,K.J. 1970 Introduction to Stochastic Con-
trol Theory, Academic Press
Âstrom,K.J.; Wittenmark.B. 1973 :"On Self-Tuning
Regulators",Automatica 9 185
Âstrom,K.J.; Wittenmark,B. 1980 :"Self-Tuning
Controllers Based on Pole-Zero Placement" IEE
Proc. 127,D 120
Âstrom,K.J.; Borisson.U.; Ljung,L.; Wittenmark,B.
1977 :"Theory and Applications of Self-Tuning
Regulators" Automatica 13 457-476
Bonvin,D. ; Rinker.R.G.; Mellichamp,D. 1980 .'"Dy-
namic Analysis and Control of a Tbular Autother-
mal Reactor at an Unstable State" Chem.Engng.
Science 35 603-612
Bonvin,D.; Mellichamp,D. 1982a "A Generalized
Structural Dominance Method for the Analysis of
Large-Scale Systerns"Int.J.Control 35 807-827.
Bonvin,D.; Mellichamp,D. 1982b "A Unified Deriva-
tion and Chritical Review of Modal Approaches to
Model Reduction".Int.J.Control 35 829-848.
Bonvin,D.; Rinker.R.G.; Mellichamp,D. 1983a-d :"
On Controlling an Autothermal Fixed-BED Reactor
at an Unstable State-I: Steady-State and Dynamic
Modelling"; "-II: Discrimination among Rival
Models to Achive Suitable Internal Structure"; "
-III: Model Reduction and Contro Strategies
which avoid State Estimation" Chem.Engng.S c ience
38 233-244; 245-255; and 607-618.
Borisson.U. 1979 :"Self-Tuning Regulators for a
class of Multivariable Systems" Automatica
15 209-215
Bortolotto,G.E.; Jergensen.S.B. 1985 : "Multiva-
riable Adaptive Control of a Recycle Reactor".
Proceedings of IFAC workshop on Adaptive Control
of Chemical Processes. ed. Η.Β.Unbehauen Frank-
furt oct. 21-22.
Box,G.E.P.; Jenkins,G.M. 1970 "Time Series Analy-
sis Forecasting and Control" Holden Day
Buchholt,F.; Clement,K.; Jergensen,S.B. 1979
"Self-tuning Control of a Fixed Bed Reactor"
5'th IFAC symposium on Identification and system
Parameter Estimation, Darmstadt, paper C3.5.
Canavas.C. 1986 : "Estimation of the Dynamic
Behavior of a Fixed-Bed Reactor Through Filte-
ring" . Precedings of IFAC Symposium on Dynamics
and Control of Chemical Reactors and Destina-
tion Columns. Brighton.
Chadeev,V.M.; NMaksimov,Y.M. 1986 : "An Adaptive
System with an Identifier for Control of the
Temperature Profile in a Tubular Reactor".
Precedings of IFAC Symposium on Dynamics and
Control of Chemical Reactors and Destination
Columns. Brighton.
Chylla,R.W.; Adomatis,R.A.; Cinar,A. 1986 : "De-
sign and Control of Tubular Autothermal Reac-
tors : An Evolutionary Approach". Precedings of
IFAC Symposium on Dynamics and Control of Chemi-
cal Reactors and Destination Columns. Brighton.
Cinar,A. 1984 :"Controller Design for a Tubular
Chemical Reactor". Can.J.Chem.Eng. 62 746-754
Clarke.D.W.; Gawthrop,Ρ.J. 1975 :"Self-Tuning
Controller" IEE Proc^ 122 929-934
Clement.Κ.;Jergensen,S.Β. and Serensen,J.Ρ.
1980 "Fixed-Bed reactor Kalman Filtering and
Optimnal Control-II: Experimental Investigation
of the Discrete Time Case with Stochastic
Disturbances".Chem.Engng.Sci.35,1231-1236.
Clement,K.;Jergensen,S.Β. 1981 :"Experimental
Investigation of a Fixed Bed Chemical Reactor
Control System Designed by the Direct Nyquist
Array Method" Proc.triennal Congr. IFAC 2743-50
Pergamon Press
Clement,K.;Jergensen,S.B. 1983 "Experimental
Investigation of Axial and Radial Thermal Dis-
persion in a Packed Bed".Chem.Engng.Sci.38,835-
842.
Clough,D.E.; Ramirez,W.F. 1976 : "Mathematical
Modelling and optimization of the Dehydrogena-
tion of Ethylbenzene to form Styrene". AIChE
Journal 22 1097-1105.
Crider,J.E.; Foss,A.S. 1966 :"Computational Stu-
dies of Transients in Packed Tubular Chemical
Reactors" AIChE J^ 12 514-522
Crider,J.E.; Foss,A.S. 1968 "An Analytic Solu-
tion for the Dynamics of a Packed Adiabatic
Chemical Reactor" AIChE Journal 14, 77-84.
Danckwerts, 1953 :"Continuous Flow Systems"
Chem.Eng.Sci. 2 1-13
Davison,E.J. 1966 "A method for simplifying linear
dynamic systems" IEEE Transac. Aut. Control AC
11 93-101
Denn,M.M. 1969 :"Chemical Reaction Engineering -
Optimization, Control, and Stability" I.&E.C.
61no.2, 46-50
Dixon,A.G. and Cresswell,D.L. 1986 ."Effective
Heat Transfer Parameters for Transient Packed-
Bed Models". AIChE Journal 32 ρ 809-819.
Doyle.J.C. 1982 :"Analysis of Feddback Systems
with Structured Uncertainties" IEE Proc.
129,D(6) 242-250
Economou.C.G.; Morari,M. 1986 : "Internal Model
Control. 6. Multiloop Design". Ind.Engng.Chem.
Process Pes.Dev. 25 411-419.
Eigenberger,G.; Butt,J.B. 1976 "A Modified Crank-
Nicolson Technique with Non-Equedistant space
steps" Chem.Eng. Sci. 31 681-691
Epple.U. 1986 : "Model Reduction for Nonlinear
Systems with Distributed Parameters". Precedings
of IFAC Symposium on Dynamics and Control of
Chemical Reactors and Destination Columns.
Brighton.
Fieguth.F.; Wicke,E. 1971 :"Der Ubergang vom zund
/Losch-Verhalten zu Stabilen Reaktionszustanden
bei einem Adiabatischen Rohrreaktor" Chem.Ing.
Techn. 43 604
Foss,A.S.; Edmunds,J.M.; Kouvaritakis,B. 1980 :
"Multivariable Control System for Two-Bed Reac-
tors by the Characteristic Locus Method" I.E.C.
Fundam. 19, 109-117
Froment,G.F. 1974 :"Fixed Bed Catalytic Reactors
- Technological and Fundamental Design Aspects"
Chemie Ing.Techn. 46 374-386
Garcia,C.E.; Morari,M. 1982 :"Internal Model Con-
trol 1. A Unifying Review and some New Results"
Ind.Eng.Chem.Process Pes.Dev. 21 308-323.
Garcia,CE.; Morari,M. 1985a,b : "Internal Model
Control 2. Design Procedures for Multivariable
Systems" and "3. Multivariable Control Law Com-
putation and Tuning Guidelines" Ind.Eng.Chem.
Process Pes.Dev. 24 472-484 and 484-494.
Garey,G.F.; Finlayson.B. 1975 "Orthogonal Colloca-
tion on Finite Elements"Chem.Eng.Sci. 30 587-596
Georgakis,G.; Aris,R.; Ammundson,N.R. 1977a,b,c :
"Studies in the Control of Tubular Reactors I:
General Considerations", "II: Stabilization by
Modal Control", and "III: Stabilization by
Observer Design". Chem.Engng.Science 32 1359-
1369, 1371-1379 and 1381-1387.
Gilles,D. 1974 :" Quasi-stationares Verhalten von
wandernden Brennzonen". Chem.Engng.Science 29
1211-1216.
Gilles,D.; Eigenberger,G. and Ruppel,W. 1978 ."Re-
laxation Oscillations in Chemical Reactors".
AIChE Journal 24 912-920.
 Fixed Bed Reactor Dynamics and Control
Grimble,M.J. 1981 :"A Control weighted minimum-
variance controller for non-minimum phase sy-
stems" Int.J.Contr. 33 751-762
Grimble,M.J. 1984 :"Implicit and Explit Self-
Tuning Controllers" Automatica 20 661-669
Goldschmidt,L.Β.; Hallager,L.; Jorgensen,S.Β.
1985 "Multivariable Fixed-Bed Reactor Adaptive
Setpoint Tracking".Proceedings of IFAC workshop
on Adaptive Control of Chemical Processes. ed.
H.B.Unbehauen Frankfurt oct. 21-22.
Goldschmidt,L.B.; Hallager,L.; Jorgensen,S.Β.
1986 : "Dynamics and Control of an Adiabatic
Fixed Bed Reactor with Varying catalyst
Activity". Preceedings of IFAC Symposium on
Dynamics and Control of Chemical Reactors and
Destination Columns. Brighton.
Hallager,L.; Jorgensen,S.B. 1981 "Experimental
Investigation of Self-tuning Control of a Gas
Phase Fixed Bed Ctalytic Reactor with Multiple
Inputs". Proc.triennal Congr. IFAC 2743-50
Pergamon Press
Hallager,L.; Jorgensen.S.B. 1983 "Adaptive
Control Of Chemical Engineering Processes" IFAC
workshop on Adaptive Systems in Control and
Signal Processing" San Francisco.
Hallager,L.; Goldschmidt,L.Β.; Jorgensen,S.Β.
1984 "Multivariable Adaptive Identification
and Control of a Distributed Chemical Reactor".
Large Scale Systems 6 323-336.
Hansen,K.W. 1973 "simulation of the transient
behaviour of a pilot plant fixed-bed reactor"
Chem.Engng.Sci. 28, 723-734
Hansen,K.W.; Jorgensen.S.B. 1974 "Experimental
Investigation of the Dynamics of a Catalytic
Fixed Bed Reactor".Adv.Chem.Ser.,133, 505-518
Hansen,K.W.; Jorgensen.S.B. 1976a,b "Dynamic
Modelling of a Gas Phase Catalytic Fixed Bed
Reactor -I: Experimental apparatus and determi-
nation of reaction kinetics" and "-II Results
from dynamic experiments":Chem.Engng.Sci.,31, 579
586 and 587-598.
Harris,T.J.; MacGregor,J.F.; Wright,J.D. 1980a :
"Self-tuning and Adaptive Controllers: An
Application to Catalytic Reactor Control".
Technometrics 22 153-164.
Harris.T.J.; MacGregor,J.F.; Wright,J.D. 1980b :
"Optimal Sensor Location with an Application to
a packed Bed Tubular reactor".AIChE J^ 26 910-
916
Hlavacek,V.; Votruba.J. 1974 :"Experimental Study
of Multiple Steady States in Adiabatic Catalytic
Systems" Adv.Chem.Ser. 133 545-558
Hlavacek,V.; Votruba,J. 1977 :"Steady state opera-
tion of Fixed-Bed Reactors and Monolithic Struc-
tures" in Chemical Reactor Theory ^ A Review ed.
L.Lapidus and N.R.Amundson 315-399 Prentice
Hall
Hlavacek,V.; van Rompay,P. 1981 :"Current Problems
of Multiplicity and Sensitivity of States in
Chemically Reacting Systems" Chem.Eng.Sci. 36
1587-1597
Hlavacek,V.; Puszynski.J. van Rompay,P. 1982 :
"Multiplicity and Propagating Fronts in Adiaba-
tic and Nonadiabatic Fixed-Bed Reactors" Am.
Chem.Soc.Symp.Ser. 196 89-96
Hofmann,H. 1979 :"Fortschritte bei der Modellie-
rung von Festbettreaktoren" Chem.Ing.Techn. 51
257-265
Hoiberg,J.A. ; Lyche.B.C: Foss.A.S. 1971 "Ex-
perimental Evaluation of Dynamic Models for a
Fixed-Bed Catalytic Reactor".AIChE J.,17, 1434-
1447
Hong,J.-C; Lee,H.H. 1985 : "Fixed bed with Cata-
lyst Deactivation: On-Line Estimation of Deacti-
vation, Control and Design".AIChE J^ 31 302-
311.
Jensen,K.F.; Ray,W.H. 1982 : "The Bifurcation
Behavior of Tubular Reactors" Chem.Engng.Science
37 199-222.
Joseph,B.; Brosilow,C.Β. 1978 :"Inferential Con-
trol of Processes: Part I. Steady state Analysis
and Design " and "Part II. The Structure and
Dynamics of Inferential Control Systems" AIChE
J. 24 485-492 and -509.
DCCR-c
23
Jorgensen.S.B.; Hansen,K.W. 1976 ."Dynamic Model-
ling of a Gas Phase Catalytic Fixed-Bed Reactor-
III: Identification Using Binary Pulse Sequence
Test Signals" Chem.Engng.Sci.,31, 473-479.
Jorgensen.S.B. ; Clement,K. 1977 .-"Experimental In-
vestigation of Reaction Zone Control in a Fixed
Bed Reactor" Chem.Eng. with Per Soltoft Danish
Technical Press, Copenhagen 69-85
Jorgensen,S.Β., Goldschmidt,L.G.; Clement,Κ.
1984 "A Sensorlocation Procedure for Chemical
Processes"Com Chem.Engng.8 195-204.
Jorgensen.S.B., Goldschmidt,L.G.; Hallager,L.
1985 -."Sparse Process Modelling for Robust Mul-
tivariable Adaptive Chemical Process Control"
Proc. 7'th IFAC symposium System Identification
and Parameter Estimation, York England
Jorgensen,S.Β., Bortolotto,G.Ε.; Kaas-Petersen,C.
1986 :"Bifurcation behaviour of a Recycle Fi-
xed-Bed Reactor"in preparation.
Jutan,A.; Tremblay,J.P.;MacGregor,J.F. and Wright,
J.D. 1977a "Multivariable control of a butane
hydrogenolysis reactor I: State Space Reactor
modelling".AIChE Journal 23 732-742. 742-750,
751-758.
Jutan.A.; MacGregor,J.F. and Wright, J.D. 1977b,c
"Multivariable control of a butane hydrogenoly-
sis reactor II: Data collection, Parameter Esti-
mation, and stochastic diaturbance Identifica-
tion" and "III: On-line Linear Quadratic Stocha-
stic Control Studies" AIChE Journal 23 742-
750, and 751-758.
Jutan,A.; MacGregor,J.F. and
Wright,J.D. 1984 "Multivariable computer control
of a butane hydrogenolysis reactor".AIChE
Journal 30 220-226.
Kershenbaum,L.; Lopez-Isunza,F. 1982 "Dynamic
Behavior of an Industrial Scale Fixed-Bed Cata-
lytic Ractor" 7'th Internat.Sym Chem.React.
Engng.Am.Chem.Soc. 196 pl09-120.
Koivo,H. 1980 :"A Multivariable Self-Tuning Con-
troller" Automatica 16 351-366.
Kozub,D, .J.'; MacGregor,J.F. ; Wright,J.D. 1986 :
"Multivariable Control of a Catalytic Tubular
Reactor Using Both Wiener-Hopf Controller Design
and Internal Model Controller Design Approaches"
Precedings of IFAC Symposium on Dynamics and
Control of Chemical Reactors and Destination
Columns. Brighton.
Kumar,S.; Seinfeld,J.H. 1978 "Optimal location of
Measurements in Tubular Reactors" Chem.Eng.Sci.
33 1507-16
Kuruoglu.N.; Ramirez,W.F.; Clough,D.E. 1981 : "
Distributed Parameter Estimation and Identifica-
tion for Systems with Fast and Slow Dynamics. A
Tubular, Fixed-Bed Reactor to Form Styrene".
Chem Engng.Science 36 1357-1364.
Lapidus,L. 1961 :"On the Dynamics of Chemical
Reactors". Chemical Progress Symposium Series 36
57 34-39
Lee.K.S.; Lee,W.-K. 1985 :"0n-lLine Optimizing
Control of a Nonadiabatic Fixed Bed Reactor"
AIChE J^ 31 667-675.
Lee.R.S.H.; Agnew,J.B. 1977a,b "Model Studies
for a Vinyl Chloride Tubular Reactor. 1.Steady-
State Behavior" and "2. Dynamic Behavior" Ind.
Engng.Process Pes.Dev. 16 490-494 and 495-
501.
Lehr,C.G.; Yurchak,S.; Kabel.R.L. 1968 : Response
of a Tubular Heterogeneous Catalytic Reactor to
a Step Increase in Flow Rate". AIChE Journal 14
627-636.
Litz,L. 1979a "Ordnungsreduktion Linearer Zustan-
dsmodelle durch Beibehaltung der dominanten
Einbewegungen" Regelungstechnik 27 80-86
Litz,L. 1979b "Praktische Ergebnisse mit einem
neuen modalen Verfahren zur Ordnungsreduktion"
Re gelungs technik 27 273-280
Lubeck,B. 1974 "Numerische Losungsmethode fur
nichtlineare partielle Modelgleichungen der che-
michen Verfahrungstechnik" Chem.Eng.Sci. 29
1320-28
MacGregor,J.F. and Wong,A.K.L. 1980 ."Multivariate
model Identification and Stochastic Control of a
Chemical Reactor" .Technometrics 22^ 453-464.
24 S. B. J0rgensen
MacGregor,J,F,; Penlidis,Α.; Hamielec,Α.Ε. 1983
"Control of Polymerization Reactors" fifth IFAC
PRP Conference, Amsterdam
Mandler,J.A.; Morari.M.; Seinfeld,J.H. 1986 :
"Control System Design for a Fixed-Bed
Methanation Reactor". Chem.Engng.Science 41
1577-1597.
McDermott,P.E.; Mellichamp,D.A. 1984 :"An Auto-
Pole -Placement Self-Tuning Controller" Int.J.
Cont. 40 1131-1147
McDermott,P.E.; Mellichamp,D.A.; Rinker,R.G. 1986
a,b : "Pole-Placement Control of a Fixed-Bed
Autothermal Reactor Part I: Single Variable
Control" and " Part II: Multivariable Control"
AIChE Journal 32 1004-1014, and 1015-1024.
McGreavy C. 1983 :"0n-Line Computer Control System
for Chemical Reaction Processes"Com Chem.Engng.
7 529-566
Mellefront,D.J.; Sargent,R.W.H. 1977 :"Optimal
measurement policies for control purposes"Int.J.
Control 26 595-602.
Michelsen,M.L. ,-Vakil,H.B. ,-Foss,A.S. 1973 "State
Space Formulation of Fixed-Bed Reactyor Dynamics"
Ind.Engng.Chem.Fund. 12,323-328.
Morari,M.; Stephanoupolos,G. 1980 :"Studies in the
Synthesis of Control Structures for Chemical
Processes. Part III: Optimal Selection of Secon-
dary Measurements in the Framework of State
Estimation in the Prersence of Persistent Un-
known Disturbances" AIChE JL 26 247-260
underwater,D.; MacGregor,J.F.; Wright,J.D. 1986
"Nonlinear Temperature Control of a Catalytic
Tubular reactor Using Self-Tuning Reactors".
Submitted to Can.J.Chem.Eng.
Padberg,G.; Wicke,E. 1967 :"Stabiles und insta-
biles Verhalten eines adiabaten Rohrreaktors am
Beispiel der katalytischen CO-Oxidation".Chem.
Engng.Science 22 1035-1051.
Padmanabhan,L.; Lapidus,L.:"Control of Chemical
Reactors" in Chemical Reactor Theory A Review"
ed.L.Lapidus and N.R.Amundson.
z
Prentice-Hall
314-399.
Prager,D.L.; Wellstead,P.Ε. 1980 :"Multivariable
Pole-Assignment Self-tuning Controllers" Proc.
IEE 128,D, 9-18
Puszynski,J.; Snita,D.; Hlavacek,V.; Hofmann.H
1981 : "A Revision of Multiplicity and Parame-
tric Sensitivity Concepts in Nonisothermal Non-
adiabatic Packed Bed Chemical Reactors" Chem.
Eng.Sci. 36 1605-1609
Puszynski,J.; Hlavacek,V. 1984 : "Experimental
Study of Ignition and Extinction Waves and Os-
cillatory Behavior of a Tubular Nonadiabatic
Fixed bed Reactor for the Oxidation of Carbon
Monoxide ". Çhe mJE ngng.Science 39 681-692
Ray.W.H. 1972 : "Fixed-Bed Reactors; Dynamics and
Control". Proc.Second Internat.Sym Chem.Reac.
Engng. A 8-1 to A 8-19. Elsevier Amsterdam
Rivera,D.E.; Morari,M. 1985 :"Internal Model Con-
trol perspectives on model reduction" Proc.Am.
Contr.Conf.,Boston MA 1293-1298.
Romagnoli,J.; Alvarez,J.; Stephanopoulos,G. 1981 :
"Variable measurement structures for process
control"Int.J. Control 33 269-289.
Root.R.B.; Schmitz.R.A. 1969 : "An Experimental
Study of Steady State Multiplicity in a Loop
Reactor". AIChE JOurnal 15 670-679
Root,R.B.; Schmitz,R.A. 1970 : "An Experimental
Study of Unstable States in a Loop Reactor".
AIChE JOurnal 16 356-358.
Schmitz,R.A. 1975 : "Multiplicity, Stability, and
Reacting
Sensitivity of States in Chemically
Systems- A Review" Chemical Rection Engine-
ering Reviews. ed. H.M.Hulbert. Adv.Chem.Ser.
148 156-211.
Schnelle,P.D.; Richards,J.D. 1986 :"A Review of
Industrial Control: Difficult problems and Wor-
kable Solutions" Procedings Chem.Process Control
III 749-802, ed.M.Morari & T.J.McAvoy, Elsevier
Seborg.D.; Edgar,T.F.; Shah,S.L. 1986 : "Adaptive
Control Strategies for Process Control: A Sur-
vey" AIChE J^ 32 881-913
Sharma.C.S. and Hughes,R. 1979a,b ."The behaviour
of an adiabatic fixed bed reactor for the oxida-
tion of carbon monoxide -I: General Parametric
Studies", and "-II: Effect of Pertubations "
Chem. Engng.Science 34 p613-624 and 625-634.
Silva,J.M.; Wallman,Ρ.H.; Foss,A.S. 1979 "multi-
bed Catalytic reactor Control Systems:
Configuration Development and Experimental
Testing".Ind.Engng.Fundam.,18, 383-391.
Sinai,J.; Foss,A.S. 1970 "Experimental and Com-
putational Studies of the Dynamics of a Fixed-
Bed Chemical Reactor".AIChE J. 16 658-669.
Serensen,J.P. 1976 "Experimental Investigation
of the dynamics of a fixed-bed reactor", chem.
Engng.Sci.,31 719-725
Serensen,J.P. 1976 "Experimental Investigation of
the optimal control of a fixed-bed reactor",
Chem.Engng.Sci.,32 763-774
Serensen,J.P.;Jergensen,S.Β.,; Clement,Κ. 1980
"Fixed-Bed Kalman Filtering and Optimal Control-
I: Computational comparison of Discrete vs
Continous Time Formulation" Chem.Engng.Sci. 35
1223-1230.
Stangeland,B.E.; Foss,A.S. 1970 "Control of a
Fixed-Bed Reactor".Ind.Engng.Chem.Fund.,9,38-48
Sundaresan,S.; Amundson,Ν.R.; Aris,R. 1980 ^'Ob-
servations on Fixed-Bed Dispersion Models: The
Role of the Interstitial fluid" AIChE J^ 26
529-536
Terndrup,H.; Clement,K.; Jergensen,S.Β. 1983
"Multivariable Fixed Bed Reactor Control using
Frequency Methods" Proc.IASTED Symp. Applied
Control and Identification p.4.12-4.17, Copenha-
gen
Tremblay,J.P. and Wrigth.J.D. 1974 "Computer Con-
trol of a Butane Hydrogenolysis Reactor". Can.
J.Chem.Engng. 52,845-47.
Tremblay,J.P. and Wrigth.J.D. 1977 "Multivariable
model reference adaptive control of a pilot
scale packed bed tubular reactor." Proceedings
IFAC symposium on Digital computer applications
to process control. North-holland publishing
company.
Vakil,H.B.; Michelsen.M.L.; Foss,A.S. 1973 "Fixed
bed reactor control with state estimation".Ind.
Engng.Chem.Fund. 12,328-335.
Van Doesburg,H.; De Jong,W.A. 1976a,b : "Tran-
sient Behaviour of an Adiabatic Fixed-Bed Metha-
nator I: Experiments with Binary Feeds of CO or
CO^ in Hydrogen" and "II: Metahanation of mix-
tures of Carbon Monoxide and Carbon Dioxide"
Chem.Eng.Sci. 31 45-51 and 53-58.
Villadsen,J.; Michelsen,M.L. 1978 Solution of
differential equation models by polynomial ap-
proximation. Prentice Hall, New Jersey
Vortmeyer,D.; Jahnel,W. 1972 :"Moving Reaction
Zones in Fixed Bed Reactors Under the Influence
of Various Parameters" Chem.Eng.Sci. 27 1485-
1496
Vortmeyer,D.; Winther R.P. 1982 :"Impact of Poro-
sity and Velocity Distribution on the Theoreti-
cal prediction of Fixed-Bed Chemical Reactor
Performance" Am.Chem.Soc. Symp.ser. 196 49-61
Wallman,P.Η., Silva,J.M.; Foss,A.S. 1979:"Multi-
variable Integral Controls for Fixed-Bed Reac-
tors" .IndJtagngJFunda^ 392-399.
Wallman,P.Η., Foss,A.S.1981 ."Experiences with
Dynamic Estimators for Fixed-Bed Reactors".
Ind.Engng.Chem.Fundam.20 234-239.
Wellstead,P.E.; Prager,D.; Zanker,Ρ.1979 ^Pole-
assignment Self-Tuning Regulator" Proc.IEE 126
781-789
Wicke,E.; Padberg,G.1968 :"Einfluss von Stoff-
und Warmetransport bei reaktionen in gas-
durchstromten Kornerschichten", Chem.Ing.Tech.
40, 1033-1038.
Wicke,E.; 0nken,H.U. 1986 : "Statistical Fluctua-
tions of Conversion and Temperature in an Adia-
batic Fixed-Bed Reactor for CO Oxidation". Chem.
.Science 41 1681-1687.
Wong,C.; Bonvin,D.; Rinker,R.G.; Mellichamp,D.
1983 " On Controlling an Autothermal Reactor
at an Unstable State-IV: Model Fitting and
Control of the Laborator Reaktor"
Chem.Engng.S c i enc e 38 619-633.
Zhao-Ying,Z.; Àstrom,K.J. 1981 :"A microprocessor
implementation of a LQG self-tuner" Department
of Automatic Control, Lund Institute of Techno-
logy, Sweden
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

DEVELOPMENTS IN DISTILLATION
EQUIPMENT AND SYSTEM
DESIGN - IMPLICATIONS FOR
DYNAMICS AND CONTROL
R. J. P. Brierley
Imperial Chemical Industries PLC, Engineering Department North East Group,
PO Box 6, Billingham, Cleveland, TS23 1LD, UK

Abstract. In the last decade or so trends in distillation equipment and

system design have been established which are likely to have significant
effects on dynamic behaviour and control requirements. For example packings
are now being used in preference to trays in a number of large columns with
high purity products. Also coupled columns with heat integration are often
considered as a means of energy saving. Examples are used to compare the
liquid inventories of packed and trayed columns to demonstrate the limited
applicability of vapour recompression.

Keywords. Chemical industry; column packings; controllability; distillation;

dynamic response; system design; trays.

INTRODUCTION ii) improved understanding of liquid distri-

The aim of this paper is to show the relevance
of research on distillation column dynamics
and control in two areas of distillation tech-
nology which are currently changing. The two
areas are (1) the use of packings (in which
the invention of structured packings has
played a significant role) and (2) develop-
ments in system design aimed at reducing the
energy costs of distillation. In various
ways, these trends involve the industrial app-
lication of long-standing technologies as much
as new inventions.
TRENDS IN COLUMN INTERNALS
butor design, leading to better designs and
giving increased confidence in scale-up (eg
Zanetti and others, 1985),
iii) large scale testing, and
iv) accumulated experience.
From the point of veiw of dynamic performance
there are two major differences between
packings and trays. The first is liquid
inventory, which is discussed further below.
The second is pressure drop - packings have one
fifth to one tenth of the pressure drop of
trays.

For many years trays have been the preferred Liquid Inventory - trayed and packed columns
type of internals for large scale industrial compared
distillation columns. Until the middle of
the 20th century, bubble cap trays were stan- To establish a simple basis for comparison, a
dard, but since then many alternative tray total reflux operation is considered. Appendix 1
devices have been invented. Two types now gives details including the calculated column
share the bulk of the business: sieve trays sizes and liquid inventories. The latter are
and valve trays. Packings were little used summarised in Table 1, together with the liquid
for several reasons: residence times based on liquid rate and total
inventory.
a) older types had poor capacity and poor
efficiency compared with well designed trays, TABLE 1 Column Inventories and Total
Residence Times
b) scale-up was unreliable, probably due to
inadequate liquid distributor design, and 3
Inventories Trayed Packed
(m liquid) column column
c) packings were expensive, especially when
compared with sieve and valve trays. Reflux drum 4 4
Trays 11
Over the last decade a combination of factors Packing 2
has led to the successful use of packings in Liquid distributors 1
large scale distillation duties: Column base 4 4
Total inventory 19 11
i) new packing designs, both structured
packings (eg Meier and co-workers, 1979) and
random packings (eg Strigle and Porter, 1979), Total residence time (min) 25 15

25
26 R. J. P. Brierley
The table shows that in this example, which
is reasonably typical, the liquid inventory
in a packed column system is little more than
half that in a trayed column system. Further-
more, looking specifically at that part of the
system in which mass transfer is occurring (ie
the trays and the packing), the contrast is
even more striking: the trays account for some
60% of the total inventory, whereas the packed
beds hold only 20% of the total inventory.
These numbers will vary considerably from case
to case, but the general message will remain.
Liquid inventory is the major capacitance
affecting the dynamic performance of distilla-
tion systems. In packed columns, the lags
resulting from those parts of the liquid inven-
tory which are not directly involved in the
mass transfer process are especially signifi-
cant. This refers to liquid in the column
base, reboiler, reflux drum, pipework, etc.
As a result, control systems for packed disti-
llation columns are often conservative. Typic-
ally, the heat supply to the column is set at
a fixed rate and such a column must be using
more energy than necessary in order to ensure
that product specifications are always met.
These lags can be reduced but factors other
than the design and performance of the control
system under normal operation must be consid-
ered. For example, is the reflux drum really
necessary? In normal operation it almost cer-
tainly is not, but problems can arise in upset
conditions, and start-up may well be more diff-
icult if that liquid inventory at distillate
composition is not available.
Hydraulics
Another area where trayed and packed columns
perform differently is in their hydrualic be-
haviour. The term "hydraulics" in this con-
text refers to the behaviour of the vapour and
liquid phases on the trays or in the packed
beds. It has already been mentioned that the
pressure drop of packings is much less than
that of trays. More relevant is the varia-
tion of liquid inventory with variations in
liquid and vapour rates. To a first approxi-
mation, in distillation, the liquid/vapour
ratio at a given point in a column will be con-
stant in normal operation.
In packed beds the process is vapour phase con-
tinuous with liquid streams flowing down the
packing surface. The liquid inventory falls
with reduced rate; at half design rates the
liquid inventory is likely to have fallen by
something like 30%.
On trays, the variation of liquid inventory
with rates is much more complicated than in
packed beds. At low rates the process is
liquid phase continuous - discrete vapour
bubbles pass through a relatively quiescent
liquid layer on the tray. At higher rates
the vapour does not pass through as discrete
bubbles, but creates a turbulent froth. At
still higher rates little liquid exists on the
tray floor; most of it is blown up into the
vapour space in the form of droplets. These
characteristic conditions are known as the
bubbly, mixed froth and spray regimes. Hofhuis
and Zuiderweg (1979) described a fourth
regime - emulsion flow. The regime obtained
depends not only on flowrates, but also on
tray design. At pressures around atmospheric
it is normal to have the mixed froth regime
with a spray in the vapour space above it.
Moving from the bubbly regime to mixed froth and
then to spray, the liquid inventory on the tray
decreases - especially between the last two.
Within a given regime it is possible - though
design correlations rarely predict it - that
liquid inventory varies inversely with rates.
TRENDS IN DISTILLATION SYSTEM DESIGN
Since the oil crisis of 1973-4, much thought has
been given to energy saving through the develop-
ment of sophisticated system design and analysis
techniques. Whatever happens to oil prices, it
is likely that these techniques will continue to
be used and will continue to produce designs
which are more sophisticated and energy effi-
cient than those of 15 years ago. These des-
igns have significant effects on the dynamic be-
haviour and control requirements of distillation
systems, especially when substantial amounts of
heat integration are used. Control may not nec-
cessarily be more difficult - it may be easier.
Several areas in which developments have occurred
or are possible are discussed below.
Sidestream Columns for Chemical Separations
Sidestream columns have been used in the oil in-
dustry for many years, but the separations requ-
ired there are relatively sloppy and are not
based upon a separation of identifiable chem-
icals, but on the achievement of effects in the
various products - eg vapour pressure, cloud
point, pour point. The outline flowsheet shown
in Fig.l is typical of many chemical processes -
reactor product passing through two columns, the
first to remove light impurities and the second
to remove heavy impurities. Depending on the
detailed process requirements, it is often poss-
ible to achieve both separations in one column,
by withdrawing the product as a sidestream.
It is not generally possible to get a high rec-
overy of the byproducts into the top and bottom
streams, but since they will be present in the
feed in small concentrations high recoveries are
often not necessary. In such circumstances,
considerable energy savings (as much as 50%) may
be achieved by using a sidestream column. But
how should it be controlled? With three product
streams there are in principle three composition
specifications to be met.
A feed stream which is to be split into two main
product streams, but with a small byproduct
stream of intermediate volatility, presents
another situation where a single sidestream col-
umn can be used. In this case the byproduct
sidestream will not be pure but will contain
significant proportions of the two main products.
But since the sidestream rate is small the loss
of products may well be justified by the savings
achieved with a single column.
Vapour Recompression
Vapour recompression (Fig.2) is a technique des-
igned to drastically reduce the energy consump-
tion in distillation by compressing the overhead
vapour so that it will condense at a higher temp-
erature than the column base. Thus the heat of
condensation may be recycled. This system,
which appears to be highly attractive at first
sight, has not been extensively used. The main
reason for this is that in many cases the
pressure ratio required for the compression of
the overhead vapour would result in an expensive
 Developments in Distillation and System Design
multi stage machine which makes the scheme
look unattractive.
Appendix 2 extends the example of Appendix 1,
assuming the separation of toluene from meta
xylene, to calculate the compression ratio
required for both a trayed column and a
packed column with structured packing. The
calculated values of compression ratio are 3.9
for the trayed column and 2.8 for the packed
column. Neither case looks promising, but
the benefit of packing is clearly demonstrated.
Even with no pressure drop and no temperature
difference in the condenser/reboiler, a com-
pression ratio of 2.1 is required. Therefore,
it can be seen that vapour recompression is
likely to be viable only with systems of low
relative volatility. Vapour recompression has
been applied to such systems in commercial
plants, for example ethylene/ethane splitters
on olefine units.
Heat Integration
Vapour recompression may be regarded as an ex-
ample of heat integration in which the integra-
tion is confined to a single column and its
ancillaries. Another form of heat integration,
which has been in use for many years and is
similarly confined to the column and its ancill-
aries, is feed/product heat interchange and the
use of condensation heat to preheat the feed.
The most extensive use of these techniques has
been in crude oil distillation.
Much attention has been given in recent years
to the heat integration of distillation columns
with the rest of the process (Linnhoff and co-
workers, 1983). Many chemical processes in-
volve not just one distillation step but sev-
eral. The heat requirements of distillation
are often large compared with those of the
rest of the process and the most obvious scope
for heat integration is in linking columns
together so that the condensation heat from
one column is used to reboil another. We all
know that distillation column control is not
easy - putting two columns in series is surely
asking for trouble. This has been the atti-
tude in the past to the linking of columns by
heat integration. Where such systems have
been used, various steps have been taken to
avoid undesirable interactions:
a) adjacent columns not linked unless with
intermediate storage,
b) either or both columns operated at con-
stant reboil rate,
c) a trim condenser installed in parallel
with the linking condenser/reboiler so that
disturbances in the higher temperature column
are not passed to the other, etc.
The use of linked columns for energy saving is
an important area which is currently con-
strained by an inadequate understanding of the
dynamic performance of such systems and how
they can best be controlled.
Complex Columns
A simple column is one with only two product
streams, distillate and residue. For a bi-
nary feed a simple column will in principle
achieve any desired separation of the two com-
ponents. This discussion of complex columns
27
is based on consideration of a feed mixture con-
taining three components. For a more detailed
discussion and evaluation, consult Doukas and
Luyben (1978).
The simplest complex column is a sidestream
column, discussed above. As was noted, the
simple sidestream column is of limited value
since it is not possible to produce three pure
products, and the applications described were
cases which do not have this requirement. Three
pure products can be achieved by the addition of
either a sidestream stripper (Fig,3) or a side-
stream rectifier (Fig,4). In both cases the
feed to the sidestream column contains only two
components. The sidestream column may have its
own condenser and reboiler but it is more usual
to omit the condenser (stripper) or the reboiler
(rectifier) as shown.
An alternative approach is to do a partial sepa-
ration in a prefractionation column - the
distillate being a binary mixture of components
A and B, while the residue is a binary mixture of
components Β and C. These may then be separated
in two further columns as shown in Fig.5. Clearly
these two columns may be combined. The pre-
fractionator may have its own reboiler and con-
denser or it may be directly linked to the main
column as shown in Fig.6. Although this may
show substantial benefits on the basis of a
steady-state design, control problems may render
it inoperable. Dynamic simulation seems to be
the only way to solve the control problems or to
demonstrate conclusively that they are unsolv-
able.
CONCLUSION
This paper has aimed to demonstrate that in at
least two areas of development in the industrial
application of distillation technology, current
developments in the simulation of column dynamics
and in the design of control systems could prove
very helpful and timely. Of course this will
hopefully be in addition to progress with the
many long-standing and well-known problems of
distillation system operation and control.
REFERENCES
Doukas, N. and W.L. Luyben (1978). Economics of
alternative distillation configurations for
the separation of ternary mixtures.
Ind. Eng. Chem. Process. Pes. Dev., 17,
272-281.
Fair, J.R. (1973). Gas-liquid contacting. In
R.H. Perry and C H . Chilton (Ed.), Chemical
Engineers Handbook, 5th ed. McGraw-Hill,
Kogakusha, Tokyo. Chap. 18, pp 3-25.
Hofhuis, P.A.M. and F.J. Zuiderweg (1979).
Sieve plates: dispersion density and flow
regimes. 3rd. International Symposium on
Pistillation. IChemE Symposium Series
No.56. Pp 2.2/1 - 2.2/26.
Linhoff, Β., H.A. Punford and R. Smith (1983).
Heat integration of distillation columns
into overall processes. Chemical Engineering
Science, 35, 1175-1188.
Meier, W., W.R. Hunkeler and Ρ Stocker (1979).
Performance of the new regular tower packing
"Mellapak". 3rd International Symposium on
Pistillation. IChemE Symposium Series
No.56. Pp 3.3/1 - 3.3/18.
Strigle, Jr, R.F. and K.E. Porter (1979). Metal
Intalox - a new distillation packing. 3rd.
International Symposium on Pistillation.
IChemE Symposium Series No.56. Pp 3.3/19 -
3.3/34
28 R. J. P. Brierley

Zanetti, R., H. Short and A. Hope (1985). APPENDIX 2. Vapour recompression

Newsfront: boosting tower performance
by more than a trickle. Chemical This appendix extends the example of Appendix 1,
Engineering, 92, May 27, 22-27. assuming the separation of toluene from
metaxylene and calculating the required pressure
ratios for the compression of the overhead
vapour.
APPENDIX 1. Liquid inventory

As a basis for comparison an atmosphericc Column Trayed Packed Notes

total reflux operation with toluene is Top pressure 0.56 bar 0.054 bar (i)
considered : Base pressure 1.56 bar 1.054 bar
Base temperature 156 °C 141 °C (ii)
Pressure 1 bar Condensing temper- 166 °C 151 °C (iii)
Temperature 383 Κ ature
Flow rates 10 kg/s Condensing pressure 3.9 bar 2.8 bar (iv)
Molecular weight 92 3 Pressure ratio 3.9 2.8
Liquid density 780 kg/m 3
Vapour density 3 kg/m
z
Liquid viscosity 0.25 mN s/m Notes :
Theoretical plates 45
i) Tray pressure drop = 8 mbar
For the author's convenience, proprietary cal- Packing pressure drop = 3 mbar/m
culation methods were used for column design, In both cases the pressure drop could be
but similar answers should be obtained with reduced to 30% of these figures with appro-
published methods (eg Fair, 1973). priate design changes.

ii) Bubble point of metaxylene at base pressure.

Trayed column design
iii) 10°C temperature driving force for
Column diameter 2 m condenser/reboiler.
Outlet weir height 50 mm
Clear liquid height on each tray 50 mm iv) Dew point of toluene at condensing
Tray efficiency 65 %
3 temperature.
Liquid on 70 trays 11 m

Packed column design

Two types of packing are considered: random

packing (50mm Pall rings) and a structured
packing of similar hydraulic capacity.

Packing Random Structured Notes

Column diameter 1.74 m 1.74 m (i)
HETP 0.65 m 0.4 m (ii)
Packed height 30 m 18 m
Liquid hold-up
3
3 vol%
3
5 vol%
2.1 m 2.1 m
No. of dist- 5 3
ributors 3 3
Tot liq on dist- 0.9 m 0.6 m (iii)
ributors

Notes :
i) Based upon a design F factor of 2.5

F = vapour velocity (m/s) χ vapour density (kg/m )

ii) HETP = height equivalent to a theoretical plate

iii) Assumes liquid in each distributor is

equivalent to 75mm across the total
column cross sectional area.

In this example, the liquid inventories of the

two types of packing (including distributors)
are essentially equal.

Column base and reflux drum

Design practice varies widely - for this

example, the residence time in each is assumed
to be 5 min, based on liquid rate. The res-
ultant liquid inventory is 4 m .
 Developments in Distillation and System Design 29

. Light Products
ends

Turbo blower

Reactor_
Product Feed-

Condenser/reboiler

^ Distillate
Heavy
"ends —•-Residue

Fig. 1 . Conventional column sequence Fig. 2. Vapour recompression system

Ο
Feed _
A + Β + C
Feed
A + Β + C

Fig. 3. Sidestream stripper system

V3>
Fig. 4. Sidestream rectifier system

A + Β

Feed
Feed ,
A + Β + C •Β A + Β + C

Β + C Β + C

Ο
Fig. 5. Prefractionator and two simple Fig. 6. Integrated prefractionator and
product columns sidestream product column
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

CONTROL OF POLYMERIZATION REACTORS

J. F. MacGregor
Department of Chemical Engineering, McMaster University, Hamilton,
Ontario, Canada

Abstract. For the control of polymerization processes, fundamental models play a very important role. This is due to a
number of reasons: the lack of available on- line sensors, the complexity of the polymerization process, and the nonlinear
operating space of batch and semi-batch reactors. Most early research on polymerization reactor control centered around
simulation and open-loop optimal control policies using these models. Recently, significant progress in modelling, sensor
development, and nonlinear control methods has opened the way for closed- loop control of some fundamental polymer
quality variables.

This paper will highlight some of the most important past theoretical and practical developments, and will discuss some
future directions for research.

Keywords. Optimal control; Nonlinear Control; Sensors; Modelling; Polymerization Reactors.

temperature (e.g. Hamer, 1983). In heterogeneous polymerization

INTRODUCTION
There are many aspects to the control of polymerization reactors.
Industrial emphasis has centered around the sequencing and
automation of batch and s e m i - batch r e a c t o r s , and the
temperature control of the reactors. In this paper I shall
concentrate exclusively on problems of controlling polymer
quality variables, and ignore the automation and temperature
control problems except in as far as they affect quality. No
attempt will be made to give a comprehensive review of the
literature, since excellent reviews have ben presented at past
IFAC conferences on Instrumentation and Automation in the
Paper, Plastics and Polymerization (PRP) Industries (Amhren,
1977; Hoogendoorn and Shaw, 1980; MacGregor et al., 1983).
Rather, reference to papers will be made only to illustrate certain
aspects of specific control problems
The paper is outlined as follows. Given the importance of
fundamental or mechanistic mathematical models in the control
of quality variables, a discussion of the nature of these models and
their limitations is first presented. Attention is focussed
exclusively on free radical and catalyzed ionic polymerizations. A
discussion of past work on open-loop optimal operating policies, is
followed by a discussion of on-line measurements and state
estimation. Finally, past work and future directions on the
closed-loop control of continuous and semi-batch reactors are
discussed.
MATHEMATICAL MODELS
(emulsion or solid catalyzed Ziegler- Natta) mass transfer rate
limitations between the phases may sometimes be important. In
ionic Ziegler-Natta catalyzed systems the major uncertainty lies
in trying to model the active site complexes of the catalyst system.
For example, in homogeneous Z-N systems such as the solution
polymerization of butadiene, there is as yet no agreement on the
number, nature, or role of the active catalyst species. Similarly, in
heterogeneous systems little knowledge exists about the activity,
or role of the active catalyst sites.
Most of the polymerization models reported in the literature are
deterministic in nature. They in no way account for the stochastic
variability that exists in industrial processes. For free radical and
ionic polymerizations the major sources of stochastic disturbances
are varying amounts of reactive impurities. Trace amounts of
these impurities are present in industrial feedstocks, and their
l e v e l s often build up in m o n o m e r recovery o p e r a t i o n s .
Fluctuations in these impurity levels are the major reason for the
batch-to-batch variations in polymerization rate and polymer
quality, and for the drifting conversion and quality observed in
continuous reactors. Therefore, in order to be able to relate
mechanistic model predictions to observed plant behaviour, (e.g.
through state estimation), or to be able to control polymer
production rate and quality using these models, it is important
that the effect of these reactive impurities be better understood,
and accounted for in some manner in the models. Recently some
studies on modelling the effects of monomer and water soluble
impurities in emulsion polymerization have been reported (Huo et
al. 1986; Penlidis et al., 1986).

Mathematical models for polymerization reactors consist of one or
more of the following set of balances in one form or another:
overall material balances for all the major components; molecular
balances for molecular weight development; population balances
for particle size development.
Material Balances
A set of differential equations accounting for the accumulation,
inflow, outflow, and disappearance by reaction of the monomers,
solvent, initiators and catalysts, chain transfer agents, and other
ingredients such as emulsifiers or stabilizers comprises the main
part of any model. The overall balance equations are usually easy
to write down in the case where spatial mass and heat transfer is
not a problem. The uncertainties usually lie in the description of
the reactor rate t e r m s . F r e e r a d i c a l , bulk or s o l u t i o n
polymerizations at low conversions are among the easiest to
model. At higher conversions or lower solvent concentrations
diffusion controlled macromolecular reactions (Trommsdorf or gel
effect) usually become important, and bulk mass and heat transfer
in the vessel lead to spatial variations in concentration and
Molecular Weight Development
There are a number of different approaches that have been taken
to develop expressions for the molecular weight development in
polymerizations (e.g. statistical methods, instantaneous property
balances, transform methods and the method of moments). From
the viewpoint of reactor control the most useful description is
usually a set of differential equations describing the low order
moments (Qi = Σ γ ϊ Ρ ; i = 0,1,2) of the MWD. From these moment
γ
states the number and weight average molecular weights are
easily calculated.
For free radical polymerizations in which either termination or
transfer reactions dominate the models appear to be reasonably
well developed (e.g. Hamielec and MacGregor, 1983). The major
uncertainities lie in obtaining reasonable e s t i m a t e s of the
termination rate parameters (including the gel effect parameters)
and the rate constraints for the transfer reactions (to modifier, to
polymer, to internal double bonds, etc.) Obtaining good estimates
of these parameters is far from straightforward. Large amounts of
data on molecular weights are required, and often the type of data
required (GPC, NMR) is very difficult to obtain due to the poor

31
32 J. F. MacGregor
precision of the measurements or to the insolubility of the
branched polymer, etc..
For Ziegler-Natta systems the understanding of the molecular
weight development process is much less well understood. Again
this involves a more detailed understanding of the nature of the
active catalyst sites, and of the effect of impurities on them than
currently exists. Diffusional effects in solid catalysed gas phase
reactions may also play a role in molecular weight development
(Laurence and Chiovetta, 1983).
Population Balances
In heterogeneous polymerizations such as emulsion systems
where the particle size distribution plays an important role in
both the final product quality, and the rate of reaction, population
balances accounting for the birth, death and growth of polymer
particles is an important part of the model. Min and Ray (1974)
and Penlidis et al., (1986) have summarized the population
balance and age distribution approaches. Major uncertainties lie
in modelling the particle nucleation and coagulation phenomena,
particularly in the presence of non-ionic stabilizers. Usually these
phenomena are so unpredictable and occur so rapidly that little
can be done to actively control them during a polymerization. The
best that can be done is to try to avoid having to deal with them by
using seeded emulsions, and an adequate amount of stabilizer.
OPEN-LOOP OPTIMAL POLICIES
Given the lack of availability of on-line sensors, and the lack of
well developed nonlinear feedback control methods, it is not
surprising that most early research on polymerization reactor
control centered around the simulation, and the development of
open-loop optimal control policies based on d e t e r m i n i s t i c
mathematical models. Batch, semi-batch and continuous
stirred-tank reactor policies which optimized some measure of
performance, subject to certain constraints, were u s u a l l y
developed through the use of the Maximum Principle. Common
objectives were to minimize batch or start up time subject to
producing polymer with a given number average molecular
weight or polydispersity. In some simple cases, experimental
verification was obtained. For the case of copolymer composition
control, a necessary condition for an optimal policy could be
expressed directly from an understanding of copolymerization
kinetics, thereby e l i m i n a t i n g the need for the M a x i m u m
Principle. Final selection among various possible constant
composition feed policies would then have to be based on other
considerations such as minimizing branching, minimizing batch
time, or safety considerations. A review of the literature is
presented in MacGregor et al. (1984).
The use of these deterministic models to examine the existence of
multiple steady-states and periodic solutions in continuous
reactors through bifurcation theory (e.g.Choi and Ray, 1985) has
also been treated. These studies may be useful for pointing to
potential operating problems in reactors. However, the results are
highly dependent upon the structure and parameters of the model.
The predicted bifurcations often occur well outside of the
conditions under which the model has been developed, and are
difficult to verify experimentally.
In general, simulation and optimization based on deterministic
polymerization reactor models are useful mainly for the improved
understanding they generate, and for comparing alternative
operating policies. Because of the uncertainties in these models,
particularly in the molecular weight part, and because the models
ignore the effect of reactive impurities, etc., the optimal policy
results are rarely directly applicable. This is discussed again later
under the section on feedback control.
ON-LINE MEASUREMENTS AND STATE
ESTIMATION
Even the off-line measurement of many polymer quality variables
in the laboratory presents a formidable task. Molecular weight
measurement methods for highly branched homopolymers, and for
co- and ter-polymers, etc., are still in the development stage.
Polymer structure properties (e.g. tacticity, sequence length
distribution, etc.) require sophisticated NMR techniques. The
analysis of particle structure or morpholgy in heterogeneous
polymerization requires meticulous sample preparation and the
use of electron microscopes. Even with the successful analysis of
such fundamental properties, we are still often left with the
difficult task of understanding how they relate to the end-use
performance of the product.
Therefore, on-line measurements are only feasible, at this time,
for certain simple fundamental properties, and for certain
surrogate variables which relate to some fundamental and
end-use properties. Most of the sophisticated analyses are not
feasible for on-line application because of the lack of instrument
robustness to the demands of the plant environment, and because
of the need for meticulous sample preparation.
On-line conversion measurements using densitometers (both
vibrating tube, and nuclear) have proven very successful in many
applications (Shork and Ray, 1981, Penlidis, 1986). On-line
energy balances around the reactor to track instantaneous
reaction rates, monomer conversion, and heat transfer coefficients
(Harris and Rushing, 1982; Wu, 1985; MacGregor, 1986) have also
become widely used. On-line particle size analyses for emulsion
polymerization based on turbidity and dynamic light scattering
methods are now being investigated (Gossen et al., 1986).
However, the measurement of surrogate variables such as melt
index, melt viscosity or pressure drops, etc. are probably the most
effective means of monitoring many industrial processes.
Given the scarcity of on-line measurements there would appear to
be a great incentive to use state estimation methods to infer the
values of other state variables in mathematical models of the
process.
Two other papers at this conference deal with Kalman Filtering
methods, one to track the material and population balance states
in emulsion polymerization using densitometers, etc. (MacGregor
et al., 1986) and the other to estimate the molecular weight
distribution using an on-line GPC (Papadopoulou, 1986). The
former paper also stresses the importance of modelling the effect
of impurities, and of incorporating nonstationary stochastic states
into the model of the process in order to obtain realistic state
estimators that are capable of tracking an industrial process.
Observability plays an important role in state estimation for
polymerization processes. In general, the molecular weight states
are unobservable from measurements on the material balance or
the particle population balance states of the model. Since on-line
measurements for the latter states are more readily available, this
means that tracking the molecular weight development of a
polymerization process is often very difficult.
FEEDBACK CONTROL
Continuous Processes
In high volume continuous polymerization processes, regulation
or disturbance rejection is a major priority, although start-up and
grade change policies can also be very important. Open-loop
start-up or change-over policies can be developed by the
optimization procedures discussed earlier, and approaches to
feedback control during these periods would be similar to those
discussed later for semi- batch reactors. Therefore in this section I
shall concentrate on the regulation of continuous processes about
given set-points.
Importance of Reactor Design: The importance of considering
reactor stability and control at the d e s i g n stage of the
polymerization process cannot be overstated. Operating and
control policies are often severely restricted by heat transfer
limitations in the process, by the lack of provision for surrogate
measurements throughout the process, by the inflexibility of the
purification and feed system, and by the poor dynamic behaviour
of the process. Although numerous industrial examples exist to
illustrate each of these situations, I will focus here on only one
 Control of Polymerization Reactors
examples which illustrates the importance of considering the
dynamic behaviour of the process at the design stage.
In the design of continuous reactors it is usual to consider only
steady state models, and then consider the dynamics and control
problems later. In the case of emulsion polymerization in CSTR
trains this approach can lead to disastrous consequences. It has
been found that the e m u l s i o n p o l y m e r i z a t i o n of m a n y
commercially important monomers in CSTR's leads not to a
constant steady state operation, but rather to a sustained
oscillation or limit cycle behaviour in all the latex and polymer
properties. Using dynamic models, verified through extensive
experimentation, Kiparassides et al. (1979, 1980) showed
conclusively that the reason for such behaviour lay in the
discontinous particle nucleation phenomena occurring in the early
reactors.
Pollock et al. (1982; 1983) used this fundamental understanding of
the oscillation phenomenon provided by the dynamic models to
show that the oscillations could not be eliminated through use of
control theory. However, they did show that a reactor train
redesigned to include the splitting of the monomer and water feeds
between a very small first reactor and the subsequent large
reactors would eliminate the oscillations. Furthermore, by
regulating the split of monomer and water between these reactors,
and the initiator flow to the first reactor, both the monomer
conversion and latex particle size could easily be controlled in a
stable, non-oscillatory manner. Figure 1 shows a schematic of this
split-feed reactor system and Figure 2 shows some conversion
results taken from a pilot plant on the continuous emulsion
polymerization of vinyl acetate (Penlidis, 1986). In part A a single
CSTR was employed and the oscillatory behaviour is clearly
evident. At point Β the small reactor and the split feed system
were switched in, and the oscillations immediately disappeared.
A change in the feed splits was then made to effect a conversion
change at point C. Hence, the major controller problem
(oscillations) was eliminated at the design stage, and a new, much
more flexible and stable system resulted.
Regulation of Cont: AUQUS Processes. The control of continuous
polymerization processes, where on-line measurements of state or
surrogate variables is available, or where inferred values are
available from state estimators, parallels closely the control of
most other continuous industrial processes. A large body of linear
univariate and multivariate control theory is available to handle
such problems. However, there are a number of things that are
common to many polymerization processes and should be
addressed in developing a control scheme. Some of these are
discussed below.
The major source of disturbances in ionic and free radical
polymerization processes is due to the presence of varying
amounts of reactive impurities in the feeds to the reactor. The
feed monomers themselves have varying impurity levels, and
consideration should always be given to the feasibility of
controlling some of this feedstock variability. More important, in
processes with recycle of recovered monomer, is the buildup of
impurities in the recovery and recycle system. In these processes
it is not uncommon to concentrate the control effort on the
recovery system. If, in this way, one can achieve control over the
impurities entering the reactor, the need for advanced control
around the reactor itself is greatly reduced.
Another characteristic of polymerization reactor control is the
need to deal with noisy and infrequent measurements. Some of
these measurements often come from off-line analyses performed
in a quality control laboratory. Given the difficulty in getting any
measurements of polymer quality, it is important that one make
use of all the available data. Examples of such off-line
measurements useful for process control are: Mooney viscosity
and solids measurements in the production of butadiene rubbers;
melt-index and density measurements in polyolefin production;
viscosity and fiber characteristics such as dye uptake, etc. in the
manufacture of synthetic fibers. A common industrial approach to
using such off-line data is to implement statistical quality control
procedures such as Stewart, CUSUM or EWMA charts, and take
corrective action only infrequently whenever a statistically
significant deviation from target is detected. Such procedures,
however, are appropriate only if process dynamic effects are not
33
33
important at the given sampling interval, and if there is a cost
associated with taking a corrective action. In other cases, it is
more appropriate to identify discrete empirical transfer function
and disturbance models (ARMAX models) relating these output
measurements to the process inputs, and then use minimum
variance of LQG-type stochastic controllers. Examples of this
approach are treated in MacGregor and Tidwell (1980), Rushing
(1986), Kelly et al. (1986).
A final point, that is apparent to most people working with
polymerization reactors, is that these processes are highly
interactive, and usually must be treated as true multivariable
systems. A change in any one of the input feedrates (e.g. initiator,
monomer, etc) will usually affect many of the output quality
variables. It is for this reason that so much effort has been spent
in efforts to develop mechanistic models of polymerization
processes.
Semi-Batch Reactors
Much of the polymer industry is based around semi-batch reactor
technology. In particular, most low volume and specialty
polymers are made in these reactors They allow the flexibility of
producing a wide variety of products within the same equipment.
Most of the optimal open-loop control policies discussed earlier
were aimed at developing temperature, and initiator or monomer
feedrate policies versus time which would produce polymers or
copolymers with specified molecular weight, composition, or
morphological properties in a minimum batch time. These
policies are all based on deterministic models, and precisely
known initial conditions. Unfortunately, implementation of these
policies as a preprogrammed trajectory over time poses serious
problems. First, the deterministic models are often poor,
particularly with respect to MW development, and the policies are
often very sensitive to mismatch between the process and its
model. Secondly, there are unmodelled, stochastic elements to the
process, such as the concentration of impurities in the initial
charge and in the subsequent semi-batch feeds, the activity or
active site distribution of Ziegler-Natta catalysts, and unexpected
temperature variations that occur during the batch cycle. Such
uncertainties in the initial conditions and the subsequent
operation of the reactor would lead to wide variability in the final
product, unless feedback from on-line measurements is used to
correct the trajectories. This area of nonlinear, multivariable
feedback control, aimed at compensating for the effect of
unmodelled dynamic and stochastic effects on optimal trajectory
control of semi- batch reactors, promises to be one of the most
interesting and potentially rewarding areas of control research
over the next several years.
A common approach to trajectory control employed in other areas
such as aerospace is to linearize the model about the precomputed
trajectory, and apply control adjustments based on linear control
theory to try to bring the reactor states back to their open-loop
optimal trajectories. In principle this approach is feasible. It
would require local linearizations, extended Kalman Filtering,
and finite horizon LQG controllers, for example. However, even
the objective of bringing the reactor back to a trajectory that is
based on a purely deterministic model or even based on past
experience is questionable, because in the presence of stochastic
variations that have caused the reactor states to deviate from this
trajectory, the precomputed trajectory will no longer be optimal.
On the other hand, to use general nonlinear programming
methods to recompute optimal control trajectories on-line at each
control interval would appear to be computationally impractical
given the highly nonlinear, multivariable nature of the models.
Suboptimal approaches to feedback control which make use of the
nonlinear mechanistic models plus a few on-line measurements
are currently an active area of research.
Approaches to handling these nonlinear semi-batch reactor
control problems through the use of traditional adaptive control
with linear models have been suggested (e.g. Kiparissides and
Shaw, 1983; Houston and Shork, 1986). However, in this author's
opinion, such approaches could have serious problems. The
parameters of these linear mdoels often have to adapt rapidly to
account for reaction nonlinearities. Such rapid parameter
34 J. F. MacGregor

variations in general linear models could lead to periods of
unrealistic model predictions, and possible i n s t a b i l i t i e s .
Questions about persistency of excitation of the input signals also
arise. Furthermore, these models make no use of existing process
knowledge about the direction of changes that one would expect,
for example, that temperature will increase with increased
reaction. Therefore, if adaptive control is to be used it would seem
that using simple mechanistic nonlinear models, and adapting a
few meaningful parameters would hold more promise.
SUMMARY
MacGregor, J.F. and Ρ W Tidwell (1980). Modelling and Control
of Continuous Industrial Polymerization Reactors, ACS Symp
Series. 124. 251-268.
MacGregor, J.F., A. Penlidis and A.E. Hamielec (1983). Control of
Polymerization Reactors, Proc. IFAC PRP-5 A u t o m a t i o n .
Antwerp, Belgium, Pergamon Press. Also in Poly. Proc. Eng.. 2,
179-206(1984).
MacGregor, J.F., D.J. Kozub, A Penlidis and A.E. Hamielec
(1986). State Estimation for Polymerization Reactors. Proc. IFAC
Symp. DYCORD-86, Bournemouth, U.K.

In this paper a review of some major difficulties inherent in
modelling and controlling polymerization processes has been
presented, some past work on reactor control has been discussed
and some directions for future research have been suggested. This
overview is by no means intended to be comprehensive. Rather, it
represents the perspective of the author based on his experiences
in this area.
Min, K.W. and W.H. Ray (1974). On the Mathematical Modelling
of Emulsion Polymerization Reactors. J. Macro. Sci. - Rev. Macro.
Chem..Cll. 177-255.
Papadopoulou, S. (1986). Continuous Estimation of Chain Length
Distribution in a Polymerization Reactor: Kalman Filtering
Considering GPC Measurements, Proc. IFAC Symp. DYCORD-86,
Bournemouth, U.K.

REFERENCES Penlidis, A. (1986). Ph.D. thesis, McMaster University, Hamilton,

Amrehn, H. (1977). Computer Control in the Polymerization
Industry, Automatica. 13, 533-545.
Choi, K.Y. and W.H. Ray (1985). The Dynamic Behaviour of
Fluidized Bed Reactors for Solid Catalysed Gas Phase Olefin
Polymerization. Chem. Eng. Sci., 40, 2261-2279.
Gossen, P.D., T. Kourti, A. Penlidis, J.F. MacGregor, and A.E.
Hamielec (1986). On-line P a r t i c l e Size and Conversion
Determination in Continuous Latex Reactor Trains. Proc. IFAC
PRP-6 Automation. Akron, Ohio, Pergamon Press.
Hamer, J.W. (1983). Ph.D. Thesis, University of Wisconsin,
Madison Wisconsin.
Ontario Canada.
Penlidis, Α., J.F. MacGregor and A.E. H a m i e l e c (1986).
Mathematical Modelling of Emulsion Polymerization Reactors: A
Population Balance Approach, J. Coatings Tech.. 58,49-60
Penlidis, Α., J.F. MacGregor and A.E. Hamielec (1986). Effect of
Impurities on Emulsion Polymerization: Case I Kinetics. J. Appl.
Poly. Sci. (to appear).
Pollock, M.J., J.F. MacGregor and A.E. Hamielec (1982).
Continuous Polyvinyl Acetate Emulsion Polymerization Reactors:
Dynamic Modelling of MW and Particle Size Development and
Application to Multiple Reactor System Design. ACS Symp.
Series. 197. T. Provder, ed., 209-222.

Hamielec, A.E. and J.F. MacGregor (1983). Modelling

Copolymerizations - Control of Chain Micrsostructure, Long Pollock, M.J. (1983). Ph.D. thesis, McMaster University,
Chain B r a n c h i n g , C r o s s l i n k i n g and M o l e c u l a r W e i g h t Hamilton, Ontario.
Distribution. Polymer Reaction Engineering. Berlin Int.
Workshop, K.H. Reichert and W. Geisler, editors, 21-71, Hanser Rushing, E.V. (1986). Control of Solvent Extraction from Acrylic
Publ. Munich. Fibers. Proc. IFAC PRP-6 Automation. Akron, Ohio.

Hoogendoorn, K. and R. Shaw (1980). Control of Polymerization Shork, F.J. and W.H. Ray (1981). On-line Monitoring of Emulsion
Processes, Proc. IFAC PRP-4 Automation. Ghent, Belgium, 623- Polymerization Reactor Dynamics, ACS Svmp. Series. 165. 505-
639, Pergamon Press. 514.

Huo, B.P., J.D. Campbell, A. Penlidis, J.F. MacGregor and A.E.

Hamielec (1986). Effect of I m p u r i t i e s on E m u l s i o n
Polymerization: Case II Kinetics. J. Appl. Poly. Sci. (to appear).

Houston, W.E. and F.J. Shork (1986). Adaptive Predictive Control

of a Semi-Batch Polymerization Reactor. Proc. Amer. Cont. Conf.,
Seattle, Oregon, June, 1986.

Kelly, S.J., J.F. MacGregor and T.W. Hoffman (1986). Control of a

Continuous Polybutadiene Reactor Train. Can. J. Chem. Eng.. (to
appear).

Kiparissides, C , J.F. MacGregor and A.E. Hamielec (1979).

Continuous Emulsion Polymerization: Modelling Oscillations in
Vinyl Acetate Polymerization. J. Appl. Poly. Sci.. 23,401-418.

Kiparissides, C , J.F. MacGregor and A.E. Hamielec (1980).

Continuous Emulsion Polymerization of Vinyl Acetate (3 parts).
Can. J. Chem. Eng.. 58,48-71.

Kiparissides, C. and S.R. Shaw (1983). Self-tuning and Stable

Adaptive Control of a Batch Polymerization Reactor. Automatica.
19,225-235.

Laurence, R.L. and M.G. Chiovetta (1983). Heat and Mass

Transfer During Olefin Polymerization from the Gas Phase,
Polymer Reaction Engineering. Berlin Int. Workshop, K.H.
Reichert and W. Geisler, editors.
 Control of Polymerization Reactors 35

MONOMER 70% - 9 0 %
WATER

RECYCLE

INITIATOR

SMALL
SEEDING
REACTOR

FIRST LARGE REACTOR

IN TRAIN

Figure 1 : Split-feed reactor system

Figure 2: Conversion results from split-feed reactor train

Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

DIAGNOSING AND PREVENTING TRAY

DAMAGE IN DISTILLATION COLUMNS
W. R. Ellingsen
Ε. I. du Pont de Nemours & Co (Inc), Engineering Department, Engineering
Service Division, Wilmington, Delaware, USA

Abstract. Little attention has been given to diagnosing and preventing tray damage in
distillation columns in the distillation control literature. However, tray damage can
impose a severe operating cost. When tray damage results in a complete shutdown or
limited production operation, especially of high throughput processes, the cost
penalty can be severe. Costs include both the direct expense to dismantle and repair
the columns and also the resulting loss in production. Based on the results of a
survey of 17 incidents of column tray damage, causes of tray damage are identified.
Along with discussing the cause and prevention of distillation column tray damage, the
problem of diagnosing the nature of column damage is discussed in some detail. This
is important since certain column behavior, such as foaming, can appear to be the same
behavior caused by tray damage and/or pluggage. Column instabilities due to flow
changes to and from the column and pressure disturbances can also appear to be caused
by tray damage. It is obviously very important to isolate the cause of the problem
since the solution can either be trivial or very costly. Also some recommendations
are given for column instruments and controls that can aid in stabilizing column oper-
ation and diagnosing the cause of the problems. "It can be very disappointing for a
control engineer who has spent considerable time and effort to devise and implement an
improved distillation column control strategy that might be saving $100M or more per
year, for that savings to be lost because during a start-up, upset, or shutdown, the
column trays were damaged."

Keywords. Distillation; distillation column; tray damage; distillation control;

column instability; foaming in distillation columns.

INTRODUCTION problems, and (2) damage caused by mechanical

In the past, internal damage has occurred in a
number of distillation columns in the Du Pont
Company resulting in repair costs and significant
process downtime. It is believed that abnormal
operating conditions cause most of the tray
damage. The objective of this study was to spec-
ify instrumentation, control strategies, and
operating procedures to minimize the probability
of damaging a column.
The study was initiated by conducting a survey of
column damage at several Du Pont plants. The
survey included a relatively small number of
incidents (17), but they probably represent the
general distribution of the types of column
damage that has occurred in the Du Pont Company,
as well as throughout the chemical industry.
This conclusion is also supported by the expe-
riences of two Du Pont distillation specialists,
Frank E. Rush and Noel G. O'Brien, Engineering
Department, Engineering Service Division,
Wilmington, Delaware.
failure. Of those in the category where damage
occurred because of operating problems, 72?ό were
related to a high liquid level in the bottom of
the column covering a number of trays. In 21% of
the cases, trays ruptured due to a vacuum that
was present in some sections of the column. In
1% (one case), the damage was caused by a high
vapor rate.
In the mechanical failure category, one incident
was due to fatigue failure (No. 8) that could
have been associated with virbration; and the
other two, Nos. 3, 6, and 8, were due to poor
installation of the trays. One incident, No. 4,
was due to corrosion.
Assuming this survey was a fair representation of
all tray damage, preventing tray damage associ-
ated with high liquid in the bottom of columns
provides an opportunity for reducing tray damage
caused by operating problems by 72% and reducing
total incidents by 76%

CAUSES OF TRAY DAMAGE - SURVEY OF

COLUMN TRAY DAMAGE INCIDENTS
The results are tabulated in Fig. 1. There are a
total of 17 cases, but 18 separate incidents.
(Cases 8 and 12 have both types of failure.)
Notice that there are two general causes of tray
damage: (1) damage resulting from operating
Most of the recommendations mentioned later deal
with the high liquid level in the bottom of the
column. The two most obvious recommendations for
the best protection against tray damage are:
1 ) do not turn on the steam to the column while
liquid is above the calandria, and 2) liquid
should be pumped out to a normal calandria level
prior to starting steam to the column.

37
38 W. R. Ellingsen

No. Diameter Operating High Poor

Case of of Pressure High Liquid Vapor Fatigue Tray
No. Trays Col, In. Vacuum in Bottom Rate Failure Installation Corrosion

1 X
2 X
3 30 X
4 72
5 20 96 X
6 144 X
7 20 72 X
8 30 72 X and X
9 20 X
10 30 120 X
11 30 120 X
12 96 X and X
13 20 X
14 60 90 X
15 17 X
16 20 X
17 20 X

Note: Case No,. 12 is counted twice,, but No. 8 once.

Breakdown in Causes of Damage

For Operating For Total No.

Error Only - 14 Total of Incidents - 18 Total

72% High Liquid in Bottom 56% High Liquid in Bottom

21% Vacuum 17% Vacuum
7% High Vapor Rate 5% High Vapor Rate
22% Corrosion and Installation

Fig. 1. Summary of Case Histories of Column Tray Damage - Cause of Damage

COLUMN TRAY DESIGN
The industrial tray design standard for all sizes
of distillation columns is for a minimum tray
differential pressure of 0.24 psi (6.6 in. of
water). Because of the high competition among
the tray manufacturers, most of them design for
the 0.24 psi loading. (A minimum tray spacing of
18 in. is also an industrial standard. This is
partly for ease of maintenance.) For small dia-
meter columns, the cost to increase the tray
strength is usually proportionally less; in fact,
the common materials used usually result in tray
strength greater than the minimum standard of
0.24 psi.
The differential pressure standard is for ver-
tical stressing in both directions. When the
column is operating, the vertical differential
pressure caused by the vapor flow partly counter-
acts the downward force of the tray liquid
loading.
Another general industry standard is for the gas
pressure drop across a tray to be on the order of
1 to 3 in. of water. Therefore, if the vapor
rate is providing 3 in. of upward force gas drop
across each tray and the liquid level on each
tray is 3 in. of water providing a downward
force, the net force on the tray would be zero.
Now consider the following two cases where the
trays are dry or flooded.
Dry Trays
If the vapor rate provides 3 in. of dry gas pres-
sure drop, then the upward force will be 3 in. of
water, which is half the design force. For trays
with fixed openings such as sieve and bubble cap
trays, doubling the vapor rate will increase the
pressure drop by a factor of four. This is
because the flow rate is approximately equal to a
constant times the square root of the pressure
drop.
For the valve trays, the resistance to flow
remains nearly constant until all valves are
completely open. Doubling the vapor flow might
only slightly increase the pressure drop. In
either case, if the flow is increased by the
square root of two (for the case of fixed opening
in the trays), the pressure drop will double from
say 3 to 6 in., which is still within the design
maximum load. For valve trays this is even more
conservative and, as mentioned, the flow can be
close to double. The vertical pressure drop will
remain close to 3 in., or 50% the maximum
designed vertical load.
Flooded Trays
At a given nominal gas rate providing an upward
force of 3 in. across the tray and, with the tray
flooded to 9 in., the net downward force is
6 in., which is near the maximum design load.
Since spacing between the trays is 18 in., if a
tray were to flood to 18 in., then there would be
a downward force of 15 in., which exceeds the
design load by more than double.
The importance of the above scenario is that no
matter what the cause of flooding or of a tempo-
rary draining of trays, especially in the case of
flooding, it is relatively easy to double the
designed force on a tray. Often when flooding
occurs and the column is generally upset, an
operator might greatly reduce the steam to the
column and, therefore, reduce the boilup rate.
 Diagnosing and Preventing Tray Damage
This would increase the downward hydraulic force
by the amount of reduction in gas pressure drop
due to the reduction in boilup. Thus, for a
column with 18 in. tray spacing, flooded 18 in.,
and the vapor rate reduced, the downward hydrau-
lic force could be close to three times greater
than the maximum design load for the tray.
What might be concluded from the above analysis
is that preventing flooding of the trays should
prevent most of the tray damage. This, however,
does not seem to be true since most of the dam-
age, especially when related to having a high
liquid level in the bottom of the column, is for
the trays to rupture in the upward direction.
However, not all of the column damage is being
examined carefully and, if done so, it might
reveal that some of the damage is being caused by
downward rupturing of the trays. Only about 3 of
the 17 incidents surveyed appeared to be asso-
ciated with downward rupturing of the trays. It
seems prudent, however, in any case, to prevent
flooding of the trays. (Note that in the above
analysis, the hydraulic pressure would be less in
proportion to the degree of gasification on a
tray. For example, if 50% of the volume were
bubbles, then the liquid would be 50% less and so
the hydraulic loading would be 50% less.)
Cause of Tray Damage When the Bottom of the
Column Is Flooded With Liquid
In most cases of column damage, where the steam
is added to the column with the bottom trays
flooded with liquid, it cannot be reasoned that
the damage was due to the high vapor rate even if
it were two-phase flow. Therefore, there prob-
ably are vertical hydraulic forces, such as the
water hammer effect, that cause the damage. It
appears that these damages would be especially
dominant during the very initial introduction of
steam to the column. Therefore, boiling up very
slowly to begin with, until vapor equilibrium is
established in the column (in other words until
vapor begins condensing at the top of the
column), should reduce the chance of damage
occurring.
DIAGNOSING COLUMN TRAY DAMAGE
Some Causes of Column Misbehavior
1. Tray damage.
2. Fouling or plugging.
3. Foaming.
4. Measurement errors in key column variables,
such as pressure, temperature, level, flows, and
concentration.
5. Instability caused by improper controller
tuning and/or poor control strategy (such as not
providing sufficient control rangeability and/or
decoupling of variables that interact).
6. Instability caused by uncontrolled external
variables such as feed rate changes to the
column, feed composition changes, and variations
in steam quality and/or steam supply pressure
(with no steam flow controller), and others,
depending on the nature of the separation and
equipment being used.
Determining Cause of Column Misbehavior
Items 1 to 4 are manifested by poor separation
and can cause instability.
DCCR-D
39
Items 5 and 6 will cause instability but probably
exhibit adequate average separation.
Poor Separation. If the separation becomes
poorer during normal operating conditions and
if the column is stable, then tray damage or
an error in one or more measurements could be
the cause. For example, if the steam flow-
meter erroneously reads high, then the boilup
is lower than it appears. Also, if the reflux
flowmeter erroneously reads high, then there
would be less reflux. (Note that an unknown
change in feed composition would also appear
as a change in separation under normal column
operating conditions.) Poor separation can
also be caused by fouling or plugging. Foam-
ing may cause a poor separation because of
entrainment. However, usually instability is
also apparent. Instability, if minor cycling
is occurring in the key variables, usually
will not result in poor separation. On the
other hand, large instability such as dumping
of the column affects the separation.
By maintaining an operating performance
history, any suspicion of poor separation
could be verified by comparing current
operating parameters with a similiar or
identical set of earlier operating conditions.
A mathematical model previously matched to a
normal operating column could also be used in
comparing with current operating conditions,
to determine if the separation has degener-
ated.
Instability. Minor instability, such as con-
tinuous cycling of one or more variables, is
most likely due to improper controller tuning.
Continuous cycling can also be caused by a
continuous cycle or a disturbance in a primary
variable such as the column feed rate, pres-
sure, and others. Severe cycling, especially
when the period of the cycle is greater than
the natural period of any of the key column
variables, can be due to tray damage, fouling
or plugging of the trays, or foaming.
Foaming. Isolating foaming as the cause of
column misbehavior from other causes is very
important since the solution to the problems
are quite different. It is very costly and
frustrating to shut down a column and examine
it for tray damage and find none.
It can be very difficult to isolate column
behavior caused by foaming from those caused
by plugging or tray damage. However, there
are some guides to the cause of foaming which
are listed as follows:
1. Particulates in liquid on trays can cause
foaming. (This was observed in a solvent
column where water and solvent were being
separated. A small amount of sodium chloride
impurity was precipitating and causing foam-
ing. An antifoam agent was added, and it
eliminated the problem.)
2. Two liquid phases where it becomes a single
phase. Foaming may occur in the region of the
column where this occurs.
3. Cannot have foaming with two phases always
present. When two phases are present, the
phase with the lower surface tension acts as
an antifoaming agent.
4. When separating two components, where the
component being removed from the bottom of the
40 W. R. Ellingsen
column has a lower surface tension than the
component removed from the top, foaming will
not occur unless (1) or (2) exist. This is"
called the Marangoni effect. An example of
this is the separation of Dimethylacetimide
(DMAc) and water. This cannot foam because
water being removed from the top of the column
has a higher surface tension than DMAc being
removed from the bottom and vice versa.
Foaming can usually be eliminated by adding
antifoaming agents in the 10-100 parts per
million range, which is usually acceptable.
Effect of Foaming in Columns
1. Sometimes foaming causes a reduction in
separation due to reduced tray efficiency.
This is caused by foam entrainment from tray
to tray.
2. Foaming can cause tray flooding by reducing
the density of liquid in the downcomer by slow
disengagement of bubbling in the downcomer.
3. Foaming on a tray will produce a dam or a
partial blocking of the entrance to the down-
corner. (This has been observed in test
installations.)
GENERAL APPROACH TO TROUBLESHOOTING
Although troubleshooting a distillation column
has to be tailored somewhat to each column, there
are three general levels of problem isolation
that can be used on any column. These are:
1. Assure that the measurement instrumentation
and control loops are functioning properly.
2. In the case of column instability, determine
if disturbances are being introduced into the
column externally such as by variations and feed
rate, reflux rate, steam rate, reflux temperature
variations, and other external variables.
3. If the items mentioned in 1 and 2 are not the
cause of the column problem (instability or poor
separation), then the source of the problem is
internal. The troubleshooting can now focus on
whether the instability or poor separation is due
to plugging, foaming, or tray damage.
Aids in Diagnosing Column Tray Damage
and Other Problems
Aids used in diagnosing column problems can also
be used in preventing tray damage. Probably the
most significant is to measure differential
pressure and temperature across sections of the
column. The ideal situation would be to measure
the pressure and temperature and composition on
each tray. A compromise, in order of priority,
would be to measure the pressure drop across
sections of the column, say 10% or at least 25%
sections. This will help locate the region in
the column where a problem has developed. A
temperature profile consisting of at least five
temperature measurements up the column would
greatly enhance the analysis of the state of the
column. The ability to draw liquid (or vapor)
samples from sections in the column for analysis
could greatly enhance diagnosing problems because
the degree of separation would be known from an
analysis of the samples.
Providing a liquid level measurement for the
bottom 25% of the column (differential pressure)
would aid the operator not only in verifying that
liquid is up in the column (in other words, veri-
fying that the calandria level measurement is
correct), but aid him in deciding how to elimi-
nate the liquid. If the liquid level is above
one or two trays, the operator might elect to
very slowly boil that liquid out. Whereas, if
the liquid level is up 25% of the column, then
the operator would elect to pump the liquid out
before introducing steam to the column.
Obtaining a Column Density Profile
Scanning a column empty and during normal and
abnormal operating conditions with a nuclear
radiation scanning system can isolate column
problems. This approach cannot be used in
preventing tray damage, but is very useful in
diagnosing tray damage and other column prob-
lems by showing the liquid mass profile in the
column. However, installing the instrumenta-
tion mentioned above would, in most cases,
help in preventing the tray damage, as well as
in diagnosing the cause of tray damage and
other column problems.
PREVENTING COLUMN TRAY DAMAGE
The incentive to preventing distillation column
tray damage has to be placed in an economic
prospective. If we want to spend the money, we
can design distillation columns that will never
be damaged under the most severe operating condi-
tions. Thus, a tradeoff exists among reducing
capital investment (which increases the proba-
bility of tray damage), the attendant cost of
repair, and loss of production due to downtime.
A related question to be asked is, "Can we use
the industry minimum design standard column con-
struction, but eliminate tray damage by improved
instrumentation and control?" The cost tradeoff
becomes minimizing capital investment in column
construction, but adding some capital for addi-
tional instrumentation and control. (Note that
enhanced instrumentation and control could lead
to more economical operation of a column under
normal operating conditions.)
The answer to simultaneously minimizing capital
investment cost and operating cost is not simple.
The answer for an existing column versus a new
column would be different. Based on the analysis
of the 17 incidents of column damage and the
apparent causes of the damage, some recommenda-
tions that can be used as a guide in reducing the
probability of column tray damage are given
below.
Liquid Above the Bottom Tray
With a column that has liquid above the bottom
tray, the best procedure is to pump the liquid
out until the calandria level is normal; then
proceed to turn on the steam to bring the column
to normal operating conditions. This would
eliminate what appears to be over 70% of the
source of distillation column tray damage. Since
this will not always be practical, a compromise
procedure would be to:
1. Ramp the steam flow rate to the column to
approximately 25% of the design rate over 30 min-
utes to one hour.
2. Hold at that rate until it is obvious that
vapor is being condensed at the top of the
column.
3. Ramp to the minimum design rate over a period
of 30 minutes to one hour.
 Diagnosing and Preventing Tray Damage
4. Hold at that rate until the liquid has
reached a normal calandria level. During the
second phase, the reflux should be set to the
minimum acceptable rate until the liquid in the
column reaches the normal calandria level.
Bottom Tray Liquid Seal
Construct the bottom tray seal to be especially
strong. This is a low-cost item, but if the
bottom seal is broken, especially on trays with-
out inlet weirs, vapor may bypass a large number
of trays by going up the downcomers. A partial
bypass of vapor up a damaged bottom seal may
cause periodic flooding and dumping of the above
tray which might propagate up a number of trays.
Construction of Bottom Trays
For columns prone to tray damage, construct the
bottom 25% of trays to be stronger than the above
trays. The design standard is approximately
6 in. of water differential pressure, which, if
it were increased to 18 in., would be consider-
ably more capable of handling stresses produced
by high vapor velocity or the liquid flooding
load. Since the usual trace spacing is 18 in., a
completely flooded tray with no counter vapor
force would not rupture the tray in the downward
direction.
Measuring Liquid Level in the Column
A liquid level differential pressure measurement
should be provided for the bottom 25% of the
column. This would give the operator critical
information needed as to the degree of flooding
in the column and help him in deciding the best
method to use in reducing the liquid level. For
example, if the liquid were only above the bottom
tray, a reasonable approach would be to ramp the
steam to the column as the means for reducing the
liquid level. On the other hand, if the liquid
is up 25% in the column, the course of action
would be to turn off the steam to the column (if
not already off) and to pump the liquid out of
the column to the normal base level.
Pumping Out Excessive Liquid
Provision should be made for easy diversion of
bottom liquid to either the feed supply tank or
to other temporary storage, so that the operator
will freely use this method of reducing the
liquid in the column to its normal level. This
diversion or pump-out system flow should be
actuated from the control room.
Flow Control of Steam to Column
The steam to a column should be automatically
flow controlled. The pressure of steam upstream
(of the flow measuring device) should be pressure
controlled so that a constant pounds-per-hour of
steam will be sent to the column. The set point
to the steam flow controller should have a ramp
function so that when the operator increases the
steam rate to say 10%, it will increase to 10%
over a given period of time. The ramp rate
should be adjustable. A set point rate limiter
should also be provided if the controller is used
in a cascade control loop. This will prevent an
automatic controller, such as a temperature
controller or level controller, from rapidly
increasing the steam rate. If the equipment is
available, the rate of change of steam flow
should be measured and alarmed. A bypass should
be available on the ramping function so that the
operator can initiate an emergency shutdown of
steam. The maximum valve opening to correspond
41
to the maximum vapor flow design rate should be
limited directly on the control valve. In
addition, the maximum flow set point on the steam
flow controller should be limited.
It should be noted that tray damage, in most
incidents in the survey, occurred when the steam
flow to the column was within normal or reduced
rate. This observation shows that limiting the
steam flow rate to the column is not sufficient
protection in preventing tray damage in a column
with liquid up in the lower tray section.
Inlet Weir
An inlet weir should be provided on the bottom
25% of the trays to prevent vapor bypass up
downcomers and possible flooding and dumping
propagation should one of the bottom trays be
damaged. Inlet weir are necessary even for valve
trays since they do not completely seal. It is
especially important for sieve trays. The bubble
cap trays do not need inlet weirs since they
retain liquid on the tray.
Fatigue Failure Due to Tray Vibration
Fatigue failure due to tray vibration usually
occurs at low vapor rates (near the dumping rate)
at pulses of 1 to 50 Hz. This does not seem to
be a significant problem. Educating the oper-
ating personnel as to the nature of the problem
should be adequate. If needed, instrumentation
can be installed for analyzing frequencies of
vibration.
Additional Measurements for Analyzing
Column Problems
Install five or more sample taps and thermowells
at various locations on the column. The loca-
tions should be determined by the temperature or
composition profile. The sample taps should be
capable of being used to both measure pressure
and to sample the liquid on the tray. The tem-
perature profile should be measured and at least
be available on demand.
Depending on the particular situation, the
sectional differential pressures can be measured
continuously and recorded in the control room; or
when a problem occurs, that a portable differen-
tial pressure instrument can be used on a one-
time basis to measure a sectional differential
pressure. Also, a liquid sample could be
obtained and analyzed for use in diagnosing an
apparent column problem.
Nuclear Radiation Scan
On critical columns, especially the large-
diameter columns (greater than 5 ft), a nuclear
radiation scan to obtain the density profile up
the column would be useful in diagnosing problems
such as tray damage, pluggage, or foaming.
Ideally the column should be scanned empty and
again scanned when operating under normal condi-
tions. Then, if and when a problem occurs, an
additional scan can be run to help isolate the
nature and region of the problem. It is
important that the problem be defined because of
the large cost of a shutdown.
The most frustrating problem is that of foaming
in the column. This can easily be interpreted as
tray damage or plugging which would call for a
shutdown. However, the solution to foaming is to
add an antifoaming agent or eliminate the cause
of the foaming.
42 W. R. Ellingsen
CONCLUSIONS
The survey indicated that about 12% of the inci-
dents of tray damage that are operational based
(as distinguished from mechanical failure or
corrosion), are related to having calandria
liquid level that extends above the bottom trays
in the columns. High liquid level in the bottom
of a column is an easily recognized condition,
and can be remedied if dealt with carefully.
A careful diagnosis must be made to determine the
cause of column misbehavior since some causes,
such as control loop instability, foaming, and
others, may exhibit the same phenomenon as caused
by tray damage. Installing some additional
column instrumentation will aid in diagnosing
problems but, in addition, will provide useful
operating information.
ACKNOWLEDGMENTS
Frank E. Rush, Du Pont Company, Wilmington,
Delaware, was especially helpful in providing
cause and effect information on foaming in
columns. He also reviewed the paper and made
valuable suggestions. Noel G. O'Brien, Du Pont
Company, Wilmington, Delaware, provided infor-
mation on column construction and references on
tray failure due to vibration and other problems.
REFERENCES
O'Brien, N. G., and Curran, B. F., Du Pont
Company, Wilmington, Delaware, Engineering
Department Report No. 15576, January 1978,
"Mechanical Failures of Seive Trays - Survey
of Tray Vendor Experience".
Morehead, P., "Pulsation Studies", Glitsch, Inc.,
November 1980, Report No. 1-80.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

A DISTILLATION PLANT WITH AN

INDIRECT HEAT PUMP FOR EXPERIMENTAL
STUDIES OF OPERATION FORM, DYNAMICS
AND CONTROL
+
L. Hallager , B. Toftegârd, K. Clement, and S. B. j0rgensen*
Instituttet for Kemiteknik, The Technical University of Denmark,
DK-2800 Lyngby, Denmark

ABSTRACT. A 19 sieve tray pilot plant distillation column has been designed and
erected at Instituttet for Kemiteknik, The Technical University of Denmark. The
plant utilizes a heat pump to recirculate the energy from the condenser to the
reboiler. The column is equipped with exchangeable trays, in order to be run in
either a conventional continuous or a periodic cycling mode.

The purpose of the process system is to perform: Control studies of conventional

continuous distillation and of periodic cycling distillation. Comparative studies
of the two operation forms with respect to separation efficiency and energy effi-
ciency .

Two types of distillation separations will be investigated: First a binary distil-

lation using methanol and isopropanol, with the primary purpose of evaluating the
properties of the separation process, using a thermodynamically simple system.
Later an azeotropic distillation separating isopropanol and water using toluen as
entraîner. This system is used as an example of a thermodynamically more demanding
distillation.

In this paper the the column and heatpump set-up is described and results from
steady state and transient experiments are shown and discussed.

INTRODUCTION Advanced control.

Distillation processes are responsible for appro-
ximately 50% of the capital investment in chemical
plants and for about 3% of total energy consump-
tion in USA in the mid seventies. Especially the
latter number has been reduced somewhat through
the energy conserving efforts of the last decade.
However the order of magnitude of these numbers
are probably still representative for the western
world. Consequently there is a significant incen-
tive to reduce the capital investment and to im-
prove the energy efficiency of this expensive
separation process.
One or both of these goals may be at least
achieved by two means :
The control of continuous distillation is a sub-
ject which is extensively covered in the littéra-
ture, see e.g. Waller (1982) and To H i v e r (1980).
However, a number of aspects of the present pro-
ject to our knowledge has been covered only to a
limited extent in the open littérature. These
aspects include:
ο Control of columns with energy feed-back from
top to bottom.
ο Control of columns which, in order to save
energy, operate close to minimum reflux.
partly
ο Control of azeotropic columns.

ο improving the process control , taking the
inherent multivariable character into account.
ο improving the efficiency of the procès itself.
The primary purpose of the present project is to
make theoretical as well as experimental research
into each of these areas and into their combined
effect.
The purpose of this paper is to provide a presen-
tation of the equipment along with experimental
results showing the behaviour of the column and
the energy recycle loop, and the interaction be-
tween these.
+ present address: CELL Automation A/S
Djursvang 7Β
DK-2620 Albertslund
* to whom correspondence should be addressed.
Due to the multivariable nature of the distilla-
tion process, and the couplings introduced by the
energy feedback, single loop controllers will not
be able to give satisfactory disturbance rejection
or setpoint tracking. Rather a multivariable
scheme must be employed and it is planned to
investigate multivariable adaptive control, the
design of which only requires qualitative know-
ledge of the process dynamics. Other types of
multivariable control schems will also be tested
once the necessary process model is available.
Periodic cycled operation.
The conventional method of distillation is based
upon the continuous flow of vapour and liquid. In
plate columns this leads to a considerable degree
of effective backmixing and thereby to reduced
separation efficiency, Lewis (1936). One method
to overcome this problem, and for which the pre-
sent equipment is prepared, is to operate the

43
44 L. Hallager et al.

column periodically. Periodic cycled distillation

was first introduced by Cannon (1961). In perio-
dic cycled operation the flows are varied periodi-
cally. The cycle period can be divided in a vapour
flow part (VFP) and a liquid flow part (LFP). In
the VFP vapour is introduced into the column and
the liquid is retained on the plates by the vapour
flow. In the LFP the liquid is flowing to the
tray below, preferably without any mixing.

Other periodic cycled operation forms have been

proposed e.g. by Baron et.al. (1980). By defining
a column efficiency as the necessary number of
trays for conventional distillation divided by
those for periodic cycled distillation for same
separation and point efficiency it can be shown
theoretically that the column efficiency can be
doubled. (McWhirter and Lloyd, 1963).

PROCESS EQUIPMENT
The three parts of the plant: column section, heat
pump section and tank park are shown schematically
in Fig. 1 to 3. The main dimensions are given in
tables 1 to 3.
Column section.
The column section is constructed in stainless
steel (18-8) to reduce corrosion problems when the
system contains water. This section is insulated.
During binary operation feed is pumped from two of
the tanks WA, MET, ISOP, BOT and TOP by means of
the pumps PF1 and PF2, through the preheater
(HEFS) and into the column (COL). The reboiler
produces a vapour flow and some liquid may be
removed as bottom product through the water cooled
heat exchanger (HEBW) and into the bottom product
tank (BOT). The vapour is condensed at the top of
the column in HECOND, pumped to the accumulator
(DEC) by PC. The pump PT then returns reflux to
the column and top product through the water co-
oler HETW to the top product tank (TOP). A small
fraction of the top vapour flows to the effluent
condenser (EFFCOND) where inerts are removed to
the venting system.
Fig. 2. Heat pump section
Compressor section.
The compressor section has been delivered and
installed by Brd. GRAM A/S. Two 8 cylinders piston
flow compressors are installed. Freon 114 is the
medium utilized in the heat pump. This equipment
is designed to run between approximately 55 and
105 deg.C corresponding to app. 5.5 and 16 bar
abs. The tubings and vessels of the compressor
section have been insulated before start up, in
order to prevent the potential problem of liquid-
faction in the compressors. The heat pump is
essentially a standard design for this type of
equipment. The compressors (COMP) compress the R-
114 vapour. Most of the vapour is condensed in the
reboiler (HERB). The extra condenser (HESCOND)
condenses an amount corresponding to the heat
introduced by the compressors. The liquid passes
on to the receiver (REC) and through a heat ex-
changer (HECI) to the freon evaporator (HECOND).
GLSEP acts as a demister, preventing liquid to
pass on to the compressors and also enabling HE-
COND to operate as a thermosiphon reboiler to

In the ternary case the entraîner is pumped to the

decanter (DEC) by PE. In the decanter two liquid
phases are formed. The light organic phase is
returned to the column by the pump PR, and the
heavy watery phase is by the pump PT mainly retur-
ned to the column. The remaining fraction is
pumped to TOP as top product.

Fig. 1. Column section. Fig. 3. Tank park and pump station.

 Distillation Plant with an Indirect Heat Pump 45

improve heat transfer. In HECI the vapour is

superheated before it enters the compressors in Preheater (HECI):
order to prevent condensation during the compres- one pass tube and shell type
tube length 3.03 m
sion cycle. ACOOL is an aircooled heat exchanger internal tube diameter .015 m
which at steady state transfers an amount of heat tube wall thickness .002 m
equivalent to the compressor power to the sur- number of tubes 108
roundings . total external heat transfer area 16. m**2

Secondary condenser (HESCOND):

Tank park. two pass tube and shell type
tube length 3.03 m
internal tube diameter .015 m
The tank park contains seven tanks buried 1.5 m tube wall thickness .002 m
below the ground level. Two of these reservoirs number of tubes 70
serve as bottom and top product tanks (BOT and total external heat transfer area 11. m**2
TOP), four contain the rawmaterials for the dis-
tillations (ISOP, MET, ENT and WA), and one serves
as an extra tank (EXT), providing a general dump TABLE 3 Tank volumes.
facility connected to the security system.

The pumps associated with the tanks are surrounded

Bottoms product tank (BOT) 4.0 m** 3
by a battery of on/off valves (BV1 - BV31) by Top product tank (TOP) 4.0 m**3
which it is possible to connect the tanks to the Isopropanol tank (ISOP) 4.0 m**3
feed pumps and thereby feedtubes, and to the pro- Methanol tank (MET) 2.5 m**3
Extra tank (EXT) 2.5 m**3
ducttubes. There are three feedpumps : a small
Entraîner (toluen) tank (ENT) 1.5 m**3
metering pump for the entraîner (ΡΕ), and two Water tank (WA) 1.5 m**3
regular feedpumps (PF1, PF2) for the 'heavy' and
'light' component respectively. A fourth pump (PM)
is connected to all of the tanks, and is used to
transport the contents from one arbitrary tank to
another. The above pumps are contained in a semi BASIC INSTRUMENTATION
open outdoors pump grave.
The process is instrumented in accordance with the
project purpose, where especially the dynamic and
TABLE 1 Main dimensions for the column section. control investigations are demanding in terms of
direct measurements and actuators. Since the con-
trol aspects include sequencing operations a rela-
Column (COL):
height 10.5 tively large number of on/off actuators are also
internal diameter .47 required. Towards the end of this section one of
distance between trays .50 the many possible layouts of the single control
number of sieve trays 19
loops for this plant is given.
Trays :
hole diameter of trays .006 m
fractinal free area .05 Measurements.
hold-up area/tray .013 m**2
weir height .05 m
Temperature measurements are performed on each
Thermosiphon reboiler (HERB): tray in the column, at several places in the
one pass tube and shell type tubing and at both sides of almost all heat ex-
tube length 3.00 m changers and at several locations in the heat pump
diameter .85 m
internal tube diameter .015 m
section. Temperature measurement is mostly per-
tube wall thickness .002 m formed using iron/constantan thermocouples, with
number of tubes 846 specially designed amplifiers in order to give an
total external heat transfer area 128. m**2 acceptable accuracy. A number of Pt 100 sensors
Condenser (HECOND):
are used to check the thermocouples.
one pass tube and shell type
tube length 3.00 m Flow measurement is performed on the two feed
internal tube diameter .015 m
streams, the two product streams, the two possible
tube wall thickness .002 m
number of tubes 442 streams from the decanter and at one of three
total external heat transfer area 67. m**2 possible positions in the reboiler section. All
flow measurements are turbine meters, in ord-.i; to
Steam feed heater (HEFS), product coolers (HEBW, H E T W ) :
obtain a wide measuring range with a small pres-
one pass tube and shell type
tube length 2.50 m sure loss, irrespective of the magnetic properties
internal tube diameter .015 m of the fluids.
tube wall thickness .002 m
number of tubes 70
The essential pressures in the column and heat
total external heat transfer area 16. m**2
pump section are measured. In total 14 pressures
Accumulator (Decanter) (DEC) : are measured this number includes up or downstream
internal diameter .394 m pressures for most pumps in the plant.The diffe-
internal length .985 m
rential pressure is measured over the reboiler,
volume .120 m** 3
condenser, top plate and the middle of the possi-
Flow rate ranges : ble feed plates.
reflux, binary .25 2.0 m**3/h
reflux, ternary .4 3.2 m**3/h
1.0 m**3/h
Two capacitance level meters are installed in the
feed .15
decanter. All tank levels in the tank park are
also monitored. Provisions for automatic sampling
TABLE 2 Main dimensions for the heat pump sec- of the top, bottom and feed compositions as well
as of selected tray compositions are included in
the design. A high speed gas chromatograph which
Compressors (COMP): together with the liquid phase sampling will be
number controlled by a separate micro computer, is pre-
number of cylinders sently under development. At the present stage of
electrical motor drives 99
development the analysis time is less than 20 s in
cooling power 560
the three component case.
46 L. Hallager et al.

Actuators. DATA EQUIPMENT

The plant is supplied with 45 on/off valves (mark-
ed BV) in order to select the tanks to be pumped
from or into, and to set up the desired flowcir-
cuit, such as feed tray location etc. Three fast
butterfly valves (VI to V3) are mounted on the
tubing to and from the reboiler and at the vapour
outlet from the top of the column. These valves
may by used to turn off the vapour flow in the
column for use in periodic cycling.
The computer control equipment is organised in a
hierarchial structure as illustrated in Fig. 4. A
commercially available distributed micro computer
system is used as the process interphase and for
single loop control. A general purpose computer is
used for implementation of high level controls,
data storage and retrieval, and for handling ope-
rator communications.

For the purpose of control 9 pneumatic and 1 Control objective

electrical control valves are included (CV1 to
CV10). Furthermore 6 pumps are of variable speed General
type. 5 of these are centrifugal pumps (PF1, PF2, 4 ΜΙΜΟ fc-
purpose
computer
PC, PT and PR) and 1 is the entraîner metering
pump (PE). The compressors may be operated with an
even number of cylinders, thus the capacity of the
heat pump may be varied in 8 steps. The aircooler
(ACOOL) is a three fan unit, and each of the fans
may be switched on or off.

D P I
Single control loops. <
H H

The desired pairing of controlled and manipulated

variables may in several cases depend upon the
particular control problem at hand. Since this
pairing relatively flexibly may be changed with
the distributed computer system, described in the PROCESS
next section, the following pairing may only be
taken as an example.

The seventeen actuators are at present manipulated Fig. 4. Hierarchial contorl structure. The gene-
using eighteen control loops, two of which are ral purpose computer hoste a multivari-
cascaded, as shown in Table 4. A loop is used for able controller, which may gives set-
each of the feed flows and for the feed tempera- points to single-loop PID controllers in
ture. The bottoms product flow rate is controlled the process computers.
by the effective reboiler level and the distillate
flow rate to the receiver (DEC) by the condenser
level. The column venting is controlled in loop 17 Process interface system.
by a temperature in the effluent condenser. Three
loops are used to control the receiver (DEC) le- The process interface is taken care of by a STELLA
vel, reflux and product flow rates. The two first system, supplied by S.T.Lyngsoe. It is a micro
of these loops are cascaded at present for opera- computer based process control computer system
tion at relatively high reflux ratios. In the case consisting of a number of possibly distributed
of azeotropic distillation the light phase hold-up processors lablled alpha-units and a host labelled
in the decanter will be controlled by PR. The heat a beta -unit. Each of the alpha-units can handle
pump is controlled using three loops: The capacity 30-32 digital or analog in or outputs. The commu-
control ( loop no 16 ) , i.e. the number of cylin- nication is handled by the beta unit which uses an
ders in the compressors, the high pressure (loop ether net type controller. The beta-unit is equip-
8) which is controlled by the cooling water flow ped with a CRT for display of actual process
i.e. CV8, and the freon evaporation pressure which values in graphical or alphanumeric form. The
is controlled by loop 9. The remaining five loops process may be operated entirely from the beta
(no 2,3,6,7 and 13) are at present used in manual unit , but there is only a very limited storage
mode for setting the appropriate outputs to con- capability for data.
stant levels.
In order to handle the relatively large number of
In the control studies the output from a multi- temperature measurements and digital inputs and
variable controller may be used as setpoint for outputs a special multiplexing system was deve-
some of the loops as shown in Fig. 4. loped and programmed into two of the alpha-units.
In this way up to 64 thermocouples and 128 digital
inputs and an equal number of outputs can be
handled.
TABLE 4 Control loops and their status during a
start-up. All variables are scaled. Two types of single loop controllers are available
in the alpha-units. A three point controller was
the first type available to this project the out-
from this controller is integrated on analog
PROCESS
1 FM4 CV1 MAN. - 378 integrators, which has a limited integration time
2 INTG (CV2) -•> CV2 MAN. 0 - 820 to produce the desired output current. At a later
3 INTG (CV3) CV3 MAN.
4 FM 3 CV4 MAN. 0 stage a digital PID contro lier was developed. At
5 TS2 - CV5 AUTO 2390 f,700
795
present this type of controller is implemented on
6 INTG (CV6) CVé MAN. 0
7 INTG (CV7) - , CV7 MAN. 0 800 five of the loops.
8 P10 CV8 AUTO 290Θ - 125 125
9 P8 - C79 MAN. 0 64
10 FMI -.·•• PF1 MAN. 0 794
-if PF2 MAN. 0 121
11 FM2
212
Background computer system
12 DP4 -: PC MAN. 0
13 INTG (PE) -·>• PE MAN. 0 820
14 FM5B PT AUTO 2986 - 331 319
15 L2 PR MAN. 0 616 The background computer system consists of a PDP
16 P8 - , CAP MAN. Û 11/73 multitasking computer which handles communi-
17 TEC1 -VCV10 MAN. 1 000 0
18 LI PT AUTO - 331 0 cation with the stella system, with the operator
 Distillation Plant with an Indirect Heat Pump
and with peripheral PC's which are used for gra-
phical presentation of the process status and for
time plots of selected measurements and actuator
values. In addition the PDP also has storage fa-
cilities for handling relatively large amounts of
experimental data. The basic software to handle
the jobs of the background system has been deve-
loped as part of this project and is described in
Clement and Hallager (1985). The basic philosophy
behind the software package is that a frame should
be provided within which the future user can load
or program a specific task with a very limited
effort. Such a task could be a high level multi-
variable controller or a calibration routine.
This software has been applied for sampling the
results shown later.
SECURITY CONSIDERATIONS
The plant is built to satisfy the requirements of
the fire and labor inspection authorities. Due to
the magnitude of the system considerable efforts
have gone into obtaining a relatively safe system.
Passive security measures.
The plant is physically spread over several loca-
tions. The column section ( corresponding to fig.
1) is placed in a 16 m high tower building ap-
proved as a fire hazardous area, using complemen-
tary equipment. The compressors, which do not have
explosion proof motors are placed in a separately
ventilated room, serving as a sound and fire bar-
rier .
Security valves are placed in the column section
(see Fig. 5 ) : in three positions around the top of
the column system and in one position at the bot-
tom. The outlets of these valves are connected to
a common venting and draining system, which vents
to the environment above roof and drains to the
extra tank. Furthermore the drain system of the
floor is connected to this system, in order to
prevent liquid losses into the public sewage sys-
tem. The heat pump section is equipped with secu-
rity valves to the environment (see Fig. 6) from
both the high and the low pressure side of the
compressors.
Fig. 5. Safety valves and venting system in the
column section.
Active security measures.
The plant operation is monitored by a commercial
Programmable Logic Controller. The strategy behind
the active security system contains two lower
levels of requirements to be satisfied before
allowing entering energy or material into the
process :
47
ad
1 1
Fig. 6. Safety valves in the heat pump section,
ο The basic security requirements must be full-
filled before anything may be done. On this
level pressures, concentration of explosive
gasses, ventilation of the compressor enclosure
and cooling water to the compressor system are
supervised.
ο If the basic requirements are ok, most of the
'passive' equipment along with the heat pump
system may be operated. The 'passive' equipment
consist in valves opening the different pas-
sageways of the system, but not allowing heat
into the system. Tank levels are monitored and
valves leading to a tank may not open if the
level is too high.
ο The 'active' parts (pumps, steam valve) of the
equipment only may be operated if not blocked
downstream. This means that a pump is not al-
lowed to run if the downstream tubing is not
open to a larger vessel, and that the steam
valve only may open if one of the feed valves
of the column (BV40-BV44) is open.
The signals for this system are measurements of
the actual equipment state and not just monitoring
of the control signal. If a condition is violated
only the necessary part of the system is closed
down ,e.g. if a tank is about to overflow, only
the appropriate valve and pump is turned off, or
not allowed to be turned on.
This security system has proved to be very valuab-
le in the operation of the plant both in limiting
the effects of human error and especially in shut-
ting down the plant in case of mechanical or
electrical malfunctions. The latter has indeed
occurred at a small handful of occations.
EXPERIMENTAL EXPERIENCE
At present the plant has been the subject of
several initial test runs in order to ensure a
reliable operation of the different components.The
initial tests have been performed with only the
simple PID controllers working, and with a limited
amount of automatic data storage.
The plant has been started on several occations by
means of the preheater (HEFS) and kept running at
total reflux only by means of the heat pump. How-
ever many of the simple loops are highly nonline-
ar, necessitating tuning upon change of operating
point. Consequently they are obvious choices for
some type of adaptive or automatically tuned con-
trollers .
48 L. Hallager et al.

Results from a pressure variations experiment are

presented in figure 7 a-e. The components in the
column were methanol and some water. The venting
to the atmosphere was almost shut off. The high
pressure in the compressor circuit, measured by
P10, is controlled using CV8. The reflux, feed
flow, and top product flow were held constant. No
bottom product was taken out, and as the feed was
pure methanol the water amount in the column was
constant.

All times are specified as time after start of

logging. At 1 hour 50 min the setpoint to the P10
controller was changed from 10 Bar to 11 Bar. At
3 hour 16 min the setpoint was changed back to 10 3600.0 7200.0 10800.0 14400.0
Bar. The response of P10 is shown on Fig. 7a. Fig. 7d. Column bottom pressure, PI, in mBar vs.
The signal to CV8 is shown on Fig. 7e. During the seconds from start of logging.
increase the controll valve is saturated. The
control is obviously not tuned as good as possi-
ble, as less overshooting would be prefered. The
low pressure in the compressor circuit, P8, shown
on Fig. 7b, follows P10 closely. The column top
pressure P3, Fig. 7c, and the column bottom pres-
sure PI, Fig. 7d., also follows P10 quite closely. 0.8 J

12.0 r_ 0.6 J

0.4 J
3600.0 7200.0 10800.0 14400.0
Fig. 7e. Control signal to control valve CV8 vs.
seconds from start of logging.

This experiment shows that the control of the high

pressure in the compressor circuit also controls
3600.0 7280.θ 1Θ8ΘΘ.0 144ΘΘ.0 the pressure in the column. The change in pres-
Fig. 7a. Compressor high pressure, P10, in Bar vs. sure in the column dos not change the vapour flow
seconds from start of logging. significantly. The vapour flow is determined by
the number of active cylinders in the compressors
and the valve position of CV9 (Fig. 2 ) .

ACKNOWLEDGEMENTS

The authors would like to thank: The Danish Energy

Department (Energiministeriet), the National Coun-
sel of Technical Research (Statens Teknisk-Viden-
skabelige Forskningsrâd), and the participants of
the industrial support group for finançai support.

REFERENCES

Baron,G., S.Wajc and R.Lavie (1980),"Stepwise

periodic distillation",Chem.Eng.Sci.,35,859-865
Fig. 7b. Compressor low pressure, P8, in mBar vs.
seconds from start of logging. Cannon,M.K.,(1961),"Controlled Cycling Improves
Various Processes",Ind.Eng.Chem.,53,629

Clement and Hallager, (1985),'Software for Data

Acquistion and Control.',IFAC-IFIP Workshop on
Real-Time Programming, Purdue University, West
Lafagette, Indiana, Oct 6-8.

McWhirter,J.R. and W.A.Lloyd, (1963), "Controlled

cycling in distillation and extraction.",Chem.
Eng.Progress.,59,June,58-63

Tolliver (1980), Presented at Distillation Dyna-

mics and Control Course (may) at Lehigh Universi-
ty, Bethlahem PA.

7200.0 14480.0 Waller (1982) "University Research on dual compo-

sition controll of distillation: A review", in
Fig. 7c. Column top pressure, P3, in mBar vs. se- Seburg,D.E. and Edgar,T.F., Chemical Process Con-
conds from start of logging. trol 2., Engineering Foundation and AIChE.

48
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

MULTIVARIABLE CONTROL OF INDUSTRIAL

FRACTIONATORS
A. Shakouri
Koninklijke I Shell-Laboratorium, Amsterdam (Shell Research B.V.), Badhuisweg 3,
1031 CM Amsterdam, The Netherlands

ABSTRACT

The multivariable Nyquist Array method offers a concept which enables the classical
single-input/single-output Nyquist control design methods to be extended to multi-
variable systems. This concept is based on partial decoupling using a compensator net-
work to achieve what is known as a "diagonal-dominant" structure, whereby single-loop
controllers can be designed independently for the different loops. The method described
here offers an approach by which a compensator network is designed automatically, a
function minimization algorithm being used to obtain dominance over a specified
frequency range by minimizing the ratio of moduli of the off-diagonal terms to the
diagonal terms of the appropriate open-loop transfer matrix. Once a suitable compensa-
tor has been found, the design of an overall control system is completed by designing
single-loop controllers for the different loops separately.

The applicability of the method is demonstrated, with the aid of a dynamic model of a
high-vacuum distillation unit, by the design of the viscosity control configuration
for that unit. Further, attention has been paid to incorporating such a control confi-
guration into modern, distributed-control systems such as Foxboro Spectrum.

KEYWORDS. Cascade control, quality estimation, dynamic compensator, frequency domain

control design, integrated control system, multivariable control system, industrial
control, quality control.

INTRODUCTION For implementation in integrated control systems it

The process instrumentation in current use has
reached a level of sophistication such that
advanced process control schemes are now being
introduced more and more to improve process ope-
ration. Increasingly, on-line quality measuring
instruments, distributed-control systems and
process computers are becoming standard refinery
equipment, so that a higher level of process
optimization is becoming feasible. If we bear in
mind that distillation processes are responsible
for approximately 40 % of refinery energy consump-
tion, the significance of the improved control of
these particular processes becomes apparent. The
objective is, by means of improved control, to be
able to produce much closer to specification,
reducing quality give-away and hence also energy
consumption. The successful application of advanced
process control schemes, including multivariable
control, depends on the availability of a reliable
control design method.
The multivariable Nyquist Array design method
(Rosenbrock 1969, 1974) forms a powerful tool in
the design and analysis of multivariable control
systems. This method tends to break down the multi-
variable problem into a number of almost non-inter-
acting single-input/single-output feedback problems
which can be solved using classical control design
techniques. Reduction of the interaction is
achieved by using a compensator network in such a
way that the modified open-loop transfer matrix is
diagonal-dominant. Without dominance the multi-
variable Nyquist Array method neither confirms nor
denies stability of the closed-loop system.
is most desirable, in view of the limited availabi-
lity of arithmetic blocks, to achieve the reduction
in the interaction by making use of static or low-
order dynamic compensators. The methods available
to achieve this reduction (Rosenbrock, 1969, 1974;
Leininger, 1979a, 1979b) require a high degree of
involvement of the designer and most are only
suitable for the design of static compensators.
In this paper a method for designing a dynamic
compensator will be introduced that is based on
the Direct Nyquist Array (DNA) method. This method
utilizes a conjugate direction minimization
technique to adjust automatically the compensator
parameters so as to satisfy the required dominance
condition.
The applicability of the resulting control method
is demonstrated via the control design of a high-
vacuum distillation column. Due to the particular
structure of the distillation column, models have
been used that can be divided into a fast dynamic
part, relating process inputs to intermediate
variables (e.g. temperature and pressure), and a
slower part, relating to the product qualities.
This justifies the use of a cascade control system
(master-slave control). A dynamic pre-compensator
is used in the slave control loop to reduce the
process interaction, which allows single-loop con-
trol in the master control loop. In the situation
where the quality measurements are subject to
relatively large analysis times, Smith prediction
can be combined with DNA (Smith, 1957). In these
cases, very much the same performance is achieved
as with computer-based time-domain multivariable

49
50
methods (e.g. linear quadratic control: Kwakernaak
and Sivan, 1972; Ten Hacken and Van Wijk, 1984).
This paper is organized as follows. Firstly the DNA
design method is summarized, after which the
decoupling algorithm for the compensator design is
presented. The process under consideration is des-
cribed briefly and the off-line simulation results
are presented. Finally, the performance improvement
achieved is evaluated and a number of conclusions
are drawn.
DNA CONTROLLER DESIGN
The notation used in this section corresponds to
that adopted previously (Rosenbrock, 1969, 1974).
The control configuration considered is given by
Fig. 1. The open-loop transfer function matrix Q(s)
is defined as
Q(s) = P(s) C(s) = L(s) G(s) K(s) C(s)
with G(s): η χ η process transfer function matrix
L(s): η χ η post-compensator transfer function
matrix
K(s): η χ η pre-compensator transfer function
matrix
C(s): η χ η diagonal controller transfer func-
tion matrix.
-J^S)-^ C(S) - { - » . K(S) L(S)
Fig. 1. Feedback structure of multivariable
control system using DNA method
The design of a controller by the DNA method is
split into two phases. Firstly, the interaction in
the process model G(s) is reduced over a pre-speci-
fied frequency region with the use of a pre-compen-
sator and/or post-compensator. In the next section
this compensator design is discussed in more
detail. The decoupled process P(s) can then be
controlled by a diagonal controller C(s). The
stability of the closed-loop transfer function
1 _1
H(s) = (I + Q(s))" Q(s) = R(s) Q(s)
with R(s): η χ η return difference matrix
can be investigated via the following procedure.
Use is made of the property of diagonal dominance,
defined by
|l+Q ..(s)| > E | Q - - ( s ) | -dc.(Q(e))
JJ
1 IJ J
I =
for all columns and over the pre-specified
frequency range.
Equation (2) considers "column diagonal dominance";
"row diagonal dominance" is defined in a similar
way
|l + Q..(s)| > Σ |Q ..(e)| = dr.(Q(s))
again over the pre-specified range.
A. Shakouri
The diagonal dominance can be graphically inter-
preted via an Array of the Nyquist plots of the
diagonal elements Qi^(s) with, in each frequency
point, a circle with a radius equal to dcj(Q(s))
or drj(Q(s)) for column and row dominance respec-
tively; these bands of circles are called the
Gershgorin bands. If the point (-1,0) does not lie
within the Gershgorin band, diagonal dominance is
obtained. In case R(s) is (either row or column)
diagonal-dominant the Nyquist stability criterion
can be used for evaluation of the stability of the
closed-loop system by application of the following
rule
η, - η (5)
h q
i=l
with n-j_: the number of encirclements of the
Gershgorin band of Q ^ ( s ) of the point
(-1,0) in the Nyquist plot
(1) n^: number of unstable poles of H(s)
nqi number of unstable poles of Q(s)
If Q(s) is not diagonal-dominant, the above-men-
tioned rule cannot be applied. The Gershgorin bands
and the Ostrowski bands can be plotted in the
frequency plots to evaluate the system interaction,
for the open-loop and closed-loop process respecti-
vely. For the design of the controllers use is made
of classical criteria like gain and/or phase
margin.
THE DECOUPLING ALGORITHM
The reduction of interaction attainable, and
thereby ultimately the controller performance,
is dictated primarily by two factors
. the structure of the pre- and/or post-compen-
sator;
. the ability of the designer to manipulate the
compensator parameters with respect to some
interaction criterion.
The order of the compensator is a compromise: on
the one hand the order should be kept limited for
implementation reasons; a high compensation order,
on the other hand, gives a better reduction of the
interaction and allows better control performance.
This emphasizes the need of a method for automatic
generation of compensator matrices, resulting in a
diagonal-dominant system for a given compensator
structure. Thus, different compensator structures
can be evaluated in a fast and efficient way,
(2) enabling the above compromise to be found. The
algorithm proposed therefore utilizes a numerical
optimization method to minimize the system inter-
action, defined as a function of the compensator
parameters.
The compensator design procedure in its most gene-
ral form is characterized by the following phases
. specification of frequency region {0, w
. structure determination and parameter
m }a x
(3)
initialization
. parameter optimization
. evaluation of the interaction reduction
. design and implementation considerations.
In the first step use is made of an automatic pro-
cedure to detect the maximum frequency of interest
(Leininger, 1979b). A good initial pre-compensator
is the (pseudo) inverse of the open-loop process
model gains. The third step is performed automatic-
ally by making use of the numerical minimization
(4) method of a function of the interaction measure.
In order to be able to implement the compensator
determined in an integrated control system, the
following additional constraints must be imposed
on the compensator elements
 Multivariable Control of Industrial Fractionators 51

. each element must be stable In the optimization the parameters are determined
. the order of the denominator must be greater which minimize the following criterion
than or equal to the numerator order
. each element is minimum-phase u
K 1 JJ J
. the ability to specify only real poles and/or F (a
J J
) = max
Ω i=l
Σ
Κ : < J") I / :< J > I (8)
zeros in the elements must be present.

These constraints follow immediately from the

definition of the different functions in integrated with Ω = {0,
k = 0,
U>
3
M} A X
control systems, such as Foxboro Spectrum and
Honeywell TDC-2000. J = 0, 1, ..., n.

For reasons of simplicity, in this paper we only
consider the case of the design of a pre-compensa-
tor K(s) in order to reduce the column interaction
of P(s). L(s) is taken as the unity matrix. Reduc-
tion of interaction in the columns, as opposed to
reduction of interaction of the rows, is less
complex in that it limits the number of parameters
in the optimization procedure. The columns of the
compensator can be designed independently. Other
cases, like post-compensator design, follow concep-
tually the same procedure.
The following parametrization is used for each
compensator element Kjj(s) to satisfy the above
requirements
a . . (1 + a . .s)
n 0
κ
) (1 +ea. . . s I +L aL -. . s
- _ o y
0 PROCESS DESCRIPTION
Given a compensator structure for a column, the
parameters which minimize the criterion are deter-
mined. If the interaction reduction is not suffi-
cient the order is increased until a satisfactory
reduction is found. The procedure is then repeated
for all columns, after which one can commence with
the controller design. The success of any numerical
optimization method depends on the shape of the
contours of the performance function and the
convergence properties of the algorithm employed.
In an attempt to reduce the impact of the contour
irregularities the variable-matrix algorithm is
used (Powell, 1981). By using this method the
constraints on the parameters are translated into
some analytical function to eliminate the need to
use constraint optimization.
CO)

with a £j > 0, k - 1, 2, 3. The process under consideration takes place in a

k high-vacuum distillation unit (HVU), producing
The latter constraints follow from the criterion distillates as feedstock for the manufacture of
from Routh (Morris, 1966). The above parametriza- luboils. Figure 2 gives a simplified flow scheme of
tion allows complex poles and zeros. If, however, the unit with the main controls. The feed is long
only real poles and zeros are allowed the following residue produced in a crude distiller. The product
parametrization is used streams are: gas oil (GO), spindle oil (SO), light
machine oil (LMO), middle machine oil (MMO) and
K () S a ..(l + a
0 3 si) j short residue (SR). The main objective is to keep
(7) the viscosity of the distillates to within narrow
I J = T T T , 3)(l a ..s)
U J +2 limits in order to arrive at constant luboil
viscosities.
with a
kj i> 0, k = 1, 2, 3.

^ VACUU
M
""set

ΓΓ55]

lQ3l

Γθ4ΐ

Fig. 2. Simplified scheme of the HVU

PROCESS CHARACTERISTICS AND CONTROL F, > T, -> Q,

STRATEGY
From a number of cases of application in practice
it has been found that a general structure for a
distillation columns exists, which can be expressed
as follows:
> T
0
Here the relations between manipulatable variables
F (for instance reflux flows, reboiler duties,
etc.), intermediate variables Τ (column tray
52 A. Shakouri

temperatures) and qualities Q are given, with The model used for the control design is given by
and Q 2 representing the product qualities SO and the following equations:
LMO, respectively (see Fig. 2 ) . Together with the
column pressure, tray temperatures determine the /T (s)\
product qualities via vapour/liquid equilibria.
3 0
ÎT (s) =
/*24<·>\
G(s) ( ' 2 6 ( e ) )
\t
3 4
This structure proves to have a number of charac-
3( S7) / \F 7(s)/
2
teristics influencing both the modelling procedure /t o(S)\
and the final control scheme. The first part has
H(s) 3
\Q2<s>/
relatively fast dynamics and considerable inter-
action, in contrast with the second part, which is The exact model parameters are given in the
characterized by slow dynamics and/or time delays appendix.
and relatively little interaction.

To control a process with the above characteristics RESULTS

the following structure is suitable. The inter-
action in the first part of the process can be
reduced by making use of a pre-compensator, thus
allowing the use of single-input/single-output PID
controllers. To overcome the adverse effect of time
delays on the control performance Smith predictors
or quality estimators can be used (Smith, 1957;
Alevisakis and Seborg, 1973); intermediate
variables are used to predict the actual qualities.
The interaction in the system can be found by
inspection of the column interaction indices Fj_
(i = 1, 2, 3) as defined by equation (8). Figure 4
shows a decoupled first column; Fj close to 0 over
the entire frequency region of interest. The second
and third loops, however, have a high interaction
index.

In this particular application, however, a cascade

control structure was chosen. This structure is
believed to be more robust with respect to model-
ling errors in the delays between the temperatures
and the qualities. The final control system
configuration is given by Fig. 3.

—I I I I
0 0.25 0.5
FREQUENCY, rad/min
~-<ψ *j PI (D) [
Fig. 4. Column interaction of the
( D ) uncompensated process
H&Hpi f
With the use of the method presented a pre-
compensator is designed; the resulting interaction
indices F 2 and F 3 are given by Fig. 5 and the para-
meters in the appendix. A considerable reduction of
the interaction is obtained.
Fig. 3. Control strategy of control system 1

CHOICE OF MANIPULATABLE VARIABLES AND

CONTROLLED VARIABLES

The first step in the modelling procedure of the

process is to determine the manipulatable
variables, disturbances and controlled variables.

All process variables which have a direct influence
on the controlled (output) variables are defined
as: manipulatable variables if they can be manipu-
lated, and disturbances if they cannot be manipu-
lated. In this case the controlled variables are
the properties of the products mentioned above.
a in <F d
a ent n
The choice of the manipulatable variables, the
reflux flows F 2 4 , F 2 6 27 controlled
variables, the viscosities of SO and LMO, are
mainly based on available dynamic models
(Ten Hacken and Van Wijk, 1984).
A dynamic model between the refluxes and the
viscosities showed long dead times which would
result in poor controllability. For this reason
intermediate variables (temperatures), through
which the model can be split into a relatively
fast and slow part, have been opted for. From
measurements it has been found that the draw-off
temperatures T 3 0 and T 3 7 correlate well with the
qualities and Q 2 , and that the draw-off temper-
atures correlate well with the reflux flows. The
temperature T 3 4 is added as a controlled variable
to give a square system.
0 0.25 0.5 0.25 0.5
FREQUENCY, rad/min FREQUENCY, rad/min
(a) (b)
Fig. 5. Open-loop interactions for (a) loop 2
and (b) loop 3 with compensator
The settings of the controllers were determined
off-line using frequency analysis, by making use of
phase and gain margin; the final tuning was done on
the basis of step responses. The control settings
were slightly conservative for reasons of robust-
ness. Figures 6-8 give the results of the off-line
tuning via the responses of the system to set point
changes in Qi, T 3 4 and Q 2 respectively.
The different elements of the compensator and the
controllers are implemented in control blocks of a
Foxboro Universal Control Module (UCM). Special
attention has been paid to a correct initialization
procedure, needed for smooth switching of the
control system from manual to automatic mode; the
outputs of the controllers must be determined from
the manual flow settings via the inverse of K ( 0 ) ,
the static gains of the pre-compensator.
 Multivariable Control of Industrial Fractionators 53

S TE P T
O I(NR) E L . are necessary to enable the compensator to be
implemented in an integrated control system; these
S TE P T
O I(NR) E L . are defined.

The applicability of the design method as a whole

has been demonstrated by the design of a system for
the control of product stream qualities in a high-
vacuum unit. The variations in product quality were
reduced by a factor of 2.5.
_J
0 0 2 40 36 2 04 0 36
SAM E P.NLo SAM E P. N
Lo REFERENCES
F I. Ga6
Alevisakis, G., and D.E Seborg (1973). An extension
of the Smith predictor method to multivariable
S TE P T
O I(NR) E L . linear systems containing time delays. Int. J.
Control, £ , 541-551.
Ten Hacken, G.V. , and R.A. van Wijk (1984). Use
S TE P TO I(N
R) E L . of multivariable control on a high-vacuum dis-
tillation unit. Proc. IFAC 9th World Congress,
Budapest, 2» 116-121.
Kwakernaak, Η., and R. Sivan (1972). Linear Optimal
Control Systems. Wiley & Sons, New York.
Leininger, G.G. (1979a). Diagonal dominance for
multivariable Nyquist Array methods using
J I I I I function minimization. Automatica, 15, 339-345.
2 04 0 36 2 04 0 3Leininger,
6 G.G. (1979b). New dominance character-
SAM E P.No
L SAM E P.No
L istics for the multivariable Nyquist Array
method. Int. J. Control, 30, 459-475.
Morris, J. (1966). A simple derivation of Routh's
stability criteria. Control , 10, 358.
S TE P T
O I (NR) E L . Powell, M.J.D. (1981). Some Properties of the
S TE P T
O I (NR
) EL. Variable Metric Algorithm. Harwell Subroutine
Library, England.
/ Rosenbrock, H.H. (1969). Design of multivariable
control systems using the inverse Nyquist
/ Array. Proc. IRE, 116.
Rosenbrock, H.H. (1974). Computer-Aided Control
/
1/
.30
/ T System Design. Academic Press, New York.
tT 34 Smith, O.J.M. (1957). Closer control of loops with
dead time. Chem. Eng., 53, 217-219.
2 04 0 36 2 04 0 36
SAM E P. N
Lo SAE M P. No
L
APPENDIX

Elements of the process models G(s) and H(s):

Figs. 6-8. Closed-loop responses in (a) quality
and (b) temperature to step increases (0.00462 s - 0.014) Exp(-1.5 s)
in the set point of: Qj_ (Fig. 6 ) ,
G (.)
u =
s + 3.305 s + 0.289
T 3 4 (Fig. 7) and Q
Sample interval 30 s
2 (Fig. 8 ) .
(-0.000112 s - 0.0216) Exp(-0.5 s)
G (s)
The control system is now operational. In compari-
1 2 s + 1.65 s + 0.385
son with the previous situation, in which the flows
were adjusted manually to obtain the correct quali- -0.00039 Exp(-2.5 s)
G (s)
ties, the temperature and quality variations are 13 s + 0.0347
reduced considerably, as Table 1 demonstrates.
G (s) = 0
2 1
TABLE 1 Improvement of Variation of the r ( \ (-0.00122 s - 0.000102) Exp(-3 s)
Controlled Variables 22
s + 0.1669 s + 0.0216

G (s) = 0
Variable Variât ion: Variât ion: 2 3
manual automatic
G (s) = 0
control control 31
T
T30. 2°c 6.5
4.0
1.2
1.7
G (s)
3 2
37. ,*C
Ql, mm /s 0.2 0.1
Q 2 , nmvVs 0.6 0.2 χ 0.00001 χ Exp(-3 s )
G ()s -0.0015
33 s + 0.0823
CONCLUSIONS
0.00491 Exp(-20 s)
H (s) -
A method is presented for extending the use of u s + 0.113
function minimization algorithms to the design
of dynamic compensators, which allows a better H (s) = 0
decoupling of the separate parts of the process
1 2
and thus improved control performance. Constraints H (s) = 0
21
54 A. Shakouri

n K8
u / \ _ 0.022 Exp(-15 s)
22 ' s + 0.159

Compensator matrix K(s).

1 0 -0.65 - 4.73 s -0.46 - 4.13 s

1 + 1.38 s 1 + 25 s

0.0 1.0 0.0

η η -0.14 - 6.74 s 1.0

1 + 21.3 s

Slave controllers

controller Τ, : C,As) = -45(1 + -τ^—)

30Λ 11 11 s

controller : C ( s ) = -25(1 + y^-)

00
IL 4 S

controller T · C ( s ) = -250(1 + — )
37 QO
33 12 s

Master controllers

controller Q : C ( s ) = -5.5(1 + yj-^)

n

controller Q : C ( s ) = -1.2(1 + -^j)

2 22
 Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

OPTIMAL OPERATION OF A DISTILLATION

COLUMN WITH TIME-VARYING
ENERGY COSTS
A. R. van Heusden and G. H . Heusinkveld
Department of Chemical Engineering, Twente University of Technology,
Enschede, The Netherlands

Abstract. It is demonstrated that optimization of a total chemical plant results in a se-

quence of stationary operations, which have to be realised according to a strict schedule.
The non-linearities in the incremental costfunction of utilities lead to a switching stra-
tegy which consists of alternating lower and higher feedrates. Conditions, which have to be
fulfilled, are summed up. A smooth transfer between the successive stationary modes of
operation, obviously is an important restriction. In order to study the workability of this
strategy, experiments were done by means of an extractive distillation process, carried out
in a bench- scale packed column. Relatively simple feedforward control, added to the normal
feedback control of the distillation column is found to be adequate.

Keywords. Optimization, energy management, control system synthesis, load regulation, step
response, modelling.

INTRODUCTION bility for making the mode switches. The fol-

There is an ongoing trend in the process indus-
try towards generating utilities for one*s own
use. Even cogeneration of steam and electricity
becomes more and more a normal situation. Usu-
ally several parallel units have been installed
(sometimes historically determined). When opti-
mizing such powerhouses, it is quite obvious to
choose the demands (dictated by the process-
management) as independent variables. Optimiza-
tion of the load-distribution between the par-
allel units is then possible, and 2 to 5 per-
cent saving in primary energy is attainable.
For relatively small utility plants, non-linear
programming is suitable (Cho and Blevins, 1981
or Draga, 1983). If the structure of the utili-
ties is more complicated, linear programming
(LP) techniques are more adequate (Boenink
et.al. 1984), in order to reduce the amount of
computational effort. The choice of units "in"
or "out" use also can be made intuitively. If
not, a mixed-integer LP-routine is necessary,
with higher costs of computation (Papoulias,
1983). The optimization of a total factory (all
the procès units and the utility units) is even
more complicated, but with LP-techniques a
reasonable possibility. When optimizing a whole
plant, there may be reasons that lead to depar-
ture from the "static" optimization, as will be
stated in the next section.
lowing examples will be illustrated below:
i. periodical changement of electricity costs
ii. steam-boilers in parallel
iii. cogeneration of steam and electricity.
i. Periodical changement of electricity
costs.
If the process is connected with the national
grid, there may be periods of increased rate
e.g. day/night or week/weekend. Lowering the
consumption during increased rate and visa
versa leads to optimality, see Fig. 1. The mean
value (F) of the feedrate of the processes has
to meet the planned amount. Restrictions are
twofold (this also goes for ii and iii).
a. There must be a certain overcapacity
<F > F ) .
X
b. Storage capacity of products and intermedi-
ates has to be adequate.

COSTS OF
ELECTRICITY
QUASI-STATIC OPTIMIZATION

This term should be explained: We mean an opti-

mization by changing operating conditions such
as feedrate, without taking into account the
switch-over period. The response-time is assu-
med to be short with regard to the duration of
the steady-state operation, so there is no
evidence for dynamic optimization. All non- Fig. 1 Changing of production-level (F) of a process when costs of
linearities in the costfunction of the utili- electricity are periodically.
ties may lead to savings, if there is a possi-

DCCR-E

55
56 A. R. van Heusden and G. H. Heusinkveld

ii. Steam boilers in parallel.

When operating with several (N) boilers, from a
fuel consumption point of view, it is best to
operate with Κ boilers in use (Κ < N) and to
hold the other cold-standby.
If the Kth boiler (see Fig. 2) is only partly
loaded, the best operation is to use K-l and Κ
boilers alternately.
SUBOPTIMAL
. PRODUCTION OF STEAM
FIG. 2 Fuel costs as a function of the amount of full-loaded
boilers (K) in use.
iii. Cogeneration of steam and electricity.
We have some experience with LP-application of
a large industrial utility plant (Boenink et
al., 1984; Smit, 1984). At that location, there
is quite a difference in accounting the used
electricity from the grid and the supply back-
wards to the grid. Fig.3 shows the results of
the LP-model for two different configurations
(in the LP-program determined by Boolean vari-
ables) .
DESCRIPTION OF THE APPARATUS
In order to study continuous distillation, a
small column (height 2.1 m, inner diameter 0.05
ra) consisting of double-glass elements has been
used. For packing material small rings were
used, consisting of twisted metal wires with
fine meshes. The height of the three packings
were about 0.5 m, see Fig. 4. For more details,
see Knol (1986).
As mixture for the extractive distillation,
acetone (b.p. 56.6 °C) and methanol (b.p. 64.6
°C) have been chosen. In order to come through
the azeotrope at 80 wt.% acetone, water was
also feeded, which has more affinity to metha-
nol than acetone (Carlson et.al, 1954).
The control of the four temperatures, viz. top-
bottom- and both feed-temperatures, was carried
out by means of a microcomputer with appropria-
te interfaces. Furthermore, six control actions
could be generated by the computer: two feed-
rates (pumpsteering), reflux-ratio (time-modu-
lation of the solenoid valve) and three heat-
inputs (electronic regulation of the current to
the electric heaters).
The compositions were analysed by means of a
GLC, combined with a data acquisition system on
a separate microcomputer. The application pro-
gram was written in Pascal and a smal part in
the assembler language.
COOLING WATER

The optimal points (A and B) are to be used

preferrably for the total optimization, so
switching between A and Β implies switching
between feedrates of one or more steam-consu-
ming process units.

ENERGY
COSTS

g—[vvw;

, PRODUCTION OF
LP-STEAM

BOTTOM
Fig. 3 Energy costs as a function of steam production, PRODUCT

at constant demand of electricity.

COOLING WATER

Fig. 4 Lay-out of the distillation column.

 Distillation Column with Time-varying Energy Costs 57

STATIONARY OPERATION
TABLE 2 Parameters of the transfer function
For the lab-scale process two stationary opera-

<3 <3
ting prints have been chosen: One point with operating point 1 Κ
high and one with low energy consumption rate. (°C.min/ml)
Because product qualities are fixed and have to
be maintained within limits, the feedrate is temp, bottom/feed -0.614 2000 470
also different for both operating points. The rate
condition and concentration of both feedrate temp, feed/feed rate -0.412 160 40
and solvent are fixed. The solvent rate is temp, solvent/solvent -0.827 100 25
minimized thus minimizing the effort to sepa- rate
rate the solvent from the bottom product. The
two operating points are listed in Table 1. (°C/W)

temp, bottom/power 0.0423 1400 230

reboiler
TABLE 1 The operating points of the column temp, feed/power feed 0.675 180 70
preheater
input quantity operating point temp, solvent/power 0.238 100 70
solvent preheater
1 2

feed rate (ml/min) 27 42 operating point 2 (°C.min/ml) (s) (s)

solvent rate (ml/min) 42 76
power feed preheater (W) 30 45 temp, bottom/feed -0.148 300 240
power solvent preheater (W) 140 230 rate
power reboiler (W) 550 850 temp, feed/feed rate -0.466 100 40
reflux ratio distillate (-) 2 2 temp, solvent/solvent -0.493 60 10
rate

(°C/W)
The operating points with product concentra-
tions and rates are reproducible. Now we like temp, bottom/power 0.0277 1300 60
to switch-over between the two points from time reboiler
to time. A few conditions have to be met: the temp, feed/power feed 0.247 130 108
time intervals between switching-over has to be preheater
much longer than the time interval in which temp, solvent/power 0.182 115 25
dynamic effects of the switch-over extinguish. solvent preheater
Secondly no variations in product quality are
allowed during switch-over. The first condition
allows a quasi static optimization and the
second has to be realised by using an adequate
control strategy.

DYNAMIC MODELLING

The dynamic behaviour of the column was

investigated. The responses of the different
automatically measured temperatures and of the TABLE 3 Relationship between controlled and
product qualities were measured after changing input variables
the energy consumption rate of the various
heaters. The responses of the temperatures were controlle5d bottom top feed sol-
fitted into a transfer function, consisting of temperatiires vent
a time delay cascaded with a first order
element. In Table 2 the responses are given in input signals
term of the parameters of the following
transfer function: power strong weak zero zero
Ts reboiler fast slow
Hs " d
power feed weak weak strong zero
< >= ( 1 % T ) (D
F S preheater slow slow fast
From the dynamic analysis we learned among
other things: power solvent weak weak zero strong
preheater slow slow fast
- Temperatures of feed and solvent are strongly
and fast disturbed after changing feed and
solvent rates. feed rate strong weak strong zero
- The system is non-linear: time constants fast slow fast
dependent on flow rates can be expected.
solvent rate weak weak zero strong
Another observation is the difference in gain
for disturbances in different directions fast fast fast
(around an operating point). The non-linear
reflux ratio weak weak zero zero
behaviour was not taken into account.
slow fast
FEED-BACK CONTROLLERS

The column is controlled by six input signals.

We can use these to control four automatically
F r o
measured temperatures. The control scheme is m st e
ht eiaw bdd le
o et
c u eist control
dh e
derived from Table 3. scheme as shown in Fig. 5.
58 A. R. van Heusden and G. H. Heusinkveld

SWITCH-OVER WITH HELP OF FEED FORWARD CONTROL

Switch-over strategy can also be derived from

the dynamic model. In the previous paragraph
the energy consumption of the bottom heater is
remained as a degree of freedom. We now change
this quantity stepwise and we will correct the
deviations in the controlled temperatures by
means of a feed forward scheme. By controlling
these temperatures the product qualities will
be kept close to their desired values.

If a temperature Τ is dynamically determined by

two input quantities x^ and x according to
2
ôT(s) = H (s).ôx (s) + H (s).6x (s) (4)
1 1 2 2
and we apply a know disturbance δχ^, then the
correction δ χ can be computed according to
Fig. 5 The adopted control scheme. 2
l+τ^ s -S(T, ,-τ, )
v 9 9
(5)

Some remarks: The correction consists of a time shift:

- The temperature at the top of the column is
influenced by both the reflux ratio R and the
solvent rate Φ · These influences are small
3
however, because the relationship between
temperature and concentration is not strong
in this case. Therefore the temperature in
the top of the column is not controlled.
- We would like to keep the energy consumption
of the bottom heater as an independent quan-
tity. In that case we can change it stepwise.
Since the bottom heater is the biggest energy
consumer we can control the energy consump-
tion in this way.
- The solvent rate <J> is adjusted linearly with
the feed rate
g
according to
2.3 <|> - 19.3
f (2)
From the dynamic model we can derive the para-
meters of the three controllers. The chosen
controllers minimize the quadratic loss func-
tion
Jm 2dt
t[" <yi - y )
Where = setpoint and y = controlled variable
This derivation was done for both operating
points. The most cautious controller was chosen
for the experiments in order to guarantee sta-
bility. The results are given in Table 4.
i-t^ 2 correction of x after a change
' ' in x^;
2
<0: correction of x before a change
in x^
2
and a lead-lag filter with time constants, that
are derived from the time constants of the
first order elements in the dynamic model. The
switch-over strategy is now as follows (compare
with Table 4; average values are taken).
t=t :
0
Adjustment of the power consumption of the feed
preheater with a lead of 165 sec, a lag of 130
sec and a gain of 0.0667 W/W
t=t + 6 sec:
0
Adjustment of the power consumption of the sol-
vent preheater with a lead of 120 sec, a lag of
20 sec and a gain of 0.555 W/W.
Switch-over of the bottom heater and adjustment
3 of feed rate with a gain of 0.0533 (ml/min.W)
without applying the lead-lag filter.
<>
The switch-over procedure thus consists of a
correction of the feed and solvent preheaters
in advance by using a lead-lag filter. After
that the bottom heater and the feed rate (to-
gether with the solvent rate) are switched
simultanuously to their new values.

The feed-forward scheme has to be combined with

the feed-back control scheme by adding the
control actions of both.
TABLE 4 Optimal controller parameters of a PI
algorithm
RESULTS
gain reset
In order to assess the feed-forward strategy,
(ml/min.W) time (s)
switching between the operating points has been
temp, bottom/feed rate 5.7 2000 carried out with and without the feed-forward
superposed on the feed-back control. Fig.6
(W/°C) shows the results for the bottom temperature
without (a) and with (b) feed-forward control
temp, feed/power feed 3.4 335 when switching from high to low column load.
preheater The reverse switch is shown in Fig.7 (a) and
(b).
temp, solvent/power 5.8 172
solvent preheater
 Distillation Column with Time-varying Energy Costs 59

In spite of the noisy signal (due to some prob-

Fig. 6a
lems with the A/D converter) we may conclude
that the upsets are quite acceptable. The time-
scale of the error signal is found to be much
shorter in the case of the feed-forward con-
trol. So the whole process reaches steady state
much sooner.

Simular results were obtained with respect to

the other three temperatures. The variations in
the composition of the distillate were not
1000 2000 3000 4000 5000 eooo detected with the gas chromatograph.
Time (sec)
DISCUSSION AND CONCLUSIONS

By using the above mentioned control strategy

S we can switch-over between the various opera-
ting points, without upsets in the controlled
temperatures and in the temperature in the top
of the column. After analysis the concentra-
8
ί Μι tions of the products appeared to be within
narrow limits.

Fig. 6b The same parameters for the correction have

been used for switching-over to both operating
points. In the course of the experiments these
have been adjusted slightly to improve dynamic
behaviour.

500 1000 IS 30 2000 As far as switching-over is concerned, we have

Time (sec)
Fig. 6 Switch-over behaviour from high to low
load of the bottom temperature without
(a) and with (b) feed forward control.
succeeded in changing energy consumption step-
wise. The strategy given in this paper enables
us to adapt energy consumption to the availa-
bility or to an optimal level. In this way we
can save money from our energy bill.
REFERENCES

Boenink, H.J. et.al. (1984): Optimum energy

utilization in total chemical plant operation.
Journal A, 25_ 168 - 176.
f
Carlson, C.S., Smith, P.V. and Morrell, C.E.
Fig. 7a (1954): Separation of oxygenated Organic Com-
pounds by Water Extractive Distillation, Ind.
Eng. Chem., 46, 350.

Cho, C H . and Blevins, T.L. (1981): Energymana-

gement in the paper- and pulpindustry, PT/Pro-
cestechniek, 36, 365 - 369 (in Dutch).

Draga, E.R., (1983): Implementation of indus-

trial energy management systems; Foxboro Soest,
2000 3000
The Netherlands.
Time (sec)

Knol, R.J.F. (1986): Optimization of the opera-

tion of extractive distillation, Master Thesis
Twente University of Technology (in Dutch).

Papoulias, S.A. and Grossman, I.E. (1983): A

Structural Optimization Approach in Process
Sunthesis, Computers and Chemical Engineering,
Jj 695 - 706.

Smit, J.J. (1984): Optimization of Utility

Systems. Master Thesis, Twente University of
Technology (in Dutch).

1000
Time(sec)

Fig. 7 Switch-over behaviour from low to high

load of the bottom temperature without
(a) and with (b) feed forward control.

59
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

CONTROLLING AND OPTIMIZING

INTEGRATED FRACTIONATION PROCESSES
IN A PETROLEUM REFINERY
M . Sourander* and S. Gros**
^Process Automation, Neste Technology, Porvoo, Finland
**Process Technology, Neste Technology, Porvoo, Finland

Abstract In this paper computer control of integrated fractionation

processes in a petroleum refinery is discussed. The implementation
of on-line optimization of product yields versus energy usage is
shown. The combination of on-line economic closed loop descission
making with constraint controls, representing system limitations in
fractionators and heat exhangers are also illustrated. A comparison
of the situation before and after the computer control implementation
is shown in terms of obtained control quality as well as yield
improvements.

Keywords. Computer control

On-line operation
Optimisation
Decoupling
Models
Oil refining

UNIT CONFIGURATION OF THE NAANTALI REFINERY
A brief introduction of the Naantali refinery is
needed before a detailed discussion about
optimization strategies can be presented.
The refinery has two atmospheric crude units. The
long residue from these are further fractionated
in a vacuum unit.The vacuum short residue is
termally cracked in a visbreaker unit to obtain
heavy fuel oil. The light and heavy vacuum gasoil
is the main feed for the catalytic cracker unit
besides the heavy gasoil (HGO) from the crude
units. The light and medium atmospheric gasoils
(LGO and GO) along with the corresponding frac-
tions from the catalytic cracker are hydrotreated
to form the diesel and light gasoil pools.
Kerosine is treated and used as jet fuel or used
for blending. The heavy naphtha is hydrotreated
and reformed to obtain motor gasoline while the
light naphtha is stabilized in debutanisers
and further treated in naphtha splitter,
deisopentanizer and depentanizer columns. It may
be emphasized that the product cooling, pumparound
and feed preheat systems of crude unit 1, crude
unit 2 and the vacuum unit form a complex
integrated heat exchanger network.
In the computer control system implemented at the
Naantali Refinery 14 fractionation units were
involved. In this paper, however, the attention
is focused mainly on the two crude units, the
vacuum unit and the light naphtha splitter system.
Valmet's Damatic systems, each of which is also
connected to the process control computer. The
process computer system was delivered by Nokia
Oy from Finland and consists of Digital Equipment
Corporation's VAX 11/750 computer using VMS 4.2
operating system, Nokia PMS process management
system software, instrumentation interface and
M/M communication hardware and software.
During the modernization all control functions
of the refinery were reconstructed to support the
refinery profit maximization objective. This was
done by the Neste Technology Group in close
co-operation with the operating personnel of the
Naantali Refinery. As a general rule it was
established that all basic controls, that are
needed to run the process units without excessive
manual operations have to be implemented in the
digital instrumentation systems. Because the
instrumentation systems are all redundant this
improves the security. On the other hand, control
functions that require large amounts of
supporting calculations, optimization or special
control algoritms were realized in the computer.
To plan, implement and commission the the
computer controls 18 manyears of work were
required. In this paper no distinction is made
between controls in the instrumentation system
and in the computer because of their functional
similarity. Similar control projects, as the
one discussed in this paper, have been analyzed
erlier by (Nasi, 1983; Rinne, 1982; Sourander,
1984).

Instrumentation
During the modernization period 1984 - 1985 the
Naantali refinery analog instrumentation was
completely replaced by digital instrumentation.
The manufacturer is Valmet Oy from Finland. The
digital instrumentation consists of four of
Control objectives
The economic performance of a given petroleum
refinery is governed by interactions bewteen
yield and product quality infractionation units,
by interaction between feed quality, process
conditions and product quality in conversion
units and by blending characteristics in sales

61
62 M. L. Sourander and S. Gros
product finishing. This means difficulties in
the control mechanisms but also possibilities
for additional benefits from interunit
manipulations.
Hence also a refinery wide, closed loop optimi-
zation system, developed within Neste OY, was
installed in the Naantali computer, but is not
described in this paper.
The control objectives for the refinery section
which is described here are as follows:
1. Maximize recovery (or yield) of every
distillate product against its most
critical quality property
2. Allocate energy optimally to emphasis
separation vs. energy interaction
3. Operate within the equipment
limitations and drive against the
constraint which represents the biggest
profit
4. Maximize control responsiveness by
use of proper control structure,
implicit decoupling and dynamic element
5. Retain control robustness to emphasis
the predicted process condition
variations
CONTROL STRATEGY
Product quality control
In fractionation of petroleum, different column
products represent distinctive cuts of the
feedstock which are measured by (standard
laboratory testing) distillation curves. The
split between two adjacent cuts is determined by
the cutpoint between them. The cutpoints can be
calculated by using column direct measurements
and they are used for every petroleum fraction
as a primary quality property to control the
quality of the lighter product heavy end.
The quality control consists of yield and
balancing control, cutpoint control and analyzed
property control as cascade type logic as shown
in the Fig. 2. The sidedraw product yields (in
percentage of column feed) are stabilized using
dynamically compensated yield controllers and
feedforward controllers to decouple the
propagation of disturbances further than to the
adjacent product below. Cutpoints are controlled
to targets by manipulating the yield controller
setpoints in order to maintain constant product
fraction width. The cutpoints in their turn are
adjusted to keep constant analyzed product
qualities (Fig. 2 ) . These controls are signifi-
cantly slower than the cutpoint controls because
of analyser and process delay times. The two
crude units are identical in control structure.
The overhead product of the atmospheric crude
units is further distilled to a number of
fractions in downstream units and has no
distillation range requirement.This fact, the
overhead system in two stages and the overhead
noncondensible gas being charged to the vapor
recovery compressor gives the oportunity to
minimize the overhead light naphtha for the
benefit of heavy naphtha without the ordinary
drawbacks of light overhead product. Therefore
the overhead temperature minimizing constraint
control is used. Note that light naphtha is fed
to the debutanizers, light naphtha splitter,
deisopentanizer and depentanizer in this order.
The reboilers of these use the pumparound heat
of both crude columns and thus propagate dis-
turbances in the overhead system and the
subsequent fractionation columns back to each
crude column mid section.The lowest side draw
product in any column, heavy gasoil or cutter
stock is maximized against a number of
constraints using special logic in the constraint
controller. This is because of the sensitive
contamination of the product by entrained residue
when overflash approaches nil. Maximization is
economically justified because of the substantial
value difference between sidedraw and residue.
Pumparound control
Since vac. overhead is returned to crude unit 1
pumparound as an additional feed, also
disturbances in each crude unit sidedraws and in
vacuum unit are propagated back to the crude unit
1 midsection. Similar disturbances enter from
other units to crude unit nr 2. The above
underlines the importance of stabilization and
optimization of the pumparound section in the
column. The special features of pumparound loops
make the conditions of stabilization difficult.
The stabilization of the pumparound duty, for
example, does not stabilize the internal
conditions of the column. Therefore the
calculated internal flow rates at some trays
above the pumparound return are controlled to
targets by manipulating a pumparound flow
controller setpoint and a bypass flow
simultaneously. The total pumparound duty is
stabilized by adjusting the heat returned to
crude preheat when changes in other pumparound
heat exchangers (light naphtha fractionators
reboiler) or elsewhere occur.
In order to assure operation inside feasible
range of the operating window, internal flow
rate min/max constraint control features are
added to internal flow controllers.The
constraints are, besides the physical equipment
limits, heat exchanger maximum duties, valve
positions, temperature limits etc, also the
column flooding and downcomer back-up limits.
Pressure control
Pressure minimizing constraint controller is
employed to ride the constraints affecting
pressure (and internal flow rates). During
constrained operation, pressure setpoints are
relaxed dynamically in the fastest way, while
the internal flow rate controller reacts slowly
to the constraints that affect both the vapor
rate and the pressure. The longer term economic
impacts that need interunit calculations is then
performed in the energy optimizer, which
calculates the optimum internal flow rates and
column feed temperature simultaneously. Observ-
ations show that in our cases the flow rates
have been maximized within other limitations,
while sacrificing the pressure. Our steady state
simulations, however, show that this somewhat
unique observation is no more necessarily valid
if a vacuum unit operating at minimum floating
pressure is used instead. Then the optimum
operation would also be much more sensitive
to the relative product values and the value of
fuel gas. This underlines the need to use energy
optimizer, capable of taking into account the
refinery structure and economics.
Energy versus yields
In a crude distillation column the separation
must be understood as a function of the
distillation curves. This is illustrated in
Fig. 3. The figure shows the crude oil true
 Integrated Fractionation Processes in a Petroleum Refinery
boiling point distillation curve and the
corresponding product TBP-curves. If the
separation was complete the product distillation
curves should coincide with the crude oil curve
for the distillation product in question. A low
separation gives distillation "tails" containing
high boiling point fraction (which for the product
for instance LGO will result in a low cloud
point). A higher separation gives a smaller
"tail"(and a better cloud point) which in turn
enables a better yield on account of the product
yield below.
PARAMETERS AFFECTING THE SEPARATION
It is important to know which parameters are
possible to manipulate in order to affect the
separation efficiency. These parameters are the
variables which are manipulated when the objective
function is maximized. In the following a crude
distillation system is analyzed.
The column pressure affects the separation, due
to the increased relative volatilities at
decreased pressures. In our case the optimal
operation is minimization of pressure against
constraints as for instance the maximum control
valve position, flooding etc.
The stripping stream decreases the hydrocarbon
partial pressure and thus has the same effect
as the above discussed column pressure. The
optimum stripping steam ratios in crude columns
and side strippers have been investigated through
steady state simulations. The hydrocarbon to
steam ratios are maintained through ratio
controllers.
The internal flowrates are important parameters
when trying to affect the separation. Increased
internal flows leads to a better separation.
The crude columns have one pumparound and one
top reflux with both warm and cold reflux, see
Fig. 2. If the pumparound duty is decreased
the top reflux duty will automatically increase.
This will lead to increased internal flows in
the upper part of the column and subsequently
also to better separation in the upper part.
The separation below the pumparound is not
affected. This shows that if higher yields of
lighter (and in our case more expensive) products
are desired the column should be cooled mainly
by top reflux. A high feed temperature also
leads to higher internal flows, and better
separation.
OPTIMIZATION OF ENERGY RECOVERY AND YIELDS
The objective function to be optimized is the
sum of product income and net energy cost.
High feed temperature results in an increased
energy consumption. Cooling in the upper part of
the column using cold reflux means that energy
is transferred to water instead of preheating
crude, having the same effect.
To be able to calculate how much extra energy is
needed, rigorous simulations of the heat exchanger
network have been carried out (see Fig. 4 ) .
Also rigorous distillation simulations have been
used to adjust the functions describing the
yields change with changing pumparounds, top
reflux and feed temperature.
In crude and vacuum units (with no reboiler) all
the primary energy for fractionation is entered
into the column as feed enthalpy and its source
63
is the absorbed heat in crude heater fired with
fuel gas or oil. The feed flashes and generates
internal vapor flow, the volume of which expands
while meeting internal reflux and becoming
lighter at every tray. The vapor rate can
decrease only by removing heat from the column
at the pumparounds. The column geometry requires
some heat to be removed and also the overhead
condensers are sized to match this.
Optimization of the pumparound duty
Generally the pumparound duty can widely vary
without exeeding column constraints. A decreased
duty causes increased vapor flow in the upper
part of the column, which in turn improves
separation and the yields of the lighter products
above the pumparound at the expense of the lower
products. The pressure effect will cause an
increased long residue yield and the preheat
changes of the heater feed will cause an
increased fuel gas usage. Interaction to vacuum
pumparound exchangers will increase their duty
affecting favourably the heavy vacuum gasoil
yield. Steady state modelling technique with
feedback information from the unit to adjust
the models for successively increased accuracy
is used. The final feedback on long term is the
product laboratory distillation curves which
are used on daily basis.
Optimization of column feed temperature
In these columns the other possibility to affect
energy usage is to adjust the column feed
temperature. This optimization is even more
complex because it is dependent on the factors
described earlier, overflash recovery in crude
units and in vacuum unit and vacuum unit heater
operation credit and the optimum value of the
pumparound duty. The optimum column feed
temperature is calculated and implemented via
feed temperature constraint controllers.
Vacuum unit optimization
The vacuum unit has total draw-off trays for
products and the unit is operated against a
fixed minimum overhead product temperature.
Light and heavy vacuum gasoil are blended and the
heater operates at its maximum metal
temperature.Therefore, in our case, the energy
optimization is straightforwardly maximization
of heavy gasoil pumparound duty (at the expense
of light gasoil pumparound duty) against
operating constraints. Note that Hy.VGO is
maximized against its colour and other
constraints by manipulating the washflow back
to the column, whereas during normal operation
the overflash is returned to the feed.
Light ends fractionation optimization
In light ends fractionation chemical compounds
can be analyzed and defined as light and heavy
key components and treat the column as pseudo
binary distillation. For example, in debutani-
zer i-pentane and in deisopentanizer n-pentane
is the heavy key component. In all these columns
composition control hard target is the distillate
product heavy key composition, which is
controlled according to specification, by
indirectly shifting the column material balance.
The soft composition target at bottoms product
is allowed to have more variations and is
indirectly controlled to a long term optimum
value by using energy optimization which
determines optimum column internal flow rates.
64 M. L. Sourander and S. Gros

These are implemented through a min/max. TABLE 2. Computer Control Results

constraint controller by adjusting the bottom
reboiler duty. Since the reboiler energy is
derived from the crude unit pumparound enthalpy Product quality before after
this calls for a comparison of the relative
usefulness of energy in light end columns, crude Standard deviation
columns or crude unit preheat.
Hy Naphtha 90 % pt. 2.5 1.0
CONTROL PERFORMANCE AND RESULTS Kerosene density kg/m N.A. 0.7
Lt gasoil cloud pt. °C 1.5 0.7
The computer control system has been operating Gasoil viscosity 0.4 0.2
successfully for about a year. The availability Hy.gasoil colour 1.6 1.0
of the control functions in the digital
instrumentation system has been virtually 100 %
due to the redundant equipment (measurement and
control valve faults not included). The computer Product yield
control availability has been 99,6 per cent, the improvement per cent
floating point accelerator card being the only of crude oil charge
observed fault. Naphtha base + 0,8
Kerosene + 1,0
The control performance and optimization effects Lt gasoil - 0,5
have keenly been followed by the operating Gasoil - 2,1
personnel and compared to the time prior to Hy.gasoil - 1,0
computer control. The economic improvements long residue - 0,2
attributable to the computer control show that
the payback time has been about one year. The
result are shown in Table 1.
REFERENCES
Quality control
Sourander, M. & al."Control and Optimization of
The product qualities are clearly more stable olefin- Cracking Heaters", Hydrocarbon
than before the computer control. The yields of Processing Vol 63, No 6, 1984
more valuable products have increased due to
optimized pumparound duties and column feed Nasi, M. & al. "Experience with Computers Control
temperatures, while keeping the same quality of an Ethylene Plant", Hydrocarbon processing,
targets. This is due to less pronounced "tails" Vol 62, No 6, 1983.
in the distillation curves of the products.
Rinne, R. & al. "Experience with Distillation
Energy allocation Unit Computer Control", Hydrocarbon Processing,
Vol 61, No 4, March, 1982.
In the product value situations of the past
year, despite of the tremendous changes in the
crude oil price, the product value differences
have not changed directionally. Also the optimum
strategy that the computer has implemented seems
to be stable in nature. In all cases both crude
units' pumparound duties have been at maximum
limits and the column feed temperatures at maximum
limits, both determined by column and feed heater
design geometry factors.
 Integrated Fractionation Processes in a Petroleum Refinery

SOLVENTS
NAPHTA
LT SPLITTER

1 11
SYSTEM GASOLINE CATALYTIC MOTOR
HDS REFORMING GASOLINE
HY. N A P H T A
JET FUEL
KERO

GASOIL DIESEL
LGO
HDS DOMESTIC
GO FUEL OIL
α LGO GO
HGO
3
OC VAPOR CAT PROPANE
RECOV. POLY.
CATALYTIC BUTANE

CRACKER
LONG
RE S I 0 SULFUR FUEL G A S
LVGO REC.
SULFUR
HVGO

VIS -
SHORT RESID HEAVY
BREAKER
FUEL
OIL
ASPHALT

Fig. 1 Simplified unit configuration of the Naantali refinery.

2. Neste Oy Naantali refinery atmospheric crude unit nr 1 and vacuum unit control scheme
 Fig. 3. Illustration of separation efficiencies in a crude distillation system.

GO P U M P A R ODU N

Ο». Ο
TO
CRUE D 1
H E A TRE

TO
CRUE D 2
H E A T RE

GO
P U M P A R O DU N

Fig. 4. The integrated heat exhanger network of Crude Unit 1, Crude Unit 2 and vacuum
feed preheat trains.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

CONTROL STUDIES OF A COMPLEX GAS

PHASE REACTOR SYSTEM
D. B. Aldren and P. J. Bujac
ICI Pic, Engineering Department, Β runner House, Winnington, Northwich,
Cheshire, UK

Abstract. A dynamic model of a continuous gas phase reactor has been developed. The
purpose of this model has been to improve the understanding of a highly complex exo-
thermic multi-stage reaction system with the objectives of increasing both plant
throughput and reaction selectivity. The model has been used to examine the dynamic
relationships between the main variables; feed rates, operating temperatures and heat
inputs.

The simulation studies described in this paper are a significant advance in explaining
the behaviour and control of the reactor. Methods of improving this control and of
increasing plant output are indicated.

INTRODUCTION

The plant being studied is a continuous gas phase
reactor system which includes a fluidised bed
followed by reactors in series. The plant was
designed, built and commissioned in a very short
timescale with only limited data available from
pilot plant trials. During the post-commissioning
period difficulties were experienced in maintaining
steady state operation and realising design produc-
tion rates. The optimisation of the complex system
proved difficult due to the large number of plant
variables which could be adjusted. A dynamic model
was built to study the interactions of the process,
plant and control systems. The primary objective
was to improve the product make by improving the
understanding of the system and developing operating
strategies. This paper outlines the development of
the dynamic model and its subsequent use to assist
plant operation and examine alternative control
systems.
Reaction 1 requires a catalyst and is carried out
in a fluidised catalyst bed. Reactions 2 & 3 take
place both in the fluidised bed catalyst disengage-
ment space and in a downstream secondary reactor.
Figure 1 illustrates the overall reactor system.
The use of a secondary tubular reactor is to in-
crease the yield of product Y. For sequential
reactions it is known that yield is encouraged by
a plug flow system. Figure 2 indicates the in-
crease in yield compared to that for a single stage
system.
The main control sytems for the reactor include
a independent flow control for the feeds A, Β and
C, and diluent Nitrogen.
b control of the fluidised bed temperature using
circulating heat transfer oil.

THE PLANT AND PROCESS
For commercial reasons, the specific details of the
process described in this paper cannot be given.
However the process involves sequential gas phase
reactions; the dominant reactions are
1A + B + C = X + P
2X + B = Y + P
3Y + B = Z + P
where X is a reaction intermediate
Y is the desired product
Ζ is an over-reacted by-product
Ρ is an inert by-product
Other reactions occur with both further intermediates
and by-products being formed. Components X,Y and Ζ
account for some 75% of the total products. The
reactions have also been simplified by ignoring under
reaction for 1, ie all A is reacted with X, and by
considering compounds with similar extents of re-
action together. This assumption is defended on the
grounds that the purpose of the model was to predict
the dynamic variation in Y made rather than to
predict the complete product spectrum.
c control of heat input, to the transfer pipe and
secondary reactor walls, using electrical
tracing.
Control of reaction 1 in the fluidised bed had
been good, however, control of reaction 2 and 3 in
the reactor headspace and secondary reactor had
been poor. The reaction is exothermic with the
heat being removed by temperature increase of the
gases, which include diluent nitrogen, and heat
losses to ambient. This heat balance which
essentially governs the temperature and extent
of reaction is sensitive to changes in reactor
feed rates, bed temperature and trace heating
inputs. With relatively small reaction exotherms
and significant heat capacities in the equipment
walls, there are several long lags in the system
and steady state control is difficult to achieve.
To improve the control and to allow operation at
higher rates, a better understanding for the
system response was required. To achieve this
understanding for a complex interactive system,
it was decided to develop a dynamic model of the
reactions occurring downstream of the fluidised
bed.

67
68 D. B. Aldren and P. J. Bujac
DYNAMIC MODEL
The dynamic simulation language used to model the
1
plant is the 'Advanced Continuous Simulation
Language , ACSL (ref 1 ) . The language is a high
level, FORTRAN based, user friendly package which
may be used interactively to give on-line control
of the simulation. The models are prepared from an
equation or block diagram description. Libraries
of commonly used model elements have been created
within ICI enabling complete models to be easily and
1
quickly assembled using a building block approach.
The "lumped modelling technique, which assumes
homogeneous elements, has been used extensively; with
distributed systems being modelled by dividing into
a finite number of lumped elements in series. The
dynamic elements of the model include the reactor
headspace, transfer pipe and the secondary reactor.
Main Assumptions
As with all models, there is a compromise between too
simple a system description and too much complexity
which leads to excessive computation time and over-
specification of parameters. To achieve a time and
cost effective model, certain assumptions have to
be made. In this model the following assumptions
were made
a A simple non-dynamic and empirical model was
chosen to describe the fluidised bed. This model
was based on previous laboratory and kinetic studies
and describes the distribution of major reaction
components leaving, the fluidised bed. The model
related the concentration of 'X' to the primary
parameters, bed temperature, residence time, and
'Β/Α' feed ratio. For a given 'X' concentration,
the distribution of Ύ ' and 'Ζ' were calculated.
b Reactions 2 and 3 are described as first order
Arrhenius type expressions
d(Y) = - djX) = A2 exp (-E/RT) (X) (B)
dt dt
d(Z) = - d(Y) = A exp (-E/RT) (Y) (B)
dt dt
3
where (Β), (Χ), (Y) and (Z) are molar concentrations
Ε is activation energy, assumed the same for
both reactions
Τ is absolute temperature
R is the universal gas constant
c The reactor sytem was described as a series of
continuous stirred tank reactors (CSTRs) or well
mixed vessels. The flow pattern in the reactor
disengagement headspace will be complex; a simple
2 cell model was however considered to be adequate
on the basis of previous plant data. Although the
transfer pipe and secondary reactor are more properly
described by a plug flow model, most of the reaction
is complete before the gases reach this section and
no significant loss in accuracy should occur if a
CSTR model is used for each unit. This simplification
allows the definition of a basic model element (the
cell) to describe the various reaction stages.
The basic cell
The basic cell with typical input and output is shown
in Fig 3. The model consists of the following
specific elements (MACROS) and standard elements from
the ICI libraries.
1. REAC: describes the reaction kinetics for a CSTR
for reactions 2 and 3. In addition the MACRO
calculates the heats of reaction and component
concentrations.
2. ΗΡΑΝ: describes the heat and mass balance for
the gas flow reactor wall, insulation and heat
losses to the surroundings. The heat balance for
the gas flow assumes a well mixed CSTR. The
MACRO has been configured to allow additional
N2 and 'B' flows at specified inlet tempera-
tures. A heat flowrate term has also been
included to allow for catalyst carry over, say
from the fluidised bed into headspace. In
practice this facility was not used.
3. TRANI0: models a temperature transmitter as a
first order lag with a selectable time constant
This was used for temperature measurements
throughout.
4. C0NTB: models a three term controller with
selectable tuning parameters. This was used
for temperature control of the heat tracing
on the transfer pipe and secondary reactor.
5. MOTOR: models a valve actuator with selectable
stroke time for the Nitrogen control, valve.
6. GVALV: models a gas valve with selectable
characteristics. This was used for the addi-
tional Nitrogen flow control to the fluidised
bed. The valve sizing gives a maximum
Nitrogen flowrate of 100 mass units/hour
Additional features included in the model are:
a Continuous calculation of the component distri-
bution from the fluidised bed allowing the bed
and feed flowrates to be changed during
simulations.
b The inital section of the model calculates,
the Arrenhius constants, the initial concentra-
tions for each cell, the average specific heat
of the gas, and converts flowrates from mass
to molar.
The main inputs to the model include
Flowrates for A, B, C and Nitrogen
Bed temperature
Wall set point temperatures
Typical outputs include for each cell
Product distributions
Gas temperatures
Wall temperatures
SIMULATION STUDIES
The model has been used for both steady state
and dynamic simulation studies. Steady state
simulations may be carried out by effectively
ignoring the large time constants of the reactor
walls and insulation eg by setting wall and insu-
lation densities to small values. Steady state
results have already been used to determine stable
operating points for the plant to give improved
product yields.
This report is concerned with dynamic simulation.
Dynamic simulation runs were carried out
a on the existing plant control system to obtain
a better understanding of the response of the
reactor to changes in the main control
variables,
b on alternative control systems and potential
plant developments
For these runs, the model was used to describe the
system response, from steady state, to a step
change in one of the main operating parameters.
 Complex Gas Phase Reactor System
The base case operating point used in the majority
of the simulation runs is described by the following
Flows to the fluidised bed A 1 0 0 Mass units/hour
Β 3 5 0 Mass units/hour
C 1 0 0 Mass units/hour
N2 1 5 0 Mass units/hour
Fluidised bed temperature 3 5 0 degC
Transfer pipe wall temp set point 3 5 0 degC
Secondary reactor wall temp set point 3 5 0 degC
This base case corresponds to a final reactor exit
gas composition of
X 3% molar
Y 50% molàr
Ζ 20% molar
Others 27% molar
Extensive sensitivity trials have been undertaken
to examine both over and under reaction. The
simulations included in this paper give an indica-
tion of the plant responses and the main interactions
Gas Temperature Control
The headspace gas temperature has no direct closed
loop control system on the existing plant. The
temperature is manipulated by adjusting the flow
controller setpoints.
The transfer pipe and secondary reactor control
systems were originally designed to control the
gas temperatures by manipulating the heat input to
the respective walls by means of electrical heat
tracing. The model was used to confirm that the
control systems were impractical due to the large
difference between the gas residence times and the
thermal time constants of the walls eg for the
transfer pipe the gas residence time is typically
4 seconds whereas the thermal time constant of the
wall is approximately 1 2 0 0 seconds. The model was
used to investigate the effect of reconfiguring
the control sytems to maintain constant wall
temperatures. Installation on the plant has pre-
vented damage to the heat tracing from high
temperatures and stabilised the plant by eliminating
oscillations in the wall temperatures. This
modification is considered part of the existing
control sytems throughout the rest of the study.
Response to 'Β' flowrate changes
As both these simulation runs and plant experience
indicate that the system is particularly sensitive
to changes in the 'Β' flow, this response will be
examined in some detail. The 'Β' flowrate was
reduced from 3 5 0 to 3 3 0 units/hour at 0.5 hours
simulation time. Figs 4 and 5 shows the rapid
reduction in the gas temperatures throughout the
plant eg headspace temperatures drop by 3 0 deg C
in a few minutes. The gas temperature continues
to fall at a reduced rate after 8 hours. Fig 6
shows the headspace wall temperature responses
confirming the long time constants. Fig 7 and 8
show the gas component concentrations exit the
secondary reactor. The initial rise in product
concentration is followed by the slow reduction as
the simulation continues. The results demonstrate
the difficulties experienced by plant operators
in attempting to maximise product make at accept-
able by-product concentrations.
Responses to reactor wall temperature changes
The transfer pipe has been chosen to demonstrate
the effect of changing the wall temperature,
achieved by changing the controller setpoint. With
69
a 'Β' flowrate of 3 3 0 units/hour the transter
pipe wall temperature set point was increased
from 3 5 0 to 3 9 0 degC. Fig 9 shows the response
of the gas temperatures in the transfer pipe and
secondary reactor increasing by 2 0 degC in 0.4 hrs
Fig 1 0 shows the effect on the product concentra-
tions in the transfer pipe and exit the secondary
reactor. The transfer pipe wall temperature
takes approximately 0.3 hours to reach the new
setpoint. The wall temperatures must be optimised
as part of the overall plant operating strategy.
Alternative Control Systems
Sensitivity trials using the model indicated that
the headspace gas temperature could be controlled
by several methods, with improvements in the over-
all control of reaction. The most practical
and simple method is to control the flow of
diluent Nitrogen to the bed. The advantages of
controlling the nitrogen are; it does not affect
the stochiometry of the reactions, it is readily
available a reasonable cost and presents no post-
processing problems. The model has been used to
examine the effects of controlling the upper head-
space gas temperature by adjusting the nitrogen
flowrate between 0 - 1 0 0 units/hour above a base
flowrate of 1 2 5 units/hour, required to fluidise
the bed. The controller set-point has been
initiated to approximately .the value found in
previous results for the same 'Β' flowrates. A
simulation run with 'Β' increased from 3 3 0 - 3 5 0
units/hour has been undertaken. The controller
tuning parameters used were
Proportional gain = 1 0 . 0
Integral action = 4 0 seconds
Derivative action = 4 0 0 seconds
Fig 1 2 shows the exit product concentration and
the main waste by-product concentration increasing
by 0.5 and 1% respectively. The simple closed
loop control system is difficult to tune to give
improved response. Some of the factors limiting
the performance of the control system are; the
measurement time constant of 1 5 sees, the valve
stroke time of 1 0 sees, the lag of the nitrogen
flow through the fluidised bed (not modelled) and
the limited heat capacity of the nitrogen entering
the headspace at the bed temperature of 3 5 0 degC.
The lags are significant when compared to the gas
residence time of the headspace which is typically
4 0 sees. The simulation demonstrates the head-
space temperature can be controlled to improve
product yield. Further developments such as
introducing nitrogen at a reduced temperature
directly to the headspace have been simulated
with improved control. However, no plant data on
the effects of the mixing in the headspace are
available. The simple control system would
require minimum plant modifications, whereas the
more complex schemes would require significant
and costly changes.
CONCLUSIONS
The model explains the observations from the
plant that the reactor system is difficult to
control ie it is not easy to maintain optimum
product concentrations in the off gas. The
simulation runs indicate the highly sensitive
and non-linear response of the *B' flow, the
rapid response of the reacting gas system and
the long lags of the reactor walls. With
variations in feed rates, general "noise", and
limited off gas analyses (only every few hours)
it is not surprising that good steady control is
difficult to achieve.
Plant experience and the model results have
already shown that the original strategy
70 D. B. Aldren and P. J. Bujac

of controlling the secondary gas temperatures by Further work is continuing to improve product
adjusting the electrical heat input into the yield with several developments including those
transfer pipe could be improved. This work has outlined in this paper being investigated in
shown that maintaining a constant environment in the more detail.
transfer pipe and secondary reactor, by controlling
wall temperatures, is more appropriate and that REFERENCES
change has already been made. This study has also
shown that the control of the reactor could be 1 ACSL User Guide/Reference Manual, MGA
improved by headspace temperature closed loop control Associates USA
Further simulation runs and plant validation work
are underway to confirm these advantages. 2 Dynamic Model Libraries Vol 1, PANMAC &
CEGMAC, User Guide/Reference Manual,
The dynamic modelling has formed part of a wider Control Applications, ICI Engineering Dept.
de-bottlenecking and development study. The results
of the overall study have been that plant perfor-
mance has improved from operating at 70% of design
capacity to operating for sustained periods at 200%
design.

Fig. 1. The Reactor System

EXTENT OF REACTION

Fig. 2. The Product Distribution

 Complex Gas Phase Reactor System
71

CELL ( i )

ADDITIONAL FEEDS

FROM CELL ( i - l )
TO CELL(i+1)
X .

REAC
T+1
REACTION
"i-l
KINETICS i+1

ADDITIONAL
GAS FLOW

TOTAL
GAS FLOW GAS FLOW
FROM CELL(i-l)

QC .

WALL

INSULATION

VQLOS

Fig. 3 . The b a s i c modelling cell.

IS8L_. Tgft BEL - - -

w
Οι
τ450.ο ο 450.
σ-
•450. G
σ-
Ό αι αι
43C. S 430, .30. Ξ
ο
OJ 00
-· 410. -η 410. •410. ^
Ή
Φ •Η Ή U
Ο αι h i 01
390. = .1390. ° 390. -390.
α α. Ο.

£ § 370. -370. f
~ 370. - ω
350. J350. ro
3.0 4.0 Ç.0 6.0 3.0 CD
,ό
3.0 4.0 5.0 8.0 *
.* 350. Time ( h o u r s ) Time (hours)

Fig. 4. T h e h e a d s p a c e gas temperatures. Fig. 5. T h e t r a n s f e r pipe & secondary

reactor tempera^tur.

DCCR-F
72 D. B. Aldren and P. J. Bujac

IPi. -XI.
, 360. τ380. u ς 20. 53.0
355. 375. S w 1G 52.5
35€>. j- -370. χ -
52.0
345. Ρ -365.^ >
c 12. 51.5
340. ο c
U 51.β ο
335. • 'Τ5 8.
y- 50.5 %
330. • -350. Ι ·*-
50.e *J
i 4. c
325. • 7 υ 43.5 ϋ
32©. <4θ. ή
ο 0. 43.$ c
3.0 4.0 ç.e> .<? 8.<? 3.e 4.0 5,e ο
Time (hours) .β ο
Time (hours)

Fig. 6. The h e a d s p a c e wall temperatures. Fig. 7. Exit concentrations Χ,Υ.

TBffc.

16
Ν ·

>-
C
<H 4.

3.? 4.<?
me ( h o u r s 0.20 0.30
Time ( h o u r s )

Fig. 8. Exit concentration λ . Fig. 9. The transfer pipe & secondary reactor
gas temperatures.

J350. <u
0.20 0.3O
0.20 0.30
Time ( h o u r s )
Time ( h o u r s )

Fiq. 10. Product concentrations Fig. 1 1 . The h e a d s p a c e & s e c o n d a r y reactor

gas temperatures.

S Y 3

^ 54.0r
3ϋ 5 3 . 5 -
m
53.0- Χ
>- 52.5 -
-6.
c
ο
c 52.0-
•4. TJ
•2 51.5 - m
J 51.β-

£ 5^.5-
* 50. « 0.20 0.2
c 0.00 Time (hours)

Fig. 12. Exit concentrations.

Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

AN ADAPTIVE CONTROL ALGORITHM FOR

A THERMAL CRACKING FURNACE
C. McGreavy and D. Odloak
Department of Chemical Engineering, University of Leeds, Leeds LS2 9JT, UK

Abstract. A nonlinear first order model is used to represent the dynamic behaviour of
a pyrolysis furnace. In particular it is capable of being readily identified using
either control or load change variables. The proposed nonlinear model enables the
furnace dynamics to be fitted very satisfactorily. This is because the dynamic
behaviour of hydrocarbon pyrolysis furnaces changes with time as coke is deposited on
the internal tube surfaces, so that it is necessary to continuously adjust the model
parameters to the current state, good performance can be expected from adaptive
control. To do this, it is necessary to develop control algorithm based on the
proposed nonlinear model. In particular, it is possible to adapt the classical PID
controller by continuously updating the controller settings as new model parameters
become available. Investigations covering typical operating conditions encountered in
an industrial furnace show that significant product and energy savings are possible
using the proposed adaptive strategies.

Keywords. Adaptive control; thermal cracking; PID controller; nonlinear model; penalty
function.

INTRODUCTION effective (White, 1976), although the computational

The olefins plant is an important link between the
refining of crude oil and the chemical industry
since it provides a means of cracking the petroleun
hydrocarbons into the basic building blocks of the
petrochemical industry. The amount of material
processed annually is very large and since the
feedstock represents a very significant fraction of
the total operating costs there is a need to
optimize the performance of the plant continuously.
Indeed, even marginal improvements, beyond
analytical detection, have a major impact on
incremental profitability. The principal factor
which influences high performance is the control of
the temperatures in the pyrolysis furnaces since
they ultimately determine the product distribution
and hence yield from the process. A typical unit
is illustrated in Fig. 1.
efficiency is not very high. Other methods such as
minimum variance (Davidson, 1973) and extended
Kalman filter (Joshi, 1973; Jazwinski, 1970) which
allow for model errors in the covariance of the
measurement noise are more refined but add further
to the complexity of the resulting algorithms.
Moreover, the optimal control law is formulated on
the basis of the model being an exact
representation of the system (Sage, 1968) and so
limits the effectiveness of the approach.
An alternative method of addressing the problem is
to use the PID algorithm and tune the settings
on-line. This can be done by using a supplementary
objective function to optimize the controller
settings (Gill, 1974). In fact, this objective
function can be chosen to be identical with that of
the optimal state variable controller.

While it is possible to build a model of the The problem can now be restated in the following
process which is capable of predicting the way:
behaviour reasonably accurately, the detail which
has to be included make it unsuitable for direct Given a plant represented by the discretized
use in computer control. Some simplication is approximate equation
necessary and a trade-off has to be made in terms
of the improved performance due to better control
against the comprehensiveness of the model. It is χ . = f(x , a , u )
generally not feasible to apply optimal control -k+1 ~k -k k
laws derived from state variable methods based on
the large non-linear models used in simulation where χ is the state vector and u the
studies. control and the process parameters^are represented
by at
However, an approach based on a suboptimal policy time kT (T is the sampling time). The measurements
can give good control and can usually be are given by
implemented very easily. For example, a linear
second order model with transport delay can be used
to approximate the effect of the fuel flow rate on
the exit temperature. Such models have been found
to approximate high order processes reasonably well
and lead to simple algorithms for the control law, where ν is the measurement noise.
while the transport lag offers an extra degree of
freedom without increasing the order of the system. The values of a_ for the model are chosen so as to
To use it effectively, however, requires the model minimize a function of the form:
parameters to be adaptively updated on-line.
J = ) +
This technique has been widely used and the I ^ Σ P H
k k
evidence seems to point to direct search methods k=o
k
based on least squares estimates as being the most

73
74 C. McGreavy and D. Odloak

where ρ, ζ > 0 . Gas volume

Where load changes occur, a model based on the T +b

simplified second order model with delay is prone
*1 * g l ^ v(H
to failure because the effect of this disturbance F - +H ( T g) )
on the parameters is to prevent the identification
- Q
algorithm from converging. In fact, the model is p= 0 (3)
only really suitable for servo-loop control
(Odloak, 1980). For regulatory control a more
detailed and realistic model is called for which T = 1.48236 Τ + 10.1847 W - 0.61058 Τ
Τm ρ ρ ο
explicitly includes the effects load changes. For
the pyrolysis furnace, this means it is necessary
- 25.7892 (4)
to account for lags arising from the refractory
walls and the tubes, as well as the dynamics of the
combustion zone. The consequence of this is a
Q = 3.80878 Τ + 345.2445 W - 0.44438 Τ
higher dimensionality model but one which, by Ρ Ρ Ρ
careful evaluation, can be reduced to manageable
form to enable advantage to be taken of some of the - 2725.06 (5)
concepts derived from the state variable methods
(Odloak, 1980).

In discrete form, the control law becomes:

CONTROL PROBLEM

u(k+l) = u(k) - K
While the advantage of state variable control c {[T (k) - Τ (k-1)] -
methods is attractive in offering the prospect of
achieving high performance from industrial
- &
processes, it is often difficult to gain e <e
fc> -
t w * - » ] * -
acceptability with operating staff. Moreover, the
full benefit is only realizable when the system is Κ ΔΤ
) T [ sT] ep (
t k( k )
well defined. On the other hand, the classical
approach using three term PID control tends to be - - f r " "
more easily assimulated by operators, but the κ τ () k 2( )Tk _+τ2 1
performance tends to be non-optimal because tuning
" ~ i r & " ο<- )] (6)
is essentially based on approximate dynamic models P P ρ
derived empirically from limited testing.

Clearly, a blend of these two philosphies has some

attractions where the optimality properties of the
state variable formulation is used for the control
algorithm while the basic simplicity of the three
term algorithm is retained as the interface with
the user. In particular, there is scope for Κ ΔΤ Κ Τ,
employing algorithms which will minimize the loss Let = T. and *f
1
= T, where ΔΤ is the sampling
of product quality during transient load changes Τ£ ΔΤ d
without operators having to continually adjust the period
controller settings. For example, this could be The problem consists of determining the values Κ ,
accommodated by including in a performance index, T. and Τ, such that it minimizes a performance
the control effort or energy spent to bring a index of the form:
system back to a desired set point as well as a
penalty for degradation of the product due to any Ν 2 r
off-set from the desired conditions. S = J {qpT (i) - Τ ] + pu(i) }
^ ρ set-J
The simplest approach is to use a conventional PID
controller and set the parameters by carrying out a It should be noted that the difference between the
three dimensional search in order to minimize the model output temperature and the set point is
selected performance index. To follow this weighted by a factor q, but the fuel flow rate (and
procedure a minimum realizable set oi equations not the difference) is weighted by p. The choices
describing the furnace is required. This needs to of ρ and q depend on the relative costs of energy
account for the changes which occur in the furnace and degraded product. Ν is the number of time
and tube walls as well as the gas space. A intervals the system is expected to take to settle
suitable approximation is given by (Odloak, 1980): down. This value will obviously affect the
settings of the controller and the overall
Refractory wall performance of the loop. In general Ν should be
dT selected so as to be compatible with the setling
H + time of the system. Too small a value of Ν would
T? U L L
Α —f- = GS(2) Σ Τ + SS(1,2) Σ Τ _ 3στ_ (1) overweigh the proportional gain Κ while too large
sat g I K Ν would exaggerate the integral time T^.

Tube wall It is convenient to use a non-linear search method

to determine the best controller settings in
4 conjunction with variational procedure for
(88(1,1) - Α Ε ) Σ Τ * + SS(2,1) Τ £ + GS(1) Σ Τ - estimation of the model parameters.
ΤΤ Σ

- V ° (2)
 An Adaptive Control Algorithm

The first variation of Eq. ( 3 ) gives From Eq. (1)

3 3 3
1
4a ΣΤ° ΔΤ + 4b ΣΤ°3 ΔΤ_ + [
g g Ρ-
Η H ( T ° ) ] δα
Α —
dt
= 4GS(2) Σ Τ ° ΔΤ + 4ASS(L,2) Τ ° ΣΤ^+
g g Τ Τ
3
ο 3Η
U i
ΔΤ + 2 * ΔΤ = 0 + 4σ3τ° Δ Τ Κ
3T g SS(2,1) 3Τ ρ
G Ρ

and Δ Τ _ = 3 Δ Τ
Τ 2 ρ 0
03 Η 9 3 3 3
4Bj A T ° a
4a,
ι
ΣΤ
g
- «°
3Τ
ΔΤ
g
+
2+ 4 4ASS(l,2) T ° a
2+ 4Σ3Τ£ 0 ] Δ Τ
2 Ρ
g
9 Q o 3 3
+ [4GS(2) A T d
, SG(2) p
δΤ [H - H(T°)] δη
g+ 4A3T° d ]
9 δη

SS(2,1) 3 T
P +
F

Then
So

δΤ = C. δΤ + d, διι
g 1 Ρ 1

where _PF _
H ( T dt
a
3 δ Τ ρ+ d 3δη - ^
t d
["F -g>L
/ ΓΤ,ο ο 3H
4a, a T - u
g
From Eq. ( 3 ) The variational form of Eq. (4) is:
3 3
4 [S(L,L) - A ^ J ΣΤ° ΔΤ
Τ+ 4SS(2,1) Τ°
Σ ΔΤ
Κ+ J
ôu(k+l) = ou(k) - {rT°(k) - T°(k-l)1
Ρ Ρ
3Q
+ 4GS(1) ΣΤ°3 ΔΤ - ΔΤ = 0
g g °Τ ρ

3
Υ ν Υ ν
4SS(2,1) Σ Τ ° Δ Τ „ + 4GS(1) ΣΤ° ΓΟ,ΔΤ + D,6u] +
K K g ' Ρ ι
3 - Κ
c
ΔΤ (K) + Κ
ρ c
ΔΤ
ρ
(K-1)
'
4 [SS(1,1) - Α ^ ] Σ Τ
Γ °°
Τ ( ) kT k T
)
3 Q
- &p - set< ^ i " T?6T (k) -
P ΔΤ = 0
p
3Τ~ Ρ
Ρ
- [Τ°0Ο - 2T°(K-L) + T°(K-2)] Δ Τ
3
-1
ΘΤ
Η ,ο3
4 G S ( 1 ) AT° Cj+ 4(SS(1,1)
- τ" δτ (k) + 2RR δτ (k-i) - τ" δτ (k
4SS(2,L)AT' d ρ
d p d p
3
Α Ε ) ΣΤ° - ^ ΔΤ
ΤΤ
3
+ 4GS(1) Σ Τ ° d, δη
g ι

ΔΤ
Κ - C
0 Δ Τρ
2
+ d
06u
2
76 C. McGreavy and D. Odloak

du (k+1) = <5u(k) - {[T°(k) - Tp(k-l)] -

T ()k
" C set " W ^ ^ c "
T T( k ) 6X(k+l) = AôX(k) + BoX(k-l) + CÔX(k-2) + ϋδρ

" C p« " set >i

0 J
- |T (k) - 2T°(k-l) + T°(k-2)"î ÔT, -
i- ρ ρ ρ α where ρ is the parameter vector.
Τ++ T ( )6k T +
For a deviation from a known initial state
" Κ ι J
P
+ ΓΚ° + 2Τ°]δΤ (k-1) - Τ° δΤ (k-2) = 6u(k) yo)
c d-i ρ d p

U(0)
- ÔK
cr- iν26Ίί - r ôT
3 d- SjôT (k)
p
- s ôT (k-l) - s 6T (k-2) Eq. (7) can be solved to give
2 p 3 p
S 1
In Eq. (5) approximated by 6X(k) = W(k) δ Ρ

u (k+1) - ou(k) δΧ(0) = 0

ΔΤ
and becomes
W(k) is a matrix that can be determined by the
relat ion :
d T (k) T a+ +U d +
d
2
~ à — 3 p « 3 « Γ~ΔΤ

{r.OK + r ôT. + Γ„δΤ, + s.oT (k)

l e 0 ι
2 3 d 1 ρ W(k+1) = A W ( k ) + Β W(k-l) + C W(k-2) + D

+ β δΤ (k-1) + β δΤ (k-2)}
20 ρ 30 ρ

If u(k) is the fuel flow rate in the interval: k to

(k+1), the above equation cna be integrated to with W(k) = 0 for k < 0.
obtain:
The performance index now takes the form:
û i
δΤ (k+1) = e a 3 δΤ (k) + -L Ν T
Ρ " Ρ a s = ι [x(k)
k=l
- x s e] t <a[i(k) - x s e] t
{ d r 6 K +Ôr T +
where
3 **Μ ^ΔΤ
+t l c 2 i

+ r-δΤ, + s,6T (k) + β δΤ (k-1) + q ο

3 d 1 Ρ 20 ρ ρ set
and X

+ s ÔT (k-2)]} ο ρ
3 p
The variation set of equations can then be put in
the form:
Since X(k) = X (k) + X(k), X (k) is the trajectory
obtained with the values Κ , Τ . and Τ , for the
c i d
δΤ (k+1) A(l,l) A(l,2) ÔT (k) controller parameters.
Ρ p
Imposing the condition for a minimum:
ou(k+1) A(2,l) A(2,2) 6u(k)
Ν
X Ν
T
I W (k)Q [X°(k) - X (k)] = - I W (k)Q W(k)ôp
B(l,l) 0 "δΤ (k-1) " " C(l,1) 0" q e t k=l
Ρ
+
k=l

B(2,l) 0 δη (k-1) _ C(2,l) 0

δΤ (k-2) D(l,l) D(l,2) D(l,3) δκ

Ρ c

6u(k-2) LD(2,1) D(2,2) D(2,3)

δΤ,

N
δ {
τ _ ι Ν
= " l W (k)QW(k)} { I
ρ k=l k=l
An iterative procedure can thus be developed by
assuming the values K , 1\ and and using the
£
above relations to find an up-dated set based on
the variations δΚ δΤ. and δΤ„.
c i d r
For the case where α = 59, 3 = 15.81, methods based
on the reaction curve do not give acceptable
settings because the system response cannot be
approximated adequately by a first order plus time
delay transfer function. The settings given by
these methods are too conservative and cannot be
taken as a basis for comparisons. Consequently, it
is necessary to use the concept of maximum
proportional gain and ultimate period for the
closed loop response, and employ the Ziegler and
Nicholas method to set the PID controller which
results in Κ = 0.15, T. = 0.05 and Τ, = 0.12 being
i he preferred
i d
controller values.
The loop response obtained with these controller
settings can be compared with the controller that
minimizes the quadratic performance index using the
proposed algorithm. In searching the 3 parameter
space, the range of convergence of the proposed
method proves to be very narrow and the algorithm
diverges even using the Ziegler and Nichols
settings as the initial values of the parameters.
The method proposed here is much more stable and
has a broader range of convergence while only
requiring a search in two directions with one
parameter kept constant. Figure 2 shows the
response of the loop based on Ziegler and Nichols
settings as well as that using the approach
described above, with T^ being taken as constant.
There is an improvement in the performance index
from 15.75 to 15.51 showing that, for this
particular case, the empirical adjustment procedure
approaches an optimal strategy. For this case,
with a weighting matrix
1 0
Q =
0 0.05
12 iteractions are necessary for convergence.
Figure 3 shows the behaviour of the parameters and
performance index at each iteration. Although the
number of iterations necessary for convergence is
quite large, the computational effort is not great
and the solution can be obtained rapidly and
certainly real time, so there is no difficulty
using the algorithm for on-line control.
To gain some impression of the circle of
convergence for this algorithm, instead of assuming
the Zielger-Nichols settings (K = 0.15, T. = 0.05,
T, = 0.12), significantly different initial values
(K = 0.05, T. = 0.02, T - 0.12) are used. Figures
c d
3 and 4 show \ h e trajectory for the parameters and
the performance index in both cases. It can be
seen that all converge to the same values as in the
previous case. This is representative of the
behaviour and more extensive tests confirm that the
algorithm has excellent convergence properties.
The best controller parameters for load changes and
Q prove to be K = 0.135 and T = 0.0767.
£ i
To have acceptable performance, it is essential to
verify that the optimum PID controller has been
achieved for set point changes. Using the
An Adaptive Control Algorithm 77
Ziegler-Nichols algorithm there is no change in the
controller settings because the stability limit for
w'Ck^Q
the closed loop is the same for both load
disturbances and set point changes, so the same
values are used. The best settings would not be
expected to be the same as for load changes because
the system is non-linear. In fact for a 2 C
increase in the set point, the best controller
parameters are Κ = 0.1324 and T. = 0.01469 (with
T kept at 0.12). For this case, the advantage of
d
using the proposed setting algorithm over the
traditional method is even greater. The
performance index for the proposed method is 12%
lower than the with Ziegler-Nichols method. This
means that during the settling time, if the
weighting matrix is:
1 0
0 0.05
12% less product and energy have been wasted. This
is a considerable saving in the operation of
pyrolysis furnace. Since large amounts of energy
and products are involved, any improvement in the
control policy results in very significant savings
per year.
Because the relative weight of product quality and
energy spent in the control can change depending on
economic factors, it is useful to verify the effect
of the ratio r - Q(2,2)/Q(l,1) on the performance
of the optimum PID controller. Figure 5 shows the
closed loop response for r = 0.01, 0.1 and 0.5.
The loop response becomes more damped as the
relative importance of energy spent on control is
decreased, as would be expected.
CONCLUSIONS
It has been shown that a control policy can be
advantageously defined using an adaptive model of
the pyrolysis furnace. When used in conjuction
with a PID controller tuning can be by means of the
conventional empirical methods or by a direct
search of the controller parameters which seek to
minimize a quadratic performance which takes
account of product quality and energy consumption
index. An algorithm has been proposed which has
fast convergence characteristics, and is very
effective in updating the proportional gain and
integral action time of the controller. It has a
large circle of convergence, which makes it
suitable for on-line updating when the model
parameters change significantly with time. In all
cases the performance index which is to be
minimized is less than can be achieved by taking
the stability limit values prescribed by the
Ziegler-Nichols method, and has resulted in savings
of up to 12% in the cost index.
NOMENCLATURE
aj Gas absorption coeffienct
a2 Proportionality constant
A-j, Tube wall area per unit length of
the convection
b j Tube absorption coefficient 2
GS Gas-surface direct exchange area (m )
Ρ Weighting factor for loss of product in
performance index
q Weighting factor for cost of control in
performance index
Qp Heat flux (kw/m )
Z
SS(i,j) Surface i - surface j direct exchange area
(m )
P^ derivative time (s)
Τρ Furnace temperature (K)
Tg Flue gas temperature (K)
TR Refractory wall temperature (K)
Τ-ρ. External tube wall temperature (K)
78 C. McGreavy and D. Odloak

Ziegler and Nicholt (Kc » 0.15. Ti » O.QSJd Ό.12)

Sampling time (S)
time (s) PI opitmum l xo. ^ dβ (
(Kc «0.1278, Ti - 0.0716. Td«0 12 )
Gas velocity (m/s) *C Index 15.51 kg/3

GREEK SYBMOLS

Fractional disappearance of the reactant

Surface absorptivity
Adjustable parameter for fitting furnace
model (radiation) ^
γ Rate of reaction (Kmol/m s) 2 4
σ Stefan-Boltzmann constant (Kw/m s Κ )
ε Surface emissivity
8
REFERENCES

Davidson, J.M. (1973). Ph.D. Thesis, Texas A & M

University.
Gill, P.J. (1974). Ph.D. Thesis, University of
Leeds.
Sage, A.P. (1968). Optimum System Control,
Prentice Hall.
Odloak D. (1980). A Study of the Dyanmics and
5.0 7.5
Control of an Ethylene Cracker, Ph.D. Thesis,
TIME (min)
University of Leeds.
White, D. (1976). Ph.D. Thesis, Princeton.

Fig. 2 PID controller performance - load change

Convection f^° fl Steel structures

„ o o o0 /
•faΟ Ο oy
oo ο

-Burners

-Peepholes
20
Cracking! radiantl ITERATION
coil

Fip;. 3 PI controller optimal parameters for

Refactory,
D • two different initial conditions

Fig. 1 A typical modern pyrolysis furnace.

Fig. 4 Optimal PI seeking algorithm for two

different initial conditions
 An Adaptive Control Algorithm

. 5 Closed loop response with PID controller

effect of the relative value of fuel and
product
r = fuel consumed/product rate
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

DYNAMIC SIMULATION OF RESIDUE

HYDRODESULFURIZATION TRICKLE-BED
REACTOR
Yung-Cheng Chao, Jhy-Shyong Chang and Hsiao-Ping Huang
Department of Chemical Engineering, National Taiwan University, Taipei,
Taiwan

Abstract. An adiabatic residue hydrodesulfurization trickle-bed reactor packed with

a porous catalyst undergoing deactivation has been simulated numerically in order to
examine the dynamic behavior of this specific reaction system. One dimensional pseudo-
homogeneous model incorporating the effects of mass and heat dispersion is used in
mathematical derivation. The parameters used in the simulation are primally based on
the experimental data by Shoji Kodama and other correlations. The method of orthogonal
collocation is used to obtain the solution of coupled mass and energy balance equa-
tions. The catalyst deactivation model adopted was proposed by Shoji Kodama to include
the interaction of demetallization and coking reaction on a catalyst. The performance
of the reactor during start-up period is examined. In the meantime, step-wise changes
of feed composition, feed rate, inlet temperature on the dynamic behavior of the
reactor are investigated. It is concluded that the study of the whole simulation will
be helpful in operating an existing commercial residue hydrodesulfurization reactor.

Keywords. Distributed parameter system, Dynamic response, Hydrodesulfurization, Oil

refinery, Trickle-bed reactor.

Introduction

Direct desulfurization has become increasingly
important in meeting the demand for low sulfur
fuel oils that has been created by man's desire to
maintain a high quality environment. Residue
hydrodesulfurization (HDS) in trickle-bed reactor
is one method capable of providing low-sulfur
fuels that will meet restrictions on sulfur oxide
emmisions. The two reactions of primary interest
in the residue HDS process are desulfurization
and demetallization of residuals. Since the feed
to HDS trickle bed reactor is a mixture of gas
and liquid, the problem of dynamics of vapor-
liquid phase with catalyst deactivation is more
complicated than that of a usual packed-bed
reactor with homogeneous feed. Only very few
information on the dynamics of vapor-liquid phase
reactors are found in the literature (Yan, 1980).
However, there is still no information in the
literature on the dynamic behavior of residue HDS
reaction system. In the same time, to face the
possible change of operating variables such as
inlet feed temperature, feed rate etc., the study
of dynamic of residue HDS reactor is indispensa-
ble, and this is the main objective in this
study. To set up a reactor model for suitable
description of the dynamic behavior of HDS reac-
tor system, the following key factors are consi-
dered.
(1) A suitable catalyst deactivation model which
cover explicity the possible variations of feed-
stocks and HDS catalysts. Shoji Kodama (1980)
proposed a deactivation model which meets the
requirement and will be used in this simulation.
(2) A suitable model for describing the flow
characteristics in this simulated bench scale HDS
trickle-bed reactor (reactor inside diameter,
2.8cm; reactor bed length, 42cm). Among the
possible models are: (a) the axial dispersion
model of Mears (1971) and Shah (1975), (b) the
holdup model of Henry and Gilbert (1973), and
(c) the effective catalyst wetting model of Mears
(1974). Montagna (1975) concludes that the axial
disperison model fits the experimental data for
the hydrodesulfurization of 367o and 53% reduced
Kuwait crude better than the other two models in
the bench scale trickle-bed reactor.
(3) Suitable descriptions of mass and heat transp-
ort phenomena in the three phase operation set-up
of HDS trickle bed reactor. This will be consi-
dered in the derivation of mathematical model.
Mathematical model
The conventional residue HDS unit is a trickle-
bed reactor, that is, a cylinder containing a
fixed bed of catalyst, through which a mixture of
gas and liquid flows concurrently downward. The
reactor is adiabatic, so there is no radial
transport of heat. Since the ratio of reactor
length to diameter (42/2.8) is large and by
Montagna's conclusion (1975), the liquid distri-
bution within the catalyst bed is assumed to be
quite uniform and all the catalyst particles are
effectively wetted in this bench scale residue
HDS reactor studied. The fluid pattern can be
approximated by plug flow incorporating the
effects of mass and heat dispersion. To simplify
the model, the following assumptions are made.
1. The main reactions and the parallel poisoning
reactions occur within the porous catalyst
particles can be described by power-law rate
expression shown in Table 1 (digested from Sho j i
Kodama, 1980).
2. The time constant to reach flash equilibrium
between liquid and gas phase is much less than
reactor space time, therefore hydrogen concentra-
tion in the residue is assumed to be affected
only by reaction pressure and temperature of the
reaction system and the solubility of hydrogen is
given by the Eq. (4) in Table 1.

81
82 Yung-Cheng Chao, Jhy-Shyong Chang and Hsiao-Ping Huang

Table 1 Rate Expressions for the Reaction System axial distance and temperatures. If suitable
and Solubility Relationship of Hydrogen physical properties were available, the inclusion
in the Residue. (Shoji Kodama, 1980) of a variable velocity term could not cause any
complication.
Rate equation for d e s u l f u r i z a t2i o n
Table 4 Equations for the Calculation of Axial
A Ε _ K C, C (1-1) Mass and Heat Dispersion Coefficients
V <°cat os fs s h s
Rate e q u a t i o n f o r v a n a d i u m r e m2o v a l e m
r = V, d /D .
J P 1a (4-1)
r = Q ^ A E
ν £ Kν C,h Cν
c a t ov fv
(1-2) J | j = s,v
Rate f o r coke deposition (4-2)
re q=u a t i o n r
A K . . - / 0 A K C , q (1-3) J
c Peat oc c l cat ο oc2 n e Ρ .1 = 0.042R ,
0.5
(4-3)"
emj . el
|j=s,v
S o l u b i l i t y of h y d r o g e n i n t h e r e s i d6u e
Ρ =P (4-4)"
3
C - ( 8 . 9 1 10-6) Ρ + (4.16 ΙΟ" ) eH U
em
(T-273)-1.40 1 0 " (1-4) * Eq.(4-3) is the correlation of Hochman and
f o r 50<P<180, 573<Τ<373 Effron (1969)
** Eq.(4-4) is cited by Shah and Paraskos (1975)
3. The d e a c t i v a t i o n of t h e c a t a l y s t is adopted
from S h o j i Kodama's d e a c t i v a t i o n model g o t h e r e d
in Table 2 with the r e l a t e d parameters being With the above assumptions, the following unsteady
l i s t e d i n T a b l e 3. state mass and energy balance equations for the
reactor can be formulated.
T a b l e 2 C a t a l y s t D e a c t i v a t i o n Model
( S h o j i Kodama, 1980) Material balance for the external fluid phase
A = A (1-a W ν -a c W c ) (2-1) 2
OS ο ν
a c_ V
A
ov
= A (1-jS Ψ -jS Ψ )
ο
(2-2)
*
1
d _c
- = D 3 C d
' V V ' c c 2 " l -
dt
A = A (1-7 Ψ ) (2-3) dχ ax
oc ο 2
' V νV Ψ
D = D V (2-4) - p"\ ρ .A · E · e · Κ . C, · C (1)
s so Ci-*' )
1 cat os
r c s h s
fs
¥max

Table 3 Parameter Estimation (Shoji Kodama, 1980) for the Reaction System (Table 1)
and Deactivation Model (Table 2)

Correlation between Parameters and Properties

Parameter Values for "D" Catalyst/Iranian Heavy V.R.
of Catalyst/Feedstock
6 3 3
D 2.10X10"6 3 1.57x10' ?Xexp(-5X10 /RT)
7 2
so
1.57X10" exp(-1.75 10" /RT) 4.58* 10" Xexp(-4.32X10 )X P
D L 1
vo +6.27X10 X P -13.1XC -16.9-5X10 /RT)
10 3 ν 3 A
As
k 1.22xl0 1xe x0p ( - 3 1 . 3 x1 0 3/ R T ) k *χ exp(-30 103/RT)
x x A
s
1. 28 χ 1 0 2χ exp( -24. 2 1 0 3 k * exp(-30 10 /RT)
k X /RT) vo V * X
kV 3 . 0 0 x l 0 " x5e x p ( - 2 0 . 0 1 03 / R T ) 3 10 /RT)
X (Ρ 1 0 -0.0261) exp(-20
x x x
kc l 5.35xl0" Xexp(-10.0xl0 /RT) 5 . 3 5 χ 1 04
" χ exp(-10310 /RT)
c2
X
0.261(31.4wt7o) 7.4X10" X(3.48X 10 -1/P )X(0.32-C )
^ max ν A
As
« v/ « c 1.60/1.80 1.6/1.6
1.60/1.88 1.6/1.6
βν' jSc
rv 1.60 1.6 3 4
0.885 2.9X10 X(2.06xl0" +P )x(0.407-C )
ν A
As

*k /k : Values obtained-from Initial Activity

so vo

4. The reactor operates adiabatically.

5. The flow pattern is described by the axial
€
1 1 = D
V a V, 1
dispersed plug flow. Radial dispersion effects Λ t 2
of mass and heat are negligible. The related
dispersion coefficients are listed in Table 4.
l 2
6. Resistances to transport of mass and heat
r
-p' O
r -A - E · € -K -C.-C (2)
1 cat
r
ov f v c v h ν
between the external fluid phase and the particle
surface are neglected, but mass diffusional with the following boundary conditions:
resistance within the particles is taken into At bed inlet
consideration. The temperature of the catalysts
is assumed to be the same as that of external
dC ,
fluid at the same position (Yan, 1980).
V (C .-C I + ) = -D (3)
7. The fluid velocity, density, and axial disper- n n
sion coefficients are taken to be independent of 1 ei si x=0 as |
χ9 χ += 0
 Residue Hydrodesulfurization Trickle-bed Reactor 83

where p ^ is the density of solid catalyst; p^

v,(c ,-e +) = (4)
1 vi ν x=0 ax
n x=0
is the density of liquid fluid dissolved with
hydrogen; €^ is the volume fraction occupied by

At bed exit solid catalyst in the bed; C , C ., C are the

ps' pi' pg
heat capacities of solid catalyst, liquid phase
= 0 (5) and gas phase individually; Τ is temperature;
D , D„
d x x=L T1t Hg are the axial heat dispersion coeff i-
Hl'
cients of liquid ans gas phase respectively; r
is the reaction rate of desulfurization shown in
s
= 0 (6) Table 1; Hs is the reaction heat of desulfuriza-
dx x=L tion (Shah, 1975); Δ H and R are the heat of
' ' ν e
vaporization and the rate of evaporization respec-
where C , C , C are the concentrations of sulfur,
s ν u
η tively; is the superfical velocity of liquid
vanadium and hydrogen in the liquid; t is process
fluid dissolved with hydrogen.
time; χ is the distance from the reactor inlet;
f are the volume fractions occupied by
c For an individual particle to include the intra-
liquid and catalyst aparticle
e r Θ in the bed respec- particle mass diffusional resistance
tively; PqH> J Ρcat *"^ densities °f liquid
oil and catalyst respectively; D , D are the "oil s^
' a s ' av Ρ -A -K -C.-C' (12)
2 , s
J
axial dispersion coefficients for sulfur and r dr dr cat os s h s
vanadium respectively; A , A
' os' ov
are the effective r
surface areas of catalyst under use for desulfuri- 1 d , 2 _ oil \
—τ — ( r -D A -K -C,-C ' (13)
zation and vanadium removal respectively; Ef > 2 , ν cat ov ν h ν
g r dr ,
dr
E.. are the effectiveness factors for sulfur and
fv
vanadium inside the catalyst particles; k , k^ with boudary conditious:
are the intrinsic rate constants; is the
C. = c.
superfical velocity of liquid fluid; subscript i J (14)
denotes the inlet condition; 0 is the position
just after inlet surface of the reactor; L is
the position just before the outlet surface of
= 0 (15)
the reactor.
s
r=o v
Heat balance for the external fluid phase j =>

dT where D , D denote the effective diffusivities

s s sp l 1p
1 Η 1 ν c
of reactant inside the catalyst particle; r is the
at χ χ
99 radial position in the catalyst particle; R is
the radius of catalyst particle.
c
R - Δ Η --£-(λ,·0 · ν . Τ ) - r ·(- H )· (7)
e v ^ x ^ l p l lΊΊ s e
s c For the catalyst activity balance
T
e a a = lOOr lp (16)
% 0 .C
P% Pg^
= (D
γ Hg u
x) + R .λ H
e^ ν at
r α t aχ
a a
r
- l _ ( p -c · ν ·τ) (δ)
g pg g t_ c a c
a χ = lOOr / p (17)
a
T
Combination of Eqs.(7) and (8) gives Eq.(9): t
cr y a with the initial conditions:
s s ps 1 <pl pl gg Pë ^ t
Q (x,0) = f (x) (18)
T
a
λ
Hg Q (x,0) f (x)
c
(19)
ax ax
where Q , Q are the quantities of contaminants,
1 ν c
--^-((P^ C . V 0-C .VjT)-r (-AH).f^ (9) sulfur and vanadium on the catalys respectively;
ax
w l 1 ng+ |
"pi n f gt
pg ;
r and r are the rate expressions of vanadium
ν c
removal and coking reaction respectively shown in
with the boundary conditions:
Table 1.
At bed inlet
a Τ Following Shah (1975), the dimensionless groups
V, P C -.(Τ.-τΊ +) =-(D +D ) (10) of the above equations were defined, and Eqs. (1)
1 n u lu A
x=0 to (19) were reduced to a set of dimensionless
1 1 pl ι x=0 HI Hg
coupled second-order partial differential equa-
0
At bed exit tions and the problem is of boundary value type.
The method of orthogonal collocation (Villadsen,
1967) was used to tackle this distributed
parameter system. In the present work, eight
(11) collocation points for the external field and
ax x=L seven for the spherical partical were used. The
resulting 64 ordinary differential equations were
84 Yung-Cheng Chao, Jhy-Shyong Chang and Hsiao-Ping Huang

Table 5 Physical Properties of HDS Catalysts (Shoji Kodama, 1980)

Pore Surface Equiva- Packed

Catalyst Type 3
Vo1ume 2
Area lent Dia. 3
Density
(cm /g) (m /g) (mm) (g/cm )
A Ni-Co-Mo 0.440 280 1.16 0.740
Β Ni-Co-Mo 0.500 180 1.41 0.650
C Ni-Co-Mo 0.558 161 1.22 0.650
D Co-Mo 0.601 155 1.34 0.660
Ε Co-Mo 0.786 341 1.59 0.528

Table 6 Properties of Feedstocks (Shoji Kodama, 1980)

Khafj i Iranian Iranian Iranian

Kuwait Khafj i Boscan
Heavy A.R. Heavy A.R. Heavy V.R.
A.R. A.R. V.R. Crude
(S-l) (S-2)

Specific Gravity (15/4°C) 0.9734 0.9518 0.9624 0.9587 1.0379 1.0389 0.9978
Viscosity at 50°C, cSt 617.4 186.0 462.4 312.9 - - 5,612
at 100°C, cSt - - - - 2,890 6,401 200.4
Carton Residue, WT7O 11.5 9.1 9.7 9.5 21.6 23.8 15.9
Asphaltene Content, WT7O 5.9 2.6 3.7 2.8 8.2 13.0 11.5
Sulfur Content, WT7O 4.09 3.65 2.57 2.62 3.67 5.39 5.18
Nitrogen Content, WT7O 0.26 0.22 0.45 0.35 0.76 0.47 0.59
Metals, Vanadium, ppm 74 49 142 120 270 165 1,130
Nickel, ppm 27 12 54 38 92 53 106
Sodium, ppm 35 2 3 48 4 4 45
Iron, ppm. 2 3 3 14 13 13 4

Table 7 Reaction System's Parameters and Base Conditions

for the Start-up of the Reactor

Parameter Values: 3 Opera tin g Conditions

-ΔΗ
8 1.833xl0 kcal/kg T.(x,0) 658°K 2
0.7493 kcal/kg °K Ρ 140 kg/cm . G
Si
c 3.516 kcal/kg °K C (x,0) 0 kg/kg
pg s
c 3
26.76 kcal/kg 3°K C (x,0)
r ' 0 kg/kg 2
ps 0.66xl0 4kg/m 3 C . 2.62x10"3 kg/kg
si
PEAT 0.266xl04 kg/m 3 0.12x10" kg/kg
C
V.
Pc 0.596xl0 kg/m VI 0.135 m/hr
3 3 l
pv 0.79x10 kg/m (at system Ρ & T) H /oil 1000
Pi 3
2
3
5.056 kg/m 3 system Ρ & Τ)
(at
Ps 3.99xl0 kg/m
R 1.987 cal/g mole °K
8
0.5621
€
Cc
0.1654
€s
0.3967
€l
0.4379
S

Table 8 Summary of Step-wise Change of Case Studies

^^Qj? era ting T C

LHSV1 C . H /oil Ρ
^^•xcondi tions i si VI
9
(hr" ) (°K) (wt%) (wt%) (ratio) kg/cm G
Case study

Fig. 2 0.818 6 4 3 - * 653 2.62 0.012 1000 140

Fig. 3 0.818—»1. 241 643 2.62 0.012 1000 140
Fig. 4 0.818 643 2 . 6 2 — 3 . 62 0.012 1000 140
 Residue Hydrodesulfurization Trickle-bed Reactor 85

then solved on a CDC NOS 2. computer using Summary and conclusions

DEGEAR's variable-step integration algorithm with
Jacobian calculated internally by finite differ-
ences provided in the Package of IMSL.
Results and discussion
Simulation was performed first, choosing Iraian
Heavy A.R. as the feedstock and catalyst "D" as the
residue catalyst while suitable parameters given in
Table 3, 5, 6 in order to observe the transient
response of start-up period of the simulated
reactor with fresh catalyst. Base on the condi-
tions in Table 7, Fig. 1 shows the sulfur con-
centration and temperature profiles along the bed
at various dimensionless time. A maximum of 75%
desulfurization efficiency is observed, where the
hot spot of the bed moves from inlet to outlet
gradually, and a maximum Δ τ of 34°C is detected.
Since desulfurization catalyst generally has long
cycle time, catalyst deactivation is not notice-
able for a short period after the start-up of the
reactor. This phenomenon signifies that residue
HDS process is of slow dynamic in catalyst deacti-
vation. The same phenomenon was also observed
previously by Hanika (1981) in his experimental
studies of dynamic behavior of a laboratory
trickle-bed reactor. This transient temperature
profile is useful in estimating the proper loca-
tion of intra-stage quenching (Shah, et al.,
1976; Mhaskar,1978) provided that the appropriate
catalyst temperature limitation is given. After
the start-up period of the reactor, it is observed
to operate at pseudo-steady state conditions. At
time fo 20 Γ ( l r equals to 1.222hr), step wise
changes of inlet feed temperature, oil feed rate,
sulfur feed concentration were performed respec-
tively to examine the dynamic behavior of the
simulated reactor. The individual change of
above operating variables is given in Table 8.
Fig. 2 shows the dynamic profiles of sulfur
concentration and temperature along the bed after
a step change of oil feed temperature. It denotes
that high inlet feed temperature increases the
desulfurization efficiency while promotes much
higher bed temperature distribution along the bed
which may be ruinous to general catalyst reaction
system. Fig. 3 shows the dynamic profiles of
sulfur concentration and temperature along the
bed after a step change of oil feed rate (or LHSV
(Liquid Hourly Space Velocity)). It is observed
as the feed load is increased, the desulfurization
efficiency reduces. But total heat release of
desulfurization reaction increase to promote the
temperature distribution to a higher level along
the bed. This is one of the reasons to apply
intra-stage quenching technique to protect the
catalysts in the bed from the phenomenon of
sintering. Fig. 4 shows the dynamic profiles of
the sulfur concentration and temperature along
the bed after a step change of sulfur feed concen-
tration. From this figure, the most detrimental
effect is detected. Due to high load of inlet
sulfur content, a large quantities of reaction
A mathematical model has been developed to
simulate the dynamic behavior of an adiabatic,
deactivating residue HDS trickle-bed reactor.
Using this simulator, the dynamic performance of
the start-up period operation, step-wise change
of operating variables were examined. From the
phenomena observed, it will be helpful in acquir-
ing the possible dynamic characteristic of this
specific reactor system.
References
Henry H. C. and J. B. Gilbert (1973). Scale up
of pilot plant data for catalytic hydroprocessing.
Ind. Eng. Chem., Process Pes, Dev., 12, 328-334.
Hochman, J. M. and E. Ef forn (1969). Two phase
concurrent downflow in packed beds. Ind. Eng.
Fund., 8, 63-71.
Mears, D. E. (1971). The role of axial dispersion
in tricks flow reactors. Chem. Eng. Sci. , 26,
1361-1366.
Mears, D. E. (1974). The role of liquid holdup
and effective wetting in the performance of
trickle-bed reactors. Chem. Reac. Eng. II, ACS
Monograph Ser., 133, 218-233.
Mhaskar, R. D. and Y. T. Shah (1978). Optimum
quench location for a hydrodesulfurization reactor
with time varying catalyst. 2. cases of multiple
gas quenches and a liquid quench. Ind. Eng.
Chem., Process Pes. Dev., 17,27-33.
Montagna, A. A. and Y. T. Shah (1975). The role
of liquid holdup, effective catalyst wetting, and
backmixing on the performance of a trickle bed
reactor for residue hydrodesulfurization. Ind.
Eng. Chem., Process Pes. Pev. 14, 479-483.
Shah, Y. T. and J. A. Parakos (1975). Criterion
for axial dispersion effects in adiabatic trickle
bed hydroprocessing reactors. Chem. Eng. Sci. ,
30, 1169-1176.
Shah, Y. T. and R. P. Mhaskar (1976). Optimum
quench location for a hydrodesulfurization reactor
with time varying catalyst activity. Ind. Eng.
Chem., Process Pes Pev., 15, 400-406.
Shoji Kodama (1980). S imulation of residue hydro-
desulfurization reaction based on catalyst deacti-
vation model. J. Japan Petrol. Inst., 23, 310-329.
Villadsen, J. V. and W. E. Stewart (1967). Solu-
tion of boundary value problems by orthogonal
collocation. Chem. Eng. Sci., 22, 1483-1501.
Yan, T. Y. (1980). Pynamics of a trickle bed
hydrocracker with a quenching system. Can. J.
Chem. Eng., 58, 259-266.

heat are released. Maximum is equal to 48°C,

which is a very harmful operation mode in any
existing commercial residue HDS reactor. Still,
the effects of step-wise change of H2/oil at
different conditions on the dynamic behavior of
the reactor had been studied but not include in
this content for the sake of brevity. Further, a
similar set of dynamic performance of the simulated
reactor by the same variations of the operating
variables used in the simulation of Fig. (2) -
(4) were collected but at different initial condi-
tions (dimensionless time is set to 80 r instead
of 20 τ , I t equals 1.222 hr). It is concluded
the whole simulation work is acceptable.
86 Yung-Cheng Chao, Jhy-Shyong Chang and Hsiao-Ping Huang

Fig. 1 Dynamic profile of sulfur concentration Fig. 3 Dynamic profile of sulfur concentration
and bed temperature during the start-up and bed temperature after a step change
period (operating condition shown in of oil feed rate (or LHSV) (operating
Table 7) condition shown in Table 8)

1.0 r "i 1 1 1 1 1 1 1—~T 1.2 I I I 1 I 1 1 1 1

t = 20 Τ
t
o= 2 0 T e
Curve 1 At = 0 T Curve 1 At = 0 T
1.0 2 2T ~
2 27
3 4.5 Τ
4 7T
0.8 5 8T
6 9.5 Τ
ο> T= 1.222 hr. -
Ο>

V/'
0.6

- 2 -
0,4

0.2 - 6
5 43""""
700

—
680 h
< ^ /
/ ——•

1
l_--r========
660h

640 I 1 1 1 1 1 1 1 1
0.2 0.4 0.6 0.8 1.0
X/L

Fig. 2 Dynamic profile of sulfur concentration Fig. 4 Dynamic Profile of sulfur concentration
and bed temperature after a step change and bed temperature after a step change
of oil feed temperature (operating condi- of inlet sulfur content (operating condi-
tion shown in Table 8) tion shown in Table 8)
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

FUZZY CONTROL OF A PILOT BATCH

POLYMERIZATION REACTOR
G. Rong and J. C. W a n g
Laboratory for Industrial Process Modelling and Control, Zhejiang University,
Hangzhou, PRC

Abstract. The computer control system of a pilot scale batch polymeri-

zation reactor has been built up, where a fuzzy control algorithm is
used to automate nonlinear and poorly-defined propylene oxide polymeri-
zation process. Satisfactory regulation of pressure and temperature of
the reactor has been obtained in real-time runs, as the result of adopt-
ing different fuzzy linguistic rules during different stages of the batch
reaction. Fuzzy methodology is applied to detect which stage the process
is in. Furthermore, an expert fuzzy controller for temperature control
is discussed and applied to the reactor's model in simulation runs.

Keywords. Computer control; fuzzy logic controller; chemical variables

control; batch polymerization reactor; expert fuzzy control; nonlinear
system.

INTRODUCTION

The main purpose of this paper is to apply fuzzy
logic control techniques to batch polymerization
reactor. There has been considerable number of
studies on control of batch polymerization reactor
and applications of fuzzy logic controller.
Amrehn (1977) reported on process computer appli-
cations in the polymerization industry.
Hoogendroorn (1980) presented a survey on the
control of commercial continuous and batch polym-
erization plants. It is easy to see that the auto-
mation of batch reactor's operations such as
metering, feeding, process monitoring and logging
can be put into practice without substaintial
obstacles as the wide applications of process
computer. But the control problem of reaction con-
ditions (temperature or pressure) remains unsucce-
ssfully solved because the polymerization process
is often nonlinear, time-variant and, sometimes,
poorly-understood. Most of the modern control
theories are adequate for the control problem only
when mathematical model of the reactor is availa-
ble. This is found in studies of early investiga-
tors (Nakagawa, 1 9 7 0 ; Maroquin and Luijben, 1 9 7 3 ;
Ham and Liemburg, 1 9 7 5 ) · It is often not the case
in practice to obtained a dynamic model of a batch
polymerization process. Thus, Kiparissides and
Shah (1983) used two adaptive control techniques,
in simulation runs, to get better temperature
tracking and minimum batch time despite of the
nonlinear and time-varying characteristics of the
commercial scale PVG batch reactor. To some extant
adaptive control is a solution to the problem.
However, adaptive methods based on identification
cannot show their advantages if little information
about polymerization kinetics and model structure
is available.
Sometimes, a batch reactor is simply controlled by
human operator. An alternative approach to the
control of reaction conditions is to investigate
the control strategies employed by human operator
and design fuzzy logic controllers. Since Mamdani
and Assilion (1975) presented the original paper
DCCR-G
on fuzzy logic controller, much interest has been
shown in the application of similarly structured
fuzzy logic controller to varieties of processes
(Kickert and Van Nauta Lemke, 1 9 7 6 ; Tong and co-
workers, 1980; Ray and Dutta Majumder, 1 9 8 5 ) . King
and Mamdani ( 1 9 7 7 ) first applied fuzzy logic cont-
roller to temperature control of a stirred tank,
part of a batch reactor process. Liu (1985) desig-
ned a fuzzy temperature controller for a realistic
batch sulphonating reactor. Vaija and co-workers (
I985) reported their work on failure detection and
safety control for a hypothetical polymerization
by using fuzzy methodology. Review papers (Tong,
1 9 7 7 ; Tong, 1 9 8 ^ ; Maiers and Sherif, 1985) gave
much information about theories and applications
of fuzzy controller. The idea of expert fuzzy
controller (EFG) was mantioned "by Tong ( 1 9 8 4 ) , and
some structures of EFG also suggested.
In present work, the microcomputer control system
of a pilot batch polymerization reactor has been
considered. A fuzzy control algorithm for pressure
-temperature control of the reactor is employed to
automate the nonlinear and time-variant process.
This algorithm is derived from a set of hierach-
ical fuzzy linguistic rules. It consists of a pre-
ssure fuzzy controller, a temperature fuzzy contr-
oller and a fuzzy selector to decide which lower
controller to use from stage to stage of the poly-
merization. The implement of the algorithm gives
satisfactory regulation of pressure and tempera-
ture in real time runs. Design of EFG for temper-
ature control of the reactor operating under
isothermal condition has been given and evaluated
in simulation runs.
STATEMENT OF THE PROBLEM
Discription of Control System
Figure 1 . shows a schematic diagram of the micro-
computer, control system under consideration. The
APPLE-2 microcomputer with clock board and 8-bit

87
88 G. Rong and J. C. Wang
A/D—D/A converter has been used not only to imple-
ment various control modes "but also to automate
batch reactor's operations such as process monit-
oring and experimental data handling. The plant is
a pilot scale ( 2 0 0 0 ml) sealed vessel of stainless
steel, equipped with agitator, pressure and temp-
erature sensors,cooling coil and electric heater.
The major control variables are voltage of elec-
tric heater (V) and cold water flowrate to cooling
coil (F). The output variables, which can be mea-
sured and controlled directly, are pressure (P)
and temperature (T) of the reactor. Negative ion
polymerization of propylene oxide (sometimes,
ethylene oxide) produces polyether product which
is a kind of useful chemical auxiliary. The pilot
batch reactor is designed for laboratory scale
test of development of new polyether products and
for experimental investigation of kinetics of
polymerization processes.
Pressure-temperature Control of the Reactor
Usually the rate of polymerization can be contro-
lled indirectly by changing temperature of the
reaction when neither concentration of monomer nor
conversion rate of monomer can be measured in real
-time enviornment. In the reactor, pressure of
vapor phase is directly proportional to tempera-
ture and concentration of the volatile components.
Hence changing pressure is another way to control
reaction rate. Volatile monomer propylene oxide
will evaporate and pressure of the reactor will
rise as the temperature increases during stage of
heating initiation of batch polymerization. When
the pressure reachs desired value (pressure set
point), monomer reacts to produce polyether and
heat is released at a rate of 2 0 kcal/mol. Press-
ure should be maintained at its set point as long
as possible if the temperature does not exceed
prescribed high limit. In latter stage of the
polymerization, the temperature will be higher
than its limit in order to prevent pressure from
decline. Temperature control should be introduced
and take the place of pressure control as soon as
the temperature is higher than its limit.
Isothermal Operation and Temperature Control
Polymerization of propylene oxide is carried out
under isothermal condition for investigation of
kinetics of the reaction. Depolymerization of
polymer product has been observed in isothermal
polymerization process. This phenomenon is inde-
cated by the anomalous restoration of pressure and
concentration of monomer. Thus the task of temper-
ature control includes removal of heat when reac-
tion releas heat and supplement of heat during
stage of initiation and stage of depolymerization.
DESIGN OF FUZZY CONTROL ALGORITHM
FOR PRESSURE-TEMPERATURE CONTROL
Early investigators (King and Mamdani, 1 9 7 7 ? Liu,
I985) reported their applications of fuzzy logic
controller to temperature tracking in batch react-
or process and achieved successful control results.
Motivated \sy their work, a fuzzy control system
for tracking reaction conditions like pressure and
temperature has been considered. The closed-loop
system is shown in Fig. 2 . Three parts of the
algorithm are fuzzy selector, fuzzy temperature
controller and fuzzy pressure controller. The
algorithm's inputs are error of pressure (EP),
rate of its change (CP), error of temperature (ET)
and rate of ET's change (CT). The outputs of the
algorithm are flowrate to cooling coil (F) and
heating voltage (V).
Fuzzy selector is derived from fuzzy linguistic
rules ( 1 - 7 ) » given in Appendix. These rules are a
summary of human operator's decision making
strategies adopted to decide which control modes
will be needed during different stage of polymer-
ization. If we treated these rules by using the
compositional rule of inference as described by
Zadeh ( I 9 6 5 ) , a huge decision making look-up table
would be obtained. It is not efficient to imple-
ment the table in a microcomputer control system
due to large number of data involved. So we simply
, but effactually, implement the selecting rules
by translating them into normal logic decision
process. The translation is based on look-up table
relating fuzzy subsets representing fuzzy value of
inputs to quantised input values. The look-up
table has the same form as described by King and
Mamdani ( 1 9 7 7 ) . Fig. 3 shows how the fuzzy selec-
tor works.
Fuzzy linguistic rules (20-28) in Appendix are
used to design pressure fuzzy controller in the
form of look-up table. Human operator uses EP and
CP to affect flowrate of cold water. Only absolute
value of flowrate is considered in the controller
because the flowrate is changed in such a way that
the normally-closed valve of cooling coil's entry
is opened in a part of the time interval of samp-
ling and control. Positive EP is eliminated by
cooling. Negative EP will be eliminated by the
exothermic reaction itself when EP is small, or by
selected heating action if EP is negative-big.
The fuzzy controller takes similar structure as
described by King and Mamdani ( 1 9 7 7 ) . The number
of elements in the universe of discourse about EP,
CP and F are 14-, 1 3 and 7^espectively. For example
, 7 elements in the universe of discourse about F
represent 7 levels of control action, and these
7 values are assigned to the grade of membership
to get the definition of four fuzzy subsets desc-
ribing control action, as shown in Table 1 .
Similar definitions of fuzzy subsets about EP and
CP can be written out.
TABLE 1 Look-up Table Relating Fuzzy Subsets
to quantised Control Values
0 1 2 k 6
3 5 7
PB 0 0 0 0 0.1 0Ά 0.8 1.0
PM 0 0 0.2 0.7 1.0 0.7 0.2 0
PS 0 Λ 1.0 0.8 0 Λ 0.1 0 0 0
0E 1.0 0.5 0 0 0 0 0 0
note ι PM (positive medium^ P3 (positive small)
PB (positive big) and 0C (nil) are fuzzy
subsets about control action.
According to the procedures of design (King and
Mamdani,1977; Liu, 1 9 8 5 ) , a look-up table, Table 2
, has been established, which represents a map
from levels of EP and CP to levels of control (F).
Similarly, fuzzy linguistic rules (9)-(iç;) given
f
in Appendix, are used to build up fuzzy controller
of temperature, which is another look-up table.
EXPERT FUZZY CONTROLLER
General Idea of Expert Fuzzy Controller
Tong ( 1 9 8 ^ ) pointed out that expert fuzzy contro-
llers (SFC) differ from fuzzy logic controllers in
 Fuzzy Control of a Pilot Batch Polymerization Reactor 89

TABLE 2 Look-up Table of Pressure Controller high quality of temperature tracking throughout
the whole polymerization process.

EP The parameters of FLG are quantized factors Ke

-3 -2 -1 -0 0 1 2 3 k 5 6 and Kc and output factor Ku. If we choose a set of
of parameters that is adaptable to overcome over-
0 0 0 0 2 shoot of the temperature, it will cause untolér-
-6 0 0 0 1 1 2
able negative error during depolymerization stage
-5 0 0 0 0 0 0 1 1 2 2 2 and in the latter period of batch reaction. This
means that different control rules or, at least,
-4 0 0 0 0 0 0 2 2 2 2 3 FLG with different parameters will be required to
-3 0 0 0 0 0 0 2 3 3 3 3 obtain excellent temperature control throughout
the process. EFG can take over the task of control
-2 0 0 0 0 0 0 2 3 if it has hierachical fuzzy control rules.
-1 0 0 0 0 1 1 2 k 5
The EFG applied to reactor's temperature control
CP 0 0 0 0 0 1 1 2 5 6 7 has the following high-level rules
1 0 1 1 1 1 2 2 k 5 6 7 if <Polymerization .. Stage Sl>
2 0 1 1 1 2 2 3 5 6 7 then (Use .. Rule Set Rl)
3 0 1 1 2 2 3 3 5 6 6 then if < Polymerization .. Stage S2 >
0 1 1 1 2 3 5 6 7 7 then (Use .. Rule Set R2)
0 1 1 2 2 3 5 6 7 7
6 0 1 1 2 2 3 4 5 6 7 7 then if < Polymerization .. Stage S4>
then (Use .. Rule Set H4)
where j < . . > some fuzzy proposition
at least two important ways. EFG contain more (..) non-fuzzy proposition
complex knowledge about process control, and they 51 stage of initiation
use this knowledge in more complex way. 52 first isothermal stage
53 stage of depolymerization
Typical structure of linguistic rules of fuzzy 54 second isothermal stage
logic controller may be
The form of high-level rules implies that S1-S4
if < Process..State > then < Control..Action > appear in turn. S1-S4 can be considered as fuzzy
subsets in the universe of discourse of measurable
where < .. > represents some fuzzy proposition. state of the polymerization such as temperature,
This kind of rules is no more than a map with non- pressure and time of reaction. Membership function
linear gains from process states to control ac- relating measurable states to subsets S1-S4 has
tions . been built up with trial and error method, utili-
sed to detect which subset the polymerization
Various structure of EFG had been suggested by process belongs to in terms of maximum degree of
Tong ( 1 9 8 4 ) . One of the most interesting forms of membership principle.
EFG is the EFG that permit strategic knowledge to
be included and have high-level rules to decide Low-level rules, i.e. control-level rules, are
which low-level rules to apply, i.e. designed to meet the different requirement of tem-
perature control from stage to stage. Low-level
if <Process..State 1> then <Use..Rule Set 1> rule set Rl mainly deals with normal logic control
of the electric heater's voltage. R2,R3 and R4
if <Process..State 2> then <Use..Rule Set 2> have the same fuzzy linguistic rules used to build
up FLG for eliminating error of the temperature,
if <Process..State N> then <Use..Rule Set N> but have different parameters of the FLG to adapte
time-varying characteristics of the batch process.
The low-level rules may be fuzzy linguistic rules
for design of fuzzy logic controller, or non-fuzzy
control policies. RESULTS AND DISCCUSION

For instance, the fuzzy control algorithm for pre-
ssure-temperature regulation mentioned in previous
section can be considered as a kind of EFG with
hierachical structure. The high-level rules are
linguistic rules for fuzzy selector and the low-
level rules include on-off control policies for
heating and fuzzy linguistic rules for pressure
and temperature control.
Expert Fuzzy Control of Reactor's Temperature
During isothermal operation of the batch polymer-
ization reactor, it is a hard work to keep temper-
ature at its desired value because of the non-lin-
ear and time-variant characteristics of the batch
process. Fuzzy logic controller (King and Mamdani,
1977? Liu, 1985) has been applied to eliminate
error of temperature of the polymerization reactor
by Rong (1986) and found that fuzzy logic control-
ler (FLG) with constant parameters cannot give
Results of Pressure- temperature Control
The implement of the fuzzy control algorithm for
regulation of pressure and temperature in the
pilot batch plant, where a poorly-understood ion
polymerization of propylene oxide takes place,
gives pressure-temperature tracking profiles as
illustrated in Fig. k. Though only on-off manipu-
lated modes of cooling water flowrate is available
in the plant, the performance of the fuzzy control
algorithm has met the technical needs, and it is
better than the performance of simple on-off
control and manual control. The fuzzy control
algorithm can also obtain satisfactory results, as
shown in Fig. 5» when initial conditions (recipe
of reactants) are changed. Robustness of the
algorithm is not difficult to find according to
Fig. 5 and Fig. 4 . Because of on-off mode of cool-
ing conventional PID algorithm cannot be imple-
mented; and by lack of mathematical model of the
process no bang-bang control algorithm can be
90 G. Rong and J. C. Wang

designed. Thus no attempt has been made to compare TABLE 3 Results of Simulation Runs
fuzzy control algorithm with PID controller or
bang-bang control. The advantages of fuzzy control
algorithm lie in two aspects: it is the sole algo- Digital control Isothermal batch reaction
rithm simply designed, satisfactorily performing algorithm
and easily implemented in the microcomputer cont- Overshoot Variance Error-M
rol system we considered; its robustness makes it
possible to use one control algorithm to automate FLCs Ke=1.2
polymerization reactions with different recipe. Kc=2.0 0.54 I.06 -1.18
Ku=0.l7

Simulation Results of EFC PI: Kp=0.5

A mathematical model of the pilot batch reactor
has been built up by Hong ( 1 9 8 6 ) , which includes
kinetic behaviou of the polymerization of propy-
lene oxide. Continuous regulation of the flowrate
of cold water is available in the model so that
the fuzzy logic controller (FLC) designed by King
and Mamdani (1977) can be directly used to elimi-
nate temperature error of the reactor under iso-
thermal operation. Well turned digital PI contro-
ller is also applied to control of temperature of
the plant. Expert fuzzy controller based on fuzzy
stage recognition and FLCs with different control
parameters has been applied to the model of pilot
batch polymerization reactor. Table 3 lists the
simulation results of FLC, PI controller and EFC.
PI controller gives less overshoot tnan FLC, but
pessimistic negative error of temperature because
of the nonlinear and time-variant characteristics
of the process. FLC has been turned to ensure the
stability of the system and decrease overshoot.
When the FLC works in the closed-loop system,
negative error of reactor's temperature is not
avoidable after initial stage of the polymeriza-
tion. EFC gives the best tracking of the reactor's
temperature. The reason for EFC eliminating nega-
tive temperature error effectively is that EFC
adopts differently turned FLCs and necessary
heating policies in different stages of the batch
polymerization process.
CONCLUSION
Particularly designed fuzzy control algorithm for
regulation of pressure and temperature of the
pilot batch reactor has obtained desired pressure-
temperature tracking performance and shown its
robustness. Using different control rules as the
polymerization progressing is a way to overcome
time-variant characteristics of the process.
REFERENCES
Amerehn, H. (1977). Computer control in polymeri-
zation industry. Automatica. 13» 533-5^5·
Ham, A. A. and J. H. Liemburg (1975;. Comparison
of five temperature control systems for a
exothermic batch reactor. JournalA, 16, 73-
78. ' "
Hoogendroorn, Κ. (I98O). Control of polymerization
processes. Proceedings of the 4th IFAC Conf-
erence on PRP~Automation. ICC, Ghent, Belgium.
Kickërt, W. J. M. and""!. R.~Van Nauta Lemke (1976).
Application of a fuzzy controller to a warm
water plant. Automatica, 12, 301-308.
King, P. J. and Ε. H. "Mamdani "(1977).
The application of fuzzy control system to
industry processes. Automatica, Γ), 235-242.
Kiparissides, C. and S. L~. Shan (1983)·
Self-turning and adaptive control of a batch
polymerization reactor. Automatica, I 9 ,
225-235.
Liu, M. (I985). A case design of fuzzy temperature
control for chemical reactor. Control and
Instruments in Chemical Industry, 12, 2.* 20-
0.28 1.24 -I.29
Ti=1.0 min
EFC: use rule ζ-Λ
Rl if té [0, 41 ]
R2 if té [41,386 ] 0.24 0.87 -O.61
R3 if té [386,442]
R4 if té 1422,600]
note:
t—reaction time in minute
Overshoot— (*C)
Error-M—the worst negative ex*ror after
2 reaction
initial stage of the
Variance^—- ^ [T(i) -Tsp]
In all runs the sampling time is 0.5 min.
26 (in Chinese).
Maiers, J. and Y. S. Sherif ( 1 9 8 5 ) · Applications
of fuzzy theory. IEEE Trans. Syst. Man Cybem.
t t
. 1 5 . i» 175-189.
Mamdani, Ε. H. and S. Assilian ( 1 9 7 5 ) ·
An experiment in linguistic synthesis with a
fuzzy logic controller. Int. J. Man-Machine
Stud. , 7 , 1 - 1 3 ·
Maroquin, G. and W. L. Luijben ( 1 9 7 3 ) · Practical
control studies of batch reactors using reali-
stic mathematical models. Chem. Eng. Sci. ,
28, 993-1003.
Nakagawa, M ( 1 9 7 0 ) . Large PVC manufacturing plant
and computor control system. CEER, lg?°jL 3,
1 2 - 1 6 and CEER, 1 9 7 0 , 4 , 27-32" ~
Ray, K. S. and D.~Dutta Majumder ( 1 9 8 5 ) ·
Fuzzy logic control of a nonlinear multi-
variable steam generating unit using decoup-
ling theory. IEEE Trans. Syfit.>Man, Cybern.,
1 5 . 4, 539-558.
Rong, G. ( I 9 8 6 ) . Microcomputer control of a pilot
batch polymerization reactor and investigation
of fuzzy control algorithms. Master Thesis,
Zhejiang University. (in Chinese)
Tong", R-M-.M^BocFand A. Latten (1980). Fuzzy
control of the activated sludge waste water
treatment process. Automatica, 1 6 , 695-697·
Tong, R. M. (1977).A control engineering review of
fuzzy system. Automatica, Γ3, 559-569·
Tong, R. M. ( 1 9 8 4 ) . A retrospective view of fuzzy
control systems. Fuzzy Sets & Systems, 14,
199-210.
Zadeh, L. A, (1965). Fuzzy sets. Inform. Control,
8, 338-353.
APPENDIX
Linguistic Rules for Pressure-temperature Control
The universe of discourse about decision has three
values, i.e. ms=(1,2,3 } , and three fuzzy subsets
Heat (heating)
MS=< CTRL-P (control pressure)
. CTRL-T (control temperature)
Let ms = l means MS=Heat; ms = 2 does MS = CTRL-P;
and m s = 3 does MS = CTRL-T.
The fuzzy subsete describing error and its change
 Fuzzy Control of a Pilot Batch Polymerization Reactor 91

axe PB (positive big), PM (positive medium),

PS (positive small), PO (positive nil), GO (nil)
NO (negative nil), NS (negative small),
NM (negative medium), NB (negative big).

Selector rules.
if ET»(NB or NM) and EP»(NB or NM) then MS»Heat
(1)
if ET-(NB or NM) and EP=NS and CP~(NB or NM)
then MS=Heat (2)
if ET-NS and GT-(NB or NM) and EP«NS and
GP»(NB or NM) then MS«Heat (3)
if ET=(PS or PB or PM) then MS-CTRL-T (4)
if ET<PO or NO) and CT*(PS or PM or PB)
then MS-GTRL-T (5)
if ET-NS and GT-(PS or PM or PB) and CP«(NS or NM
or NB) then MS=CTRL-T (6)
if M S * Heat and M S * GTRL-T then MS = CTRL-P (7)

Heating rule.
if ms»l then V»200 otherwise V*0 (8)
Fig. 1 Microcomputer control system
of the batch reactor
Temperature control rules. When MS=CTRL-T
if ET-(PB or PM) and GT=(GO or PS or PM or PB)
then F=?B e (9)
if ET«(PB or PM) and CT NS then F-PM (10)
if ET-P3 and GT-(GO or PS) then F-PM (11) on-off
if iST=(NO or PO or PS) and GT»(PB or PM) controller
then F«PM (12) pilot
Tsf> fuzzy mi=3l fuzzy Τ
if ET=NS and CT=(PB or EM) then F=PS selector! controller plant
(13)
if ET=NS and GT=(NB or NM or N3 or GO)
then F-OG fuzzy Ρ
S (14) controller
if ET-(PO or NO) and CT (NB or NM or NS or GO)
then F=OG (15)
if ΕΓ"(ΡΟ or JÎO) and GT-PS then F-PS (16)
if ET-PS and GT-(NS or NM or NB) then F-OG (17)
if ΕΓ*(ϊΜ or PB) and CT-(NM or NB) then F=PS (18) Fig. 2 Fuzzy pressure-temperature
control system
if ET=(NM or NB) then F=OG (19)

Pressur control rules. When MS=CTRL-P

if SP»(PM or PB) and CP«(PS or PM or PB)
S
then F=PB (20)
if EP (PM or PB) and GP-(G0 or NS) then F-PM (21)
if EP*(PM or PB) and GP-(NM or NB) then F=PS (22)
if EP=PS and GT=(PM or PB) then F=PM
(23)
if EP=PS and CT«(C0 or PS) then F=PS
(24)
if EP=PS and GT=(NS or NM or NB) then F=0G
(25)
if EP=(P0 or NO) and GT=(PS or PM or PB)
then F=PS (26)
if EP=(PO or NO) and CT=(C0 or NS or NM or NB)
then P-OC (27)
if EP-<NS or NM or NB) then F=0G (28)

MS—GTRL-T 1 j MS-CTRL-Pl | MS-Heat

Fig. 3 Fuzzy selector

92 G. Rong and J. C. Wang

Θ 48 112 176 24Θ

time in minute

Fig. 4 Performance of real-time run of

fuzzy control algorithm

Θ 48 112 176 24Θ

time in minute
P sp z 6.0 (kg/cm*)
e
T sp Ζ 120 ( C) ; recipe (30% of catalyst KOH) changed

Fig. 5 Performance of the fuzzy control algorithm

Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

EARLY DETECTION OF HAZARDOUS STATES

IN CHEMICAL REACTORS
R. King
Institut fur Systemdynamik una Regelungstechnik, Universitàt Stuttgart,
Pfaffenwaldring 9, 7000 Stuttgart 80, FRG

Abstract. Extreme reaction conditions, such as high temperature or high

pressure are characteristic features of chemical processes. Unknown para-
meters, unknown parallel or consecutive reactions and the highly nonlinear
and therefore parameter sensitive dependence of the reactor behaviour on
temperature and concentrations sometimes make safe operation difficult.
Existing measurements e.g. temperatures, indicate only part of the process
state. Hence the onset of dangerous modes may be detected too late. With
Kalman filters the whole state of a plant is reconstructed in order to de-
tect hazards at an early stage. Criteria are formulated for safe operation
depending on the complete process state. In addition, the future development
of the process is predicted. Finally, not only one but several fault modes
are included in a detection system. A separate filter is designed for every
possible fault. Adequate countermeasures with a discrimination between diffe-
rent fault modes can be applied. These methods, including experimental results,
are illustrated by the examples of a batch and of a semibatch reactor.

Keywords. Alarm systems, Bayes methods, fault detection, Kalman filters,

nonlinear filtering, parameter estimation, pattern recognition, prediction,
varying structures.

INTRODUCTION step, the future development of the process is

Even slight changes in process parameters can have
a significant influence on the global behaviour of
a chemical plant, due to the highly nonlinear de-
pendence of the reactor behaviour on the process
state. This is very important, when substances are
handled or produced which are thermally unstable
or toxic. The difficulties will further increase
when multiple material and energy fluxes between
process units are introduced to meet higher demands
on quality and productivity.
predicted by applying a fast simulation based on
the actual state estimate. Finally, multiple fil-
ter methods are used to discriminate between dif-
ferent faults of the plant or the sensors. Fault
detection and diagnosis are well known, e.g. in
space and aircraft systems, in conventional and
nuclear power plant, in turbines etc. A review is
given in Willsky (1976). Himmelblau (1978),
Gilles and Schuler (1982) were the first to suggest
the application of these methods in chemical
plants.

Physical properties, easily measured, such as tem-

peratures, are used as control variables of a
plant. Therefore, it is necessary to ask, whether
as control variables they are able to detect
faults. It is also important to know if sensors
partly or completely fail.
The early detection of hazardous states in chemical
systems can be a difficult task. Because
1. A chemical process can only be described by a
high number of state variables, such as tempera-
tures, pressures, concentrations etc. Only part
of this 'process state' can be measured on-line
with an acceptable time delay.
2. The processes are very complex and strongly
nonlinear.
3. To judge the safety of a process, the detection
system must be based on the history of the pro-
cess as well as on the actual measurements.
This paper will show, that model-based measuring
techniques including pattern recognition will help
to overcome characteristic problems of chemical
engineering. It will be possible to detect hazard-
ous states at an early stage using a Kalman filter
to estimate unmeasured concentrations. In a next
CRITERIA
Various criteria can be applied to describe danger-
ous behaviours. The question, as to the suitabili-
ty of the criteria, can only be answered in view
of the specific conditions of a process. When both
the first and second derivative of the temperature
with respect to time are positive, this criterion
(Hub, 1978), popularly used in industrial applica-
tion, indicates the dangerous state of a reactor.
The balance of enthalpy, a function of the whole
state (i.e. temperature, pressure, total mass and
concentration c^ of compound Α ·) is analyzed in
Ί
order to avoid repeated differentiation of the
measurement process.
= f(T,p,m,c c ,...) (1)
p2
This analytical expression for the first derivati-
ve could not be used up to now, because the concen-
trations are difficult and expensive to measure.
Now, they are replaced by the estimates of a Kal-
man filter. Differentiation of (1) with respect to
time gives an analytical expression for the second
derivative as well. The main operation in a filter
is integration, which gives signal smoothing, in-

93
94 R. King

stead of differentiation, which gives signal delling errors it can give only a rough estimate.
roughing. Therefore, better results are expected.

Example Fig. 2 shows predictions for two experiments of

the sulphonation reaction performed again in an
To demonstrate the operation of the proposed de- adiabatic vessel. Both experiments are started at
tection system, the following example of technical 11.00 h. In one case the runaway occurs at 12.46 h
interest is considered. An aromatic compound, A in the other at 13.52 h. The predictions of time
is sulphonated by sulphur trioxid, A , to yield
l9 to runaway are inaccurate at the beginning due to
the desired monosulphonic acid A :
2 the sensitivity of this example and because the
3 temperature decreased in one of these adiabatic
experiments for unknown reasons. After some time
A
1 + A
2 li A
3 + (-AH )
R1 (Rl)
the predictions become quite stable and give sa-
From numerous heat storage tests, a formal kinetic tisfying results, although modelling errors and
is found to describe the measured temperature-time disturbances of the process can possibly cause
readings. considerable errors in the simulation.

r = k expi-E^R/T) c« z\ (2)
1 1Q MULTIPLE FILTER METHODS
In the next step of the reaction the remaining In Schuler (1982) it is shown, that the use of
sulphur trioxid, A , will form polysulphonic acids, model-based measuring techniques for process moni-
2
e.g. disulphonic acid. toring can be expanded considerably. Knowledge of
the complete state vector enables us to identify
a certain mode of disturbance or to discriminate
between different modes causing hazardous states.
Furthermore, two monosulphonic acids A will be
3
formed out of one disulphonic acid A and one aro- The principle procedure can be explained by means
4
matic compound A in reaction (R3). Therefore, the of a sulphonation reactor (King, Gilles, 1984).
1
yield of A will be small as long as (Rl) is not Sulphonation of aromatic compounds is usually
finished.
4 performed in a cooled semi batch-reactor. Some of
Τ the possible operating modes are given in Table 1.
A
4 + A
1 Λ 2 A
3 + ( - A H
R) 3 (R3)

The kinetics to describe (R2) and (R3) are similar TABLE 1 Working modes of a sulphonation process
to Eq.(2). Reaction (Rl) to (R3) include all de-
sired conversions of this example. No Mode Countermeasures

In process industry, the heats of reaction are 1 normal operation -

removed very carefully. It is known that the sul- 2 exothermic decom- + increase cooling
phonic acids A and A may decompose in an unde-
3 4
sired highly exothermic reaction to form product
position
A (Regenass, 1984, Grewer, 1974). 3 penetration of stop cooling,
5 Χ cooling water stop reaction
by other means
(A , Λ ) Λ A
3 4 5+ (-AH )
R4 (R4)
4 exothermic decompo- ->· increase
sition and then cooling and then
Reaction (R4) can be described by an autocatalyti-
cal model of the form penetration of + stop cooling ...
cooling water
r = k exp(-E /R/T) ( c c ) * +
4 k 40 4o 3 +4
4i exp(-E /R/T) ( c c ) c (3) Totally different countermeasures have to be taken
41 3 +4 5 into account comparing modes 2 and 3. Penetration
of water into the vessel will form with S0^ new
For all following considerations it is supposed H S 0 under almost instantaneous release of heat.
that an appropriate model is available for each 2 4
In the most dangerous mode 4, the emergency strate-
specific process. It includes the balance equa- gy is changed during operation. Increased cooling,
tions for enthalpy, mass and individual compounds as a result of a detected decomposition reaction
A based on this reaction scheme. An extended Kal- is replaced by a shut down of the cooling.
±
man filter reconstructs the whole state of the
plant using only temperature measurements. This example demonstrates the importance of
distinguishing between different faults. A possible
Results warning system is shown in Fig. 3. For every mode
a special filter is designed. To discriminate bet-
The above mentioned criterion to characterize ween rival models, the error covariances of the
hazardous states is used on-line in adiabatic filters are used. With this statistical information
batch experiments. The alarm variables for two a probabilistic statement about which filter de-
different experiments are given in Fig. 1. The scribes reality best can be given. The mode of the
arrow with index a) indicates the time of alarm filter associated with the highest probability will
creation. The accelerated release of heat is de- then be assumed to represent the actual mode of the
tected earlier, than an operator could observe it. process.

The following chapters will briefly summarize the

PREDICTION fundamental equations for model discrimination.
Then, the example introduced above will be con-
If an alarm is created using a criterion it might sidered.
be important to know how the reaction will pro-
ceed. The process behaviour can be predicted by
means of fast simulation based on the present
state estimate and the process model. The time of
runaway is of particular interest. Because of mo-
 Early Detection of Hazardous States in Chemical Reactors 95

Theoretical considerations Example

Bayesiian Classification. The discrimination bet- In the example of the sulphonation process chree
ween different models M different model-based estimation schemes have to
i5i=0, 1, ..., Ν is based
on Bayes Rule, Eq.(4) (Peterka, 1981). be implemented in a computer for
r Mr
P W : k - i > P( il k-l) a) Normal operation mode
P(Mi|r. (4)
normalizing factor b) First fault mode: decomposition of
products A and A
It relates a priori model probability p(M |r c) Second fault mode:
to a posteriori model probability
± decomposition and
p(Mj|r ).The likelihood-function P ( y J Ν ^ , Γ ^ ) water penetration.
k
of the last innovation y is c a l c u l a t e d ^ â
k
Kalman filter based on model M where the innova- Penetrating water leads to a rapid increase in
is
tion sequence r is supposed to be a zero mean temperature. At high temperatures the decomposi-
v
white gaussian process. tion starts. Therefore, both faults are included
in filter c ) , which estimates the amount of pene-
If any of the model probabilities becomes 0 it trating water as well.
will retain this value for the whole time due to
the absorbing property of Bayes rule. Hence, if a A reformulation of the problem is necessary to
switching environment problem like fault detec- apply Bayes rule and the theory of Markov pro-
tion is considered, the probabilities have to be cesses, because both assume disjunctive models.
'artificially' bounded from below by a small num-
ber. The structure of filter a) differs from b) and c)
due to the absence of R4. Hence, a) is disjunc-
However, simulations of the sulphonation process tive to b) and c). For no water penetration mo-
show that a meaningful alarm creation is not ob- dels b) and c) are the same and therefore not dis-
tained for all initial model probabilities during junctive, see Fig. 5.
the first part of the reaction, because the like-
lihood functions of the different models are al- The new disjunctive classes are as follows, Fig.6.
most the same.
Μ = M_ desired reaction steps only
οΛ a
To avoid both problems encountered with the sing- Nil = M. class representing simultaneous
le use of Bayes rule a Markov process is intro- appearance of desired and un-
duced in the decision stage additionally. n desired reactions
Markov Process. Every operating mode is asso- M =M M. class representing appearance
2 of desired, undesired reactions
cTatecTwith a certain probability p . Transition
±
probabilities p.. take into account the possibi- with water penetration.
l i t y of t r a n s i t o n s from class i to class j ,
Fig- 4. Markov process in the example. With a Markov
process the development of model probabilities p. 1
False alarms are avoided when a priori knowledge are calculated between two measurements. By e x a -
3 can be
mining subsystems the transition rates a..
is included in the transition probabilities. If
it is known, for example, that undesired consecu- determined.
tive reactions take place only at high temperatu-
res the transition probability from normal opera- For a binary Markov process M «-> Μ., the follo-
wing relations are valid:
i
tion mode to decomposition mode should depend on
the temperature. Mathematical formulation leads = a +
to a so-called Markov process for the probability PiW - ^ j i J ' P i W ji (8)
p of mode i, Eq. (5)
± stationary probability
r a^
Pi(t) = -( I
a..) p . ( t )J+ Pii s a.. + a ..
j=o
! (a.. pJ. ( t ) ) , J7i
J
(5) and time constant τ.. . of Eq. (8)
j=o
The elements a - of the matrix A (the so-called 1
i (10)
infinitesimal operator) are related to the matrix a. · + a..
π of the transition probabilities p by Eq. (6)
(Bharucha-Reid, 1960)
±j
In on-line experiments the transition rate a
01
from normal operation mode to decomposition mode
1 im n(t) (6) was modelled proportionally to the amount of auto-
catalytical substances A + A , i.e.
t-+o 3 4
a = n
01 ( )
Combined Bayes/Markov Decision Stage. With Bayes
formula on-line measurements can be processed. 1ikewise:
A priori information can be used in a Markov pro- a = 1 ( A+ 12
cess to avoid false alarms. To combine the advan-
tages of both methods the model probability p ^ t ) 10 ^ÎÔ 3 V ' <)
of the Markov process is updated by Bayes rule, Simulation studies show that for some value of
Mr t = k A + A the system is governed by the decomposition
3 4
reaction. Hence, the stationary probability for
P ( Y k l i ' k - l ) Pi( -)
P-jV ; no rmàTi ζ i ng T a c tor *' normal operation mode should be small for this
+ value, e.g. l.E-02. The maximum time constant of
(10) depending on the concentration A + A is cho-
3 4
sen to determine the unknown coefficients a i and
0
a . T h e estimates of A + A are determined in fil-
10 3 4
96 R. King

ter c) which includes all dynamical effects. Simi-
lar to this subsystem the remaining transition
rates a-j_j are found. The transition rate to model
M tion of polysulphonic acids (R2). Hence, with nor-
2 will be proportional to the estimated amount
of penetrating water.
Bayes rule in the example. When measurements are
processed an update of the model probabilities is
possible. To use the statistical informations of
the three filters in the new disjunctive classes,
Bayes rule must be rewritten. The normalizing
factor Ν in Eq. (4) is the total probability of
the innovation y , see Fig. 6.
k
2 r M r
Ν I p( ^ V k-1> P( il k-i>
i=0
seen for experiment n , it is run with a great ex-
cess of A . Therefore, the increasing temperature
2
can be explained for a long time with the forma-
mal operation mode. Only at higher temperatures
the normal operation mode model is not able to ex-
plain the temperature development. Its probability
will decrease and an alarm will be given. This
example shows that the multiple filter algorithm
is able to identify different modes in real time
experiments as wel1.
REFERENCES

Σ ^ P(M (13) Ackerson, G.A., K.S. Fu (1970). On state estima-

i=0
1 'k-1' tion in switching environments. IEEE Tr.Atm.
Control, AC-15, 10-17.
which is the same as Bharucha-Reid, A.T. (1960). Elements of the
Theory of Markov Processes and their Applica-
Ν = P tion. McGraw Hill, New York.
aP t M j r ^ ) p p(M |r _ )
+ c c k1 (14)
GiHes7"E.D., H. Schuler (1982). Early detection
see Fig. 5. Either model M or model M must be of hazardous states in chemical reactors.
correct, therefore
a c Ger. Chem. Eng., 5, 69-78.
Grewer, Th. (1974). Thermal stability of reaction
P(M |r ) * 1 - P i M j r ^ ) (15) mixtures, Proc. Loss Prevention and Safety Pro-
c kel motion Proc. Ind., the Hague/Delft, 2/1-277.
M r + H i mmelblau, D.M.TI978).
(1978). Fault Detection and
(16) Diagnosis in Chemical and Petrochemical "Pro-
<Pa-Pc> P( a! k-l) Pc cesses. Elsevier, Amsterdam.
Hub, L. (1977). On-line hazards identification
The model probabilities can now be updated with during chemical processes. Proc. Loss Preyen-
Eqs. (17), writing M instead of M , p (t=k ) = tion and Safety Promotion Proc. Ind. DECHEMA
P t M i l V i ) and ( t = K ) = ρ ( Μ ] Γ ) .
a ±
P i + ± κ TTg7S]7Trankfurt, 265-277:
King, R., E.D. Gilles (1984). Dynamics and runaway
P a P 0( t = k J of chemical reactors. Preprint AIChE's Annual
P (t=k ) Meeting, San Francisco, USAi
0 +
( P - P ) P (t=k_) + p Peterka, V. (1981). Bayesian approach to system
a c 0 c aidentification. In P. Eykhoff (Ed.), Trends
P b Pi(t=kJ n d progress in system identification.
Px(t=k +) (17) Pergamon Press, 239-304.
(P -P ) P ( t = k J + p Regenass, W. (1984). The control of exothermic
a c 0 c reactors. Symposium on protection of exother-
p (t=k ) 1 - P ( t=k, Pl(t=k ) mic reactors, Chester, UK.
2 + 0 + SchuTër,TL (1982). Frlihzeitige Erkennung gefahr-
licher Reaktionszustande in chemischen Reakto-
Eqs. (17) are now used in the decision stage to ren. Fortschr -Ber. VDI-Z., Reihe 8, Nr. 52.
discriminate between different models. Will sky, A.S. (1976). A survey of design methods
for failure detection in dynamic systems.
Results. Every filter tries to estimate its pro- Automatica, 12, 601-611.
cess state with its inherent model. Fig. 7 shows
the temperature reading and the estimates of the
different filters for a temperature-controlled
semi batch reactor. In the example below cooling is
not strong enough to maintain the reactor at an
isothermal behaviour. The probability ρ for nor-
mal operation mode will soon decrease.

The probability
lPof the decomposition mode in-
creases to almost one. If, moreover, water pene-
T[°C]
trates at time t=3600, the temperature rises even
faster. Comparing the qualities of estimates show
that only filter c) can interprète the measure-
ments well. The alarm for decomposition only is
removed and replaced by an alarm for simultaneous
breaking-in of water and decomposition (p =l). If
2
the cooling is not stopped, a runaway of the reac-
tor will occur as shown in Fig. 7.

The multiple filter method is used as well for

batch experiments, shown as well in Fig. 1. Only
two filters are used: one for normal operation
mode and one for decomposition mode. An arrow with
index (b) indicates the time when the probability
for normal operation mode is less than 0.5.
Fig. 1. Alarm creations for two adiabatic
Fig. 1 compares alarm creations based on the cri- experiments based
terion mentioned above and the results with a mul- a) on a criterion
tiple filter method. A significant difference is b) on a multiple filter method
 Early Detection of Hazardous States in Chemical Reactors 97

13,00 ^ Time 14,00 Fig. 5. Non-disjunctive regions of the models

a), b) and c).

Fig. 2. Prediction of time to runaway

(experiments).

Process ^^^^^

normat operation
mode

Filter 1 Fig. 6. Disjunctive regions in the state space.

Decision
1. fault mode
Stage

15
Filter 2

2.fault mode

Fig. 3. Multiple filter method for fault

discrimination..

Fig. 4. Markov Process.

Fig. 7. Alarm creation in a semi-batch reactor

with runaway (simulation).
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

AN ANALYTICAL APPROACH TO
APPROXIMATE DYNAMIC MODELING OF
DISTILLATION TOWERS
N. Kapoor and T. J. McAvoy
University of Maryland, Chemical and Nuclear Engg. Dept., College Park,
MD 20742, USA

Abstract. This paper presents analytical expressions for predicting distillate and bottoms
composition transient responses due to changes in manipulative and disturbance variables of a
tower. The authors in an earlier paper showed that towers are inherently recyle structures and
linearization techniques should be applied at perturbed conditions of tower models to evaluate
realistic tower time constants. In this earlier paper, a numerical approach to estimating time
constants was presented. The numerical approach is extended in this paper to an analytical
approach that requires only steady state and design information of a tower. The analytical
approach is relatively simple to use and much less time consuming compared to dynamic
simulation. Further, the analytical technique gives insight into w h y towers respond nonlinearly.

Keywords. Distillation; dynamic modeling; modelling; approximate models; transfer functions.

REVIEW OF PAST WORK

INTRODUCTION
Kapoor et al. (1986) carried out a detailed study of six tow-
This paper presents analytical expressions for predicting
ers, discussed by Fuentes and Luyben (1983). The essen-
distillate and bottoms composition transient responses due
tial features of these towers are shown in Table 1. The
to changes in tower manipulative and disturbance variables.
time constants shown in the last column of Table 1, were
To date, published work on analytical techniques relates to
obtained from numerically calculated frequency responses
predicting transient behavior of product compositions of
using a stepping technique, developed by Luyben (1973).
an absorber section only (Lapidus and Amundson(1950),
The numerically evaluated time constants, relate response
Amundson (1966) and K i m and Friedly(l974)). This paper
times of changes in distillate composition x d t o changes in
deals with a complete distillation tower and presents simple
feed composition χ ρ at design conditions. From the actual
analytical methods that m a y be used for short-cut dynamic
responses obtained from the non-linear simulation of tower
modeling of high purity towers. The analytical expressions
models, the gain, time constant and dead time were esti-
given in this paper require only steady state information to
evaluate time constants and dead times of product compo- mated and are listed in Table 2. The time at which the
sition responses for forcing in manipulative and disturbance
variables. Further, they provide information necessary for
identifying the key sections of a tower which are the major Table 1: Essential Features of Luyben's Towers
contributors to their dynamic behavior.

For predicting the transient responses of the product com-

Case# Feed a x b = NT Ns L
D
Θ (mm)
positions Wahl and Harriott (1970) presented a numerical Kr g
n o Us
(1 - xD)
technique that utilizes the design conditions of a tower.
Their predictions of time constants of product responses
are unrealistically large for high purity towers. A n earlier 1 65.8 4 .05 13 7 .53 10
paper by Kapoor and McAvoy(1986) discusses the cause of 2 61.7 4 .001 26 13 .73 450
such unrealistically large time constants. In this earlier pa- 3 61.7 4 .00001 40 20 .78 45400
per a numerical approach to estimating time constants is 4 36.3 2 .05 18 9 2.1 30
presented. In the present paper an analytical approach is 5 31.8 2 .001 40 20 2.4 1630
taken. To put the analytical approach into perspective the 6 31.8 2 .00001 60 30 2.6 170000
earlier numerical work is briefly reviewed here.
Θ = Time constant

99
100 N. Kapoor and T. J. McAvoy

T a b l e 2: R e s u l t s f r o m O p e n Loop Responses

Case# Forcing: : xp =.i

6 Forcing: =.4
G12
Gain Θ r Gain θ τ

Vyi
1 .281 10.0 8.0 1.51 8.0 8.0
2

4
.007

.00006

.359
23.0

25.0

27.0
15.0

22.0

12.0
1.99

2.0
13.0

10.0
26.0

36.0 3—
1.54 22.0 10.0
5 .007 25.0 22.0 G
6 .00006
1.99 23.0 50.0
8
28.0 28.0 2.0 30.0 85.0

LX V
T N G7
yNE+1
Θ = T i m e constant = Dead time

r e s p o n s e s h o w s a n o t i c e a b l e c h a n g e f r o m t h e initial s t e a d y

s t a t e c o n d i t i o n s is u s e d as a n approximation to t h e d e a d

time. T h e t i m e required t o r e a c h 6 3 % o f t h e final c o n d i -

t i o n s is a p p r o x i m a t e d as the time constant. T h e g a i n is

given b y the change in i d divided b y the change in χ ρ

. T h e highest purity t o w e r s r e s p o n d much faster t h a n t h e

VyNT+1
t i m e c o n s t a n t s , s h o w n in T a b l e 1, indicate. O b s e r v e also in
L'Xnt
T a b l e 2 t h e a s y m m e t r i c b e h a v i o r o f h i g h p u r i t y t o w e r s 3, 5,

a n d 6. T h e g a i n s a n d t i m e c o n s t a n t s t h a t a p p r o x i m a t e t h e

xd r e s p o n s e s for i n c r e a s e s in χ ρ a r e significantly different

from t h e gains a n d t i m e constants that a p p r o x i m a t e t h e

B X
B
xd r e s p o n s e s for d e c r e a s e s in xp. T o explain t h e a s y m m e t -

ric b e h a v i o r o f h i g h p u r i t y t o w e r s a n d also t o e x p l a i n t h e

discrepancy between t h e t i m e c o n s t a n t e s t i m a t e s listed in

Fig. 1. B l o c k t r a n s f e r f u n c t i o n
T a b l e 1 a n d t i m e c o n s t a n t s that a p p r o x i m a t e t h e non-linear

r e s p o n s e s ( T a b l e 2 ) , a simplified d y n a m i c m o d e l of a tower

is a n a l y s e d . F i g u r e 1 s h o w s t h e b l o c k d i a g r a m o f a distilla-

tion c o l u m n consisting o f four sections o f a t o w e r namely,

c o n d e n s e r , e n r i c h i n g section, s t r i p p i n g s e c t i o n a n d reboiler. & ι ο £ ? ι ι » a p p r o a c h e s u n i t y . T h e y also s h o w e d t h a t t h e g a i n s

T h e v a r i o u s t r a n s f e r f u n c t i o n s t h a t c a n b e d e f i n e d for feed of G 4 G 5 a n d G\qGχι rarely a p p r o a c h u n i t y in m o s t tow-

composition changes are s h o w n in F i g . 1. T h e transfer ers. H o w e v e r , t h e l o o p g a i n o f t h e HeHs transfer f u n c t i o n ,

f u n c t i o n s relating Dx~ô a n d Bx~b t o Fxp c a n b e derived K e K s , a p p r o a c h e s u n i t y in a v e r y s m a l l r e g i o n a r o u n d t h e

f r o m Fig. 1 a n d a r e given a s e q n s . 1 a n d 2. design conditions of high purity t o w e r s a n d indeed around

a n y s t e a d y s t a t e w h e r e h i g h p u r i t y is a c h i e v e d a t b o t h ends

o f a t o w e r . F i g u r e 2 a is a plot o f K e K s v e r s u s xp for t h e

high purity tower 3. A s c a n b e seen f r o m Fig. 2 a near

FxW H e H s J
(
G
" ) (i)
t h e d e s i g n c o n d i t i o n o f xp = .5 t h e K e K s loop gain a p -

p r o a c h e s 1.0. It w a s s h o w n in K a p o o r et al. ( 1 9 8 6 b ) t h a t for

K e K s a p p r o a c h i n g unity t h e d y n a m i c response of a tower

BxW = 2 will b e e x t r e m e l y s l u g g i s h . F i g u r e 2 a illustrates t h a t K e K s
Fxp {l-HEHs) ° ( l - G
4G 6)
(2)
is close t o 1.0 in a v e r y s m a l l r e g i o n n e a r xp = .5. T h e r e -

fore w e e x p e c t the towers d o m i n a n t t i m e constants to b e

where
G$GqGu l a r g e o n l y in this s m a l l r e g i o n . H o w e v e r , o u t s i d e this s m a l l
He = G? 4- (3) r e g i o n , i.e. a t slightly p e r t u r b e d conditions of t h e tower,
1 — G i o ^ n
the loop gain K e K s d r o p s off drastically f r o m 1.0 result-
and
G2G3G5 ing in a m u c h smaller a n d realistic t i m e constant. This

Hs = G 1 + (4)
result s u g g e s t s t h a t o n e s h o u l d u s e a p e r t u r b e d v a l u e o f xp
1 - G4G5 t o e s t i m a t e effective t o w e r t i m e c o n s t a n t s f r o m a linearized

model. T h e a u t h o r s accordingly used a perturbed steady

state t o e s t i m a t e t i m e c o n s t a n t s using t h e s a m e approach

Since t h e g a i n s o f all G » ' s a r e positive, F i g . 1 indicates as F u e n t e s a n d L u y b e n ( 1 9 8 3 ) . T h e s e e s t i m a t e s a r e plotted
that there are a number o f positive feedback loops in a in F i g . 2 b for c a s e 3 a t v a r i o u s v a l u e s o f χ ρ . Figure 2 b
tower's d y n a m i c structure. T h e 1 — HeHs term in e q n . s h o w s t h a t t h e t i m e c o n s t a n t s a r e e x t r e m e l y large a t d e s i g n
1, 1 — G i q G h term in e q n . 2 and 1 — term in e q n . c o n d i t i o n s o f xp = . 5 . H o w e v e r , outside a very small region
3 i n d i c a t e positive f e e d b a c k l o o p s resulting f r o m t h e r e c y - the time constants are m u c h smaller a n d reasonably con-
cle s t r u c t u r e o f a distillation column. I n fact, e a c h tray s t a n t for xp > .5 a n d xp < .5. F o r t h e transients shown
i n v o l v e s a positive f e e d b a c k loop. B y lumping t h e enrich- in F i g . 1, l a r g e χ ρ c h a n g e s w e r e u s e d t o force t h e t o w e r s
i n g a n d s t r i p p i n g s e c t i o n s o f a t o w e r into He a n d Hs it and the towers operated a t xp v a l u e s substantially differ-
is p o s s i b l e t o e l u c i d a t e t h e u n d e r l y i n g r e a s o n for t h e large ent f r o m t h e design values. Since time constants change
time constants w h i c h result from a linear analysis. T h e drastically w i t h χ ρ , time constants m u s t b e evaluated at
a u t h o r s in a n o t h e r paper (Kapoor et al. (1986b)) showed perturbed conditions of a tower. Fig. 2 b s h o w s that t h e
that extremely large t i m e c o n s t a n t s will result if t h e g a i n time constants change in a fairly linear f a s h i o n w i t h re-
o f a n y o f t h e positive f e e d b a c k l o o p s , i.e. H e H s , G4G5 or s p e c t t o xp a t p e r t u r b e d c o n d i t i o n s . If a n a v e r a g e v a l u e o f
 Approximate Dynamic Modeling of Distillation Towers
LOOP GAIN V S XF
Fig. 2a. K K vs χ ρ
ES
TIME CONSTANT V S XF
τ—j—r
θ 50 θ 60
Fig. 2b. Bvsxp
the time constants is estimated over the linear zone then a
better approximation to the simulation results can be ex-
pected. Such an average value of the time constants can be
obtained approximately half way between the design and
perturbed conditions. For example, if the tower models are
forced from xp = . 4 to .5 then the time constants should be
estimated at xp =.45. The results presented in this paper
are obtained by linearizing tower models at conditions half
way between the initial and final steady state values.
To this point this paper has presented a review of the nu-
merical technique of Kapoor et al. (1986). Our objective
is to develop analytical expressions for all the composition
tranfer functions of a tower for all forcing variables. Further
we want to reduce the final complicated transfer functions
(for example eqns. 1 and 2 for xp changes) to a first order
with dead time model by the method of moments(Gibilaro
and Lees(1969)). A reduction in order of final transfer func-
tions will provide simple analytical relationships between a
tower's dynamic model and its physical variables. The next
section presents analytical expressions for x d respones for
forcing in xp . Additional analytical transfer function ex-
pressions for other forcing variables are given in Kapoor
and McAvoy (1986).
101
REDUCING T H E FINAL TRANSFER FUNCTION
To obtain a simple dynamic model for the x d response due
to xp changes we will make the following assumption. A
first order model with dead time approximates the product
composition responses for xp changes. Equation 1 is the
transfer function of the x d response due to χ ρ changes and
it can be divided, into 3 terms. The first term of eqn. 1 is
H s / ( l — H e H s) and it gives the change in the composition
of the vapor entering the enriching section due to a change
in x f . The second term in eqn. 1 is G q . Referring to
Fig. 1, G q gives the change in the composition of the vapor
entering the condenser to a change in the composition of
the vapor entering the enriching section. The third term in
eqn. 1 is G i 2 / ( l — GioGu) and it gives the change in the
x d response due to a change in the composition of the vapor
entering the condenser. The effective time constant of the
x d response for xp changes will therefore be a combined
effect of all three terms. To develop low order transfer
functions the first step is to reduce the order of GVs, H e
and H s in eqn. 1. After substituting the reduced forms
of Gt's, H e and H s in eqn. 1 the moments of eqn. 1 are
equated to the moments of a first order transfer function
with dead time, the form of which is given as:
K -T {x )s
Dxp e DF (5)
Fxp e(x )s+1
F
To reduce eqn. 1 to eqn. 5, the gain K, the dead time
and the time constant B(xf)
t d [ x f ) are related to the pa-
rameters in eqn. 1 by the method of moments. A detailed
derivation for Κ, θ(χρ) and is given in Kapoor( 1986).|
t
d{ x f )
The final expressions are:
K = 7 K9 RR+1
1 — ΚχηΚι (6)
( l - K K )
E S
Loop l(Enr. Sect.) Loop 2
1 — Λ10Λ11
KeKs — K\K>j = l\ (8)
t d { * f ) = ( t
d) 9 (9)
The expressions for the gains uTi's, TVs and t d q are given in
Table 3. Equation 7 shows that the time constant of the x d
response due to χ ρ changes is comprised of three parts. The
first term, labeled Loop 1, is due to the recycle structure
formed by integration of the enriching and the stripping
sections. The Loop 1 time constant refers to the response
time of the composition of the vapor entering the enriching
section due to a change in xp. The Loop 1 time constant
can be extremely large if the variable Ιχ which is the product
of the enriching and stripping section gains, K e K s i is very
close to 1.0. I n Fig. 2a it is shown that at design conditions
of high purity towers li — • 1.0 and therefore the Loop 1
time constant in eqn. 7 is extremely large. The Loop 1
time constant is the major contributor to the time constant
of x d in the small neighborhood near steady state. As can
be seen from Fig. 2a, the loop gain l\ drops off drastically
from 1. at perturbed conditions, implying that the Loop 1
time constant is not necessarily the major contributor to the
time constant of the x d response at perturbed conditions
of a tower. This behavior agrees with our numerical results.
102 Ν. Kapoor and T. J. McAvoy

The time constant To. refers to the response time of yi due Table 3 : Gains, Leads and Lags
to a change in vn+i (See Fig. 1). The time constant Tg
is usually much larger than the Loop 1 time constant at
perturbed conditions of a tower. The third term in eqn. 2
7, labeled Loop 2, is a result of the feedback loop formed (l + (a-l)xF) RR+% axpRR
Ae =
by integration of the condenser with the enriching section.
s+1a RR+1 1++
xpRR
1
A s shown in Kapoor et al. (1986) the loop gain I2 rarely A% - 1 DE = A£« - 1
approaches unity. Therefore, the contribution by the Loop
AST- - 1 A^-AE
2 time constant to the effective time constant of the x d K7 =
response will be approximately of the order of the sum of Ds De
(AE- I)
the condenser time constant, Tc, and T7. AN S + 1 N
S AK9 = 7
1
Ds DE
Equation 9 is the expression for the dead time of the x d Ae * - 1
response for xp changes. The dead time is a sum of all A»** -As ΚIQ =
DE
but the largest lag of the enriching section. All the lags Ds
2 RR
of an enriching section can be evaluated using analytical 1 1
Kn = RR + 1
B(l + ( a - l ) x )
expressions developed by K i m and Friedly(l974) and are t"·" Va
B 1
presented in Table 3 . Equations 7 and 9 were used for the Κ \2 =RR + 1
κ6 1-K5
evaluation of time constants and dead times of the x d re-
sponses for xp changes. Results obtained from eqns. 7
and 9 for tower 3 are compared with the the non-linear
simulation results in Fig. 3 . The tower model starts at
From K i m and Friedly's analysis:
a perturbed initial steady state corresponding to xp = .4
and the response to the design condition corresponding to H 1
x = .5 is shown. A s can be seen from Fig. 3 the agreement Lagj — L C O S
F
between the analytical responses and non-linear simulation + (J$T)
results is very good. W e studied all the six towers illus- (Enr.Sect.)
trated in Table 1 in detail (Kapoor (1986)) and compared H 1
Lead* = y=i,...,i\r-i
the analytical expression results with the non-linear sim-
ulation results for forcing in χ ρ from design to perturbed
'«*(#)
conditions and back. Both positive and negative pertur-
(Enr.Sect.)
H 1
bations in xp were considered. The analytical expressions Lagi = i = 1,...,JV
compared very well with the non-linear simulation results
for all the cases except for forcing m χ ρ from design to a (Str.Sect.)
decreased value. A significant error in the dead time was H
Leadi L
1
i = l,...,JV-l
obtained for this case. This error m a y not be a serious
' (l+^)-2V^cos(f)
problem since one is normally interested in determining re-
sponse times of towers responding from upset conditions (Str.Sect.)
back to their design specifications.

X RESPONSES FOR X CHANGES

Equations 6-9 are analytical expressions that relate the B F
gains, time constants and dead-times of the enriching and
stripping sections to x d response times for forcing in xp. The transfer function relating x b responses due to changes
Table 3 shows that the gains, time constants and dead- in xp is given as eqn. 2. Equation 2 can be reduced by the
times of the enriching and stripping sections can be related method of moments, to a simplefirstorder transfer function
to the physical variables of a tower. Further, the relation- with dead time following the same approach used for the
ships presented in Table 3 provide a method for relating the x d case. Thefinalexpressions for the dynamic parameters
product composition response time to the physical variables
such as holdup on a tray, in the condenser and reboiler, of x b responses for forcing in xp are given as:
flow variables, relative volatility, etc.. The expressions for
x b responses to changes in xp can be derived in the same
manner as for the x d case and are presented in the following Loop l(Stripp.Sect.)
section.
(10)
& x
b( x f ) = — —
2
+T +
1-U

(H)

and
t b { x f ) = {td)2 (12)

The Loop l(Stripp. Sect.) time constant in eqn. 10 gives

the effect of thefirstterm of eqn. 2. The Loop 3 time
constant gives the effect of the loop formed by the reboiler
and the stripping section on the x b response time. The
time constant T2 gives the response time of x n
t due to
changes in χχ\ Equation 12 gives the dead time of the
x b responses due to changes in xp . The analytical ex-
TI>C <KINUTES)
pressions are used to evaluate the time constants and dead
times of the towers when forced from increased values of
x b to design values. Figure 4 gives the x b response for
Fig. 3 . x d response for xp changes
case study tower 3 . Analytical expression results for all the
 Approximate Dynamic Modeling of Distillation Towers

Table 3 continued

N
S N
E
T La
T
7 = (Σ Η
i=i
i = E 9i
Ne-1
t=L — ^2 Leadj)
Ns-1 3=1
— Leadi) 1 11 1
(Lagx)
i=l ! ' I '
tee β
ΐ5
T = (Lag!) Ne
2
N
S {td)q •(^Lagj) TIME «MINUTES
)

{TD)2 = {J2La )
9i
3=2
3=2 He
Ts = T
R V
V + B/a

Hr = Reboiler
He = Condenser
Holdup Holdup

Π ΜΕ CMTNUTE3
)

Fig. 5. x h x
D B responses for V changes

Fig. 4. x
B response for xF changes

cases listed in Table 1 are compared with non-linear sim-

ulation results in Kapoor (1986). The agreement between ΠΜΕ <KINUTES>
the simulation and the analytical expressions is good for all
the cases studied.

Since space does not permit the presentation of additional

analytical expressions for the product composition responses|
for forcings in V and Β and L and JD, they are given in
Kapoor and M c A v o y (1986). However, results for tower 3
for forcing in V and L are given in Figs. 5 and 6. Figures 5
and 6 give a comparison of the approximate xd and χ β re-
sponses determined from the analytical expressions and the
non-linear responses. The agreement between the analyti-
cal expression results and the simulation results is excellent
for the V case and very good for L case. In Kapoor (1986)
analytical expression results for all the six towers and a TIME CMINUTES)
number of towers with non-ideal and multicomponent fea-
tures are compared with non-linear simulation results. A
reasonably good comparison is obtained for most of the
cases studied. A s a result the analytical expressions derived
in Kapoor (1986) present an approximate and short-cut ap- Fig. 6. xjd & 2jb responses for L changes
proach to modeling towers in a timely manner. Addition-
ally, the expressions provide insight for high purity towers
where the existing linearization techniques fail.
DCCR-H
104 Ν. Kapoor and T. J. McAvoy

CONCLUSION W A H L , E . F . A N D HARRIOTT, P . , ( 1 9 7 0 ) . I E C P R O C . P E S . D E V . .
9, P P . 396-407.

T H E ANALYTICAL E X P R E S S I O N S PRESENTED I N THIS P A P E R ARE A N

EXTENSION OF THE ANALYSIS TECHNIQUE E M P L O Y E D I N K A P O O R A N D
M C A V O Y (1986). I N THIS EARLIER P A P E R , IT I S S H O W N THAT D I S -
TILLATION TOWER M O D E L S C A N B E M O D E L E D B Y BLOCK TRANSFER NOMENCLATURE

FUNCTIONS INTEGRATED TO EACH OTHER I N A RECYCLE PROCESSING

CONFIGURATION. I N THIS P A P E R ANALYTICAL E X P R E S S I O N S FOR THE
VARIOUS BLOCK TRANSFER FUNCTIONS ARE DEVELOPED A N D REDUCED AE EFFECTIVE A B S O R P T I O N FACTOR
TO FIRST ORDER TRANSFER FUNCTIONS W I T H OR WITHOUT D E A D T I M E S Β B O T T O M S FLOW
B Y THE M E T H O D OF M O M E N T S . T H E S E REDUCED ORDER TRANSFER D DISTILLATE FLOW
FUNCTIONS ARE U S E D TO PREDICT R E S P O N S E T I M E S OF TOWERS FOR F F E E D FLOW
FORCING I N M A N I P U L A T I V E A N D D I S T U R B A N C E VARIABLES. T H E RE- Gi TRANSFER FUNCTION
SULTS O B T A I N E D F R O M THE ANALYTICAL E X P R E S S I O N S ARE C O M P A R E D He TRANSFER FUNCTION FOR ENRICHING SECTION
W I T H THE NON-LINEAR S I M U L A T I O N RESULTS. T H E A G R E E M E N T B E - Hs TRANSFER FUNCTION FOR S T R I P P I N G SECTION
TWEEN THE RESPONSES O B T A I N E D FROM THE ANALYTICAL E X P R E S - Ke G A I N FOR E N R I C H I N G SECTION
S I O N S A N D THE NON-LINEAR RESPONSES I S GOOD FOR A N U M B E R OF Ks G A I N FOR S T R I P P I N G SECTION
CASES S T U D I E D I N K A P O O R ( 1 9 8 6 ) . I N THIS P A P E R , RESULTS FOR A L R E F L U X FLOW I N T H E E N R I C H I N G SECTION
H I G H PURITY TOWER ARE PRESENTED. T H E H I G H PURITY TOWER I S V REFLUX FLOW I N THE S T R I P P I N G SECTION
FORCED F R O M P E R T U R B E D CONDITIONS B A C K TO ITS D E S I G N S P E C I - N T R A Y S I N T H E E N R I C H I N G SECTION
E TRAYS I N THE S T R I P P I N G SECTION
FICATIONS. T H E REASON FOR S T U D Y I N G THIS FORCING I S THAT FOR N
CONTROL A N D OPERABILITY S T U D I E S O N E I S INSTERESTED I N DETER-
s REFLUX R A T I O
RR
M I N I N G THE R E S P O N S E T I M E S OF TOWERS FOR FORCING FROM UPSET EFFECTIVE S T R I P P I N G FACTOR
Ss
CONDITIONS B A C K TO T H E NORMAL OPERATING SETPOINTS. V A P O R FLOW
V
B O T T O M S M O L E FRACTION
xb
xF F E E D M O L E FRACTION
T H E ANALYTICAL E X P R E S S I O N S PRESENTED I N THIS P A P E R REQUIRE xd DISTILLATE M O L E FRACTION
ONLY STEADY STATE A N D D E S I G N INFORMATION OF T H E TOWER TO
EVALUATE TOWER T I M E CONSTANTS. FURTHER, THE ANALYTICAL E X -
PRESSIONS HELP I N D E T E R M I N I N G W H I C H SECTIONS OF A TOWER, GREEK LETTERS
N A M E L Y , CONDENSER, E N R I C H I N G A N D S T R I P P I N G SECTIONS A N D RE-
BOILER, ARE THE M A J O R CONTRIBUTORS TO T H E R E S P O N S E T I M E S .
FINALLY, THESE E X P R E S S I O N S P R O V I D E INSIGHT INTO THE EFFECT OF
a RELATIVE VOLATILITY
A CHANGE I N D E S I G N O N THE R E S P O N S E T I M E S OF A TOWER. TIME CONSTANT
Θ
DEAD TIME

R E F E R E N C E S

AMUNDSON N.R.,(1966) . MATH. METH. IN CHEM. E N ^ R . P R

TICE HALL.
P. N

FUENTES C , A N D L U Y B E N W . L . , ( 1 9 8 3 ) . I E C PROC. P E S . D E V . J
22, 361-366.

GIBILARO, L A N D L E E S F., (1969). C H E M . E N G G . S C I . . 24, P P .

85-93.

K A P O O R , N . , A N D M C A V O Y , T . J . , ( 1 9 8 6 ) . S U B M I T T E D TO

AUTOMATICA.

KAPOOR, N., M C A V O Y , T.J.,AND MARLIN, T . E . , MARCH(L986).

A I C H E JOURNAL P P 4 1 1 - 4 1 8 , V O L . 3 2 , N O . 3 .

KAPOOR, N., (1986). A NUMERICAL A N D A N ANALYTICAL A P -

PROACH TO A P P R O X I M A T E D Y N A M I C M O D E L I N G OF DISTILLATION
TOWERS. P H . D . THESIS, U N I V E R S I T Y OF M A R Y L A N D .

KIM C , A N D FRIEDLY J . C , (1974). I E C PROC. P E S . D E V . .

13, 1 7 7 - 1 8 1 .

LAPIDUS, L., AND A M U N D S O N , N.R.,(1950). I N D . ENGG. C H E M .

42:1071.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

ANALYSIS OF THE DYNAMIC BEHAVIOR OF

DISTILLATION TRAINS
R. Gani, M . Albujo-Cuadrado, C. A. Ruiz
and M . Nazmul Karim*
Instituttet for Kemiteknik, Denmarfa Teknishe H0jskole, 2800 Lyngby, Denmark
*Plapiqui, UNS-CONICET, 8000 Bahia Blanca, Argentina

Abstract. A detailed study on the dynamics of distillation trains has been carried out
by simulating the behavior of multiple distillation column processes ( with and without
recycle streams ). By analyzing the dynamic behavior, potential design, operability and
control problems have been identified. Based on the dynamic analysis, design procedures
by which nonlinear behavior of the columns can be eliminated and suitable control stra-
tegies applied, are presented. Numerical results illustrating the analyis of above and
the application of the proposed procedure are also presented.

Keywords. Dynamic model? distillation; multiple columns; transient response; openloop;

closedloop; nonlinear behavior.

INTRODUCTION

Chemical processes, and in particular, the separa-
tion units of these processes usually have more
than one distillation columns arranged in differ-
ent configurations in different processes. The
design and control of distillation columns has
been an important area of research for more than
a decade. Most of the work however, has been limi-
ted to single distillation columns. At the same
time, significant progress has been made in the
development of algorithms for flexible process
design and robust and resilient control of proces-
ses ( Grossman and Morari, 1984, Morari, 1985,
Arkun, 1985 ).As a result, identification of fac-
tors sensitive both to process design and control
has gained importance.
The purpose of this paper is to show how, by stud-
ying and analyzing the dynamics of distillation
trains, the factors important for process design
and control can be identified and how these infor-
mation can then be used to obtain better design
and operation of the process by employing suitable
control algorithms.
The dynamic behavior is generated by simulations
with the generalized dynamic model of Gani and co-
workers (1986). Multiple distillation column conf-
igurations with and without recycle streams has
been considered. In this work, the nonlinear resp-
onse is considered as that which can be represent-
ed by second order ( or higher ) transfer functio-
ns with or without time delays.
Of the three separation schemes studied, the ben-
zene, toluene and xylene ( BTX ) separation process
and the styrene separation process represents more
complex configurations and therefore more difficult
design and control problems. As the objective of
this paper is not to develop the final control str-
ategies, only a simple example of the closedloop
response is presented ( for the BTX process ).
PROBLEM FORMULATION
Given, the details of the distillation train, study
f drum dynamics, reboiler dynamics and effects of
Chem. Eng. Dept.,CSU, Fort Collins,80523 Co, USA
and analyze the dynamic behavior of the system
from a control, operability and process design
viewpoint by simulations of the operations of the
process.
The required details of the process are, the col-
umn configuration details, mixture information,
individual column data ( number of plates, plate
geometry etc. ) and the conditions of operation
( for example, feed data, reflux rate, heat duties,
product specification ). Knowing these details,
simulations are performed to obtain the steady
states corresponding to the specified conditions
of operation. This steady state will be the refer-
ence steady state ( RSS ) and all subsequent stu-
dies are made with respect to this reference state.
Generation of Transient Responses
At the reference steady state, disturbances are
introduced in the conditions of operation ( one
variable at a time ) and the corresponding open-
loop transient responses are simulated through the
dynamic model. The variables that are disturbed
are the feed flowrate, feed enthalpy ( that is,
the feed state ), feed composition, reflux rate
( or distillate rate ) and reboiler heat duty ( or
vapor boilup rate ) of the upstream column ( that
is, column 1 ). Column 1 will be regarded as the
column where the fresh feed enters.the process.
Openloop transient responses for the corresponding
column and all the downstream columns ( that is,
column 2, column 3, etc. ) is obtained through
the dynamic model.
Simulation Model
The dynamic model of Gani and coworkers (1986)
incorporates the effects of simultaneous mass and
heat transfer as well as detailed hydraulic beha-
vior of the plates. The model is flexible enough
to allow the study of single column or multiple
column processes, including recycles to upstream
columns. The model allows simulations with vari-
ous degrees of complexity. The most complex vers-
ion considers detailed plate hydraulics, reflux
mass and heat transfer. The most simple version
only considers the effects of mass and heat trans-

105
106 R. Gani et al.
fer. For multiple column processes it is possible
to simulate the different columns with different
degrees of complexity. The dynamic model has built-
in thermodynamic submodels based on equations of
state ( Soave, 1972 ) as well as group contribut-
ion methods ( Fredenslund and coworkers, 1975 ).
Thus, a very wide variety of mixtures can be stud-
ied by the dynamic model. Also, the dynamic model
has a builtin numerical methods submodel which pro-
vides robust, reliable and efficient numerical te-
chniques ( Cameron and coworkers, 1986 ).
Details of the Problems Studied
Three multiple column distillation processes have
been studied ( Murphy, 1976, Doukas and Luyben,
1981, Hess and Holland, 1976 ). These processes
were earlier employed by Gani and coworkers (1985)
to test the dynamic model and, for a preliminary
study of their dynamic behavior.
The first process involves a butanes/pentane sepa-
ration scheme. The distillate of column 1 is the
feed to column 2, where, i-butane and η-butane are
separated. Pentane is removed through the bottoms
of column 1. The second process involves a benzene
toluene and xylene ( BTX ) separation scheme. The
distillate and bottoms from column 1 enter as feed
into column 2. Besides the distillate and the bot-
toms, column 2 also has a side product. The third
process is a styrene separation unit. The feed
mixture ( entering column 1 ) contains toluene,
ethylbenzene and styrene. The distillate of column
2 is recycled back to column 1. Table 1 lists some
of the necessary details of the three processes.
It can be seen that the first two configurations
do not contain any recycle streams while the third
process has a single recycle stream. Of the first
two processes, the butanes/pentane process has a
much simpler configuration.
ANALYSIS OF THE OPENLOOP DYNAMIC RESPONSE
As mentioned above, the analysis of the openloop
dynamic responses is made by perturbing one var-
iable at a time and observing the resultant res-
ponse with respect to the reference steady state*
The variables that are perturbed are the variables
that are usually manipulated for control purposes
or those variables which are designated as input
( or design ) variables. In case of configurations
without recycle, only variables of column 1 is
disturbed. For the styrene separation process,
variables of both the columns are disturbed ( only
one variable at a time ).
Configurations Without Recycle
Table 2 shows the reference steady state and the
steady states obtained from simulations tor tne
butanes/ pentane separation process. These results
correspond to disturbances in the reboiler heat
duty and the feed composition. Table 3 shows the
same results for the BTX separation process but
for disturbances in the reflux rate and feed com-
position. Several observations can be made from
the results of Tables 2 and 3. The disturbance
has greater effect on the downstream column than
in the upstream column. However, the effect on the
BTX process is greater than on the butanes/pentane
process. The reason for this is the following : in
column 1, only one variable is perturbed while in
column 2, the feed that enters, may undergo chan-
ges in flowrate, temperature and composition. The
BTX process undergoes greater changes because both
the feeds are products of column 1. As will be
shown later, this is also the cause for nonlinear
responses in this process. The steady state gains
also show a larger interaction between the two
columns in the BTX process than in the butanes/
pentane separation process ( see Table 4 ).
Figure 1 shows some of the composition profiles
of the three processes. The changées of the compo-
tion profiles due to the disturbances are illusta-
ted in Fig. 2. From these steady state composition
profiles, the location of the sensitive regions of
the columns were determined by applying the method
of Brignole and coworkers (1985). It was found
that column 2 always had a greater sensitivity
than column 1. Table 5 shows the location of the
sensitive points and the corresponding values of
the sensitivity indices ( Brignole and coworkers,
1985 ) for all the columns. The locations of the
sensitive points were found to vary ( in some cas-
es ) with respect to the dynamic response ( that
is, these locations varied with time before sett-
ling down to a steady value ).
Figures 3 and 4 illustrate some of the dynamic
responses for the butanes/pentane separation pro-
cess and the BTX process respectively. In all ca-
ses, for the butanes/pentane process, the respon-
ses from both the columns can be represented by
first order transfer functions ( in some cases
with delay ). In the BTX process, while column 1
behaved like any other single column ( and having
first order transfer functions ) , column2 in some
cases showed nonlinear behavior ( that is requir-
ing second order or higher transfer functions ).
Figure 5 shows the dynamic responses of column 2
for two different sizes of disturbances in the
same variable. The size of the disturbance causing
the nonlinear behavior depends on the type of var-
iable being disturbed, configuration of the colu-
mns and the type of variable being measured. Thus
design and control problems can be formulated for
which the nonlinear behavior will not exist.
The results of Figs. 3 and 4 also show the exist-
ence of time delays whose length was also found to
depend on which variable is disturbed, process
configuration and the location of the feed. These
factors are therefore important both for control
as well as for design of the process.
Configuration With Recycle
Comparison of the steady state profiles show simi-
lar disturbance effects on either columns and sim-
ilar sensitivity indices on either columns. This
is reasonable since a recycle stream is coming ba-
ck to the upstream column from the downstream col-
umn. Table 6 shows the reference steady state and
the steady states obtained due to disturbances in
the reboiler heat duty and the feed composition.
Some of the dynamic responses corresponding to the
results of Table 6 are illustrated in Fig. 6. It
can be seen that column 1 behaves in a manner sim-
ilar to the column 1 of the butanes/pentane process
while column 2 has a tendency for nonlinear respo-
nse ( note also the comparitively larger time del-1
ays in column 2 ). Again, by increasing the distu-
rbances, it was possible to obtain the nonlinear
behavior. But in this process, it was possible to
obtain nonlinear behavior for both columns. In the
results of Fig. 6, the effect of the recycle str-
eam on column 1 is almost negligible. On the other
hand, the continuous change from column 1 is imp-
ortant for column 2. By changing the operating
conditions however, it should be possible to incr-
ease the effect of the recycle stream on column 1.
One interesting aspect of configurations with re-
cycle streams is that the process might not come
to a steady state ( in the openloop ) if the dis-
turbance is too large or, if the initial state is
too far from the desired steady state. One way the
desired steady state can be reached in this case,
is to control ( at least ) the flow of the recycle
 Analysis of the Dynamic Behavior of Distillation Trains
stream ( shown in Fig. 7 ).
IMPLICATION ON PROCESS DESIGN AND CONTROL
From the results of the openloop dynamic responses
several typical behavior can be noted :
- The downstream columns undergo greater changes
than the upstream columns and because of this
reason they are more sensitive.
- Interaction between the columns depend on the
process configuration and conditions of oper-
ation.
- The sensitive location of the downstream col-
umns can vary dynamically with time.
- For some configurations, columns can give non-
linear responses ( openloop ) due to disturb-
ances in the upstream column. The amount of
disturbance necessary to cause nonlinear res-
ponses depends on the process configuration
and design variables ( related to conditions
of operation ).
- The dynamic responses in the downstream show
time delays and the length of these depend on
the process configuration, type of variable
being disturbed and on the location of the
feed.
The effect of plate hydraulics on the dynamic beh-
avior has also been studied in this work. It has
been found that the plate hydraulics are only imp-
ortant when highly nonlinear dynamic behavior is
present. The dynamic response of the height of the
liquid on the downcomer for the styrene process is
also shown in Fig. 7. Besides the above mentioned
conditions, plate hydraulics have been found to be
important during startup and shutdown operations
of distillation columns ( Gani and coworkers, 1986).
Important Factors For Design And Control Problems
Based on the observations of above, several 'acti-
ons' that can result in better design of the pro-
cess and at the same time simplify the control st-
rategy are now proposed. The objective for perfor-
ming these 'actions' are to a) reduce the sensiti-
vity of the downstream columns to changes in the
corresponding upstream column, and b) eliminate the
the possibility of any nonlinear response.
Reduction of sensitivity will also reduce the int-
eraction between the columns. Even though the int-
eractions can be 'handled' by modern control tech-
niques, the control problem will be simpler if they
are minimized. The elimination of the nonlinear
responses will also eliminate the requirement of
more complex control strategies. Also, the effect
of plate hydraulics need not be considered. Finally
the time delays should be correctly predicted so
that the control techniques can be efficiently app-
lied. The question of the elimination of the time
delays or the length of this period need to be stu-
died togather with the control technique ( or stra-
tegy ) being employed.
The most sensitive variable with respect to column
sensitivity and nonlinear behavior should be first
identified ( a dynamic simulation can be used for
the purpose ). Depending on which is possible, cha-
nges in process configuration, conditions of opera-
tion or individual column details can be considered
in order to reduce the importance of the most sen-
sitive variable ( or variables ). By dynamic simu-
lations the effects of these changes can be studied.
Development Of A Control Strategy
A simple control problem was defined for the purp-
ose. The objective is to keep the product specifi-
cation ( temperatures of the products ) at the de-
sired conditions. From the openloop dynamic respo-
nse the sensitive points of the columns and the
107
disturbance size of the variables that can cause
nonlinear behavior in any of the columns are deter-
mined. From the sizes of the disturbances, the
steady state gains and the values of the sensitiv-
ity indices, the most sensitive variables can be
identified.
If nonlinear behavior is present and by changing
the design variables, it cannot be completely eli-
minated, then control action on the stream linking
the two columns can be considered. This is shown
by a simple example. For the disturbance for which
nonlinear response was obtained in the BTX process
( see Fig. 5 ) , a simple flow control was used for
the stream linking the two columns. The resulting
response is also shown in Fig. 5. It can be clearly
seen that significant improvement is possible.
Therefore, the nonlinear behavior can be reduced
( if not eliminated ) by controlling the product
of the upstream column ( which is the feed for the
downstream column ). The control strategy for the
upstream column therfore should be ( in addition
to satisfying any product requirement ) to control
stream that is connected to the downstream column.
This can easily be done by applying the modern co-
ntrol techniques once the dynamic analysis has been
made and proper changes in the design have been
made.
If the nonlinearity of the downstream column is
eliminated by the control action of the upstream
column, then the design of any control strategy is
significantly simplified. Control techniques taking
into account time delays can easily be applied for
this column.
Closedloop response for the BTX process. For the
BTX process, two PI controllers controlling the
top and bottom temperatures of column1 by manipul-
ating the reflux rate and reboiler heat duty were
designed. For column 2, the temperature of the most
plate was controlled with a PI controller by mani-
pulating the reboiler heat duty. From the openloop
responses of Fig. 4, it can be seen that the time
delays are almost negligible. Figure 8 shows the
openloop and closedloop responses of the top and
bottom temperatures of column 1 for +1% change in
the feed enthalpy ( that is changing the feed tem-
erature ). Figure 9 shows the openloop and closed-
loop responses of the top and bottom temperatures
of column 2 for the same disturbance. The openloop
response of the top temperature of column 1 is also
plotted in Fig. 9 to illustrate the difference in
the magnitude of the changes in the two columns.
CONCLUSIONS
The dynamic model has been succesfully applied to
simulate and analyze the openloop transient respo-
nces of three multiple column processes. Important
factors related to process configuration, column
design and conditions of operation have been iden-
tified as those influencing the final design and
control of the process. Numerical results showing
how by dynamic simulations important interrelated
problems in process design and control can be sol-
ved or simplified have presented. And, finally,
although the closedloop responses for the BTX pro-
cess show promise, more work is necessary before
any general guidelines can be established for these
problems.
REFERENCES
Brignole, E.A., R. Gani and J.A. Romagnoli (1985),
A simple algorithm for sensitivity and oper-
ability analysis of separation process, Ind.
Eng. Chem. Process Pes. Dev., 24, 42-48.
Cameron, I.T, C.A. Ruiz and R. Gani (1986) , A
108 R. Gani et al.

generalized dynamic model for distillation Table 2 Comparison of the Steady States for
columns, part II : numerical and computation- for the Butane/pentane Process.
al aspects. Computers and Chem. Eng., 10 (3),
in press.
Doukas, N.P. and W.L. Luyben (1981), Control of an
energy-conserving prefractionator/sidestream
column distillation system, Ind. Eng. Chem. Column 1 + # Column 2
Process Pes. Dev., 20, 147-153. RSS* SS1 SS2 RSS SS1 SS2
Fredenslund, Aa, R.L. Jones and J.M. Prausnitz Top plate
(1975), Group contribution estimation of act- butane •2449 •2501 •2401 •4061 •3929 •4117
ivity coefficients in nonideal liquid mixt- i-butane •2981 •3046 •2935 •4294 •4208 •4303
ures, AIChE J., 2Π, 1086-1099. pentane •2470 •4453 •4664 •1644 •1863 •1579
Gani, R., C.A. Ruiz and I.T. Cameron (1985), Stu-
Bottom plate
dies in the dynamics of distillation trains,
IChemE Symposium Series, 92, 353-364. butane •0249 •0258 •0239 •0293 •0242 •0321
i-butane •0587 •0610 •0565 •0783 •0651 •0846
Gani, R, C.A. Ruiz and I.T. Cameron (1986), A pentane •9164 •9132 •9196 •8924 •9107 •8833
generalized dynamic model for distillation * RSS: Reference steady state
columns, part I : model description and SS1: Steady state corresponding to change in reboiler heat
applications. Computers and Chem. Eng., 10 (3), duty (+1%)
in press. # SS2: Steady state corresponding to change in feed composition
Grossman, I.Ε and M. Morari (1984), Operability,
resiliency and flexibility - process design
objectives for a changing world. Proceedings
of FOCAPD'83, CACHE, An Arbor, 931-1010. Table 3 Comparison of the Steady States for
Hess, F.E and C D . Holland (1976), Solve more dis- for the BTX Process.
tillation problems : part 6, Hydrocarbon
Processing, 55(6), 125-131.
Morari, M. and S. Skogestad (1985), Effect of
model uncertainty on dynamic resilience,
IChemE Symposium Series, 92, 493-504. Column 1 Column 2
Murphy T.J (1976), Control of heat integrated RSS SS2 SS3* RSS SS2 SS3
distillation columns, MS Thesis, University Top plate
of Minnesota, USA. benzene •7053 •7402 •7122 •9922 •1943 •9925
Palazoglu A. and Y. Arkun (1985), Studies on the toluene •2942 •2594 •2874 •0078 •0057 •0075
design of robust chemical plants, IChemE xylene •0005 •0004 •0005 - -
Symposium Series, 92, 457-468. bottom plate
benzene •1529 •2333 •1543 •0203 •0272 •0196
toluene •2942 •2594 •2874 •0078 •0057 •0075
Table 1 Details of the Three Processes Studied xylene •3086 •1703 •3070 •3723 •3110 •3727
*SS3: steady state corresponding to change in reflux rate (+1%)

Table 4 Location of the Sensitive Plates and

Separation process
the Corresponding Sensitivity Indices.
Butane/ BTX Styrene

Column '1 Column 2 Column 'I Column 2 Column 1 Column 2

No. of plates 7 22 19 40 20 20
Reflux rate Sensitive plate Sensitive index
(kmol/h) 48 78 50 57 453 45 Separation
Column 1 Column 2 Column 1 Column 2
Reboiler heat
Butane/Pentane 1 16 •0094 ·0538
duty (kmol/h) 358400 281600 486400 460800 384 128 BTX 20 22 •23146 «4449
Side product: Styrene 20 20 •0479 ·0516
flow rate
(kmol/h) - - - 51 - 32
plate no. - - - 20 - 14
Feed details:
No. of feed
Location
1
4
D1
11
1
9
Β2
4,24
2
5,14
B
1
9 Table 5 Steady State Gains with respect to
Type* fresh fresh
Composition
1 °Γ 1
fresh,D
2 1 Reflux Rate of Column 1 and Composi-
(mole
tion of Compnent i in Plate ρ
fraction)
Comp 1 •3 - •4 - •001 -
Comp 2 •3 - •4 - •913 -
Comp 3 •4 - •2 - •086 -
Flow rate
Separation Composition Plate no. Steady State Gain
(kmol/h) 45 - 123 - 453 -
(mole fraction/
kmol/h
D p Distillate from column i
Butane/Pentane
B.: Bottoms from column i Column 1 2 1 •0071
Column 2 3 16 •0525
BTX
Column 1 1 20 •0150
Column 2 1 22 •0139
Styrene
Column 1 2 20 •0002
Column 2 2 20 •0002
 Analysis of the Dynamic Behavior of Distillation Trains 109

0 4 8 12 16
plates ·> 0.5 1.0 1.5
time ( hours )
Description of the curves :
C1 : BTX process, SO.28, E=0.56 ( column 1) Description of the curves :
C2 : Butane/pentane process, S=0.02, E=0.42 C1 : top plate temp. (R) in column 1, S=676.96
C3 : Styrene process, S=0.842, E=0.922 E=677.42
t The values of S and Ε are given above C2 : bottom plate temp. (R) in column 2,
S=788.8, E=792.5
Fig. 1 Composition profiles in column 2 C3 : xB 1 of column 1, S=0.024, E=0.027
( unless otherwise stated ) for the C4 : xB* of column 2, S=0.028, E=0.044
three processes studied. xB.1 bottoms composition of component i
t The values of S and Ε are given above

Fig. 3 Openloop dynamic responses in columns

1 and 2 due to disturbances in col-
umn 1 ( butane/pentane process ).
C3
C4
C5
C6
C3

2 0
\

1 1

0.5 1.0 1.5

time ( hours )

Description of the curves :

YRSS : composition at reference steady state C1 : top plate temp.(R) in column 1,
YSS1 : composition at the new steady state S=648.45, E=648.70
Description of the curves : C2 : bottom plate temp, in column 2,
C1 : benzene in column 1 of BTX process S=636.31, E=636.33
C2 : benzene in column 2 of BTX process C3 : xB 1 of column 1, S=0.153, E=0.157
C3 : pentane in column 1 of butane/pentane process C4 : xB.] of column 2, S=0.019, E=0.021
C4 : pentane in column 2 of butane/pentane process
C5 : styrene in column 1 of styrene process Fig. 4 Openloop dynamic responses in col
C6 : styrene in column 2 of styrene process umns 1 and 2 due to disturbances in
column1 ( BTX separation process ).
Fig. 2 Composition profiles before and after
disturbances in column 1 ( shown as
difference between the two steady
states ).
110 R. Gani et al.

CI
g .612- cr
I C1 liquid height in the C2
C1 : -1% change ( openloop ) CO
downcomer
C2 C3
C2 : -7% change ( openloop ) 0.06
C3 : -7% change ( flow control C3
.611 column 2

I
/ '
C2 : top composition of component
.610-λ / /
C3 : bottom composition of component 2
/ /

/ 0.02 H
.609 H
i column 1
x
.608 -f\ /
/ /
. column 2

\ /
— ι 1 1 1 1 1 —
.607 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
0.0 0.-8 0.4 0.6 0.8 1.0 TIME (HOURS)
TIME (HOURS) Fig. 7 Dynamic responses from the styrene
separation process ( YSS, steady state
value of the variable, Y, instantaneous
Fig. 5 Bottoms composition ( toluene ) res- value of the variable ).
ponse of column 2 due to disturbances
in the reboiler heat duty ( column 1). 0.7
1, C2 : ooenloop - C1

|C3, C4 : closedloop C2
^ 0.6
C3
γ C1 C4

U\C4

€3S\/

bottom plate
0.5 1.0 1.5
time ( hours ) top plate
— r
0.0 0.2 0.6 0.8 1.0
Description of the curves
C1 xT of column 1, S=0.959, E=0.960 TIME (HOURS)
C2
3
xT^ of column 2, S=0.947, E=0.949
C3 : xB" of column 1, S=0.877, E=0.879 Fig. 8 Closed loop temperature responses for
C4 : Χ Β 2 of column 2 , S=0.860, E=0.870 column 1 ( BTX separation process ).
xT. : top plate composition of component i
xB"!" : bottom plate composition of component i s16 C1,C2,C3 : openloop - CI
t The values of S and Ε are given above
_ C4,C5 : closedloop C2
C1,C3,C5 : bottom plate C3
Fig. 6 Openloop dynamic responses in columns
1 and 2 due to disturbances in column & 12 _ C2,C4 : top plate C4

1 ( styrene separation process ). C5

Table 6 Comparison of the Steady States for /

/•
Styrene Separation Process. /
/
/
/
/
/
Composition (mole fractions)
/
Column 1 Column 2 /
SS1 SS2 RSS SS1 SS2
-
/
/
* (
. . . . ... _
column 1
/·
Top plate
benzene •0017 •0013 •0016 _ _ _
ethyl Β ••9500 •9494 •9396 •9345 •9346 •9203
styrene •0488 •0494 •0592 •0655 •0654 •0797 I I I I
0.0 0.2 0.4 0.6 0.8 1.0
Bottom plate
_ _ _ _ _ _
TIME (HOURS)
Benzene
Ethyl Β ·•8783 •8765 •8544 •8683 •8663 •8424
Styrene •1216 •1235 •1456 •1317 •1337 •1575
Fig. 9 Closed loop temperature responses for
column 2 ( BTX separation process ).
Deviation (%) = 100 ( yss - y ) / Ad
yss : steady state value of the variable, y :
instantaneous value , Ad : disturbance
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

AN INTERACTIVE SIMULATION PACKAGE

FOR THE DYNAMIC BEHAVIOUR OF BINARY
DISTILLATION COLUMNS
G. Guardabassi and R. Scattolini
Dipartimento di Elettronica, Politecnico di Milano, Piazza Leonardo da Vinci 32,
20133 Milano, Italy

Abstract . This paper presents a simulation package for the

dynamics of plate-type distillation columns. The model underneath
basically consists of material balance equations. The
relationship between the liquid and vapour compositions in the
flows leaving any single plate is described in terms of the
Murphree vapour efficiency and relative volatility. Dependence of
the latter on operating pressure and composition is taken into
account. A coordinate "closed loop" control of both the size of
the time step and the value of the averaging parameter used in
the implicit integration algorithm improves the performance and
the numerical robustness of the code. The package, designed to
run on a Digital VAX 11/750 under VMS operating system, is
constituted by a set of FORTRAN programs coordinated by a unique
command file.
Keywords. Chemical industry; numerical analysis; computer-aided
design.

1. INTRODUCTION

The transient behavior of plate-type distillation the second stripping section, the feed section,
columns has long been studied, both theoretically the first and (possibly) the second enriching (or
and experimentally ([l]-[5]). rectifying) section, the condenser (Fig.l). The
external (exogenous) variables are denoted as
For binary (or almost binary) systems it appears follows.
that mass balance equations provide a description
of the plant that is accurate enough for many X
F = feed flow rate
control design purposes. F concentration of the more volatile compo-
nent in the feed stream
In this paper, we present an interactive simula- w = reboiler manipulated variable
tion package by which a variety of "open loop" X
D = distillate flow rate
experiments can easily be simulated, visualized, D = concentration of the more volatile compo-
and recorded. This is extremely useful not only to nent in the distillate stream
help tuning the model on the behaviour of a real R = reflux flow rate
plant, but also to produce data best suited for X
Β = bottom flow rate
further control-oriented processing; e.g.: B = concentration of the more volatile compo-
computer-aided linearization, order reduction, nent in the bottom stream
"nicely nonlinear" modeling, etc. . An enlarged PT .= top pressure
version of the package, including an ample class PB : = bottom pressure
of control system structures and algorithms is
under development. The plates of the column (condenser and reboiler
included) are numbered bottom-up from 1 to η . The
The paper is organized as follows. The basic model variables relevant to the i-th plate are denoted
is described in Section 2 . Section 3 presents as follows.
the numerical integration algorithm. The "expert
mechanism" for adjusting both the step size and L^ = liquid flow rate to plate i-l
the value of the averaging parameter used in the V^ = vapor flow rate to plate i+1
implicit integration algorithm is described in x^ = concentration of the more volatile compo-
Section 4. In Section 5, the structure and the nent in the liquid hold-up
organization of the package are presented. y^ = concentration of the more volatile compo-
Finally, a few examples given in Section 6, serve nent in the vapor hold-up
primarily the purpose of illustrating the code, = flow rate of the liquid side-stream ente-
more than the accuracy by which a real system can ring the plate
in fact be described by the model of Section 2. z^ = concentration of the more volatile compo-
nent in the liquid side-stream entering the
plate
2. THE DYNAMIC DISTILLATION MODEL = liquid hold-up
Ej_ = vapor Murphree efficiency
In order to fit with the geometry of several real p^ = pressure
plants, without augmenting too much the number of
parameters needed to describe the column, this is Assumptions:
assumed to consist of no more than 7 uniform 1) E^ and are constant, and uniform within
sections: the reboiler, the first and (possibly) each section

111
112 G. Guardabassi and R. Scattolini

2) in each plate, the vapor hold-up is neglige- for all i = l,2,...,n-2

able (as compared with the liquid hold-up)
Finally, the vapour flow rates are supposed to be
given by

D,x
E2 Cond.
D Vi(t) = V ( t ) + μ (Vi - V )
x x
V ( t ) = V" + k
El
1 x w(t)
F,x
F ι—Γ where Vi is the value of Vi at the nominal

τ—r S2
equilibrium. This may be obtained, of course, by
running an accurate steady-state simulation pro-
gram. Since no energy balance equation is
(explicitly) included in the dynamic model, the
B,x assumption underneath the equation above is that
SI Reb.
B the profile of the vapour flow rates is constant,
and independent of the operating condition (if the
nominal profile of the vapour flow rates is not
known, set μ=0 ) .
Fig. 1 : The distillation column

Under the above assumptions, total and partial 3. THE NUMERICAL INTEGRATION ALGORITHM
material balances at each plate lead to writing:
0 L + The digital simulation of the model given in
= i+1 - Li V l " Vi + Qi Section 2 calls for the numerical solution of a
H x x L x+ v v + set of stiff nonlinear differential equations. In
i i = K + i i+i - i i i - i yi-i - i y± fact, the stiffness ratio, namely the ratio
between the maximum and the minimum absolute
+ Qi . eigenvalue of the Jacobian matrix, usually ranges
iZ from a few tens to a few thousands with values of
As for the single-plate sections, we set some hundreds common.

* reboiler (i=l) : V A standard implicit method of numerical

Q= Lx= 0 , Q
l= -Β ,
integration has been used to compute the state
zi = *ι , Ει = 1 ; increments on each time-step. In this sense, our
1
integration algorithm is a fairly straightforward
* feed plate : = F , = xp ; variation of Rosenbrock s "second programme" [ 1 ] .
A distinctive feature of our routine, as compared
* condenser (i=n): V = L =0 , Q with [11, is the closed-loop control of the main
zn xn 1L+ R n = 1-D ,
integration parameters. Any discussion of this
n = n > n = ' ^ = · important issue is however postponed to the next
section.
The relationships between liquid and vapor concen-
trations, across each plate, are described by: Here, we consider a single time-step. First, the
essentials of the implicit numerical integration
y - y. = E
± ±l ± ( y*i - y^) method used in the sequel will be recalled. Then,
the specific formulas relevant to the considered
yi = y\ class of distillation columns (Section 2) will be
given in detail.
yi = φ(χ ,ρ )
ί ί
α(χ,ρ) χ 3.1 The numerical integration method
φ(χ,ρ) =
1 + (α(χ,ρ)-Ι) χ Consider a dynamic system described by

x(t) = f(x(t),u(t))
in what follows, the relative volatility α(χ,ρ)
is supposed to be given by: and a time interval [τ,τ+ôt) . We assume that the
input is constant over the considered time
α(χ,ρ) = A ( p ) (1-x) + Aj^(p) χ ,
0 interval (Fig.2):

where in turn, for each k=0,l , u(t) = U ( T ) V t ε [τ,τ+ôt)

A (p) = A (p) (p7p)\ ·

k k If, as is commonly the case, the input is given in
sampled form, u(x) will generally take on the
A little thought reveals that the 4 parameters
value of the last sample preceding τ . Then, in
needed to specify α(.,.) are uniquely determined
case of periodic sampling,
by the values of α in 4 points of the (χ,ρ)
plane. These 4 values of α are in fact the only
U ( T ) = u(k T)
primary information needed to describe the vapour- x
liquid equilibrium of the considered mixture. The
where Τ is the sampling period and k
program, then, automatically computes the above
largest value of k such that kT < τ .
T is the
parameters on the basis of this "simple" informa-
Then,
tion.
x(x+6t) = χ(τ) + <5x , ôx = ( x )
As for the pressures, it is supposed that avôt ,

(x) (1-Θ) f(x(x),u(x)) + Θ f(x(x)+ôx,u(T)) ,

P -l(t) = p ( t ) = p
n n T ; av =
where Θ ε [0,1] will henceforth be referred to
Pi(t) = ρ
τ+ (p
B - ρ τ) (n-i-l)/(n-2) ,
as averaging parameter. The state increment ox
 Dynamic Behaviour of Binary Distillation Columns 113

is therefore a solution of the implicit (non- Θ L ( x ) (1-Ε )

linear) algebraic equation: Β ^ χ ) := E t ± +
i+1 Φχ(χΐ+ΐ(τ)·Ρι+ι(τ))
δχ = {(1-Θ) f(x(x),u(x)) + Θ f(x(x)+ôx,u(x))} ôt .
Θ LM + Hi/ôt
This could be solved by a Newton-Raphson routine: + E ± + Θ V (z)
±
i Φ (χι(τ),ρ (τ))
δχ(0) = 0 χ 1
ôx(k+l) = ôx(k) + z(k)
A z(k) = b(k) Θ L
E i (+x )1
Ci(x) :=
where Φ (Χΐ+ΐ(τ),ρ
i+1
χ ί (+
τ))
1
A := I - Θ f (x(x),u(x)) ôt
x
G (z) := L
b(k) := {(1-Θ) f(x(x),u(x)) + ± i (+x )1 x ii +
(x) - L (x) χ (τ) +
t ±
+ V -iW yi-i(x) - V M γ (τ) +
+ Θ f(x(x)+ôx(k),u(x))} ôt - ôx(k) . ± ± ±
+ Qi(x) ζ (τ)
As usual, one more simply sets ±
Now, by forward elimination, one obtains:
f(x(x)+ôx,u(x)) = f(x(x),u(x)) + f (x(x),u(x)) ox
x
in the implicit equation for δχ , thus obtaining
ô
Yi + φ 1(τ) ô y i1+= Γ ^ χ ) , i = 1,2,...,η-1 ,
the linear equation: ôy
n= Γ (χ)
η
δχ = {f(x(x),u(x)) + Θ f (x(x),u(x)) ôx} ôt .
x where :
It is easy to check that this amounts to stopping
the Newton-Raphson routine just after the first
step. Β ( τ ) - Α (τ) Φ . ( χ )
± ± ί1
3.2 Computation of the state increment for the G ^ x ) - A t(x) Γ±-ι(τ)
distillation column
B ( x ) - Α ( τ ) Φ . (τ)
When applied to the model of Section 2, the method t ± 11
above yields, for the increments, the following
set of linear algebraic equations. Thus, by backward substitution, ôy is readily
computed. Finally, for all i = 1,2,...,η ,
Ε δχ = { L
± ± i (v+x )1 x i (+x )1 - L±
(x) χ (τ)
± +
o - (1-Ei) ôyi-i
+ i-l(t) Vi-l(O " V±W y (x) + Yi
±
E
+ Q (x) ζ (τ) + ± φ χ(x i(x),p i(x))
t ±
In conclusion, the basic computation cycle is as
+ θ [L
i (+x )1 ô x i 1+- L±
(z) δχ +
± follows. Given the state x(x) , the inputs F ( x ) ,
Χρ(τ),ρ (χ),ρ (χ)Μχ),ϋ(χ), and the integration
+ V . ( x ) ô y ^ - V (z) δγ )} ôt Β ôt,0 Τ , one computes
i1 ± ± parameters first y(x) and
V(x) , then A(x),B(x),C(x),G(x),Φ(χ),Γ(χ),ôy and
y (x) = E ψ ( χ ( τ ) , ρ ( τ ) ) + ( l - E ^ ν±-χ(τ)
t t 1 1 finally δχ .
By this routine plus linear interpolation, the top
δγ = E Φ ( χ ( τ ) , ρ ( χ ) ) δχ + (1-Ε ) ôy _
± ± χ 1 1 ± ± i 1, and bottom concentrations at the sampling instants
t = kT , where Τ is the same sampling period used
where i = 1,2,...»η , while k
for the input functions, are computed and tempora-
rily recorded for possible further processing. The
α(χ,ρ) χ
length of the time step at the sampling instants
φ(χ,ρ) = is also recorded and displayed on request.
1 + (α(χ,ρ)-Ι) χ

α(χ,ρ) + χ (1-χ) ( Α ( ρ ) - Α ( ρ ) )
1 20 A. FEEDBACK CONTROL OF THE INTEGRATION PARAMETERS
Φ (χ,ρ) =
χ (1 + (α(χ,ρ)-Ι) χ ) The time step ôt is bound to an interval the
extremes of which are denoted ôtmin and ôtmax ,
1 T h e
respectively. e minimum
The r ôtmin is set equal to
α(χ,ρ) = Α ( ρ ) (1-χ) + Α ^ ρ ) χ , 0 · min · «
and 0
T min H i A L i + V i)
m n i= i
A ( p ) = A ( p ) ( ρ / ρ ) \ , k = 0,1 .
By k
solving fork δχ^ the 3rd equation, and
is the minimum plate-time-constant. On the other
substituting back in the first, one gets
+
Bx à +δ
C = G( ) x side, there is no real interest in giving ôt
values much larger than 0.1 the average settling
Ai(x) àVi-ι i < ) y± i ?ί+1 i time of the considered system. Thus, ôtmax is
often set equal to
where A^(x) := C ( x ) := 0 ; while, for all other H+
n
values of i = 1,2,...,η , W = < 1 "n)/Vi ·

(Θ L ( x ) + H /ôt)(l-E ) Within the above range, ôt is varied according to

Α ( τ ) := E t χ τi τ i
Θ V^Cx) , a number of rules meant to protect the process
Α i Φ ( ί( )»Ρί( )) against inconsistencies and numerical instability.
χ The first rule is obvious.
114 G. Guardabassi and R. Scattolini

RULE //I . If, for some i, χ (τ) + δχ ft [0,1] ,
± ±
then, if ôt > otmin, set ôt = ôtmin and
recompute δχ ; else, send a message and wait
for instructions (abort, reduce ôtmin).
The second rule helps facing a slightly less
catastrophic event. To express it, we need some
preliminary definition. Let
The so obtained step size updating rule can be
further refined by currently controlling the
maximum relative increment δχ~(χ) . As long as it
ranges between 0.01 and 0.1, we leave the updating
rule unchanged, whereas we increase the growing
rate up to 5 times its "natural" value if δχ~(χ)
is less than 0.01 , while we decrease it down even
to negative values if δχ~(χ) > 0,1 .

δ χ ^ χ ) := |ôx |/min{x (x),l-x (T)}

i i I
be the relative increment of the i-th state varia-
ble at time τ . Then,

δχ~(χ) := max δχ~^(χ)

i

In keeping with the fact, that the risk of numeri-

cal instability is known to decrease as long as Θ
gets close to 1, we have:
2
RULE #2 . If δχ~(χ) > 0.2 + 0.1 Θ , then
if ôt > ôtmin , set ôt = ôt/3 and recom-
pute δχ ; else, send a message and wait for
instructions (abort, restart with a reduced
dt/T
ôtmin, continue).

As long as the inputs remain constant, all state

Fig. 2 : The function β-^.)
and output transients of a distillation column
tend to be monotone and quasi-exponential. Let's
consider, then, for a moment, a linear time-
invariant first order system and see, on this "toy
example", how the step size should best be varied
in time. Precisely, let

Τ χ(τ) = - χ(τ) + u ;

hence,
τΤ
χ(τ) = u + (x(0)-u) θ - /

In view of the requirement that the integration

accuracy be roughly time-invariant, it should be
apparent that making ôt(x) proportional to the
inverse of the second derivative of χ with respect
'-Ί.ΟΟ -3.60 -3.20 -2.80 -2.40 -2.00 -1.60 -1.20 -0.B0 -0.40 0.00
to τ is a quite conceivable choice; then,
2 log(dx~)

kT
TT
ôt(x) = e^/T := ôt(0) e /
; ,
Fig. 3 : The function β (.)
2
x(0)-u
All the preceding discusssion is summarized by the
This is an "open loop" time step programming rule. following
In order to give it a "closed loop" form, let's
take the derivative and approximate it with the RULE #3 . If at time x+ôt(x) all input (exoge-
incremental ratio: nous) variables remain unchanged, as compared
with the immediately preceding time interval,
d the size of the time step is updated by the
— ôt(x) = ôt(x)/T formula:
dx
ôt(x+ôt(x)) = ( ^ ( Ô t ( x ) / T
[ôt(x+ôt(x)) - ôt(x)]/ôt(x) = ôt(x)/T ; m ,aô xx~ ) ) ôt(x) ,

hence,
ôt(x+ôt(x)) = (l+ôt(x)/T) ôt(x)
On the basis of this analysis, we tentatively set
ôt(0) = ôtmin
ôt(x+ôt(x)) = (l+3 (ôt(x)/T )) ôt(x)
1 max
where the function β^(.) is given the shape of
Fig.2 . The deviations of β^(.) from the identity
are justified, on one side, by the requirement
that, in order to keep the total number of steps
within reasonable bounds, the growing rate should
not be less than 10% , even for very small values
of ôt(x)/T . On the other side by the constraint:
ôt(x) < ôtmax = T
xa
where
3(v,w) := β (ν) £ M
χ 2
and the functions β (.),β2(.) are shown in
1
Fig.2 and in Fig.3 , respectively. Otherwise,
Ôt(x+ôt(x)) = ôtmin .
In connection with Rule #3, it is worth noticing
, that the sampling period of the output (and input)
variables is usually much larger than ôtmin . Of
course, the rule is in itself most effective in
computing responses to input functions which are
constant over relatively long time intervals (e.g.
step functions).
Due to high stiffness ratio, the values of ôt
produced by the updating formula of Rule #3 may
readily become quite large relative to T i . As a
mn
result, numerical instability may show up in some
state variable, with undamped or poorly damped
 Dynamic Behaviour of Binary Distillation Columns
oscillations of period roughly equal to 2 ôt . An
effective remedy is provided the next rule.
Rule #4 . Make a distinction between say "regu-
lar" and "auxiliary" time steps, the purpose
of the latter being to discharge, so to say,
the fast dynamics excited by each regular
time step. Precisely, each regular step is
followed by a sequence of Ν auxiliary time
steps, the first of which has length ôtmin ,
while the subsequent ones grow (slowly) in
size, according to Rule #3 . At the end of
the auxiliary step sequence, the length of
the next regular time step is computed again
by Rule #3, but setting now ôt(x), in the
right hand side of the step size updating
formula, just equal to the length of the last
regular time step.
Ν may conceivably range from 0 to 30 . In our code
it is set, by default, equal to 15.
Finally a word on the averaging parameter Θ . As
is well known, the smaller is Θ the higher is the
risk of numerical instability. As for the accura-
cy, we may go back to our "toy example" and easily
find that the exact value of Θ is then given by
- 01 -1
Θ° = ( l - e ^ ) - T/ôt
Since our primary interest is in the slow dynamics
and, in view of Rule #3, <5t/T
mxaε [0,1] , the
conclusion can be drawn that the most accurate
value of Θ ranges from 0.5 (Crank-Nicholson's
trapezoidal rule) to about 0.6 . On the other
hand, whenever the inputs hold constant for long
periods of time, relative to the process dynamics,
the risk of numerical instability increases as
long as the transient settles and the time step
approaches its maximum value. It is then
conceivable to make Θ slowly increasing in time,
and asymptotically equal to 1. The initial value
©o of Θ is specified by the operator, or set equal
to 0.6 by default. The time-constant of the Θ
updating mechanism is 2 T
m x a. Of course, the
value of Θ is reset to 0q as soon as there is a
change in the value of one of the input variables.
5. STRUCTURE AND ORGANIZATION OF THE PACKAGE
The interactive package SIMCOL is designed to run
on a VAX-11/750 using the VAX/VMS operating
system. The graphical data-display is presently
designed for VT100-type terminals with semigraphic
options. The package is constituted by a set of
FORTRAN77 programs coordinated by an unique
command file. This structure provides flexibility
to the management of the programs and allows easy
maintenance and upgrading of the package. As a
minor drawback, during a "simulation experiment",
some recorded data-files have to be read more than
once. However, being these data-files of small
size, the input/output operations involved do not
substantially affect the efficiency of the
package.
SIMCOL has been explicitely structured to meet
with the following requirements:
* provide a friendly interactive environment to
the engineer unfamiliar with the use of computers;
* guarantee an immediate comprehension of the
results during a "simulation experiment";
* allow the readability of the recorded data-
files concerning both process data and simulation
data and results;
* provide a final software product reasonably
portable on different computers equipped with a
FORTRAN compiler.
The total size of SIMCOL in its current form is
about 203KB. However, thanks to the segmented
115
structure of the package,the memory occupation is
30KB at most.
A "simulation experiment" proceeds according to
the following steps:
Step 1_ The program COLDIS provides the display of
a sketch of the bynary distillation column
together with a list of the main symbols used
during the simulation.
Step 2 A list is provided of the data-files
already recorded and containing data relevant to
different columns and mixtures previously
considered. If the data concerned with the problem
at hand are already stored, the corresponding file
is copied into a new temporary file. Otherwise the
program COLLEG provides the input in interactive
mode of the process and mixture data, namely:
- a label of the simulation;
- the units considered;
- the number of plates of each section of the
column and the corresponding efficiencies;
- the liquid hold-up of each element;
- the vector V and the scalars V*, μ and k^;
- the values of pressure, concentration and
relative volatility needed to specify a(.,.).
A temporary file is then created for future
manipulations. On request this data file can be
assigned a name specified by the user and
permanently recorded.
Step 3 The program COLVER allows to check and
possibly modify the data-file generated at the
previous step.
Step 4 A list is provided of the data-files
already recorded and containing different
equilibrium points previously computed. If the
data concerned with the equilibrium point of
interest are already recorded, the corresponding
data file is copied into a temporary file.
Otherwise the program LEGSTI provides the input of
the nominal values of the following exogenous
signals: top (p-r.) and bottom (pg) pressures, feed
(F), stream (w) and distillate (D) flow-rates,
feed composition (xp). The corresponding steady-
state composition in each plate is then computed.
On request the data and the results are displayed
on the screen and permanently stored into a new
file specified by the user.
Step 5 The program VERSTI allows to check the
nominal values of p^, pg, F, w, D, xp. In the case
that any modification of these values is made, the
corresponding new equilibrium point is computed
and recorded on request.
Step 6 The user can decide either to come to an
end of the "simulation experiment" or to continue.
In this case, the two possibilities illustrated in
the following steps 6a and 6b are available.
Step 6a The program EQUIL allows to compute a
number of equilibrium points corresponding to some
constant values of p<p, pg, F, w, D, xj-
interactively specified by the user. On request,
the results are recorded into a new data-file.
Step 6b First the user is requested to specify by
means of FORTRAN-type statements the form of the
exogenous signals p-p, pg, F, w, D, Χ γ . This can
easisy be done by making reference to a set of
FORTRAN functions recorded into an existing
program library and simulating a number of
"canonical signals" (step, ramp, sine, square
waves, ...) (program IN). Then the transient of
the concentrations on each plate corresponding to
the initial state determined at steps 4 and 5, and
to the given input functions is computed (program
TRA). The final results can be displayed in
semigraphical mode and permanently recorded. A
further option allows to have the transient of the
main simulation variables plotted on a CALCOMP M84
(program PLDIS).
Step 7_ The "simulation experiment" is terminated.
A new "simulation experiment" can be performed
simply moving back to anyone of the previous
steps.
116 G. Guardabassi and R. Scattolini

6. EXAMPLES [5] Mah, Seider (Eds.), Foundations of Computer-

Aided Chemical Process Design, American In-
The simulation package presented in the preceding stitute of Chemical Engineers, New York,
sections has been tested on several cases and is 1981.
currently used to help control-oriented modeling [6] F. De Lorenzo et al., "Theoretical and expe-
of industrial columns. In this section some as- rimental researches on the dynamics of plate
pects of the package performance are shown by distillation columns. Part II: Experimental
examples. Any illustration of the man-machine tests", Proc. X International Automation and
interaction has instead been avoided due to page Instrumentation Convention and Exhibition,
constraints. It has however to be stressed that Milano (Italy), Nov. 1968.
improving the friendliness of this interaction has
been a major engagement in every stage of the
conception, design and implementation of the pac-
kage.
Though discussing the accuracy of the basic model
presented in Section 2 is out of the scope of this
paper, in Fig.A computed versus experimental res-
ponses to step changes in the distillate flow rate
are shown. The plant is here a laboratory 20
plates column operating on a mixture benzene-
toluene [6]. The sampling period used for this
computation is Τ = 1. The CPU time needed to
compute the two transients is 23.03 sec. Fig.5
shows the step length as a function of time,
produced by the step length updating mechanism
(Section 4 ) . Note that the minimum (initial) value
of ôt is 0.0011 .

In Fig.6, the computed response is shown of the

same column to a sinusoidal variation of the dis-
tillate flow rate: °a.00 30.00 60.00 30.00 120.00 150.00 160.00 210.00 240.00 270.00 300.00
t (min)
ρ 0.071 , t < 20 ,
D(t) - \ Fig.4 Computed (—) and experimental Ό)
L 0.071 + 0.01 sin(t 2π/30) , t > 20 , response of χ to stepchanges in D.

starting from the same initial equilibrium and

using the same sampling period Τ = 1 . Due to the
nonconstant shape of the input functions, the CPU
time now becomes 8 min and 48.02 sec. Yet it may
be noted that, by increasing the sampling period
to Τ = 3, the computation accuracy is not
significantly affected, while the CPU time drops
to 4 min and 41.69 sec.

7. CONCLUDING REMARKS

A number of extensions of the simulation package

SIMCOL are currently underway. In particular, an
enlarged version of SIMCOL will embed some FORTRAN
*ϊ.00 30.00 60.00 30.00 120.00 1'JO.OO' Ιϋό.Οβ' 2ΐ0.0β' 240.Où' 270.00 ' 300.00
programs simulating an ample call of control t (min)
system structures and algorithms. During the
simulation, the transients of the main process and Fig.5 Transient of the stepienght.
control variables will be displayed on a TEK 4109
color graphic terminal. At any time it will also
be possible to modify both the control design
(structure and parameters of the controller) and
the values of set-points and independent exogenous ft.
signals.

REFERENCES

[1] H.H.Rosenbrock, "Calculation of the transient

behaviour of distillation columns", British
Chemical Engineering, Part I, pp.364-367,
Part II, pp.432-435, Part III, pp.491-494,
1958.
[2] R.G.E.Franks, Mathematical Modeling in Chemi-
cal Engineering, John Wiley & Sons, New York,
1967.
[3] W.L.Luyben, Process Modeling Simulation and "b.OO 30.00 60.00 ' 30.00 120.00 150.00 180.00 210.00 240.00 270.00 300.00
Control for Chemical Engineers, McGraw-Hill, t (min)
New York, 1973. Fig.6 Computed response of χ to a
[4] C.D.Holland, Fundamentals and Modeling of sinusoidal variation of D.
Separation Processes, Prentice-Hall, New
York, 1975.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

APPLICATIONS OF A DYNAMIC SIMULATION

OF DISTILLATION COLUMNS FOR A NEW
MANUFACTURING PLANT
Chiu-Nan Wei
Exxon Chemical Americas, PO Box 4004, Baytown, TX 77522, USA

Abstract. This paper presents a case study of applications of a dynamic simulation of distil-
lation columns for a new, complex manufacturing plant. The applications included process
feasibility studies, dynamic analyses during the design phase, control system development and
implementation, and operator training prior to start-up. The simulation was developed in a
plant environment with a minimal expenditure of money, but yielded significant savings in
capital and operating costs for the new plant. The plant was started up successfully with
advanced controls and produced on-specification products within ten days.

Keywords. Computer applications; dynamic simulation; computer control; control system analysis;
dynamic response; multivariable system; distillation column.

INTRODUCTION What control scheme would best maintain

In the chemical industry, dynamic simulations
have found widespread use in dynamic analysis
(McCune and Gallier, 1973), the evaluation of
alternative control systems (Cheung and Mar lin,
1982), the determination of safety shutdown
system (Ochiai, 1980), operator training
(Vervalin, 1 9 8 4 ) , and so on. However, it is
seldom that a single simulation model is
developed to aid the process engineer in plant
design, to help the control engineer in control
development, and to assist the operations
engineer in new plant start-up. In this paper we
will discuss an example of a dynamic simulation
that provided the functions described above. It
was developed with a minimal expenditure of
money, but yielded significant savings in capital
and operating costs for a new manufacturing
plant.
For a process plant expansion, a set of stacked
distillation columns was proposed to convert the
process from its original single feed with two
product streams to two feeds and three product
streams (Fig. 1 ) . The six components for
separation in the columns are A, B, C, D, E, and
F (listed from light to heavy). Four specifica-
tions are required in these three high purity
products: top product (150 ppm of components B,
C, D, E ) , middle sidestream (60 ppm of component
F maximum; 1.0 Wt% of component A maximum),
bottom product (150 ppm of components B, C, D,
E ) . Off-specification production from the top or
the bottom stream could cause overall plant
product to go off-specification, and any compo-
nent F, which is very expensive, in the middle
sidestream is considered a loss. Operability of
the columns played an important role in the
project success. A major concern of operations
was the column dynamic transients during plant
upsets. To minimize column interaction, one
intercolumn heat exchanger and accumulator system
were initially proposed for column decoupling.
Operations personnel desired to know:
If the heat exchanger system was required.
product specifications even under plant
upsets.
How the operators could be trained to
start up, operate, and shut down such a
complex process.
To answer these questions, a dynamic simulation
of the distillation columns was developed and
employed.
The mathematical model developed for the distil-
lation columns was presented at the 1984 Summer
Computer Simulation Conference in Boston,
Massachusetts (Wei, 1984). The presentation
initiated enthusiastic discussions with the
audience. Questions were raised such as how well
the model identified and resolved the potential
problems and what the results were. Therefore,
this paper will emphasize these points and depict
plant data with those predicted from the model.
Recent benefits to our operation as a result of
the simulator will also be included.
FEASIBILITY STUDY - DYNAMIC ANALYSIS
The main purpose of this study was to determine
whether an additional intercolumn heat exchanger
was needed to improve operability. The overall
tower performance was examined by running a
number of tests which were considered to be
possible upsets in the actual plant operation.
These upsets, as defined in Table 1, were
prepared by operations personnel. Since the
mathematical models describing the tower dynamics
and its related equipment, instrumentation and
control system were implemented on a mini-
computer to perform interactive on-line simula-
tion, a wide variety of upset tests could be
simulated by simply changing the value of the
particular upset variable from the CRT keyboard.
For analysis, the transient responses of key
process variables were displayed on the CRT
screen or plotted on a hard copy from a line
printer. The upset tests were conducted with
step changes from the normal operation
conditions.

117
118 Chiu-Nan Wei
Test 10 is given as an example to illustrate the
study procedures. Figure 2 shows the process
transient when the concentration of F, the
heaviest component in the heavy feed, was
decreased from 62 Wt% to 40 Wt%. As depicted in
the figure, the concentration of F component in
the bottom product stream decreased from 99.99
percent to 91 percent, and the impurity of
component F in the middle sidestream increased
from 80 ppm to 550 ppm. Both of the stream
products are off-specification. Fractionation is
inadequate. Two corrections are required to
effectively operate the tower under this upset:
1. Increase the firing rate to provide enough
heat for fractionation.
2. Regulate properly the liquid flow from C-2 to
C-l to provide enough reflux.
Corrections were tried one at a time. First, the
firing rate was kept constant. The temperature
controller, T-14, was placed in manual and the
intercolumn flow was regulated manually. As
shown in Fig. 3, the concentration of component F
from the middle sidestream decreased signifi-
cantly, but the impurity of C-l bottom increased
(i.e., F concentration decreased). On the other
hand, if the firing rate was increased by 20
percent of the design rate and T-14 was put on
automatic, the bottom product was on-
specification, but the loss of component F from
the middle sidestream increased during the
transient period of four hours by approximately
1.2 barrel more than that under normal operating
conditions. This is illustrated in Fig. 4. As
shown in Fig. 5, all products were on-
specification and the loss of component F reduced
significantly when the firing rate was increased
by 20 percent of the design rate and the
intercolumn liquid flow was manually regulated.
This implies that the present process design
configuration can handle this upset if it is
provided with a better control scheme. The
firing control scheme should respond to feed rate
and composition changes, and an analyzer
controller should regulate the temperature
setpoint of T-14 due to the shift of tray
temperature profile. Also a more sensitive
temperature control point is required.
The other tests were carried out in the same way
as the example described above. Since the
objective of the study at this stage is a process
evaluation rather than control system design, an
overall supervisory control system has not been
pursued. However, for some upsets, the columns
could not maintain smooth operation or could not
produce on-specification products; these problem
areas were investigated, identified, and cor-
rected during the design phase. Based on the
results of this study, the following conclusion
was made:
The process design configuration as shown
in Fig. 1 is capable of producing on-
specif icat ion products. No additional
intercolumn heat exchanger is required.
However, a better control system is needed
to improve the tower performance.
In addition, several shortcomings were revealed
in the study and hence, the following modifi-
cations were made to improve the process
operability and system controllability.
Modification of the flooded condenser:
The condenser would be operated at
approximately 80 percent of capacity
rather than 50 percent as described in the
vendor's design specification because the
operational tower pressure would be lower
than that designed.
Relocation of tray temperature control
point (T-14) from tray No. 36 to tray
No. 33.
Changes of pipe size, control valve, and
flow transmitter range for the middle
sidestream.
Installation of one analyzer in each feed
stream: Component A analyzer in the light
feed stream and component F in the heavy
feed stream.
CONTROL SYSTEM DESIGN
Process responses from the upset tests were used
to guide the design of the control scheme. This
is probably the easiest and the most effective
way to test alternative control schemes prior to
plant start-up. Many experiments can be per-
formed in a short time and the results are not
distorted by noise or other disturbances. Not
only the basic regulatory control system (i.e.,
properly pairing of the controlled variables and
the manipulated variables), but also advanced
control systems such as feedforward-feedback
control system, inferential model-based control
system, etc., can be designed and tested.
Furthermore, any design changes can easily be
implemented, less expensively, during the design
stage. As discussed in the previous section, we
have changed the location of tray temperature
control point (T-14). Figure 6 shows the actual
plant data; it indicates that the sensitivity of
the two temperatures is quite different even
though only three trays apart. Temperature which
had a shorter deadtime corresponded to feed
changes, also improved process controllability.
The overall control scheme as illustrated in
Fig. 7 was designed and implemented as follows:
1. Regulatory controls - Material balance
control (Cheung and Marlin, 1982):
Control Variables Regulated Variables
Pressure Reflux flow
Top sidestream spec. Top sidestream flow
T-l bottom level Middle sidestream flow
Middle sidestream Intercolumn transfer
spec. flow
T-2 bottom level Bottom product flow
Bottom produce spec. Firing rate (furnace
duty)
Advanced controls - Model-based (dynamic
reconciliation) control (Bartman, 1981):
A cascaded scheme (analyzer to temperature to
feed forward ratio to flow) ultimately sets
the sidestream flow to control the sidestream
specification as shown in Fig. 8. The
analyzer feedback controller uses a model-
based control. The technique uses the
following relation between the heavier
impurity and C-2 tray No. 77 temperature.
 Distillation Columns for a New Manufacturing Plant
C = a + b * Τ (1)
where C = impurity analysis
Τ = tray temperature
a = constant
b - steady state process gain
Feedback control is achieved by updating the
constant in equation ( 1 ) . Dynamics are taken
into account by delaying and lagging the tempera-
ture by the dead time and time constant between
analysis and temperature. Delaying and lagging
the temperature synchronizes it with the current
analysis. Using the updated constant in
equation ( 1 ) , a new temperature setpoint is
calculated. The models required were obtained
from the dynamic tests on the simulator. Those
models were also used for the plant start-up.
Using the simulator, all the control schemes were
developed and tested in both the same process
control computer (Honeywell 4500 system), and
Foxboro Spec 200/Videospec as would be used for
the actual plant. The software developed was
used for the plant start-up. It is preferred
that a test or evaluation uses the actual
hardware of the system, rather than a mathe-
matical representation of the hardware. Thus, it
avoided some of the problems that are related to
the specific implementation or to characteristics
of the process control computer system.
We did not put all the supervisory controls in
service when the plant started up even though all
the control software was ready; all the instru-
ment validity and analyzer reliability needed to
be checked first. Therefore, the flow rate was
initially on DDC (direct digital control) and its
setpoint was changed manually. As shown in
Fig. 9, when the heavy feed rate increased, the
heavy component (F) increased in the middle
sidestream due to the increase of heat input and
the failure to increase the intercolumn transfer
flow. As the operator noticed the problem, he
increased the flow, but increased it a little too
much, causing the lighter components to go down
to the bottom product. This oscillation
continued as operators had difficulty regulating
the flow correctly. After two days we added
temperature control on top of the flow control to
regulate the flow setpoint. The temperature
controller detected the heavy component going up
before it showed up in the middle product stream.
The operation would have been fine if the
operator kept alert and correctly changed the
setpoint of the temperature controller when the
feed rate changed. When the analyzer reliability
increased two days later, we added the analyzer
controller to regulate the setpoint for the
temperature controller. As depicted in Fig. 10,
the tower ran smoothly even when the heavy feed
rate changed.
The control loop discussed above is given as an
example of commissioning advanced control loops
in the plant start-up. The same procedures
applied to the other critical control loops. The
plant was started up successfully with advanced
controls and produced on-specification products
within ten days.
OPERATOR TRAINING
Since the model was also developed with
DCCR-I
119
intent to provide operator training, the program
was designed and executed in a reasonable amount
of computer time, 0.25 sec/cycle. The model was
then integrated with the system executive
software, the plant process control computer
system (Honeywell 4500 system) , and the Foxboro
Spec 200/Videospec instrumentation system to
perform an interactive simulation. As indicated
in the simulator configuration in Fig. 11,
operators were trained on the same operation
consoles as those they would use to operate the
actual plant. To provide thorough training, two
separate simulators were constructed. They are
the tower simulator and the "plant" simulator,
which integrated the tower model with other
process models such as reactors, adsorbers, heat
exchangers, steam system, etc.
The simulator allows dynamic transients from a
cold start-up condition to a normal (design)
operating condition and vice versa. A number of
possible plant upsets and malfunctions were
included and could be entered from the instructor
station. Thus, operators were given training in
troubleshooting. The simulator provided more
than ten initial conditions for different
operating conditions and was capable of making
"snapshot" recordings of the simulated plant
conditions at any time. The snapshot contained
sufficient information to allow restart of the
simulator from that point. The simulator also
provided selection of fast-mode operation (up to
ten times of real-time) for certain slow
processes.
CONCLUSION
Dynamic simulation can be used as a powerful tool
for operational analyses, process and control
design assessment, and operator training in the
manufacturing environment. This paper presented
a case study of a dynamic simulation of distil-
lation columns for a new manufacturing plant to
illustrate such applications.
As discussed in the paper, the model has been
used for wide ranging "what if" type of studies
to find the process operability. An expensive
intercolumn heat exchanger system which costs
millions of dollars was eliminated. In addition,
several design flaws were found and corrected
prior to start-up. The simulation was also used
to develop and test the control systems. To
implement model-based controls, several dynamic
tests were conducted on the simulator to obtain
simplified process models. It has not only saved
enormous time for plant testing, but also
determined the best controls for initial start-
up. Finally, the simulator was used to train
operators in the same environment as they would
operate the actual plant. The start-up team
acquired many hours of realistic hands-on
training, including cold starts and upsets. This
created a feeling of confidence and eliminated
many surprises for the plant start-up.
All the activities described above contributed
greatly to the successful plant start-up. It is
as though the simulator gives us the ability to
look into the future to run the unit before we
have even finalized its design. This helps
ensure that we are doing the right things and we
are doing them in the most cost-effective way.
the
120 Chiu-Nan Wei

REFERENCES

Bartman, R. V. (1981). Dual composition control

in a C3/C4 splitter. Chemical Engineering
Progress, 9^, pp. 58-62.
Cheung, T. F., and T. E. Marlin (1982). Regu-
latory control of distillation towers:
material balance vs. energy balance control.
Paper presented in the Lehigh University
Distillation Control Short Course.
McCune, L. C , and P. W. Gallier (1973). Digital
simulation: a tool for the analysis and
design of distillation control. ISA Trans-
actions, Vol. 12, No. 3, pp. 193-207.
Ochiai, S. (1980). Simulate safety shutdown
systems for design confidence. InTech,
pp. 33-38.
Vervalin, C. H. (1984). Training by simulation.
Hydrocarbon Processing, 12, pp. 42-50.
Wei, C. ΪΠ (1984) . Dynamic simulation of
multicomponent distillation columns: a tool
for process analysis, control design and
operator training. 1984 Summer Computer
Simulation Conference, Boston, Massachusetts,
Vol. 1, pp. 616-623.

TABLE 1 Upset Tests

A. Light Feed Rate

BOTTO
M PR00UC
T
1. Terminates completely.
2. Flows at half design rate.
3. Flows at 1.2 times design rate. Fig. 1 Simulation diagram of multicomponent
B. Heavy Feed Rate distillation columns

4. Terminates completely.
5. Flows at half design rate.
6. Flows at 1.2 times design rate.

C. Concentration of Component A in Light Feed

7. Increases from 54 Wt% to 80 Wt%.

8. Decreases from 54 Wt% to 40 Wt%.

D. Concentration of Component F in Heavy Feed j~B. COLUM

NTEMPERATUR
.COLUM
B N E (T-14
)
9. Increases from 62 Wt% to 80 Wt%. -.A COMPONEN
TFNIHEAV
YFEE
D
10. Decreases from 62 Wt% to 40 Wt%.
E. Fuel Gas Heating Value -.
EFIRIN
GRAT
E
11. Increases from 1,000 BTU/SCF to 1,200
BTU/SCF. 40 0 08 O12 0IS 020 24
12. Decreases from 1,000 BTU/SCF to 800 Ο00Ο ο
· TIM
E (MIN.
)
BTU/SCF.

Fig. 2 Response of process to a step change in

concentration of component F in heavy
feed

Firing rate: unchanged

T-14 on control (cascaded to F-15)
 Distillation Columns for a New Manufacturing Plant 121

D. COMPONEN
T FNIBOTTO
M PRODUC
T

COMPONEN
T FNIHEAV
Y FEE
D
/— .ΕFIRIN
G RAT
E

•TIM
E (MIN.
)

Fig. 3 Response of process to a step change in

concentration of component F in heavy
feed Fig. 6 Response of column temperatures to
changes of feed rate and composition of
Firing rate: unchanged middle sidestream product
Inter-column flow (F-15)
on manual operation

-.
0COMPONEN
T FNIBOTTO
M PRODUC
T
B.COLUM
N TEMPERATUR
E (T-14
)
k C.COMPONEN
T FNIMIDDL
E SID
E STREA
M
UPSE
T/
«.COMPONEN
T FNIHEAV
Y FEE
D
- .EFIRIN
G RAT
E

40 0 08 012 0 1G 20
0 24
O f fOl O ^ • E
TIM(MIN.
)
Œ ωο οω
Fig. 4 Response of process to a step change in
concentration of component F in heavy
feed

Firing rate: increased by 20%

T-14 on control (cascaded to F-15) HEAV
Y FEE
D -CÎT-T— C-l
—φ

1
0.COMPONEN
T FNIBOTTO
M PRODUC
T

- .BCOLUM
N TEMPERATUR
E (T-
l
-UPSE
T
A.COMPONEN
T FNIHEAV
Y FEE
D Fig. 7 Column supervisory controls
E.FIRIN
G RAT
E
-.
CCOMPONEN
T FNIMIDDL
E SID
E STREA
M
1G
0 0 20
- TIM
E (MIN.
)

Fig. 5 Response of process to a step change in

concentration of component F in heavy
feed

Firing rate: increased by 20%

Inter-column flow (F-15)
on manual operation
 Chiu-Nan Wei

8 S8

A. FEED RATE

D. COMPONENT F IN MIDDLE SIDE STREAM

8» ? â

Fig. 10 Plant start-up data after commissioning

supervisory controls

c = a + D χ τ

8 Control strategy for top product draw-off

SIMULATION
COMPUTER
COMPUTER SEL 32/77
ROOM SIMULATION MODELS
HONEYWELL
COMPUTER
4500

CONTROL
ROOM

Fig. 11 Configuration of simulator

9 Plant start-up data prior to

commissioning supervisory controls
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

NONLINEAR MODEL REDUCTION FOR

BINARY DISTILLATION
W . Marquardt
Institut fur Systemdynamik una Regelungstechnik, Universitat Stuttgart,
D 7000 Stuttgart 80, Pfaffenwaldring 9, FRG

Abstract. This paper will attempt to outline a new approach to model reduction
for separation processes based on nonlinear wave propagation phenomena.
Theoretical considerations on existence, stability, propagation velocity and
shape of the spatial concentration profiles will be summarized briefly, taking
a binary distillation column section as an example. These elements of non-
linear wave propagation form the basis for the proposed model reduction tech-
nique. A detailed presentation of the systematic development of a reduced model
for a special separation will be given. It will be shown that three dynamic
state variables suffice in order to describe the dynamic behaviour of one column
section. The resulting nonlinear model is given as a system of linear-implicit
differential-algebraic equations. The steady-state and dynamic behaviour of the
reduced model is compared to that of the detailed one, there being little differ-
ence between the two.

Keywords. Modelling, system order reduction, distributed parameter system,

nonlinear system, chemical industry, distillation.

INTRODUCTION microscopic modelling approach of balancing con-

Modelling of separation columns on the basis of the
laws of conservation of mass and energy, and the
principles of dynamics of mixtures results in com-
plex large-scale dynamic models exhibiting severe
nonlinearities. The problems appearing during si-
mulation, optimization and control of separation
processes can be lessened by the development and
application of reduced models. These are indis-
pensible for the design of sophisticated control
systems for columns difficult to operate (Retz-
bach, 1986), where often model based measurement
techniques (Gilles, 1986) must be applied.
The demand for fast simulation and high quality
control stimulated the development of various mo-
del reduction techniques suitable for separation
processes. They can be assigned to two main groups
- the linear and nonlinear approaches. The linear
models are mostly based on a direct transfer func-
tion approach or on linear model reduction of the
linearized balance equations. These techniques
yield models, which are only valid near a given
operating point. While among the nonlinear reduc-
tion techniques two approaches can be found. These
are the strongly mathematically oriented colloca-
tion methods, recently applied to separation pro-
cesses by Cho and Joseph (1983) and Stewart,Morari
and Lev ten (1985) and the compartmental technique
introduced by Benallou, Seborg and Mellichamp
(1982). The latter may be seen as a stage-by-stage
approach, lumping several physical stages to one
hypothetical stage. These methods include all non-
linear effects of a detailed model and allow for a
relatively accurate description of the steady-state
and dynamic behaviour. Their main area of applica-
tion is assumed to be fast simulation and optimi-
zation.
servation variables commonly practised, a macro-
scopic way of modelling has been attempted. To
establish a modelling methodology of this kind,
it is necessary to establish essential variables
describing the limited variety of the spatial
structure behaviour. These 'characteristic' variab-
les form some kind of minimum realization of the
nonlinear distributed system. Due to the physical
and system-theoretical significance of the new
state variables, an excellent performance of these
models in control system design can be expected.
A brief summary of some theoretical considerations
on the properties of the spatial structures is gi-
ven, after reviewing the phenomenological dynamic
behaviour of distillation columns. A more detailed
discussion of these problems may be found else-
where (Epple, Marquardt and Gilles, 1986). The
main part of the paper deals with the derivation
of a reduced model for one section of a binary di-
stillation column. The approach is similar to that
suggested by Epple (1986) for model reduction of
distributed parameter reaction systems. Finally,
the structural properties of the reduced model and
its behaviour in comparison to a detailed model
are presented.
DYNAMIC BEHAVIOUR
The dynamic characteristics of distillation columns
will be demonstrated by looking at two binary
towers, H and L. Though, column H produces high and
column L lower purity products, both plants are
working under similar operating conditions. The
operating and construction data is given elsewhere
(Marquardt, 1986).

None of the above methods takes into account the The concentration profiles are plotted versus
apparent spatial structures evolving in many sepa- column height in Fig. 1. They show similar spa-
ration processes. Instead of the spatially local, tial gradients, because of the comparable operating

123
124 W. Marquardt
conditions. In both columns, after a step change
in the reflux rate, the profiles are propagating
in the direction of the reboiler in both the strip-
ping and the rectifying sections. Their principal
shape is conserved during the propagation. An ob-
server moving with a coordinate system linked to
the front would hardly recognize any variations in
the profiles. The main difference in both cases
is the column's sensitivity to disturbances. In
contrast to column H, the propagation velocity of
column Li concentration profiles rapidly decreases
to zero. Complex multicomponent distillation
plants behave in qualitatively similar way (Retz-
bach, 1986).
MODEL REDUCTION TECHNIQUE
A Rigorous Dynamic Model
The principal dynamics of distillation are given
by a simple model, which ignores hydrodynamics and
energy balance. In the case of a packed column
balancing of an infinitesimal volume element
yields the dimensionless hyperbolic equations
|aft - | zf = - B-[y*(x)-yl
a postulated, that the dynamics of a finite extended
, (la)
ζ G [0,1]
t e [0,T]
& * Î £ - B V ( x > - y i (lb)
for the liquid and the vapour concentration χ and y
respectively. Β is a dimensionless mass transfer
coefficient, A is the ratio of the liquid and va-
pour streams in the column section and ε stands
for the holdup ratio of both phases. Frequently,
ε is taken to be zero due to the small holdup in
the vapour phase compared to that in the liquid
phase. The equilibrium relation is given by y"(x).
Each column section is modelled by a set of equa-
tions of type (1). Its boundary conditions are de-
termined by material balances of the column bound-
ary systems, i.e. the feed tray, reboiler and con-
densor. A model similar to the one stated was used
by several investigators, such as von Rosenberg
and Hadi (1980). The distributed parameter formu-
lation is suitable not only for the behaviour of
packed columns but also for staged columns. This
is true even where only a small number of stages
are concerned, if the mass transfer coefficient is
adjusted. This model forms the starting point for
the following model reduction considerations.
Wave Propagation Phenomena
The depicted dynamic behaviour of distillation col-
umns is termed as nonlinear wave propagation. In
mathematical physics, a nonlinear wave is defined
to be a spatially structure moving with constant
propagation velocity and constant shape along a
spatial coordinate/This phenomenon can be studied
analytically in an idealized, spatially infinitely
extended column sectign using Eq.(l). The investi-
gations include the development of general exist-
ence conditions of a wave front solution, its sta-
bility, shape and propagation velocity.
Existence. A wave front solution exists for all
physical meaningful system parameters, if the non-
linear equilibrium line y*(x) is continuously dif-
ferent! able and either strictly monotonically in-
creasing or decreasing in the concentration inter-
val covered by the wave.
Stability. Wave front solutions are invariant
against shifts along the spatial coordinate. This
property is established by analyzing the eigen-
value problem derived from Eq.(l) after lineariza-
tion around a steady-state spatial profile. One
solution is found for an eigenvalue λ = 0 and its
eigenfunction, which is exactly the first deriva-
tive of the steady-state profile plus a constant.
The convergence of any initial profile to the con-
stant travelling wave solution is not yet proven.
However, simulation results gives one to believe
the existence of this property, defined here as
shape stability.
Wave front shape. The shape of the spatial pro-
file can be determined analytically only for a
particular choice of the equilibrium line. It is a
structural property of the autonomous system and
therefore independent of initial and boundary con-
ditions.
Wave propagation velocity. It is shown, that the
velocity is only a function of the concentrations
in the streams entering and leaving the column
section. The determining equation is interpreted
to be a material balance of the infinite long
column section.
Model Reduction Principles
The wave propagation properties are used to deve-
lop a reduced model for one column section. It is
system is comparable to that of the infinite long
section discussed above.
Due to propagation of the profiles, the model
Eqs.(l) are transformed first in a ξ-t coordinate
system moving with a velocity w. Ignoring the va-
pour phase holdup, we obtain
î V
|X- || -B.[y
( 1 += w ) (x ) - y ] (2a)
| f = A»B[y*(x)-y] (2b)
The wave coordinate ξ is defined by
ξ = ζ - ζ' = ζ - / w dt , (3)
where ζ = ζ' is the ξ-t coordinate system origin.
The boundary conditions of (2) are formulated in
this moving coordinate system as
x(l-z\t) = ( t ) , y(-z',t) = y ( t ) (4)
X l Q
x^(t) and y ( t ) are the concentrations in the
Q
liquid and the vapour streams respectively enter-
ing the column section. The essential problem now,
is the determination of the shape of the propaga-
ting profiles. In a first approximation the deri-
vate ax/at is ignored in the moving coordinate
system. This assumption is close to reality, be-
cause the main part of the dynamic behaviour is
covered by the movement of the profile with w.
Hence, the concentration profiles may be calcula-
ted using the stationary Eq.(2). Indefinite inte-
gration leads to
y = A(l+w) χ - C , (5)
x
(1+W)dx + Co (6)
Βί^+ν'Ίχ) " A(l+w)x]
C and C are integration constants to be determi-
x 2
ned with Eq.(4). Hence, using Eq.(6) χ can be cal-
culated independently of y, which itself is deter-
mined by Eq.(5). Therefore, only the liquid con-
centration χ is considered further. If the inte-
gral of Eq.(6) can be solved analytically, the
concentration profile χ(ζ) is given by the follo-
wing implicit equation:
0 = ξ - F[xU),p] , (7)
 Nonlinear Model Reduction for Binary Distillation 125

which usually is not solvable for χ(ξ) explicitly. Eq.(6) represents a trial function for all kinds
The vector £ of dimension Ρ contains groupings of ?
of nonlinearities y*(x) in implicit form. This
1
the variables A,B,w, the integration constants Ci, term can be used in a model reduction technique
C and the equilibrium line parameters. The p . s only,if the integral can be solved analytically.
2
are called shape parameters of the profile. To achieve this requirement the only nonlinearity
in the integrand is approximated by another non-
2
For further illustration, a quadratic equilibrium linear function, which allows for an analytical
line y*(x) = a x + b x + c is studied. The integrand solution of the integral. Candidate functions are
of Eq.(6) is then given by all polynomial and rational functions. A curve fit-
ting analysis allows a systematic choice of a
trial function containing a minimum number of
shape parameters a priori to the model reduction
or after definition of suitable shape parameters itself. The study of typical equilibrium relations
1 reveals that a third order polynomial or a ratio-
[ρ (χ-Ρι)(χ-ρ )Γ . nal function, having one first order polynomial as
3 2 a numerator and another as a demoninator is ade-
In this case, integration leads to the following quate in most cases.
equation of type (7)
I XPi - χ(ξ) 4 Taking an equilibrium relation with a constant
0 = ξ- IN volatility α
P ( P R P ) "" U ) " P "
3 2 2 aX
/(χ) (10)
which can be solved for χ(ξ) explicitly l+(a-l)X

Pi ~ P as an example, Fig. 3 shows the results for three

χ(ξ) = p + 2 (8) different approximating functions in a concentra-
2
1 Ρ+3 (θΡ ι - Ρ 2) ( ξ + Ρ 4) tion interval [0.02,0.6] and α = 5. The quadratic
approximation is insatisfactory in this case,
The physical meaning of the shape parameters is whereas the cubic approximation is sufficient. The
shown in Fig. 2. The parameters p and p are the following rational function
1 2
concentration limits for ξ + ± °° of a profile ex-
tended over the system boundaries; p reflects the
3
maximum slope of the profile; p , which is identi-
(11)
4
cal to the integration constant C , determines the
2
location of the ξ-t-coordinate system origin. shows the best results. This is not surprising,
because Eq.(10) is a special case of Eq.(ll).
A simple first order reduced model is formed by
Eq.(6) and the differential equation The spatial trial function (7) in the case of a
relative volatility approximation of the equilibri-
um line is given implicitly by

following from the definition of w and z' in

Eq.(3). This equation causes a spatial shift of
=0
^P^W (p +p ).ln(x-p )] + p
'
[(P5+Pl) ln(PrX
-χ -

(12)
the profile. The shape parameters are determined 5 2 2 4
by algebraic relations given by the shape parame-
The shape parameters p< to p have the same mean-
ter definitions and the boundary conditions (3) 4
ing as stated for Eq.(8) above. The additional p
of the system. The validity of this kind of mo-
is a measure of the asymmetric properties of the
5
delling approach is due to the special stability
feature qf the system, i.e. the preservation of spatial profile. A cubic equilibrium line approxi-
the typical shape during transients. It should be mation yields a trial function structure similar
noted here, that the shape parameters £ vary to that of Eq.(12).
quasi-stationarily during transients, because of
their dependence on the dynamic state variable z' Determination of the Shape Parameters
in the boundary conditions (4).
To deduce equations for the determination of the
Often, strong boundary influences on the travel- shape parameters £ , the system equations (2) and
ling wave exist in real systems, especially if on- boundary conditions (4) are manipulated algebraic-
ly low purity products are aimed at. To cope with ally to give the following:
systems of this kind, Eq.(7) is generalized. The ax _9x 1 ax
structure of the function F(x,£) is preserved, aç M "afal"
but the shape parameters ρ are taken to be no τι
longer constant and determined by their defini- 1+w

it
tions. They are conceived as unknown time-varying (13)
variables. So, variations in the profile are rea-
lized by changing parameters in a structurally
constant spatial trial function. The shape para- y = y (χ) + (1+w) (14)
meters and the moving coordinate system origin
are chosen to make up the new state variables of
the nonlinear low-order model. |F(-z',t) - ( l ) | | ( - z ' , t ) =
+w
A practical application of the method sketched ab- B-{y"[x(-z',t)l - y ( t ) } , (15a)
ove requires general techniques in order to Q
choose the trial function F(x,p) in Eq.(7) and to x(l-z',t) = ( t ) , (15b)
derive equations for the calculation of £ and z' X l
or w. where m(x) = dy'Vdx is the slope of the equilib-
rium line. This manipulation leads to a differen-
Spatial Trial Function tial equation for χ(ξ,ΐ) with two boundary condi-
tions and an explicit relation for y ^ , t ) , if
Due to the possible decoupling of the liquid and χ(ξ,ΐ) and its derivatives are known.
vapour model equations, it is necessary to select
a trial function for χ(ξ) only but not for γ(ξ). The limit concentrations p
1and p 2in an infinite
126
long column section are determined by Eq.(15) to
suit the system boundary conditions. Additionally,
they are assumed to be quasi-steady, i.e. algebra-
ically coupled to the remaining state variables.
The method of weighted residuals (Villadsen and
Michel sen, 1978) is used to deduce the remaining
shape parameter state equations. This technique is
illustrated by those trial functions explicitly
solvable for χ(ξ) as is the case for a quadratic
equilibrium line approximation. The spatial trial
function χ(ξ) = f(i,p) is inserted in the partial
differential equation (13). Because of the approxi-
mative nature of f(ç,p), Eq.(13) is not exactly
equated to zero. There remains a residual func-
tion R showing a nonlinear dependence on £ and ξ
and a linear one on jp, which is due to the lineari-
ty of Eq.(13) with respect to 9x/3t.
The method of weighted residuals (MWR) provides
differential equations to calculate the remaining
shape parameters p^ as follows
1-z'
j R ( p p ^ ) ν^(ρ,ρ,ξ) dç = 0; j=3,...P . (16)
9 the result of the countercurrent flow boundary con-
-ζ'
These equations are simplified by the choice of
weighting functions Wj not dependent on the deriva-
tives £ . Appropriate weighting functions commonly
used in the literature are delta functions (collo-
cation method) and sensitivity functions of the
trial function 8f/apj (Galerkin method) or sensi-
tivity functions of the residual 3R/3pj (least
squares method). Due to the linearity of R with
respect to ρ integration of (16) always leads to a
system of lTnear-implicit differential equations
of order P-2 for the dynamic shape parameters. In
those cases, where the integral cannot be directly
solved, it is approximated analytically by a suit-
able numerical integration method, for example by
the trapezoidal rule.
Although the method is presented for a trial func-
tion explicit in χ(ξ), it is not limited to this
type of function. An extension to implicit trial
functions (for example Eq.(12)) can be easily
achieved. The derivatives in Eq,(13) are determi-
ned by differentiation of the implicit trial
function (7). They result in functions of χ(ξ) on-
ly. Therefore the residual is also a function of
χ(ξ) and the integration in (16) must be carried
out oyer χ(ξ) rather than over ξ.
Determination of the Propagation Velocity
The last question is that of the location of the
concentration front ζ'. Here, several approaches
are possible. It is known from theoretical consi-
derations, that the propagation velocity of the
profile is determined by a global material balance
Of a column section., Using a trial function χ =
f(ç,£) this balance equation can be written as
1-z'
AW**'
= (L^CX^TJ-XI-Z^.TJL+^Y^TJ-YIL-Z ,!)]
a linear-implicit differential equation for z'.
The sensitivity function integrals can sometimes
be solved analytically. If this is not possible,
an approximation using a finite sum may be used.
Another approach is the use of the simple diffe-
rential equation (9) and a physically based defi-
nition for w, given by the global material balance
of a column section infinitely extended over its
boundaries. With this assumption an explicit func-
tion of the limit concentrations pi and p results
2
W. Marquardt
for w.
y ( P J - y (p
9 - 1 (18)
In addition to these physically based methods, it
is possible to take a purely mathematical approach.
The location ζ' is considered as a special shape
parameter and is determined by another MWR equa-
tion of type (16).
The Reduced Model
The application of the propositions stated above
will lead us to a linear differential-algebraic
system of equations
K(s) · s = q(s) (19)
with a singular matrix K. The vector s_ consists of
the shape parameters £ , the front location ζ' and
sometimes of liquid concentrations at the outlet
of a column section. The algebraic equations are
ditions. This structure is the same for column sec-
tions and complete distillation columns. Powerful
numerical techniques for the direct solution of
this kind of system are available (Petzold, 1982).
REDUCED MODEL BEHAVIOUR
A Distillation Column Section
The quality of the reduced model is investigated
first with a column section. The equilibrium is de-
fined by Eq.(10) with α = 5 . The system parameters
are A = 1.34, Β = 10 and ε = 0. The inlet concen-
trations are chosen to be χ = 0.6 and y = 0.1.
χ 0
Three models are studied: model R, based on the
trial function (12), model C and model Q based on
a cubic and a quadratic equilibrium line approxima-
tion. R1 and C consist of three dynamic state variab-
les ( z , p , p ) whereas model Q uses only two dyna-
3 5
mic states ( z ' , p ) . It is found by simulation,
3
that the integration constant type ρ may be taken
identically as zero, without altering the model
behaviour in any case. In all simulations Eq.(9)
and (18) are used for the determination of z' and
the Galerkin method for the determination of diffe-
rential equations for the dynamic shape parameters.
Fig. 4 shows the steady-state behaviour of the
reduced models in comparison to a finite difference
solution (FD) of Eq.(l) with k equidistant grid
points. Fig. 4a shows, that the FD solution conver-
ges to the model R solution for increasing grid
point number k. In this case, the reduced model may
be termed better than any FD solution of Eq.(l), In
Fig. 4b the models R, C and Q are compared to an
FD solution with k = 250. The result corresponds
to the curve fitting analysis of Fig„ 3. Model Q
does not compare sufficiently to the FD solution,
1
(17)
whereas the models R and C demonstrate a high de-
gree of resemblance. In Fig. 5 the dynamic behav-
iour is plotted after a step disturbance in A and
y . The correspondence between the FD solution and
0
model C (cubic equilibrium line approximation) is
, satisfactory. In the case of stronger boundary in-
fluences in order to compute ζ', Eq.(7), (18)
should be replaced by the more complex Eq.(17) or
by a MWR equation.
The structural properties of the reduced model R
are investigated in a wide range of operating con-
ditions. As an example, the control variable A is
varied to move the concentration front from one
section boundary to the other. The steady-state
gains are shown in Fig. 6. p and p show an almost
3 5
linear dependence on A in the steady state. The
system nonlinearity is lumped in the front loca-
 Nonlinear Model Reduction for Binary Distillation 127

tion z* . The same behaviour is revealed by the
steady-state sensitivities. Their absolute values
show that the main contribution to the system be-
haviour is given by the location z' and the gra-
dient p of the front. The asymmetric parameter
3
p is only necessary for more accurate approxima-
5
tions.
Linear design methods must be used to develop con-
trol systems or model-based measurement techniques.
Hence, the properties of a linearized reduced or-
der model are of great importance. They are stu-
died in a wide range of the control variable A.
All states are locally observable and controllable,
if one measurement in the concentration front is
used. The three system eigenvalues vary only
slightly with A. They are real and lay in the
ranges [-0.02, -0.15], [-0.5, -0.7], [-3.1, -4.8]
respectively. Simulation studies show the simila-
rity of the nonlinear and the linearized model.
These results show that only a weak nonlinearity
is incorporated in the reduced-order model state
variables. Hence, linear design methods using the
linearized reduced model are expected to result in
efficient control systems.
Two Distillation Columns
The model reduction technique is further applied
to two packed distillation columns separating the
same feed to high and medium purity products r e -
spectively. The operating conditions are comparab-
le to columns H and L of Fig. 1. These examples
present a severe test for the model reduction
technique due to the strong boundary influences in
both cases.
A reduced model based on type R discussed above is
compared to a high grid point number FD solution
of a distributed parameter system model according
to Eq.(l) and equilibrium line (10). In both
steady states, i.e. at the beginning and at the
end of the simulation and during the whole tran-
sient, the solution of both models are almost the
same. The concentration profiles, computed with
the reduced model R, are plotted in Figs. 7 after
a large step disturbance of more than 10 % in the
extension of the method towards concentration
enthalpy and multicomponent column models and on
its application in the design of control systems
and model based measurement techniques.
REFERENCES
Benallou, Α., D.E. Seborg, D.E. Mellichamp (1982).
Dynamic Compartmental Methods for Separation
Processes: Basic Theory. Paper presented at
the AIChE 1982 Meeting, Los Angeles.
Cho, Y.S., B. Joseph (1983). Reduced-order Steady-
State and Dynamic Models for Separation
Processes, Part I and II. AIChE-Journal, 29,
261-275.
Epple, U. (1986). Model Reduction for Nonlinear
Systems with Distributed Parameters. These
proceedings.
Epple, U., W . Marquardt, E.D. Gilles (1986).
Model Reduction for Systems with Wave Propaga-
tion. To be published.
Gilles, E.D. (1986). Some New Approaches for Con-
trolling Complex Processes in Chemical Engi-
neering. Third Int. Conf. on Chem. Proc. Con-
trol, CPC III, Jan. 12-17, 1986, Asilomar, Ca.
Marquardt, W. (1985). Model Reduction Techniques
for Separation Columns. Proc. Int. Conf. "On
Industrial Process Modelling and Control",
June 6-9, 1985, Hangzhou, China.
Petzold, L. (1982). Differential-algebraic equa-
tions are not ODE's. SIAM J . Sci. Stat. Corn-
put. , 3, 367-384.
RetzDUcrT, F . (1986). Control of an Extractive
Distillation Plant. These proceedings.
Von Rosenberg, D.U., M.S. Hadi (1980). Numerical
Solution of Multi-Component, Packed Tower
Distillation Problems. Chem. Eng. Commun., 4,
313-324.
Stewart, W.Ê., K.L. Levien, M . Morari (1985).
Simulation of Fractionation by Orthogonal
Collocation. Chem. Eng. Sc., 40, 409-421.
Villadsen, J.V., M.L. Michelsen (1ÏÏ78). Solution
of Differential Equations Models by Polynomial
Approximation. Prentice-Hall, Englewood Cliffs,
New Jersey.

\j
feed concentration. The reduced model consists of
eight dynamic state variables. These are the front
location and two shape parameters in both column
sections and the distillate and bottom product con-
centration. Hence, this example shows, that the
number of state variables and the structure of the
\J 1

reduced model do not depend on the operating or

construction data of a distillation column, but on- 23 1?
\L
ly on the phase equilibrium of the mixture to be
separated.
23
(). χ 1
CONCLUSIONS

The model reduction technique presented here, 44

shows satisfactory results in the examples studied.
Its main strength lies in its use of 'characte-
ristic' state variables, describing the limited Fig. 1. Step response to a disturbance
variety of the macroscopic spatial structures in reflux rate.
evolving in separation columns. Therefore low-or-
der models including the main nonlinear effects of
the system are achieved. Due to the suitable struc-
tural properties of the model, it is expected to
serve advantageously in the design of control
systems using state space methods. One disadvan-
tage of this approach may claimed to be the rather
high complexity of the nonlinear low-order models,
which are tedious to write down in full detail.
However, after programming and debugging some mo-
dels based on different commonly useful equilib-
rium line approximations, they may be easily app-
lied to a wide class of distillation problems.
Fig. 2. Physical meaning of shape parameters
The future investigations will concentrate on the
 128 W. Marquardt

1,0 0,61
y'

0,6

1\ FL equilibrium line 1,38 A

OA H rational approximation -13,3
ι G cubic approximation
ι / quadratic approximation
0,2

0,0 '« -13,8

0,0 0,2 0A 0,6 • χ 1,0
Fig. 3. Equilibrium line approximations. Fig. 6. Steady-state gains
(Model R ) .

0,7

FD solutions
Model R S S ^ Y ^

k=100 Y////
k=5o /(/M
k = 20
Θ
0,0
0.0 0,5 zH 1,0
0,7
FD solution (250)
Model R
Model C
Model Q
// /

Jy ® x 1,0
1,0
0.0 0,5 „ z[-] 1,0

Θ
Fig. 4. Comparison of different models.
Steady-state behaviour.

0,7

A: 1,34-1,4

0,0

solution (250)
Model C

zl-J 1.0

-1.0
0,0 0,5 χ 1,0

Fig. 7. Dynamic response to a step

FD solution (250) disturbance in feed concentration
Model C 0.45 + 0.5.
Reduced model R

0,5 „ z[-] 1,0 a) separation to high purity

products
Fig. 5. Comparison of different models.
Dynamic behaviour. b) separation to medium purity
products
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, U K 1986

NON-LINEAR DISCRETE TIME MODELLING

AND CONTROL OF A PILOT
NEUTRALIZATION CHEMICAL REACTOR
G. Gilles, N. Laggoune, Β. Neyran and D. Thomasset
G.R.A.P. (Groupe de Recherche Rhône—Alpin en Automatique et ses Applications
à la Production), Laboratoire d'Automatique, Université Lyon 1, 43 boulevard du
11 novembre 1918, 69622 Villeurbanne, France and Laboratoire d'Energétique et
d'Automatique, INSA de Lyon 20 avenue Albert Einstein,
69621 Villeurbanne, France

Abstract. Many continuous industrial processes working inside a large operating range
can be modelled by variable parameter and/or bilinear continuous state equations. In
the case of processes with variable parameters, the state affine models allow to get a
discrete approximated representation. An original direct identification method, invol-
ving a variable parameter filter, is processed inside the whole operating range in order
to estimate the model coefficients. A dual approach leads to state affine control algo-
rithms by gathering a family of linear controllers into a unique control law. In the
case of bilinear continuous processes, their discrete model is of linear time-varying
type. A way for digitally controlling them can be a closed loop linearization which
leads to an implicit control law. An improvement can be obtained by adding integrators.
Experimental results got, in each operating case (variable parameter or bilinear system),
from a pilot neutralization reactor show the applicability of the methods.

Keywords. Digital control ; non-linear control systems ; time-varying systems ; bilinear

control ; optimal control ; pH control.

INTRODUCTION

The continuous industrial plants dynamically working through experimental results got from a pilot neu-
inside a large operating range involve non-linear tralization reactor working in each operating case:
phenomena that mostly cannot be suitably approached variable parameter system or bilinear system.
by linear models. In order to describe their beha-
viour, it is necessary to use non-linear models and/ PRESENTATION OF THE CHEMICAL PILOT PLANT
or variable parameter models. The control of such
systems can be managed from adaptive methods or mul- The chemical pilot plant (Fig. 1) is a continuous
ti-model techniques but it may be preferable to try neutralization reactor involving a strong base
to find a global non-linear model correctly descri- (Na OH) and a strong acid (HC1). For a constant
bing the system behaviour in a large operating ran- acid flow-rate Q^, the base flow-rate Qg(t) is cho-
ge and a unique control law. Moreover, the goal +
sen as control variable, the output variable being
being the control by means of one or several digital concentration C(t) of H ions obtained by means of
processors, it is necessary to build a discrete a pH-meter.
non-linear and/or time-varying parameter model.
In this paper, two methods for discrete modelling
and control are presented in the domain of conti-
nuous industrial processes which are linear with
time-varying parameters or bilinear. In the case of
continuous processes with variable parameters, the
C O M P U T RE
state affine models, linear in the state and poly- L R
nomial with respect to the extended inputs allow to
get a discrete approximated representation. An ori-
ginal direct identification method operating inside
the whole range is used in order to find the model
coefficients : it involves a least square method
and a variable parameter filter. A dual approach
allowed us to propose a synthesis method for state
affine control algorithms : the technique consists
in gathering a family of linear controllers into a
unique control law described by a state affine sys-
tem for which some closed loop characteristics can γ do
be considered as quasi-invariant. In the case of
continuous bilinear processes, we get a discrete
model which is linear with variable parameters de-
Qa+Q
pending upon the inputs. By applying the idea of
ZJLR 1
b
closed-loop linearization, we get a linearizing
control via non-linear state feedback. The corres-
ponding equation is implicitly solved during each
sampling period by means of numerical analysis me-
thods. It is shown how an improvement of the con-
Fig. 1 : Chemical pilot plant
trol law performances with respect to modelling
errors can be obtained by adding an integrator. The
applicability of those methods is illustrated

129
130 G. Gilles et al.

In order to get a model indépendant from the expe Control

rimental conditions, we consider :
C ( initial value of For each value of the reactor volume, the system is
0
c(t) = „ " ^ with C(t) linear. Then, the control law of the state-affine
c B c+ 0 Q : base concentra- model is necessarily a linear one depending on the
IS
tion measurable parameter V and it seems very difficult
Around a working point defined by Q q and c(o), the to build directly this control law. Therefore, we
B
infinitesimal mass balance leads to a dynamic model, can approximate this law by using the same approach
between the base flow-rate deviation q(t) and the in control than Dang Van Mien and Normand-Cyrot used
variation of the output normalized concentration in identification (Dufour and Thomasset, 1984).
c(t), under the following form :
For Ν values of V, denoted V , chosen in (v^jjj,
V_ dc(t) _ q(t) q(t).c(t)
- c(t) - (1) V"mAx) , the chemical pilot plant has a linear beha-
Qn dt
viour and there are many methods to design a linear
control law. We chose one of the most efficient, the
with ^0 = \ BO optimal control with integrator and reference model
reactor volume (Foulard and co-workers, 1977). For each value, V ,
the control law q(k) is given by :
Moreover, we must consider a time-delay (30 sec)
on the output concentration measurement because of
q(k) = -L(V ).w(k) - M(V ).r(k) + P(V ).s(k)
the pH-meter location.
α α α (3)
+ N(V ).z(k)
Equation (1) represents a bilinear system with ti-
where : F w(k) is the state of the process a
me-varying parameters depending on the reactor vo-
lume (as a rule, Qq is constant). In this paper, r(k) is the output of the integrator
we consider two different approaches of this pro- s(k) is the state of the reference model
cess corresponding with two different working con- z(k) is the reference input
L ( V ) , M ( V ) , P ( V ) , N(V ) are appropriate
ditions : a a
dimensional matrices.
a a
- in the case of small deviations of the control
variable around the operating point and when the
Just as we built the state-affine model of the pro-
volume is time-varying, the system is linear with
cess and the state-affine reference model, we can
variable parameters. Then, the use of state-affine
get, by means of multilinear regressions computed
model leads to a unique control law available for
from the values of the control matrices L, Μ, Ρ, Ν
all reactor volumes (section 3)
at different working points V , the variation of
- with a constant volume working condition and
these matrices with respect to the variable reactor
for large deviations of the control variable, the
volume V. So, we can write the o£ ,ΊΤ\ , Ç , matri-
system must be considered as a bilinear one and
ces under the following forms :
section 4 presents a linearizing control via non- σ σ
linear state feedback. J . J
^ S J
p ΐ
<£=Σ
2
l. V ;Î71 = Σ m. V ; Τ = Σ p, V
IDENTIFICATION AND OPTIMAL CONTROL FOR
1=0 j=0 j=0
TIME-VARYING VOLUME WORKING CONDITION

In order to get a global model suitable for all J

reactor volumes and to design a unique control law, =Σ n..V J
we use the theoretical results on approximation of j-0
non-linear discrete time systems by state-affine
models (Fliess and Normand-Cyrot,1980) ο* , σ^, σ^, σ^, regressions orders, are chosen to
minimize a performance criterion.
Identification
The following figure describes the complete struc-
The searched state-affine model is on the following re of the control.
form :
3
x(k+1) = ( A + A V ( k ) + ...+A .V (k)+B .q(k)+B q(k)
o r 3 3 o r
+...+B .q(k).V (k)) .x(k) (2)
3
[ y(k) = C.x(k)
with q(k) control variable and V(k) measurable vo-
lume which is an extended input.

2
state "parallel
The parameter estimation of this state-affine model affine
can be achieved by means of two different approa- reference
! state affine
|
ches. The first developped by Dang Van Mien and
Normand-Cyrot (Dang Van Mien and Normand-Cyrot, l-.-Tt:
1980, 1984, Neyran and co-workers, 1984) consists
in gathering in the state-affine system by means
of multilinear regressions a set of linear identi-
ro t n e
fications obtained for Ν values of V belonging to
(^MIN> ^MAx)* ^ method presented by the au-
thors (Dadugineto, 1985) is a more direct approach. Fig. 2 : Structure of the state-affine control
It allows to build the state-affine model from an
input-output sequence obtained during a variation Notice that the estimated state of the process is
of V belonging to (Vjîj , Vj^^) . The least square obtained by means of the internal state-affine mo-
N
estimation is biased in the case of an additive del issued from the identification. In the same way,
noise on the output. The use of a time-varying li- the reference model is a state affine one with res-
near filter which has the same structure as the pect to the extended inputs (z(k)-£ (k)) and V(k).
state-affine model, allows to reach and unbiased
a
estimation. Applied to the chemical process, the Application to the chemical pilot plant
direct approach gives better results than the other
(Neyran and co-workers 1985). Fig. 3 shows the output concentration evolution with
 Pilot Neutralization Chemical Reactor 131

respect to a série of reference input steps during sed loop system, we want to prescribe the system
a linear time variation of the reactor volume. In performances by setting a closed loop sampled data
order to improve the robustness of the state-affine transfert function under the 1following form :
control, we compare three control laws : two linear
optimal controls computer for V = 4 I and V = 12 i X(z) _ (1-d)z~
H(z)
and the state-affine one. Let us notice that only E(z) •1
1-dz
this control law gives good results for all reactor in order to reach the step input during the steady-
capacities. state working, the d pole setting the response time.
The corresponding closed loop state model is :
LINEARIZING CONTROL FOR BILINEAR WORKING v w c
CONDITION = dx, (1 - d)€ (6)
~k+1 ^k ' k
and, comparing to the open loop model (5), we get
When the neutralization reactor is operating under the equation :
a constant volume, and when the input range is lar-
f(u ) Ξ Y ( u ) . u
ge, it has been shown that the model (1) becomes k k k +[cj)(uk)-d]xk-(1-d)ek = 0 (7)
bilinear and can be written under the following
f orm : If we can get the solution u^ of this equation, we
χ = ax + bu + nxu (4) obtain a non-linear state feedback control law. But,
contrary to the continuous linearizing control case,
Numerous works have been devoted to bilinear sys- equation (7) does not lead to an explicit control
tems but they have been more related to the analy- law. It has to be solved, during each sampling pe-
sis of their properties than to their control. Se- riod, by numerical analysis procedures. For example,
veral types of control have been proposed : "bang-ban^' the successive substitutions method allows to solve
(Mohler, 1973), stabilization by Lyapunov functions equation (7) by means of the iterative algorithm :
(Gauthier and Bornard, 1980), immersion into ano- = F }( u[ d ^ ( u k )>] xik+ ( 1 - d ) e k
ther systems (Claude, Fliess and Isidori, 1983), (8)
k,i
model matching (Di Benedetto and Isidori, 1984),
pseudo-linearization (Mouyon, Champetier and Reboul- 1
The use of a faster convergence algorithm, like New-
let, 1984). Very rare are the papers dealing with ton-Raphson s one, is not necessary in the case of
concrete applications : let us mention a recent this scalar problem and if the sampling period is
paper (Cebuhar and Castanza, 1984) related to the small enough.
control of chemical reactors by minimizing quadra-
tic criterions. A linearizing control, which minimi- P.I. Discrete linearizing control
zes a quadratic criterion with variable weighting
matrices (Gilles, 1983), involving a non-linear Due to the closed loop state model form, the static
state feedback, has been presented by the authors error corresponding to a step input is theoretically
and applied to this neutralization reactor (Gilles equal to zero. Nevertheless, in practice, because of
and Laggoune, 1985a). Much rarer are the works rela- modelling errors, the "perfect model following" is
ted to digital computer control of continuous bili- not guaranteed and static errors can occur. That is
near processes. Nevertheless, we can mention the why it is preferable to systematically introduce an
paper of Alvarez Gallegos brothers (1982), upon di- integrator in the control law in order to cancel
gital control of fermentation processes modelled static errors.
by a larger class of systems : models linear in con-
trol ; the criterion which is minimized by the con- The introduction of integrators can be managed in
trol law is still a classical quadratic criterion. the classical way, like in linear quadratic optimal
More recently, the authors (Gilles and Laggoune, problems (Foulard and co-workers, 1977) by augment-
1985b)proposed an implicit digital linearizing ing the dimension of the state vector :
control law for multivariable continuous bilinear x =( hw i tv v + ( ex
)
systems. Based on a discrete time variable parame- k k " k
k" 4 v V * k i
ter model of bilinear systems, the method is recal- +
and recomputing the control law in terms of the pro-
led for the monovariable model (4) in the next two
ceding one. Nevertheless, by applying Zdan's method,
paragraphs and then twice extended.
we notice that we are not able to independantly set
the overshoot time and magnitude of the closed loop
Discrete model of the continuous bilinear system second order system. It is much more preferable to
control by means of a P.I. controller, within a se-
Systems modelled by equation (4) and controlled by cond loop, the preceding linearized system ( Eq. 6 ) ;
a digital computer with a zero hold have their evo- we can then apply the classical linear system theory
lution described by : 1
after the linearizing control law. Inserting the P.I.
x(t) = (a + n u ) x(t) + b u transfer function : κ + [Κί/(1-ζ" )] leads to the
k k overall closed loop transfer
with u = const, inside the sampling period 1 1 function :
kT ύ t < (k+1)T. O-dXK + K^z [i- z" /(k+K.)]
H(z) = 1
The system is then piecewise linear and its solution 1 + [(1-d) (K+K )-(1+d)]z~ + [d-(1-d)K]z"
i
involves the transition matrix : , r, , ™\ Λ
which must be compared with Zdan's form
b[<J)(u,T)-1J
k •1 (1 - Bz ')
<Ku ,T)=exp[(a+nu )TJ and y(u ,T)= , H(z) = Az
k k k a( + } 1 - Cz Dz
which leads to
k+1
the
J>(u, following
,T).x, discrete
+ y(u, ,T).u, model : (5) 6 t H
with : A=p —
which is linear with variable parameters depending ο 5θ— Logp -cos9 + —
o o p '
, C = 2p cos9
o
upon the control variable setting the operating
point. sinB
Logp +cos0 +p
ο ο ο _ 2
Discrete linearizing control , D = ρ
6 K
Logp -cos9 + —
By means of this discrete model, it is possible to o ο ρ
find a control law which allows the continuous bili-
near system controlled by a digital computer to get the conjugate poles ρ exp(+j6 ) being related to the
the pleasant property of global linearity. overshoot magnitude D and the overshoot time NT by
the relationships :
Remaining the same state variable x ^ s ^ for the clo- _ Νϊ/ϊ) and θ = π/Ν . (9)
ν ο
132
D and N being chosen, the following formulas (got
from the comparison) allow to tune the controller
parameters : 2 1 C O+S e
sin6 d-p
d=p (• v
-Logp +cos0 ),k= u ° k = ~ pQu o Po
° o 1-d ' ι 1-d (10)
We
.._ can notice that,, not only
__.ly k and k. are sset, but
d is also constrained to a numerical value.
Linearizing control of the delayed bilinear system
The neutralization reactor involving, as many in-
dustrial processes, a time delay, it is important
to use a model taking into account this time delay
for a more precise linearization.
In the case of the reactor delay equal to 2 sam-
pling periods, the state dimension is augmented to
3 and the discrete model becomes :
0 1 0 troller can be mixed in order to solve the problem.
0 0 1 = 11
-k
0 0
1
The prescribed closed loop transfer function being
now : H(z)=|(1-d)z~3I/I1-dz~ I, the corresponding
state model is :
Ό 1 0 0"
0 0 1 0
• [
0 0 d —κ 1-d
and, by identification, we get :
y(u ).u + [<}>(u )-d]x (k)-(1-d)e = 0
k k k 3 k REFERENCES
The implicit equation (7) is just slightly modified,
x (k) being 2T delayed : x (k) = s(k-2). The con-
1 3
troller which is obtained is an extension for bili-
near systems of the well-known Smith's controller.
Experimental results
All experiments are processed for an initial total
flow-rate Q = Q + Q = 5 5 £ / h and for a pH around 2.5.
0 A oB o
A reference corresponding to x =0.1 is equivalent
c Claude, D., Fliess, Μ., Isidori, A. (1983). Immer-
to a pH reference of 0.3. The sampling period is
T=15sec.
A first feedback yields to a linear system from a
bilinear one, and a second feedback realizes a P.I
control, allowing a cancellation of static errors.
Moreover, we master the response time by specifing
the overshoot magnitude D(%) and the first over-
shoot time NT. Fig. 4 shows the normalized concen-
tration evolution for two desired response times
(D=5%, N=20 or N=15). The static error is perfectly
cancelled. The obtained overshoot is greater than
the desired one, but the first overshoot time is
rather good. Fig. 5 shows the evolution of the nor-
malized concentration and the base flow-rate, with
respect to a greater step magnitude. The same re-
marks, as in Fig. 4, can be mentioned. The over-
shoot differences with respect to the theory are
due to the fact that, the tuning parameters (d, K,
KjO are calculated from the system model mich in-
volve some identification errors.
Bilinear delayed system
In Fig. 6, we implement, the "Smith non-linear al-
gorithm" (Eq. 11) on the process. The desired closed
loop response is a two sampling periods delayed
first order one. The result is rather satisfactory,
but the delay is too small (10% of the approximated
time constant) to point out a real improvement of
the closed-loop linearization with respect to the
non-delayed system.
CONCLUSION AND PROSPECTS
Several digital control methods have been proposed
for two classes of non-linear continuous processes:
linear system with time-varying parameters and bi-
G. Gilles et al.
linear systems. They have been applied on a neutra-
lization pilot reactor for the two classes of ope-
rating conditions and 9the results can be considered
as rather satisfactory.
, l
_ _
In the case of time varying working condition, it has
been shown that a state-affine control law allows to
keep a desirable closed-loop behaviour all over the
working volume. The extension of this method to a
multivariable approach of the pilot plant is under
investigation. In the case of the bilinear working
condition, it has been shown that a digital closed
loop linearization leads to a state feedback control
law which, if the model is exact, guarantees the
same dynamic behaviour inside the whole operating
range. In the practical case of modelling errors,
improvements have been pointed out by designing a
digital P.I. non-linear controller. In the case of
processes involving time delays, it has been shown
how the linearizing control law and the Smith con-
0 0 In order to improve the performances, and specially
the robustness, it would be interesting to use a re-
ference model following structure and to study the
static error cancellation by introducing a parallel
internal model.
1 0 Finally, theses works could be fruitfully extended
for systems presenting both a bilinear structure
° ] 2k
and a parametric variation in their operating condi-
tion. Again, the pilot reactor can be, in this case,
a very good testing device.
(11)
Alvarez Gallegos, J. (1982). Optimal control of a
class of discrete multivariable non-linear
systems. Application to a fermentation process,
Journal of Dynamic Systems, Measurement and
Control, vol. 104, pp. 212-217.
Cebuhar, W., Costanza, V. (1984). Non-linear control
of CSTR's, Chemical Engineering Science, vol.
39, 12, pp. 1715-1722.
sion directe et par bouclage d'un système non-
linéaire dans un linéaire, Compte-rendu à l'A-
cadémie des Sciences, série 1, t. 296, pp.237-
240.
Dadugineto, (1985). Identification de différents mo-
dèles discrets non-linéaires continus. Appli-
cation aux procédés industriels pilotes, Con-
férence internationale CNRS Automatique non-
linéaire, Paris.
Dang Van Mien, H., Normand-Cyrot, D. (1984). Non-
linear state-affine identification methods.
Application to electrical power plants, I.F.
A.C. symposium an Aut. Control on Power gene-
ration, Distribution and Protection, Pretoria,
Automatica, vol. 20, n° 2.
Dufour, J., Thomasset, D. (1984). Définition and
synthesis of state-affine control algorithms.
Application to vary-linear systems, IEEE
Conf. on Computers, Systems and signal Proces-
sy, Bangalore, India.
Fliess, M., Normand-Cyrot, D. (1980). Vers une appro-
che algébrique des systèmes non-linéaires en
temps discret, Analysis and optimization of
systems, Versailles, Lecture notes control
Inform. Sciences, n° 28, Springer-Verlag.
Gauthier, J-P., Bornard, G. (1980). Stabilization of
bilinear systems. Performance specification
and optimality, 4th INRIA international conf.
on Control and Optimization, Versailles,
France.
Gilles, G., Laggoune, Ν. (1985a). Linearizing control
of a class of non-linear continuous processes,
IFAC conf. CSTD'85, Beijing, China.
Gilles, G., Laggoune, Ν. (1985b). Digital control of
bilinear continuous processes. Application to
a chemical pilot plant, IFAC conf. on Digital
Computer Application to Process Control,
Vienna, Austria.
 Pilot Neutralization Chemical Reactor 133

Mohler, R. (1973). Bilinear control processes, Aca-
demic press.
Neyran, Β., Saade-Castro, J., Thomasset, D., Reynaud,
R. (1984). Modélisation dynamique non-li-
néaire d'un procédé chimique de neutrali-
sation, Congrès international S.E.E., Nice.
Neyran, B., Thomasset, D., Dufour, J. (1985). Di-
rect identification of a class of non-
linear systems. Application to a neutra-
lization process. IFAC Conf. CSTD'85,
Beijing, China.

f Cone

V : linear control computed at V » 4 1

#: linear control computed at V = 1 2 1

1000 200 0

Τ
Fig.3 : Comparison between linear and
state-affine control

o.i Η

STEP REFERENCE
MEASURED OUTPUT (N=20)
MEASURED OUTPUT (N=15)

600 "

Fig. 4. P-I. linearizing control : two different dynamics

134 G. Gilles et al

ΟΛΑ STEP REFERENCE

MEASURED OUTPUT

β
a)

60 0 »·

•/h
60

Fig. 5. a) Normalized concentration for greater

step magnitude
b) Control variable (base flow-rate)
evolution

Fig. 6. Closed loop linearization with delayed

response (smith non-linear controller)
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, U K 1986

SEMIBATCH REACTOR DYNAMICS AND

STATE ESTIMATION
H . Schuler and K. de Haas
BASF AG, D-6700 Ludwigshafen, FRG

Abstract. The dynamic behavior of semibatch reactors is characterized by a strongly

nonlinear, and especially an inherently nonstationary nature. These dynamic features
give rise to difficulties in reactor operation particulary with highly exothermic and
temperature-sensitive reaction . Typical operational problems are attributable to
these dynamic characteristics, and to insufficient information from the commonly
measured variables. This paper provides an outline of semibatch reactor dynamics,
and state estimation examples in industrial practice.

Keywords, chemical reactors; semibatch operation; modeling; dynamic behavior; state

estimation; Kalman filtering.

INTRODUCTION

The dynamic behavior of stirred tank reactors in
semibatch operation has been dealt with in only
a few papers in chemical reaction engineering
literature. This limited scientific interest is
in contrast to the great economic importance of
this reactor operation mode, which is widely used
for the production of polymers, fine chemicals
and pharmaceuticals. A large number of biotechno-
logical processes are also carried out in fed-
batch operation. The dynamic behaviour of semi-
batch reactors is characterized by a strongly non-
linear, and especially an inherently nonstationary
nature. These dynamic features give rise to dif-
ficulties in reactor operation particularly with
highly exothermic and temperature-sensitive reac-
tions. Operational problems during the start-up
and the feed period are attributable to the dyna-
mic behavior of the process, and to insufficient
actual information on the process state, acces-
sible from the measured signals (i.e. the tempe-
ratures and the feed flow rates).
normal and also during disturbed operation. State
estimation results obtained from industrial-scale
semibatch reactors will also be given.
DIFFERENT SEMIBATCH OPERATION MODES
A large number of chemicals are produced in reac-
tors operated in semibatch mode. Various versions
of this operation mode are commonly used: two
important examples are shown in Fig. 1.
In the first operation mode I, shown on the left
hand side of this Figure, one reactant Β is initi-
ally introduced into the reactor vessel. The other
reactant A is fed continuously from a storage tank
into the reactor untill the stoichiometric ratio
of the reactants is reached.

s t o r a eg t a n sk
To gain a better insight into semibatch proces- B*
ses, this paper will deal with the following
topics :
11J
The first part will describe dynamic features ty-
pical of isothermal and exothermic reactions
in semibatch reactors, for example retarded onset
of the reaction during start-up, and reaction
drop-out due to disturbances (i.e. poisoning) in
the chemical mechanism. These effects can lead to
an undesirable accumulation of the reactants,
followed by a temporary runaway of the reaction.
These phenomena will be discussed, using simula- Fig. 1. Two important semibatch operation modes
tion examples showing the most important effects.
Another important operation mode II is depicted on
The second part of the paper will provide exam- the right hand side of Fig. 1. In the initial
ples for on-line calculation of additional sig- operation phase, a certain amount of both reac-
nals characterizing the actual reactor state tants is fed into the reactor. The reaction tempe-
more distinctly then the available measurement rature is then raised by jacket heating, and by
signals (temperatures and flow rates). These state using the heat of reaction. In the subsequent feed
estimation techniques (i.e. Kalman filters) make period both reactants are fed simultaneously into
use of dynamic reactor models. They provide real- the reactor. This operation mode II is widely used
time information on the actual production rate, for the production of polymers: monomers and the
the conversion, and the composition of the reac- initiator are the reaction components in this
tion mixture. These calculated variables facili- special application.
tate the supervision of the reactor state during

135
136 H. Schuler and K. de Haas

The inherently nonstationary nature of these pro- Analyses of a similar elementary model have
cesses is apparent from the operation procedure been given by Hugo (1981), Hugo and Steinbach
of these semibatch operation modes: the dynamic (1982, 1985) and Steinbach (1985). The model (1-6)
behavior is interesting only during the start-up forms the basis of more sophisticated models in-
and the feed period. The reactor state has a typi- cluding complex reactions (see e.g. King and
cally transitional character during these periods Gilles, 1984, 1985),multiphase reactions, state
of limited length. An analysis of steady states equations of the different phases etc.
and stability is adequate only in special cases.
An example of such an analysis has been given by In the following some basic dynamic features of
Kogan and Mints (1979). semibatch reactors will be described using simu-
lation examples of the corresponding models.
In the following a short outline of elementary
semibatch reactor models will be given. These
models will be used to explain some dynamic pro- DYNAMIC BEHAVIOR OF SEMIBATCH REACTORS
blems of reactor operation.

Isothermal Behavior of Semibatch Reactors (Mode I)

SEMIBATCH REACTOR MODELS
A first impression of semibatch reactor dynamics
can be obtained by considering the isothermal case.
Models of semibatch reactors include the balances
Simulation results of this special case are
of the energy and the masses of species inside the
shown in Fig. 2. The depicted profiles of the mas-
reactor. Assuming a macroscopic reaction A+B
ses of reactants A and Β inside the reactor differ
(products), it is convenient to use the following
in the Damkohler number
material balances:
M. (7)
Β
Da =
dm
T m ; chem
chem
k exp (-E/RT)
m (0) (1a)
dtT FT T TO
relating the characteristic time of chemical reac-
dra, tion t hem to that of the feed operation tp. The
A
m.(0) (1b)
c
profile of the production rate of the chemical
dt~ FA RA' A AO
reaction is depicted as well.
dm,
Β
m m; 1.0 1.0-

dt~ " FB " R B

m (0)
B BO
(1c)
^\>\
ω -
" Da«1 /
The mass conversion rates ut
ut
a -
ε
M rV (2a, b) - / Ι Ο ^ - "
B JJKT"—•
ι" ι — —1 Ί 1 0.0- — ι — ι — ι — 1 1 ! Γ —
t time
include the production rate rV, and comprise the F
molar reaction rate r which is assumed to be of
second order in an elementary treatment
e pk X ( } CC
(3)
= o " RT A B

Using relations for the molar concentrations and

the volume
M
T Fig. 2. Isothermal behavior of semibatch
reactors in operation mode I.
A M V -B " M V ' V = -1 (4a, b, c)
A B
an elementary model is obtained for the isothermal The behavior of the production rate is easily un-
behavior of stirred tank reactors in semibatch derstood in the case of large Damkohler numbers,
operation. i.e. at slow feed rates, or at high reactivities.
In this case only a small amount of the reactants
In the case of exothermic reactions the energy fed in will accumulate inside the reactor. The pro-
balance of the reactor must be included in the duction rate coincides with the feed rate during
model : the feed period over a considerably long period.
C C ( TT) + C ( T )T Deviations from this quasistationary equilibrium
will be observed at the beginning and end of the
oT - ™FA P,FA FA " *FB P,FB FB "
h r V feed period: the production rate will increase from
initially zero to values in the range of the feed
(t - τ) + ( - * ) ; t(o ) = T
+ hA
c c R Q (5)
rate. The rate of this initial increase of reac-
tion depends strongly on the Damkohler number, and
therefore on the reactivity at the temperature
The heat capacity C , and the cooling surface A c
selected. The reaction is decelerated considerably
depend on the reactor charge:
at low (initial) temperatures. - At the end of the
feed period the production rate will slow down due
C = m c + W; Aç = Aq + A' ( V - V ) (6a, b)
p Q to the consumption of components initially present
inside the reactor.
The function A ^ ( V ) will depend on the geometric
shape of the reactor vessel.
Nonisothermal Behavior of Semibatch Reactors
Both operation modes depicted in Fig. 1 can be
(Mode I)
described by the elementary model. The two modes
differ in the initial conditions and in the feed
The complexity in the dynamic behavior of semi-
rates.
batch reactors increases considerably when taking
into account the Arrhenius-type dependence of the
 Semibatch Reactor Dynamics and State Estimation
reaction rate, as well as the energy release du-
ring an exothermic reaction. Dependent on the para-
meters of the specific case different characte-
ristic types of reaction courses can occur. Two
basic patterns are depicted in Fig. 3.
0.30 -r
Fig. 3. Dynamic behavior of an exothermic reac-
tion in semibatch operation mode I (simu-
lation result according to Fig. 1 ) .
The two patterns differ in the initial
condition of the reaction temperature.
The Figure shows one reaction course similar to
the isothermal behavior of Fig. 2. The reactants
are introduced hot into the reactor. The reaction
rate reaches high levels at the very beginning of
this starting phase. The components A will react
almost completely immediately after introduction
into the reactor. The reaction proceeds subsequen-
tly in a controlled manner.
A completely different behavior is observed when
the operation is start ed with cold reactants inside
the reactor. The other course of Fig. 3 shows a
distinct retardation ο f the reaction onset. The
reactant A accumulates in the cold reaction mix-
ture. The accumulated reactant mass reaches high
concentrations and wil 1 later react at increased
reaction rates. The re action temperature consider-
ably exceeeds the subs equent quasi-stationary level.
The retarded onset of the reaction in the starting
phase is influenced by most of the process para-
meters. Fig. 4 shows simulated reaction courses
obtained at different cooling temperatures.
. . . . ,. ,
Fig. 4. Dynamic behavior of exothermic reac-
tions in semibatch operation mode I
at different cooling temperatures.
At low cooling jacket temperatures, the cooling
effect is so strong that the onset of the reaction
is prevented. When very high cooling temperatures
are employed, the reaction will proceed regularly
137
reaching high finite conversions. In a transition
range of the cooling temperature the retardation
of the reaction onset described above can be ob-
served. In this parameter range an accumulation
of reactants will occur, which subsequently leads
to difficulties in reactor operation.
All simulation results clearly show that the tem-
perature dependence of the reaction rate, and the
exothermicity of the reaction, will cause prob-
lems in the operation of semibatch processes. The
retarded onset of the reaction leads to an accu-
mulation of reactants. This reactive mass can,
after some time, only be converted in a way which
is difficult to control. This temporary runaway
will affect product quality, and can - in extreme
cases - make safe reactor operation a problem.
The starting phase is normally followed by a
smooth-running feed period. This phase is cha-
racterized by a quasistationary equilibrium of the
feed and the reaction rate. This equilibrium is
inversely affected only by specific disturbances,
which are dealt with in the next section.
Operational Problems During the Feed Period
(Mode II)
The smooth operation of semibatch reactors during
the feed period can be affected by disturbances in
the chemical reaction. Such disturbances will be
caused by an inhibition of reaction steps by unde-
sired components in the feed, by catalyst poiso-
ning, or by consumption of the initiator induced
by a wrong temperature profile, etc. As a conse-
quence of such a reaction drop-out, the reactants
accumulate inside the reactor, having the effects
on product quality and on process operability des-
cribed above. Fig. 5 shows simulation results of a
semibatch reactor operation in mode II according
to Fig. 1. In the Figure the normal operation of a
temperature-controlled reactor is contrasted with
the reaction course when the initiator of a poly-
merization reaction has been damaged. This distur-
bance will cause a reaction drop-out, and a
drastic accumulation of reactants as well,
n o r m la o p e r a t ni o
PRODUCTION RATE
2 3
t i me
Fig. 5. Comparison of normal operation of
semibatch reactors operated in mode II,
and of disturbed operation caused by
temporary damage of an initiator.
138 H. Schuler and K. de Haas

When the initiator efficiency is reestablished,
the accumulated reactants are consumed in an acce-
lerating reaction. In this special case the strong
reaction is controlled by the cooling system.
The example shows that a disturbance in the
chemical mechanism will affect the operability
of semibatch reactors in the same way as the re-
tarded onset of the reaction during start-up.
It is difficult to recognize when a reactant is
starting to accumulate just from the temperature
signals. This undesired behavior can be seen more
distinctly from the production rate and the con-
tent of component A in the reaction mixture. These
variables are not available as sensor signals.
T , T , T
c F 1 p2 and the flow rates , V^,
v"steam are fed to the filter algorithm which pro-
cesses the data together with the balance equa-
tions of the reactor. The filter provides the most
probable actual value of the production rate. This
estimated signal is supplied to the plant operator.
In Fig. 6 this signal is depicted, as well: it
clearly shows the onset of the reaction. The
(calculated) production rate increases from initi-
ally zero (no reaction) to values in the range of
the feed rate. The production rate varies during
the feed period only to a small degree because the
reaction runs smootly. It should be noted that
this real-time calculation is similar to the simu-
lation result shown in Fig. 5.

STATE ESTIMATION

The detection of undesired behavior during semi-

batch reactor operation is made difficult because
only insufficient information on the actual reac-
tor state is available from the measured varia-
bles: the state vector contains unmeasured varia-
bles characterizing the process state in a signifi-
cant way. Examples of those variables are the pro-
duction rate and the composition of the reaction
mixture. The operation of semibatch processes would
be improved if these variables were available real-
time during process operation. These characteristic
state variables can be reconstructed from the
available measurement signals using well-known
state estimation techniques (i.e. Kalman filters).

The above remarks on dynamic behavior show the

excellent role played by the production rate in
the diagnosis of semibatch reactor operation. The
production rate clearly shows the onset of the re- time
action, as well as its drop-out due to disturb- ICQ
ances during the feed period. By relating the pro-
duction rate to the feed rate, the mass of reac-
tants inside the reactor, and subsequently their
accumulation can be derived. The production rate
is therefore the key variable for the diagnosis
of reactor operation.

Most papers on chemical reactor state estimation

contain expressions for kinetic rate laws. These β 50. 01. 51, 7$ 52.
rate equations are unreliable for a lot of impor- time

tant reactions, and the modeling effort is very

high for complex reactions. For this reason, an
estimation based on detailed rate laws seems to
be impracticable especially for multi-product
plants (Schuler, 1986).
From equ. (5) it can be seen that the production
rate is observable from the temperature using ex-
clusively the energy balance. This fact can be
used to design a state estimator providing the
actual production rate of the reaction. The mass
balances are used as well in order to calculate
the charge-dependent heat capacity and cooling sur-
face, as well as the composition of the reaction
mixture.
In the following some examples will be given in
which this state estimator has been applied to
production-scale reactors operated in mode II
according to Fig. 1.
Fig. 6 shows profiles of the reactor and jacket
3
temperatures during the operation of an industrial-
scale semibatch reactor of about 10 m in volume.
In the initial process operation phase the reactor
temperature is raised by jacket heating and steam
injection. After some time the plant operator be-
gins to feed in the reactants. The reactor tempe-
rature is controlled by adjusting the jacket tempe-
rature. The actual signals of the temperatures T ,
R
Fig. 6. Profiles of the (measured) reactor- and
jacket temperature, and the (calculated)
production rate and conversion.
Fig. 6 shows the (calculated) actual conversion,
and the actual reactant mass as well. These sig-
nals provide an impression of the composition of
the reaction mixture without using analytical
instrumentation.
Fig. 7 shows two other semibatch cycles charac-
terized by a different behavior during the reac-
tion onset. On the left hand side, the reaction
runs smoothly, whereas on the right hand side, a
retarded onset of the reaction is observed, fol-
lowed by a temporary runaway. The calculated sig-
nals of the production rate clearly indicate the
different behavior already at an early stage. Re-
cognition of the actual reactor state is facili-
tated significantly when this signal is provided
on-line during reactor operation.
 Semibatch Reactor Dynamics and State Estimation

NOTATION

A cooling surface
reactor temperature Tr C
n heat capacity of reactor content
Da Damkohler number
E activation energy
M molar mass
cooling temperature l
c R gas constant
T temperature
V volume
V volumetric flow rate
1 W heat capacity of the empty reactor
concentration
specific heat capacity
heatR transfer coefficient

• // ι
f (-Ah-
k
reaction enthalpy
preexponential factor
J J
production rate n
R mass
mass flow rate
production rate
reaction rate
time
conversion
density

Fig. 7. Two state estimation examples showing

different behavior of reaction onset.
Subscripts

C O N C L U S I SO N , A reactants
C cooling jacket
When processing highly exothermic and temperature- F feed
sensitive reactions in semibatch reactors there 0 initial condition
is observed sometimes an undesired dynamic R reaction
behavior of the process. These operational pro- Τ total
blems can often be explained by an accumulation
of the reactants inside the reactor due to un-
balanced feed and reaction rates. This nonequi-
librium is attributable to the dynamic characte-
ristics of semibatch processes, or to disturbances
affecting the mechanism of the chemical reaction.
Undesired accumulations should be prevented by
taking the dynamics into account in the operation
instructions, and by monitoring the production
rate and reactant content on-line during the
operation. Both strategies are dealt with in this
paper.
140 H. Schuler and K. de Haas

LITERATURE

A v r u j , A . a n d J . A . Maymo ( 1 9 7 0 ) . C h a r t s f o r s e m i - Parulekar, S.J. and H.C. Lim ( 1 9 8 5 ) . Modeling,

batch r e a c t o r d e s i g n . B r i t i s h Chemical Engi-
neering, 15, 776-777.
B a n d e r m a n n , F77 R. D u n s i n g a n d K . - F . E l g e r t ( 1 9 7 4 ) .
Optimierende Reaktionsfiihrung lebender Poly-
merer: Vorausberechnung von Molmassenvertei-
lungen b e i Reaktionssteuerung durch I n i t i a t o r -
Dosierung. Chemie-Ingenieur-Technik, 46, 615.
MS 1 0 5 / 7 4 .
B a n d e r m a n n , F . a n d R. D u n s i n g ( 1 9 7 4 ) . Optimierende
R e a k t i o n s f iihrung l e b e n d e r P o l y m e r e r : V o r a u s b e -
rechnung optimaler R e a k t i o n s v e r l a u f e zur Dar-
optimization and control of semi-batch bio-
reactors. Advances in Biochem. Eng./Biotech. ,
32, 2 0 7 - 2 5 8 .
Penlidis, Α., J. F. Mac Gregor and A. E. Hamielec
( 1 9 8 5 ) . Dynamic modeling of emulsion polymeri-
zation reactors. AIChE Journal, 31_, 8 8 1 - 8 8 9 .
Schuler, H. ( 1 9 8 6 ) . Modellgestiitzte meBtechnische
Uberwachung chemischer Reaktoren. Technisches
Messen. in press.
Staniskis, J. and D. Levisauskas ( 1 9 8 4 ) . An adap-
tive control algorithm for fed-batch culture.

_
s t e l l u n g von Polymeren mit d e f i n i e r t e n Mol- Biotechnology and Bioengineering, 2 6 , 4 1 9 -
massenverteilungen. Chemie-Ingenieur-Technik,
4 6 , 6 1 7 . MS 1 0 6 / 7 4 . Steinbach, J. ( 1 9 8 5 ) . Untersuchung zur thermischen
C o r r i p i o , A.B. ( 1 9 6 7 ) . Use of analog s i m u l a t i o n Sicherheit des indirekt gekiihlten Semibatch-
for semibatch reactor control systems. In: Reaktors. Dissertation D 83 TU Berlin.
S y s t e m s a n d P r o c e s s C o n t r o l , a CEP t e c h n i c a l Tracht, R. and R. Grunberg ( 1 9 8 1 ) . Modellbildung
m a n u a l , p u b l i s h e d by AIChE, 1 9 5 - 1 9 9 . und optimale Abtastregelung fiir einen halb-
E i g e n b e r g e r , G. a n d H. S c h u l e r ( 1 9 8 6 ) . R e a k t o r - kontinuierlich betriebenen Riihrkesselreaktor.
s t a b i l i t a t und s i c h e r e R e a k t i o n s f iihrung. 2 . Steirisches Seminar iiber Regelungstechnik
Chemie-Ingenieur-Technik, 58, in press. und Prozefiautomatisierung, SchloB Retzhof bei
Hugo" Pi ( 1 9 8 1 ) . B e r e c h n u n g i s o t h e r m e r S e m i b a t c h - Graz, 179 - 1 8 3 .
R e a k t o r e n . C h e m i e - I n g e n i e u r - T e c h n i k , 5—
3, Tsoukas, A. and M. Tirrell ( 1 9 8 2 ) . Multiobjective
107 - 1 0 9 . dynamic optimization of semibatch copolymeri-
Hugo, P. and J . S t e i n b a c h ( 1 9 8 2 ) . S i m u l a t i o n o f zation reactors. Chemical Engineering Science,
exothermic s e m i b a t c h - p r o c e s s e s under d i v e r s e 32, 1 7 8 5 - 1 7 9 5 .
c o o l i n g t a c t i c s . 1 5 t h European Symposium Young, M.J. ( 1 9 8 0 ) . Semi-batch reactions offer
CHEMCOMP ( " C h e m i c a l P r o c e s s A n a l y s i s a n d optimal yields. Processing, 2 7 - 3 3 .
Design using Computers") Antwerpen.
Hugo, P. and J . S t e i n b a c h ( 1 9 8 5 ) . Praxisorientier-
te Darstellung der thermischen S i c h e r h e i t s -
g r e n z e n fiir d e n i n d i r e k t g e k i i h l t e n S e m i b a t c h -
Reaktor. Chemie-Ingenieur-Technik, 57, 780-782.
Karimi, I . A . and G.V. R e k l a i t i s ( 1 9 8 5 ) . D e t e r m i -
nistic variability analysis for intermediate
storage in noncontinuous p r o c e s s e s . Part I:
Allowable c o n d i t i o n s . Part I I : Storage s i z i n g
f o r s e r i a l s y s t e m s . AIChE J o u r n a l , 3 1 , 1 5 1 6 -
1527, 1528-1537.
K i n g , R. a n d E . D . G i l l e s ( 1 9 8 4 ) . D y n a m i c s a n d
runaway of c h e m i c a l r e a c t o r s . Preprints
AIChE's 1984 Annual M e e t i n g , San F r a n c i s c o
1 9 8 4 , S e s s i o n N o . 1 6 1 , P a p e r 161 A .
K i n g , R. a n d E . D . G i l l e s ( 1 9 8 5 ) . M o d e r n e M e t h o d e n
d e r M e s s - und R e g e l u n g s t e c h n i k . B e i t r a g z u r
S i c h e r h e i t c h e m i s c h e r A n l a g e n . DECHEMA-Mono-
g r a p h i e n , Band 1 0 0 . V e r l a g C h e m i e , 1 4 9 - 1 7 1 .
K l e i n e r t , W. ( 1 9 8 3 ) . B e t r i e b s v e r h a l t e n v o n s t a r k
e x o t h e r m e n U m s e t z u n g e n im S e m i b a t c h - B e t r i e b .
D i s s e r t a t i o n D 8 3 TU B e r l i n .
K o g a n , G . V . a n d R.M. M i n t s ( 1 9 7 9 ) . M a t h e m a t i c a l
model of a s e m i c o n t i n u o u s r e a c t o r f o r c a r r y i n g
out two c o n s e c u t i v e f i r s t - o r d e r exothermic
r e a c t i o n s . Number, c h a r a c t e r , and s t a b i l i t y o f
e q u i l i b r u m s t a t e s . T h e o r . F o u n d . Chem. E n g . ,
J_3, 3 4 - 4 1 .
K o g a n , G . V . a n d R.M. M i n t s ( 1 9 7 9 ) . I n v e s t i g a t i o n o f
a mathematical model of a s e m i c o n t i n u o u s r e a c -
t o r during t h e o c c u r r a n c e of two c o n s e c u t i v e
exothermic f i r s t - o r d e r r e a c t i o n s . Phase por-
t r a i t s a n d l i m i t c y c l e s . T h e o r . F o u n d . Chem.
E n g . , U, 324-328.
L e e , H . L . a n d B . O . Yau ( 1 9 8 1 ) . An e x p e r i m e n t a l
reactor for kinetic studies: Continuously-fed
batch r e a c t o r . Chemical Engineering S c i e n c e ,
36, 483-488.
M a k g a w i n a t a , T . , M . S . E l - A s s e r , A. K l e i n a n d J . W .
Vanderhoff (1984). K i n e t i c s t u d i e s of semi-
batch emulsion co-polymerization of 80:20
vinylacetate-butyl acrylate. J. Dispersion
S c i e n c e and T e c h n o l o g y , 5, 3 0 1 - 3 2 2 .
M i n , K.W. a n d H . I . G o s t i n ( 1 9 7 9 ) . S i m u l a t i o n o f
semibatch emulsion polymerization reactors for
p o l y v i n y l c h l o r i d e (PVC) s y s t e m . I n d . E n g .
Chem. P r o d . R e s . D e v . , j_8, 2 7 2 - 2 7 8 .
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

CONTINUOUS ESTIMATION OF THE CHAIN

LENGTH DISTRIBUTION IN A
POLYMERIZATION REACTOR: KALMAN
FILTERING CONSIDERING GEL-PERMEATION
CHROMATOGRAPHIC MEASUREMENTS
S. Ar. Papadopoulou
Institut fur Systemdynamik una Regelungstechnik, Universitàt Stuttgart,
Pfaffenwaldring 9, D-7000 Stuttgart 80, FRG

A b s t r a c t . A Kalman f i l t e r c o m p o s e d o f t w o d e c o u p l e d f i l t e r s and c o n s i d e -
ring Gel-Permeation chromatographic,(GPC), measurements for the con-
t i n u o u s e s t i m a t i o n of the chain l e n g t h d i s t r i b u t i o n , (CLD), i n t h e c a s e
of f r e e r a d i c a l p o l y m e r i z a t i o n of styrene i s d i s c u s s e d . Although the
e x i s t e n c e o f a t i m e d e l a y , d u e t o t h e GPC a n a l y s i s , t h e c o n s i d e r a t i o n o f
t h i s measurement in t h e f i l t e r a l g o r i t h m produces t h e o b s e r v a b i l i t y of
t h e s t a t e s c h a r a c t e r i z i n g t h e CLD a n d p r o v i d e s c o n t i n u o u s a c c e s s t o t h e
CLD a n d i t s c h a r a c t e r i s t i c v a l u e s .

Keywords. Gel-Permeation Chromatography- Nonlinear filtering; Polymeri-

zation reactors; Quality control; State estimation.

INTRODUCTION f o r t h e e s t i m a t i o n o f t h e CLD, a n e w e s t i -
mator w i l l be t r e a t e d c o n s i d e r i n g sporadic
During the course of a p o l y m e r i z a t i o n r e a c - t i m e l a g g e d GPC m e a s u r e m e n t s o f t h e CLD.
t i o n , a large q u a n t i t y of polymer chains T h e c o n s i d e r a t i o n o f t h e GPC m e a s u r e m e n t s
a r e i n i t i a t e d , grow and a r e t e r m i n a t e d . i n t h e f i l t e r h a s two a d v a n t a g e s s i m u l t a n e -
The number and l e n g t h o f t h e p o l y m e r i c o u s l y . On t h e o n e h a n d i t p r o d u c e s t h e o b -
c h a i n s formed d u r i n g t h e r e a c t i o n , v a r y s e r v a b i l i t y of the s t a t e s characterizing
w i t h t h e r e a c t i o n mechanism and t h e r e a c - t h e CLD, w h i l e o n t h e o t h e r p r o v i d i n g t h e
t i o n c o n d i t i o n s e m p l o y e d . The d i s t r i b u t i o n p o s s i b i l i t y f o r e s t i m a t i o n o f t h e CLD a c -
of the polymer c h a i n s over t h e i r c h a i n tually during the operation.
l e n g t h , d e t e r m i n e s many i m p o r t a n t t e c h n o l o -
g i c a l p r o p e r t i e s of the polymeric product.
If t h i s f a c t i s t o be taken i n t o a c c o u n t , MODELLING OF FREE RADICAL
t h e c h a i n l e n g t h d i s t r i b u t i o n (CLD) , w h e r e POLYMERIZATION OF STYRENE
known, f a c i l i t a t e s s i g n i f i c a n t l y t h e c h a -
r a c t e r i z a t i o n and t h e c o n t r o l o f t h e q u a - The k i n e t i c s o f t h e f r e e r a d i c a l p o l y m e r i -
l i t y of t h e p r o d u c e d p o l y m e r . z a t i o n o f s t y r e n e (FRPS) i n s o l u t i o n a r e
w e l l known and t h u s o n e c a n t a k e m o d e l s
A t t i m e s t h e CLD c a n b e p r e d i c t e d a c c u r a - from t h e l i t e r a t u r e and a d a p t them t o t h e
t e l y from s t a t i s t i c a l c o n s i d e r a t i o n s , how- s p e c i f i c c a s e of a s u i t a b l e laboratory
ever, at other times accurate a priori pre- p l a n t . T h e FRPS i s r e p r e s e n t e d i n t h e f o l -
d i c t i o n s are not p o s s i b l e . Despite the lowing chemical equations (Henrici-Olivè,
necessary time-delay, in either case Gel- O l i v e , 1969; Odian, 1981):
P e r m e a t i o n C h r o m a t o g r a p h y (GPC) c a n b e
used as the b e s t a n a l y t i c a l procedure to Initiator dissociation-Chain start:
o b t a i n e x p e r i m e n t a l i n f o r m a t i o n on b o t h
CLD a n d m o l e c u l a r a v e r a g e s o f t h e p o l y m e r I —-—+ 2-1"+ N t . (1)
formed. 2
Rp
Model-aided measurement methods (Kalman- I%M — ( 2 )
F i l t e r ) o f f e r t h e p o s s i b i l i t y of r e c o n -
s t r u c t i n g i n t e r e s t i n g but d i f f i c u l t acces- Chain propagation: R
sible process states, (Gelb,1974; G i l l e s ,
1 9 7 9 ) , from e a s i l y a v a i l a b l e measurement * p
v a r i a b l e s . These methods have in a few
cases been applied to polymerization r e a c -
t i o n engineering (Adams,Schooley, 1969; Jo, C h a i n t r a n s f e r , t o monomer, t o s o l v e n t and
B a n k o f f , 1 976 ; S c h u l e r , 1 9 8 0 ; Kiparissides to chemical modifier:
and o t h e r s , 1 9 8 0 , 1 9 8 1 ) . However, there R
seems t o be no e x p e r i m e n t a l work p u b l i s h e d , P? + χ — > Ρ ^ + Ρ ; x = M,S,Rx (4)
±
i n w h i c h t h e s t a t e s d e t e r m i n i n g t h e CLD
are estimated during the operation. A Kal- Chain termination by c o mR
bination:
man f i l t e r w a s d e v e l o p e d b y S c h u l e r a n d
Zhang ( 1 9 8 5 ) f o r an o n - l i n e e s t i m a t i o n o f ft ft t
P. + P. . > P. . . (5)
t h e s t a t e v e c t o r , however one p a r t of this
vector i s not observable.

In t h e f o l l o w i n g , i n order t o overcome t h e
above mentioned d e f i c i e n c i e s of the a p p l i -
c a t i o n s of model aided measurement methods

141
142 S. Ar. Papadopoulou

Mass and heat balances results_of the average molecular weights

(MWs), Mn, Mw after the addition of Di-P-
The following equations are obtained for Tolyl Disulfide in the reaction mixture at
the mole numbers of the reactants and pro- 30 min, used here as a chemical modifier.
ducts with the assumption of quasi statio- The assumption of different chain transfer
nary balances of the radical species: constants in the model influences the
R V courses of the average MWs and therefore
dn(I) can cause divergence from the measured
(6)
"at " " " d « data, but has no influence upon the esti-
mated values of monomer and temperature.
^ - = - ( 2 f R 2 (7)
d W
+ V) R
V
t M
An explanation of the above considerations
dn(S) (8) offers the analysis of the local observ-
•R..C-V
"cE *ts ability of the state vector from the avail-
dn(RX) able measurements. The results show that
(9) the rank of the observability matrix is
dt
three only taking into consideration the
dn(y )
ο
[ ( +R R + ]
+ V measurements of temperature, (Τ), and/or
(10) refractive index (Rl). Therefore the sta-
dt tM tS W K '
tes characterizing the CLD are not observ-
dn(y-I) R + R+ I R+ -Ι able from these measurements.
~d t = I < tM tS W t) -V- -n=5) (H)
The above coniderations make obvious the
dn(y ) necessity for the development of a new
2
estimator taking into account experimental
information (GPC) about the CLD and its
(12) characteristic variables.

dn(P )
± SENSOR MODEL AND
+ ^ ^(i-D-d-ai^a^h-V
c τ τ ) α+ ( τ
I = ν%( κ- ρ· ρ\· ν
•r(M) -V
The reaction rates and other terms inclu-
ded in the equations are given in Schuler,
Zhang, 1985 and Schuler, Papadopoulou,1986,
in press. Additionaly a rate equation for
a chain transfer to a chemical modifier is
used :
R = k exp (-E /RT) · [RX] -V
RX RX KX
The laboratory plant used for the purposes
of this work is described in the same
papers.
SYSTEM STRUCTURE
From the signal flow diagram shown in Fig.1
it is obvious, that the system is composed
of two subsystems in a serial structure.
However the variables of the second sub-
system have no feed back influence on the
dynamics of the first one and especially
upon easily available measurement variab-
les such as the temperature or the monomer
concentration. Consequently possible di-
sturbances and model errors cannot be cor-
rected in the second part of the system,if
only these measurements are used in the
filter to estimate the state variables of
the system.
The presence of transfer reactions, for ex-
ample, can cause such errors. These reac-
tions have negligible influence upon the
conversion (Henrici-Olivê, Olive, 1969).
However, significant errors can result from
an inaccurate consideration of these
effects in the model.
Figure 2 shows estimated and experimental
MEASUREMENT EQUATIONS
(13)
When measuring the CLD with the help of
the GPC, the logarithm of the MW of each
) Δ +} ( polymeric chain can be obtained from the
- % ' calibration curve, which has been evalua-
(14) ted with polystyrene standards and fitted
with a third order polynomial of the re-
tention volume. The raw data resulted from
the GPC analysis is transferred from the
integrator to the process computer through
an interface. After the conversion and
treatment of the measured data, (Pickett
and others, 1966; Yau, Flemming, 1968),
the differential weight fraction distribu-
(15) tion over the chain length is obtained
according to the following equation:
k
W(i) = X (AREA. ) ;
FLCW*FR(i)*AEEA./I 1
i=1 Î,I (16)
FR(i) = 0.434/[i>MAV )'MC +2^2 V +3^C ^)];
R 1 R 3
i = chain length
The average MWs are obtained in a similar
way. In this paper two classes of measure-
ment equations will be treated, making
possible the access to the state vector of
the second subsystem. In the first case
the measurements of the number and weight
average MWs will be used. While in the
second case a specific number of weight
fractions of polymer chains are used as a
proportion of the weight of the polymer as
a whole, according to the measured CLD.
In both cases the rank of the observability
matrix concerning the second subsystem is
equal to the number of the measurement
equations.
The measurement equations used for the
second subsystem are:
+ r
Y = MW + r ; y = nT^—p (17)
2j 2 j 2j 2j
i*n(P.)
 Chain Length Distribution in a Polymerization Reactor 143

to: time delay between removal of sample

j = 1,2 ; i = 1,2,3,... (t^ 2+1) and the time of availability of
results of GPC analysis ( t ^ ) .
R +
The measurement equations used for the
first subsystem are:
Error Covariance:
y = Τ + r ; y, = [M] + r (19)
n n 2 12 (29)
The equations concerning the second sub-
system are special forms of the general P (1 +1| K ) P (30)
2 2 2 2( T K_ + A T 2, - "ο'
T „ )
vectorial measurement equation, taking in-
to account the necessary time delay requi-
red for analysis: P (31)
2( K 2+ I | K 2) = P
2( T K A T
22) +
Y_ = h (x (t-t ) ) + r_ ; t = time delay.
2 2 2 0 2 Q
Measurement equations:
L

2
FILTER STRUCTURE

According to the system structure the

1 1 L
5, (Κ,+1) = h (x ( t + A t ^ )
1 1 R (32)

state equation can be divided into two (33)

2
different parts, describing the dynamics
of the two subsystems.

Σ dn+H
(34)
Z(2
K 1)
2 + = h (x (t +At ))
2 (2 K +2 2
The modelling and measured errors are as-
2 -2 K- ^" - κγo'
T AT
+ ;1
T )}

sumed to be uncorrelated in time. The fol-

Update of the measurement
lowing algorithm (extended Kalman-Filter) χ( χ )
can be used in order to realise the state
1 (35)
estimation in the case of FRPS, using a Ί (Κ-,+Ί) = Ζ (^+1) - î (^+1)
Ί 1
continuous process model and discrete
γ ( Κ + 1 ) = γ ( Κ + 1 ) - 2 (1 +1) ; (36)
measurements of Τ and Rl plus discrete 2 2 2 2 2 2
time lagged GPC measurements:
γ_ (Κ +1 ) : GPC measurement obtained at
System Equations: 2 2 time t
K « corresponding
sample 2
to a
taken at time t
d x^ +1
1
F_ (x ,u ,t) ; x (0) (20) S (K.,+1) H (K.,+1) •P (K-,+1 I K ^ •H (K +1)+R
"dt" 1 1 1 1 -1o 1 T 1 1 1 1
(37)
vΤ = (n(I),n(M),T,n(s) ,n(RX))

3h
S ( 13 h+11 ) = H (1 +1)·Ρ (1
L
t)m 0) : ( 22 2 2 2 2 22+1IK 2)·Η 2(1 2+1)+R 2
9x, 1
x = (η(μ.),η(Ρ ))
^
Π2
Γ = * ^'±2' '
2 ± *2< = *2o (38)

State estimate propagation: Gain matrix:

1

(22)
K (K^ +1 ) = P (Κ,,+1 |K.,) •Η^Κ.,+Ι) - S ~ (Κ,,+1)
1 1
, )t 1 (39)
à-j =' £1 Φι ; x
1(t R )=x-, ( K J K ^
1 K (K +D
2 2 = P ( i + 1 | κ ) - H ( i + D - s ~ (1 +1)
2 2 2 2 2 2
(40)

*
2 - -2
=(
Ι γ^1^>^2'
^Γ^Ι
t]
'
* ^2K^22l2
1 - £2
Error covariance propagation:
(;
l ll

( KK
1
+;Q ( T)
^1 K
T( =) K(
( KK )
1
K(
)23) State

((24)
)2 5
1
estimate update:

x (κ,,+1 | κ + ΐ ) = χ
1 1 (Κ,,+1 1
| K ) + K
1 (Κ,,+1) · χ 1( K ^ +1 )
(41)

t2 - Q 2 £ 2 ν " £ 2 2I 2 >
3F,

= [i^ ^2 = "°
x (K +1 IK + 1 ) = x ( K + 1 ) κ ) + κ ( K + 1) · Y ( K + 1 )
; 2 2 2
£^^+1)^+1)=^ (1^+1 |K,,)
2 2 2
Κ (Κ +1) ·χ
2 2 2 2
(Ι^-ΗΊ) (43)
Ί Ί Ί (42)
*r Error covariance matrix^
A Priori estimates before the measure-
ment t
-x+1 *
P (K +1
2 2 IK+1 ) =P 2(K+11K)
2 2 2 - K (K +1 ) ·χ (K +1 )
2 2 2 2 (44)

State Estimates: Figure 3 shows the signal flow diagram of

the described filter algorithm. The reac-
(26) tion of the second filter is shown in Fig.4
considering the time delay for the GPC
χ (ΐ ι|κ )
2 2+ 2 = x (t +At -t )
2 K 2 2 o (27) analysis. According to the above described
filter algorithm it is obvious that the
X (K +1(K ) = x (t (28) system structure offers the possibility of
0 0 0 2 K

—2 2
a development of an estimator, composed of
two decoupled filters corresponding to
144 S. Ar. Papadopoulou

e a c h s u b s y s t e m . The f i r s t f i l t e r provides of D I - P - T o l y l D i s u l f i d e was added t o t h e

the reconstruction of the observable sta- reaction mixture as chemical modifier a t
t e s of the f i r s t subsystem, ( n ( I ) , n(M),T) 31 m i n u t e s . F i g u r e 7 a s h o w s t h e e s t i m a t e s
t a k e n from t h e measurements o f Τ a n d / o r R l . o f t h e CLD, u s i n g s a m p l e s a t 3 5 , 5 5 , 7 5 m i n
The s e c o n d f i l t e r p r o v i d e s t h e r e c o n s t r u c - of t h e r e a c t i o n r u n . A f t e r t h e f i r s t u p -
t i o n o f t h e s t a t e s c h a r a c t e r i z i n g t h e CLD, d a t e t h e e s t i m a t e o f t h e CLD s h o w s a b e t t e r
t a k e n f r o m t h e GPC m e a s u r e m e n t a s f o l l o - convergence t o t h e measured d i s t r i b u t i o n
w i n g : a ) T h e m o m e n t s o f t h e CLD c a n b e than t h a t obtained without the second f i l -
e s t i m a t e d from t h e measurements o f t h e ter (Fig. 7b).
average molecular weights*
b) A s p e c i f i c n u m b e r o f p o l y m e r c h a i n s
( n ( P i ) ) , (and t h e r e f o r e t h e w h o l e CLD), CONCLUSIONS
can be e s t i m a t e d from t h e measurements o f
the weight fractions, (WFs), o f t h e p o l y - In t h i s paper a f i l t e r a l g o r i t h m was d e -
mer c h a i n s a s a p r o p o r t i o n o f t h e w e i g h t s c r i b e d and implemented f o r t h e c o n t i n u o u s
of t h e polymer a s a whole. e s t i m a t i o n o f t h e CLD u s i n g s p o r a d i c GPC
m e a s u r e m e n t s , i n t h e c a s e o f FRPS. I t w a s
T o o b t a i n a n e s t i m a t i o n o f t h e CLD t h e shown t h a t t h e c o n s i d e r a t i o n of t h e
following steps are necessary: GPC m e a s u r e m e n t i n t h e d e s c r i b e d filter
a) One c h o o s e s , f o r e x a m p l e , t e n c h a i n has the following advantages:
l e n g t h s a n d t h e c o r r e s p o n d i n g WFs o f t h e a) I t p r o d u c e s t h e o b s e r v a b i l i t y o f t h e
measured d i s t r i b u t i o n . s t a t e s c h a r a c t e r i z i n g t h e CLD.
b ) T h e s e WFs a r e u s e d t o u p d a t e t h e m o l e b) A l l o w s t h e c o n t i n u o u s e s t i m a t i o n o f t h e
numbers of t h e c o r r e s p o n d i n g p o l y m e r i c CLD.
chains : c) P r o v i d e s t h e p o s s i b i l i t y o f a p e r i o d i c a l
a d a p t a t i o n o f t h e CLD e s t i m a t e t o t h e r e a -
n(P+) = n ( P - ) + K* (W - W) (45) lity.
d) R e d u c e s t h e i n f l u e n c e o f p o s s i b l e i n -
c) A p p r o x i m a t i o n o f t h e p a r t s o f t h e d i s t - a c c u r a c i e s o f t h e m o d e l o n t h e CLD.
r i b u t i o n function between the estimated e ) New p h e n o m e n a a s t h e c h a i n t r a n s f e r c a n
p o i n t s on t h e b a s i s of a s u i t a b l e p o l y n o - be o b s e r v e d .
mial function, ( G i l l e s , Knoepp, 1 9 6 7 ) , The e x i s t e n c e o f a t i m e d e l a y i n t h e m e a -
having t h e f o l l o w i n g form: surement s y s t e m and t h e c o r r e s p o n d i n g
Ο (-I/A) jumps o f t h e e s t i m a t e s p r o v i d e some d i f f i -
W(i) = F(i) = (A +A * i + A * i ) *e ; (46) c u l t i e s . Also a d e t a i l e d model, a process
Q 1 2 c o m p u t e r a n d a GPC a r e n e c e s s a r y . H o w e v e r
A = constant t h e i n t e r p r e t a t i o n o f a Kalman f i l t e r c o n -
Three e s t i m a t e d chain l e n g t h s a r e used s i d e r i n g GPC m e a s u r e m e n t s c a n b e t h e b a s i s
each time, in order to approximate the o f a p o s s i b l e c o n t r o l o f t h e CLD a n d t h e r e -
part of the d i s t r i b u t i o n function between fore of t h e polymer q u a l i t y .
two o f t h e l e n g t h s .
d ) B e t w e e n t w o u p d a t e s t h e CLD i s calcula-
ted c o n t i n u o u s l y on t h e b a s i s of a detai- NOMENCLATURE
led model.
I Initiator

EXPERIMENTAL RESULTS M monomer

S solvent
F i g u r e 5 shows t h e measured d a t a of t h e
a v e r a g e MWs, (Mn,Mw) d u r i n g a r e a c t i o n r u n RX chemical modifier
w i t h a h e a t i n g up s t e p t o a t e m p e r a t u r e o f Pi* radical chain of length i
a p p r o x i m a t e l y 85°C. The e s t i m a t e s o f t h e
t h e first t w o m o m e n t s , ή ( μ ) , ή ( μ ^ ) a r e Pi dead chain of length i
0
obtained f r o m t h e a b o v e d e s c r i b e d a l g o - Τ reaction temperature
r i t h m , u s i n g t h e m e a s u r e d v a l u e s o f Μη,Mw
each 20 m i n . The e s t i m a t e s o f t h e a v e r a g e Tk cooling jacket temperature
y
MWs f i t s a t i s f a c t o r i l y t h e m e a s u r e d d a t a . moments of t h e CLD
I n t h e same F i g u r e i s shown t h e r e s p o n s e j
o f t h e f i l t e r when an e r r o r i n t h e i n i t i a l Mn,Mw a v e r a g e MWs
c o n d i t i o n s f o r ή ί μ ^ ) , η ( μ ^ ) i s a s s u m e d . Mw W(i) weight f r a c t i o n of t h e chain i , as a
c o n v e r g e s s l o w e r t h a n Mn t o t h e m e a s u r e d proportion of t h e polymer as a whole
d a t a , b e c a u s e t h e t h i r d moment η ( μ ) o f
t h e CLD i s n o t c o r r e c t e d a t t h e u p a a t e
2 V r e t e n t i o n volume
κ
s t e p . Figure 6 shows t h e r e s u l t s o f an o n - AREAi a r e a c o r r e s p o n d i n g to chain with
l i n e estimation of the chain length d i s t r i - length i
b u t i o n . F i g u r e 6a shows t h e w e i g h t f r a c -
tion d i s t r i b u t i o n s , plotted each 5 minutes tx time of sample removal
d u r i n g t h e o p e r a t i o n and c o r r e c t e d from n(x) mole number of species χ
the measured d i s t r i b u t i o n s of t h e samples
t a k e n a t 2 0 , 4 0 , 6 0 , 80 m i n . The u p d a t e s MWi molecular weight of the chain i
corresponding to a time delay of approxi- χ state vector
m a t e l y 18 m i n . a r e o b t a i n e d a t 4 0 , 6 0 , 8 0 ,
100 m i n . F i g u r e 6b shows t h e c o u r s e s o f u vector of input variables
t h e e s t i m a t e s o f t h e mole numbers o f two χ expected value of v a r i a b l e χ
p o l y m e r i c c h a i n s w i t h 47 a n d 111 monomer
u n i t s , u s e d f o r t h e c a l c u l a t i o n o f t h e CLD. Ρ covariance matrix of (x-x )
Q
The c o n v e r g e n c e o f t h e e s t i m a t e s o f t h e Q covariance matrix of the process noise
CLD t o t h e m e a s u r e d o n e s , i n c o m p a r i s o n
t o t h e p r e d i c t e d CLD i s s h o w n i n F i g . 6 c . R covariance matrix of the measurement
F i g u r e 7 shows t h e r e s u l t s i n a c a s e of a noise.
r e a c t i o n r u n d u r i n g w h i c h a m o l e number
 Chain Length Distribution in a Polymerization Reactor 145

REFERENCES

η I Do n(M), L n(S). n|RX)o

Adams, P . G . , and A . T . S c h o o l e y (1969).
Ada-Predictive Control for a Batch
Reaction. Instr. techn., Jan., 57-62.
Gelb, A.(Edr.) (1974). Applied Optimal NLLL ,
E s t i m a t i o n . T h e MIT P r e s s , Cambridge.
MTU » FîILMJ.S
Î • Vï N(MI , ÎLP.FJTL.M.T.S.RXL'VJ N(MI) ,
G i l l e s , E . D . , and U. K n o e p p , ( 1 9 6 7 ) . Die
Dynamik v o n R u e h r k e s s e l r e a k t o r e n . JI » Fj U,M,T,S)*Vj Τ , J » 0.1,2
Teil II. Regelungstechnik, 262-269 . N(S) , NIPRL
G i l l e s , E.D. (1979). Modellbildung, ein
^Sl»Fi II.M.T.S).V
T ^P.FJU.MJ.S.RXL.VI R
H i l f s m i t t e l der Messtechnik, Teil 1. dniBa»F5(I.M.T.S.RXL*V N<RX) ,
DT S I.1,2,3...
T e c h n . M e s s e n , 6, 225-274.
H e n r i c i - O l i v é , G. a n d S . O l i v e , (1969). Second Subsystem
First Subsystem
Polymerisation. Katalyse-Kinematik-
Mechanismen. Verlag Chemie, Weinheim.
J o , J . H . , and S.G. B a n k o f f , (1976). Digi-
Fig. 1. Signal flow diagram of the system
t a l M o n i t o r i n g and e s t i m a t i o n o f poly-
merization reactors.
AIChE J . , 2 2 , 3 6 1 - 3 6 3 .
K i p a r i s s i d e s , C . , and J . MacGregor, A . E .
Hamielec, (1980). Continuous Emulsion
Polymerization of Vinyl Acetate,
Μη , Mw .fin,flw
Part I I I : Detection of reactor perfor-
mance by t u r b i d i t y - s p e c t r a and l i q u i d
e x c l u s i o n chromatography. Canad. J . of
Chem. E n g . , 5 8 , 6 5 - 7 1 .
Kiparissides, C , and J.MacGregor, A . E .
Hamielec, (1981). Suboptimal Stocha-
s t i c Control of a continuous Latex
R e a c t o r . AIChE J . , 2 7 , 13-19.
Odian, G . , ( 1 9 8 1 ) . P r i n c i p l e s of Polymeri-
zation. J. Wiley & Sons, sec.edit..
New Y o r k .
P i c k e t t , H . F . , and J.W. Cantow, J.F.John-
son, ( 1 9 6 6 ) . Column F r a c t i o n a t i o n o f
P o l y m e r s , V I I I . Computer Program f o r
Determination of the Molecular Weight
D i s t r i b u t i o n s f r o m GPC. J . o f A p p l .
Pol. Sci., 10, 917-924.
P i e r s o n , R . M . , H. Z . C o s t a n z a , A . H . W e i n -
stein, (1955). Bis type modifiers.
I. Behavior of v a r i o u s D i s u l f i d e s in
Bulk Styrene Polymerization.
J. of Pol. S c i . , 17, 221-246.
Fig. M e a s u r e d a n d e s t i m a t e d Mn,Mw w i t h
S c h u l e r , H. ( 1 9 8 0 ) . E s t i m a t i o n o f States t h e p r e s e n c e of chem. modifier.
i n a P o l y m e r i z a t i o n r e a c t o r . IFAC PRP •,x: Measurements; ,-: Estimates
4 Automation, Ghent, Belgium,369-376. K i n e t i c d a t a taken from P i e r s o n
S c h u l e r , H . , and S. Zhang, (1985). Real and o t h e r s (1955).
Time E s t i m a t i o n of t h e Chain L e n g t h
D i s t r i b u t i o n in a Polymerization Reac-
t o r . Chem. E n g . S c i . , 4 0 , 1881-1904.
S c h u l e r , Η . , and S. P a p a d o p o u l o u , (in MEASUREMENT VARIABLES
p r e s s ) . R e a l Time E s t i m a t i o n o f the -| G.P.C
PROCESS
Chain Length D i s t r i b u t i o n in a P o l y -
m e r i z a t i o n Reactor. I I . Comparison of
e s t i m a t e d and m e a s u r e d distributions.
Chem. E n g . Sci. ΪΓΪΙ-Ϊ2

Y a u , W.W., a n d S.W. F l e m i n g , (1968).

GPC D a t a Interpretation.
J. of Appl. Pol. S c i . , 12, 211-216. •MM fê^(X,.U2.T)
JiLlL

RIRSR ritrer becona Titrer ^

ACKNOWLEDGEMENT
STATE ESTIMATES

The a u t h o r w o u l d l i k e t o t h a n k Fig. 3. Signal flow diagram of the two

P r o f . E. D. G i l l e s f o r h i s c o n s t a n t in-
decoupled filters.
terest .
146 S. Ar. Papadopoulou

2
1.5. W ]

E =F T I M
*Ι2·*Ι2*1 °
sample removal for the
measurements at *Ί<2 **"k2*1
to = time delay

k k
k /k 2 2 W 2 2*i/k 2+
F i g. 4
. T i e
m (t ) t a k ne b e t w en e s a m p el r e- 0 20 40 60 80 IOOt/ib.120
2 2
m o v al
t i me of
( t^. - ^ )
t h es e
+ annd
c o d
u p d a et
f i l t e. r
( t
K ^ )
+F i g. 5. M e a s u rd e a nd e s t i m a t
d e M n , M .w
*: M e a s u r e m e n; t s— : E s t i m a t i. o n
: E s t i m a t ni o a s s u m ig n an e r r ro
in t he i n i t il a c o n d i t i o .n s
W i
' W 7T "(μι) w i4lM
Wii
i="(μι)

4,66.10

State Estimates ; WI,WI=!

.«s ntPwl/mol
7.00.10' t = 80 min

4,66.10

1
W 'n(PI)
û w - i ^ - . F 6.i a C o n t i n u so u e s t i m a t ni o of t he C L .
D W
'"FIFN.L '-NL„.L t=55min ^
... .Measurement
Estimation with GPC
Simulation

1074

F i g. 7
. E s t i m a t ni o of t he C L D
. A c h e m i cl a
m o d i f i re w as a d d de d u r i gn t he
r e a c t i o. n
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
iournemouth, UK 1986
STATE ESTIMATION FOR POLYMERIZATION
REACTORS
J. F. MacGregor, D. J. Kozub, A. Penlidis and A. E. Hamielec
McMaster Institute for Polymer Production Technology, Department of Chemical
Engineering, McMaster University, Hamilton, Ontario, Canada L8S 4L7
ABSTRACT
The use of extended Kalman Filters for state estimation in
polymerization reactors is discussed. An important factor in
making these filters work well in practice is to properly
account for the stochastic disturbances in the system through
the introduction of meaningful s t o c h a s t i c s t a t e s . In
particular, since reactive impurities are the major source of
these stochastic disturbances, it is natural to incorporate the
stochastics in the model through the reactive impurity terms.
This approach is illustrated experimentally for the emulsion
polymerization of styrene in batch and continuous reactors.
Using on-line density measurements the extended Kalman
Filter is shown to track conversion and latex particle size very
closely, and to provide good estimates of the impurity levels in
the reactors.
Keywords: State e s t i m a t i o n , Kalman F i l t e r i n g ,
Polymerization, Emulsion polymerizaton, Batch reactors,
Continuous reactors.
INTRODUCTION
One of the major problems in the control of polymerization
reactors is the absence of good on-line sensors. Few robust
sensors exist that enable one to track polymer properties such
as molecular weight, copolymer composition and monomer
conversion, or polymer latex properties such as size and
concentration of particles In recent years there have been
some promising results on the use of on-line densitometers to
measure percent solids or conversion (Abbey, 1980; Shork and
Ray, 1980; Penlidis et al., 1986a), on-line energy balances to
monitor conversion (Harris and Rushing, 1982; Gordon and
Weidner, 1981; Wu, 1985), and on-line particle size analysers
for emulsion polymerizatons (Gossen et al., 1986). Results
have been reported on the use of on-line gas chromatographs,
gel permeation chromatographs and surface tensiometers, but
these do not yet appear robust enough for on-line industrial
use.
On the other hand, there has been a great deal of success in
recent years in the area of modeling polymerization reactors
using fundamental mass, energy and population balances (e.g.
Ray, 1980; MacGregor et al., 1984; Penlidis et al., 1985) These
d e t e r m i n i s t i c models are capable of p r e d i c t i n g t h e
development of all the fundamental polymer or latex states of
the polymerization process in a disturbance-free environment.
However, industrial polymerizations are subject to stochastic
disturbances in many states that do not appear in these
deterministic models. Since these disturbance states affect
the polymerization process the deterministic models are of
little direct value in being able to track the progress of a
reaction. These models have found their main use in
providing us with a fundamental understanding of the
polymerization mechanisms, in designing reactors, and in
developing optimal open-loop control policies. However, by
taking care to introduce stochastic state variables in a
meaningful and concise manner, and combining the resulting
dynamic-stochastic model with feedback information coming
from one or more reliable on-line sensors, one can achieve
excellent on-line tracking of the observable reactor states, it
147
is the purpose of this paper to examine the choice of these
stochastic states, and to illustrate the use of the extended
K a l m a n F i l t e r by i t s a p p l i c a t i o n to the e m u l s i o n
polymerization of styrene in both batch and continuous pilot
plant reactors.
Only a few studies on the use of Kalman Filters to track the
behaviour of polymerization reactors have been reported in
the literature. Hyun and Bankoff (1976) used a mechanistic
model for the continuous solution polymerization of vinyl
acetate together with an empirical measurement equation
relating pressure drop through a capillary tube to the
conversion and molecular weight states. Time dependent
disturbance states were introduced in the reactor residence
time, in the pressure drop correlation, and in the initial
inhibition period. The parametric sensitivity of an extended
Kalman Filter estimating the drift rates in these disturbance
states was studied by computer simulation. Jo and Bankoff
(1976) also studied the use of the Kalman Filter to obtain
estimates of conversion and molecular weight average using a
refractometer and a viscometer for the same free radical
solution polymerization of vinyl acetate. They found that good
state estimation was achieved u s i n g data from their
experimental CSTR reactor system, and that the filter could
track the states even with significant disturbances introduced
into the system. Schuler ( 1980) discussed the use of Observers
and Kalman Filters to estimate unmeasured states in a
styrene polymerization reactor An extended Kalman Filter
was investigated by Kiparassides et al. (1981) to estimate the
material balance and population balance states in the
emulsion polymerization of vinyl acetate from turbidity
measurements. More recent applications of the Kalman Filter
to polymerization reactors include: its use for estimating
instantaneous reaction rates and heat transfer coefficients,
and total conversion from o n - l i n e e n e r g y b a l a n c e s
(MacGregor 1986); its use for detecting the possibility of the
reactor being in any of several r u n a w a y s t a t e s from
temperature measurements (King and Gilles, 1984); and its
use for estimating the molecular weight distribution using
on-line gel permeation chromatograph measurements (Gilles,
1986; Schuler et. al., 1986).
Observability is an important issue in state estimation for
polymerization processes. In most polymerization models the
material balance and particle growth parts of the model are
independent or only weakly dependent upon the molecular
weight development part (Ray, 1980). Hence the molecular
weight moments and branching frequencies are unobservable
from measurements of the former states. Conversely, the
molecular weight part of the model is usually dependent upon
the material and particle population balances. Hence, the
latter states are observable from molecular weight related
measurements. Pollock (1984) discusses these ideas in
studying the optimal selection and development of on-line
sensors. By changing the choice of the measurement matrix
and the measurement covariance matrix, the choice of single
s e n s o r s , or sensor c o m b i n a t i o n s which optimize the
generalized variance of the state e s t i m a t e s of interest
(obtained from the Kalman Filter) can be made prior to
developing any of the sensors.
148 J. F. MacGregor et al.

STATE ESTIMATION x
that is x + K h ( xu
(7)
In this section we consider the general structure of the state k i|k +1 = k i|k k k +1 - k + i|k' k »
+ +
equations and the extended Kalman Filter, and we discuss
where
some of the implications that arise from certain choices of the x
stochastic states. d
k +i
The State Equations and Extended Kalman Filter
k+i
Consider the general nonlinear deterministic model for the is now the augmented state vector and Kk is the usual linear
dynamics of the process arising from unsteady-state material, Kalman Gain.
energy and population balances:
d d K = H (P H P H + rl R (8)
k k t|k k k k i|k k v
d x (t) = f(x , u) (1) + +
where x is the vector of states for the deterministic part of the
d with
model. Given the current state estimate (x (tk/tkh the (9)
prediction of the deterministic state vector over the time P, = Α. Ρ , , . Α £ + R
k+l|k
interval tk to tk + i is obtained by integrating the differential P k kjk P k Hw PK
equation (1) using an appropriate numerical procedure (e.g. k + l|k + l ~ k+l|k k k + l k + l|k (10)
Runge-Kutta or Predictor-Corrector methods) to give

(2)
5
\«ΐν = Λ
Α ) +
j «* . >
+1 d u dt
In the above,
ah
d
Performing a local linearization and discretization of the state
d H,
k+I laxK+IIK
model (1) about the conditions x (tk + i/tt) = Xk + l/k and
u(tk) = Uk we get: is the linearized observation matrix, and d
(3)
R Οs
and R =
Ο A Ο R
where
are the transition matrix and noise covariance matrix of the
augmented state variable model.
d d
d
and Wk is a white noise vector with covariance matrix R . Equations (2), (5), and (7) through (10) define the extended
This latter white noise vector wk is introduced to account for
w Kalman Filter.
errors in the right hand side of (3) arising from modelling
errors, errors due to the linearization, and errors in the
d d The Importance of Nonstationary Stochastic Disturbance
d
settings of Uk and other variables appearing in A or Β . Its States
covariance matrix R is usually taken to be diagonal both for
w
c o n v e n i e n c e and for lack of i n f o r m a t i o n r e g a r d i n g The purpose of the Kalman Filter is to estimate or track the
covariances. Its diagonal elements (variances) are chosen to observable states of the process using information from
d
reflect the total possible error in each row of (3) over the measurements taken on only a few of the states. Just as in
interval tk to tk + i- Alternatively, R process control the major reason that a state estimator is
w could be estimated
from process data by minimizing a function of the prediction needed is because of the presence of disturbances. Otherwise,
errors from the final Kalman Filter. open-loop integration of the deterministic model via (2) would
provide excellent prediction. The stochastic disturbances
s cause the process to drift away from these deterministic model
In addition to the deterministic states there usually exist a d
predictions, and it is only through introducing the disturbance
number of stochastic disturbance states Xk . These can
usually be represented by a discrete linear state model: state model (4), plus the white noise variables wk and VK that

Xs s information from the outputs yk gets fedback into the state

A (4) estimates through the updating equation (7).
v - A
k +1~
s s It is quite common in the literature on Kalman Filtering to
s
where w k is a white noise vector with covariance matrix R .
omit the introduction of separate stochastic states given in

R
w equation (4). Instead, only the deterministic part of the state
wis often taken to be diagonal with variance elements
chosen to reflect the magnitude of disturbance that might be
model is considered as in equations (1) and (3), and arbitrary
white noise vectors and vk are simply added to the
expected in one sampling period (tk, tk+1). These elements are linearized state equations (3), and the measurement equation
usually the main parameters used to tune the Kalman Filter (6) to represent modeling error and measurement error
to give good tracking behaviour. The prediction of the respectively. Such an approach, although possibly adequate
stochastic states over the sampling interval (tk, tk + i) is given in nearly deterministic aerospace environments, is totally
by s x s s Ά x ) inadequate in nearly any real process environment. By ( 5
x - A x allowing only for the possibility of stationary additive white
k + l|k ~ k|k noise errors in the state equation, there is the implicit
The measurements are generally a nonlinear function of the assumption that the true process states can never drift away
for any prolonged period of time from the values predicted by
states and controls s the deterministic model. In this situation, the Kalman Filter
y = h(x^,x ,u ) + v will not contain any integration terms, and will not track any
k kk k real shift in level of the process caused by a nonstationary
where VK is a white noise vector of measurement errors with disturbance, such as, for example, an increase in reactive
covariance matrix R . This covariance matrix is usually impurities in the feed to a polymerization reactor.
v
easily obtained from a knowledge of the measurement sensors,
or from replicated measurement data. In an attempt to overcome this bias or offset arising from
Kalman Filters developed in this manner, it has often been
When a new observation yk + i becomes available at time tk + i the practice to increase the Kalman Gain matrix K, either
the state estimate can be updated by using the linear Kalman directly, or by artificially increasing the magnitude of the
Filter equations with the locally linearized state model (3), (4), state noise covariance matrix R^. This indeed will reduce,
 State Estimation for Polymerization Reactors
but not eliminate, the bias. However, it will also make the
Kalman Filter very sensitive to measurement noise in y
because of the high gain. Such misapplications have tended to
give the Kalman Filter an undeservedly poor reputation
among process control engineers.
As an aside, we point out that the above situation is entirely
analogous to the design of optimal LQ controllers. If
nonstationary disturbance models are not used in the
controller design, proportional type controllers having no
integral action will result. Only in the p r e s e n c e of
nonstationary disturbances is integral action necessary.
Hence, one of the most critical steps in designing Kalman
Filters is to formulate a realistic model for the behaviour of
the disturbances (4). If it is possible for the states of the real
process to drift away from those predicted by the deterministic
model, then it is important to identify the source of these
nonstationary disturbances and to model them appropriately.
For example, in the case of emulsion polymerization in a
CSTR considered in the following sections, the major
stochastic disturbances arise from nonstationary variations in
reactive impurities in the feedstrearns of the reactor. For lack
of better knowledge we often represent these by the random
walk model of the form
_ s (4b)
S j
MODEL FOR THE EMULSION
POLYMERIZATION OF STYRENE
Although the approach to state estimation discussed in this
paper is general to many polymerization systems, we will
illustrate it by way of a p p l i c a t i o n s to the e m u l s i o n
polymerization of styrene in a CSTR and batch reactors. The
Kalman Filter will be used to track conversion and particle
growth states (e.g. particle number, diameter, etc.) based on
measurements from an on-line densitometer. Since the
molecular weight states are unobservable from latex density
measurements, the molecular weight part of the model will be
ignored here.
Deterministic Reactor Model
Broadhead et al. (1985) developed a theoretical model for the
emulsion polymerization of styrene-butadiene in batch,
semi-batch and continuous stirred tank reactors. In this paper
we use the subset of their model equations relating to the
homopolymerization of styrene, and ignore the molecular
weight development part. The reduced model consists of a set
of eight nonlinear ordinary differential equations in the states
x = (N , P , Ni, N E , N , V , Vo, V ) T where N and P are
s s p P W s
the moles of unreacted styrene monomer and monomer bound
s X
in the polymer respectively, Ni and N E are the moles of
initiator and emulsifier in the reactor, Np is the number of
polymer latex particles in the reactor, Vp is the volume of the
particle phase, and Vo and V\y are the volumes of organic and
water phase in the reactor, respectively. A number of
algebraic equations relating to the calculation of the average
number of radicals per particle, the instantaneous reaction
rate, the micellar surface area, etc. from these states make up
the rest of the model. The monomer conversion (X) and
average particle diameter (D ) in the reactor can be calculated
p
from these states at any instant by
(11)
and D
Ν +Ρ
The reader is referred to Broadhead et al. (1985) for the
development of the model, and for the final state equations.
Stochastic Disturbances
In e s s e n t i a l l y all industrial free r a d i c a l and i o n i c
polymerizations the major source of disturbances arises from
reactive impurities These reactive impurities consume
radicals in free radical polymerizations, and poison or complex
149
the reactive ionic sites of catalysts in ionic systems. The result
is to alter the rate of polymerization which in turn affects all
the material balance states in the reactor. Furthermore, in
emulsion polymerization reactive impurities can also have a
profound effect on the particle population balance states
through their influence on the particle nucleation kinetics
(Huo et al., 1986; Penlidis et al., 1986b). Hence, if the
concentration of reactive impurities in the reactor changes
with time, then one would observe changes in essentially all
the states of the reactor. In continuous flow reactors where
there is a continuous feed of monomer, water, initiator, etc., to
the reactor one would expect to see a nonstationary drift in the
amount of reactive impurities entering the reactor. This is
particularly true in many continuous processes where the
polymerization is carried to less than complete conversion,
such as in the manufacture of styrene-butadiene rubber (SBR)
or polyvinylchloride (PVC). In these cases the unreacted
monomer is recovered by steam stripping of the latex, or by
other means, and then recycled to the reactor train. These
recycle streams usually contain a much higher level of
reactive impurities than the original feed streams, due to the
formation of reactive byproducts in the reactors and recovery
units.
It is important to incorporate the effect of these unmeasured
stochastic disturbances into the model using as much
theoretical knowledge as possible. For this reason the authors
carried out a series of fundamental experimental and
modeling studies on the effect of water and monomer soluble
impurities on the kinetics of various cases of emulsion
polymerization (Huo et al., 1986; Penlidis et al., 1986b). In the
present study on the emulsion polymerization of styrene these
impurity effects can be incorporated into the deterministic
part of the model through the addition of two new state
equations for [IMlw and [IMIM, the concentrations of water
and monomer soluble impurities in the reactor, respectively
(Huo, et al., 1986). The final deterministic state equation for
the styrene emulsion polymerization, therefore, consisted of
ten nonlinear differential equations of the the form of
equation (1) where
d T
x = ( N , Ps, N N N , V , V , V , [IMlw, [IM1 ) (12)
s If E P P 0 W M
For the case of continuous flow reactors the last two state
equations contain terms on the right hand side, involving the
inlet flow rates of water and monomer soluble impurities,
PlM ,in, and F[M ,in» respectively. These unmeasured
w M
stochastic disturbances can be modeled by stochastic state
equations of the form of equation (4). Since it is generally
known that these are nonstationary in nature, the simplest
assumption is that they follow the random walk behaviour
given by equation (4b) where
S T
= (FlM ,in> F i M , i n )
w m
(13)
The final state equations for the continuous reactor therefore
consist of the ten nonlinear deterministic state equations (1)
s
with xd defined as in (12), and the two nonstationary
stochastic disturbance state equations (4b) with x defined as
in (13). The combined linearized state equations are of the
form
, d . d (14)
A, A
Ο I
The case of semi-batch reactors closely parallels that laid out
for the CSTR above. Time varying stochastic disturbances in
the impurity states (13) of the feedstreams can again be
modelled by the random walk process (4b). However, another
source of variability occurs due to the batch-to-batch variation
in the monomer soluble impurity concentration in the initial
charge to the reactor, that is, in the initial state value [IMJMO-
(Water soluble impurity in the initial charge simply causes an
induction period that can be d e t e c t e d from o n - l i n e
temperature measurements). In this case guesses for the
initial value of this state x / and its variance (P / ) must be
0 0 00
selected, and the Kalman Filter allowed to track from that
point. Most of the other initial state estimates are usually
well defined from the initial charge material balances.
150 J. F. MacGregor et al.
For the case of a purely batch system, there are no feeds, and
hence the nonstationary stochastic disturbance states
defined by (13) and (4b) do not exist. The source of batch-to-
batch variability is mainly in the impurity concentrations in
the initial charge. However, the Kalman Filter is not well
suited to handling this problem of uncertainty in the initial
state estimates (x / )- With no nonstationary stochastic state
00
variables in the model, the Kalman Filter cannot eliminate
biases o*-* offsets that will result from an improper choice of
x / . This behaviour is illustrated in the next section, and
00
approaches for overcoming this difficulty are discussed.
However, semi-batch and CSTR systems are of greater
industrial interest, and are the major topic of this paper.
Therefore, an application of this methodology is also applied in
the next section to tracking the emulsion polymerization of
styrene in a CSTR.
EXPERIMENTAL RESULTS
Batch Reactor Runs
The batch reactor runs were carried out in a 10 liter stainless
steel reactor. The reactor temperature was maintained
constant at 55°C by a controller manipulating the coolant inlet
temperature to the jacket. Several batch runs were performed
with different known initial concentrations of a reactive
monomer soluble inhibitor (4-tert-butyl catechol or TBC). The
monomer conversion (X) was measured by densitometry and
checked by off-line gravimetric analysis. The average particle
s i z e (Dp) w a s m e a s u r e d off-line by h y d r o d y n a m i c
chromatography.
The experimental results for conversion versus time and
particle diameter versus time for three different impurity
concentrations in the initial charge are shown in Figures 1
and 2 respectively. The deterministic model predictions,
plotted as the solid curves, show that this model is indeed
capable of good prediction if the impurity states are known.
However, it is also obvious that the impurities have a
pronounced effect on the polymerization, and hence must be
estimated on-line in order to use the model to track the
progress of the reaction.
The conversion and particle size measurements from these
runs were only available approximately every 5 minutes
although the on-line instruments give results much more
frequently. Since this 5 minute data was considered to be too
infrequent for the Kalman Filter studies, data was generated
from the simulation at two minute intervals and typical
measurements were added. The results of applying the
Kalman Filter using the batch conversion measurements are
shown in Figures 3 and 4 for the case where the initial
impurity states estimate x / for the modes of monomer soluble
00
impurity is good (equal to the true value, x / = 50 ppm), and
00
where it is poor (x / = 30 ppm), respectively. As expected the
00 }
Kalman Filter estimates (dotted lines) are very good when the
initial state estimates are good. However, as shown in Fig. 4,
this "proportional" Kalman Filter has no ability to recover
from a poor initial estimate for the impurity level. The final
state estimates all show offset from their true values.
There exist a number of solutions to this problem of track ing
states in a purely batch reactor. One is to use repeated
nonlinear estimations on-line to estimate the unknown initial
impurity level. The nonlinear estimation is repeated upon
receipt of each new measurement, and at the end of the batch
the estimated curves would essentially correspond to those
shown in Figures 1 and 2. This approach has been used to
track the progress of batch pulp digesters using infrequent
Kappa n u m b e r m e a s u r e m e n t s . H o w e v e r , it i s
computationally very time consuming.
An alternative is to allow for a much larger stochastic term
(w(t)) in the
d impurity
[ I concentration
M ] state equation :
M - / χ
= _ ν [IML, η Ν /V + W(t) [ '5 )
IM Μ ρ ρ ν
2
When the variance of the stochastic term o
w is small this
equation predicts a dominantly determinstic decay of the
impurity concentration fruits initial state estimate x / . The
00
Kalman Filter then has little ability to adjust the estimate of
2
this impurity concentration up or down if the initial state is
poorly chosen. On the other hand, if the variance o is large
w
the w(t) component on the right hand side dominates, and the
equation resembles that of a random walk. The Kalman Filter
then provides an estimate of the impurity state which,
although more erratic, is able to adjust quickly to a new level
to account for an improperly chosen x / . Figure 5 shows the
00
Kalman Filter results from rerunning the case shown in
2
Figure 4 (true [IMlo = 50 ppm, but x / = 30 ppm), but with
00
o and the initial state variance increased by a factor of 100.
w
The estimates of all unmeasured states (N , D V ) now track
p p p
the true values very well, and exhibit none of the offset seen in
Figure 4.
Continuous Stirred Tank Reactor (CSTR) Runs
The CSTR runs were carried out in a 0.5 liter stainless steel
reactor again controlled at 55°C. A pulse of an unknown
amount of water soluble impurity (oxygen) was injected into
the inlet feedstream, and a known amount of monomer soluble
impurity (equivalent to 50 ppm) was added to the monomer
feedtank at another point. Details from the CSTR runs will be
shown and discussed at the conference.
REFERENCES
Abbey, K.J. (1980). Polymerization kinetics by precision
densitometry, Amer. Chem. Soc. Symp. Series, 165, 345-356.
Broadhead, TO., Hamielec, A.E. and J.F. MacGregor (1985).
Dynamic modelling of the batch, semi-batch and continuous
production of styrene/butadiene copolymers by emulsion
polymerization. Makromol. Chem.Suppl., 10/11.105-128.
Gilles, E.D. (1986). Chem. Proc. Control III Conference,
Asilomar, California, Jan. 1986.
Gordon, D.L. and K.R. Weidner ( 1981). Control of particle size
distribution through metering based on rate of conversion.
ACS Symp. Ser. No. 165, 513-532.
Gossen P., Kourti, T., Penlidis, Α., MacGregor, J.F. and A.E.
Hamielec (1986). On-line particle size and conversion
d e t e r m i n a t i o n in c o n t i n u o u s l a t e x r e a c t o r t r a i n s .
IFAC-PRP-6 Automation, Akron, Ohio.
Harris, T.J. and E.V. Rushing (1982). Inferential control of
conversion in a commercial scale polymer reactor, Amer. Inst.
Chem. Eng. National meeting, Miami.
Huo, B.P., Campbell, J.D., Penlidis, Α., MacGregor, J.F. and
A.E. Hamielec (1986). Effect of impurities on emulsion
polymerization: Case II kinetics. J. Appl. Poly. Sci.. to appear.
H y u n , J . C . and S.G. B a n k o f f ( 1 9 7 6 ) . Continuous
polymerization of vinyl acetate - II. On-line estimation of
process drift. Chem. Eng. Sci.. 31,953-958.
Jo, J.H., and S.G. Bankoff (1976). Digital monitoring and
estimation of polymerization reactors. Amer. Inst. Chem.
Eng.J.. 22,361-368.
King, R. and E.D. Gilles (1984). Dynamics of runaway
chemical reactors, AIChE Annual Meeting, San Fransisco,
Nov., 1984.
Kiparissides, C , MacGregor, J.F. and A.E. Hamielec (1981).
Suboptimal stochastic control of a continuous latex reactor.
Amer. Inst. Chem. Eng. J., 27,13-19.
/ MacGregor, J.F., Penlidis, A. and A.E. Hamielec (1984).
Control of Polymerization Reactors: A Review. Polymer
Process Eng., 2,179-206.
 State Estimation for Polymerization Reactors 151

ElffC
t ol I8C Conce
n
MacGregor, J.F. (1986). On-line reactor energy balances via
Kalman Filtering. Proc. IFAC PRP-6 Automation, Akron,
Ohio, USA, Oct.

Penlidis, Α., MacGregor, J.F. and A.E. Hamielec (1985).

Dynamic modelling of emulsion polymerization reactors,
Amer. Inst. Chem. Eng. J.. 31, 881-889.

Penlidis, Α., MacGregor, J.F. and A.E. Hamielec (1986a).

Continuous emulsion polymerization: Design and control
considerations for CSTR trains. Submitted to AIChE J..

Penlidis, Α., MacGregor, J.F. and A.E. Hamielec (1986b).

Effect of impurities on emulsion polymerization: Case I
kinetics. J. Appl. Poly. Sci.. to appear.

Pollock, M.J. (1984), Ph. D. thesis, Dept. Chem. Eng.,

McMaster University.

Ray. W.H. (1980). Dynamic modelling of polyermization

reactors, Proc. IFAC- PRP-4 Automatica, Ghent, Belgium, Figure 2: Experimental results and model predictions of
587-595. particle diameter vs. time for various initial
impurity concentrations.
Schuler, H. (1980). Proc. IFAC PRP-4 Automation, Ghent,
Belgium, 369-376.

Schuler, H. and S. Papadopoulou (1986). Real time estimation

of the chain length distribution in a polymerization reactor.
II. Comparision of estimated and measured distribution
function. To appear in Chem. Eng. Sci..

Shork, F.J. and W.H. Ray (1980). On-line monitoring of

emulsion polymerization dynamics. Amer. Chem. Soc. Symp.
Series, 165,505-514.
0.0 10,
0 20.
0 30.
0 40.
0 50.
0 60.
0 ).0 90.
0 100.
0
T I ME
Wu, R.S.H. (1985). Dynamic thermal analyser for monitoring MOLES STYRENE (BO PPM)
batch processes. Chem. Eng. Progress, Sept., 57-61.

Effec
t of IBC Concentratio
n

0.0 10.
0 20.
0 30.
0 40.
0 50.
0 60.
0 70.
0 80.
0 90.
0 100.
0
T I ME
MOLES IMPURITY (SO PPM)

s.«J
0

0.0 10.
0 20.
0 30.
0 40.
0 50.
0 60.
0 70.
0 80.
0 90.
0 100.
0
T I ME
NUMBER PARTICLES (SO PPM)

Figure 1: Experimental results and model predictions of

conversion vs. time for various initial impurity
concentrations.

0.0 10.
0 20.
0 30.
0 40.
0 50.
0 60.
0 70.
0 80.
0 90.
0 100.
0
T I ME
AVE. PARTICLE DIAMETER (SO PPM)
Figure 3: Kalman Filter estimates (dotted lines) and
DCCR-K true batch reactor states (solid lines) for the
initial impurity state estimate equal to the
true value (50 ppm).
 152 J. F. MacGregor et al.

20.0 ·
20.0

15.0 ·
15.0

10.0 •

0.0 10.0 20.0 30.0 40.0 50.0 60.0 70.0 80.0 90.0 100.0
TIME 0.0 10.0 20.0 30.0 40.0 50.0 60.0 70.0 5.0 90.0 100.0
MOLES STYRENE (SO PPM) TIME
MOLES STYRENE (SO PPM)

0.0 10.0 20.0 30.0 40.0 50.0 60.0 70.0 80.0 90.0 100.0
TIME 0.0 10.0 20.0 30.0 40.0 50.0 60.0 3.0 90.0 100.0
MOLES IMPURITY (SO PPM) TIME
MOLES IMPURITY (SO PPM)

0.0 10.0 20.0 30.0 40.0 50.0 60.0 70.0 80.0 90.0 100.0
TIME 0.0 10.0 20.0 30.0 40.0 50.0 60.0 70.0 80.0 90.0 100.0
NUMBER PARTICLES (SO PPM) TIME
NUMBER PARTICLES (SO PPM)

0.0 10.0 20.0 30.0 40.0 50.0 60.0 70.0 80.0 90.0 100.0
TIME 0.0 10.0 20.0 30.0 40.0 50.0 60.0 70.0 80.0 90.0 100.0
AVE. PARTICLE DIAMETER (SO PPM) TIME
AVE. PARTICLE DIAMETER (SO PPM)

0.0 10.0 20.0 30.0 40.0 50.0 60.0 70.0 80.0 90.0 100.0
TIME 0.0 10.0 20.0 30.0 40.0 50.0 60.0 70.0 80.0 90.0 100.0
VOLUME POL YMER PHASE (SO PPM) TIME
VOLUME POL YMER PHASE (SO PPM)

Figure 4: Kalman Filter estimates (dotted lines) for the Figure 5: Kalman Filter estimates (dotted lines) for the
same case as Fig. 3, but with x / = 30 ppm.
00 same case as Fig. 4, but using a large g \ for
the impurity state equation.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

VIBRATIONAL CONTROL OF CHEMICAL

REACTORS: STABILIZATION AND
CONVERSION IMPROVEMENT IN AN
EXOTHERMIC CSTR
A. Cinar*, K. Rigopoulos*, X. Shu* and S. M . Meerkov**
^Departments of Chemical and Electrical Engineering, Illinois Institute of
Technology, Chicago, IL 60616, USA
^Department of Electrical Engineering and Computer Science, The University of
Michigan, Ann Arbor, MI 48109, USA

Abstract. Vibrational control is a method for modification of the dynamic properties

of linear and nonlinear systems by introducing fast, zero-average oscillations in the
parameters of the system. Vibrational control of an exothermic reaction in a CSTR
modifies the dynamic behavior of the reactor and enables stabilized operation of the
reactor in the unstable steady-state region. Simultaneous vibrations in input flow
rate and concentration reduce the amplitude of the oscillations in reactor temperature
and/or increase in reactant conversion. The phase shift between the vibrations of the
two input variables affects the behavior of the system as well as the amplitude, the
frequency and the form of the forcing functions.

Keywords. Chemical reactor control; Periodic control; Vibrational stabilization of

nonlinear systems; Nonlinear control. Computer control.

INTRODUCTION

Vibrational control is an open-loop control strat- a periodic or quasi-periodic function.

egy implemented by fast zero-average oscillations
in the input variables of a system. The theory of Defining the fast time θ = t/ε, (1) can be written
vibrational control has been developed by Meerkov
(1980, 1982) and by Bellman et al. (1983, 1984).
The conditions of vibrational stabilizability and dx/de = εΧ (χ) + αΧ (θ, χ). (2)
vibrational controllability have been derived using
4 β
the method of averaging (Bogoliubov and Mitropolsky The general solution of a generating equation
1961). Some preliminary results on vibrational dx/d e = αΧ (θ,χ) is denoted as χ(θ) = η(θ,σ),
control of an exothermic CSTR to achieve asymptoti-
β
c = constant. When Χ (θ,χ) is differentiable with
cally stable periodic operation of the CSTR in the
β
respect to χ a substitution which reduces (2) to a
vicinity of an unstable steady state have been standard form is χ(θ) = h(θ,y(θ)). The resulting
reported by Cinar et al. (1984). equation in the standard form is
l
In this paper, the method of analysis of vibra- dy/dG = ε [dh/dyV X (h(e ,y)) = eY(6,y). (3)
t
tionally controlled systems will be summarized and
the application of vibrational control to a CSTR To analyze (3) further, asymptotic techniques can
with homogeneous liquid phase exothermic reaction be employed. Introducing
will be presented. The effect of vibrations in two l
input variables will be analyzed, theoretical and
experimental results will be discussed. lim [ah/ay]- X» Ch(e,y)]d G Z(y) (4)
Τ->α> Τ

METHOD OF ANALYSIS the averaged equation can be defined as

The analysis of a vibrationally controlled system dz/de = εΖ(ζ) . (5)

is equivalent from the mathematical standpoint to
the analysis of a dynamical system with parametric
oscillations. Since the detailed description of
the method of vibrational control is available
elsewhere (Bellman et al., 1983, 1986) only a brief
description will be provided here.
If the assumptions stated in Bellman et al. (1983,
1986) are satisfied whenever (5) has a locally
asymptotically stable equilibrium point ζ , (3) has
β
an asymptotically stable periodic solution y*(6)
and for positive «S as small as desired

I I x. (y* ) - χ, (z ) I I < 6. (6)

Consider a system of the form χ = Χ(χ,σ) which may e e
be represented as
Here x»(y ) is the average steady state of (2) and
χ (ζ») is the approximated average steady state of
dx/dt = Χι(χ) + (α/ε)Χ (t/ε,χ)
β (1) β
(2).
Here χ is the state, t is the dimensionless time
and σ is a fast oscillating parameter defined by THE REACTOR SYSTEM
σ = σ • (α/ε) f(t/ε), where σ is a constant, f(.)
0
is a periodic or
0
quasi-periodic zero-average The homogeneous liquid phase reaction between
sodium thiosulfate (Na S 0j) and hydrogen peroxide
vector, 0<ε<<1 and 0<ct are scalars and X (t/e, .) is
t ee
153
154 A . C i n a r et al.

(ΗβΟ )
β is used in this study. For the reactant
concentrations used in the feed stream the stoichi-
lost. A comparison of the effect of different
variables for implementing vibrational control is
ometric equation for this second order exothermic given in (Cinar et al., 1986).
reaction is
In this work, the effect of forced vibrations in
2Na S*0
8 3• 4H 0e -> N a S 0
a ee6 • Na SO. + 4H 0.
g e two input variables is studied. Simultaneous
vibrations in feed flow rate and reactant concen-
The experimental system (Fig. 1) is constructed tration are considered. When two input variables
such that feed flow rate and/or concentration are forced periodically the amplitude and frequency
vibrations can be accomplished without changing the of oscillation as well as the phase shift between
ratio of the reactant concentrations in the feed. the two oscillations can be selected. In this
This assures that there are no side reactions. study, zero-average nonsymmetric rectangular wave-
Previously a two feed stream configuration was used forms (Fig. 2) are used. Denoting the duty fraction
instead of the three feed stream configuration, by ο and the phase shift by σ (note that in what
causing the formation of hydrogen sulfide. The follows α and σ will be used to represent these
current configuration eliminates such side reac- variables and not the parameters of the Equations
tions but the injection of a pure water stream (1) and (2)), the three alternatives are α>σ>0,
necessitates the use of more concentrated reactant σ>α>0 and α>-σ>0. For this reactor system the most
solutions. Since the Na«S 0 solution started to
e3
crystallize in the feed lines when its concentra-
promising results are obtained using the third
alternative.
tion was increased beyond 2.9 mol/1, the magnitudes
of the amplitude variations were constrained for
Oscillations in feed flow rate and concentration
this test reaction.
are represented as
The Reactor Model JF nT<t<(n+a)T
M
F =F. (
The material and energy balances describing the 0 o 1 A +Ff U ) ) - \ *
( nT + a ) T < t <
CSTR (Cinar et al., 1985) reduce to the following C
set of dimensionless equations for adiabatic
reactor operation:
C = C ( 1 f +U ) =
) AJt AFM nT<t<(n+a)T
AF ÀF c jc (n+cOT<t<(n l)T
dx /dt = -x
o AFm +
t t+ Da(l-Xt )(l-xe)expCx /<l+x /Y)3
3 a
where the subscript ο denotes the average value,
dx./dt = -x
e• bpDa(l-x,)<l-x.)exp[x /(l+x /γ)] (7)
3 3 and M and m denote the maximum and minimum values
respectively. Defining the various average values
dx /dt = -λχ
3 3+ X B D a d - X t ) <i-x»)expCx /<l+x /γ)]
3 3 R =
/ FF =C
/C R
Here b and ρ denote the ratio of the stoichiometric F M m c AFM AFm
coefficients and of the feed concentrations,
F. = aF + (l-a)F
respectively, x and x» denote the dimensionless
t to M
M ' m
H Q
t 8 and Na,S tOj concentrations, Theand x is the
3 C aC + C
dimensionless reactor temperature. dimension-
ÂF/ AFM ^ AFm
less variables and parameters are:
C a= R
+ 1C
x, = <C*,-CA>/CA, , x = (C -C,)/C, , y = E/RTf
e L F F ÂF ( c -^ AFm
x = C(T-T )/T ]y , τ = V/F , Da = C»,k exp(-y)x
C o T =F 1 F t dC
3 f f 0
AF / o \ A F
λ = l/(l+m,c/pCpV> , Β = (-AH)C f Y/pc T
A ff
C = K C
Since x /γ<<1 the temperature dependent βexponential
β
3 ÂF FC AF„
term is simplified using Q = l-x /γ+χ /γ
t 3 3 οn ο
(= 1/(1+χ /γ)) for the remainder of this work.
3
Kp = [(aRp+l-a)(aR +l-a)l/[(a+a)R R -a(R +R )+l+a+a]
C c cF c F
VIBRATIONAL CONTROL OF THE CSTR the material balance of A gives

Vibrational control of the reactor was considered

using forced oscillations in one or two input var-
VdC A/dt' - - V W b e x p ( - E/RT) )
iables. When a single input variable is vibrated,
forced oscillations can be introduced in the feed F. (l+A f( 'f)){K
flow rate, in the feed concentration or in the 0 F u FCC
A ( lF+A ocf [ u' ( t ' - a T ) ] ) - C A}
coolant flow rate. When two input variables are
vibrated any two of these variables may be used.
The dimensionless equations for the reactor system
A detailed discussion of the effects of single with vibrating feed flow rate and feed concentra-
input oscillations and the results of flow rate tion are
variations are given in (Cinar et al., 1985). The
effect of forced oscillations in coolant flow rate d x ^ d t - - x 1[l+ApfUtJJ+f^œt )
is presented in (Cinar et al., 1986). When oscil-
lations in reactant feed concentrations were
+ Da(l-x )(l-x )exp[F (x )]
tested, variations in ratio of the reactant concen- 1 2 0 3
trations triggered some side reactions resulting in
production of H S.
f i Of the three input variables, d x 2/ d t = -Xgll+ApffaOJ+f^urt )
feed flow rate is the most effective variable. The
(9)
capacity of the cooling coil dampens the variations + bpDa(l-x )(l-x )exp[F (x )]
in the reactor temperature and reduces the ampli- 1 2 o 3
tude of the temperature swings. Consequently its dx /dt=X{-x n+Apf(u)t)]
effect is not as pronounced as that of the feed 3 3
flow rate oscillations. Furthermore, the economic
+ BDa(l- )(l-x )exp[F (x )]}
advantage of eliminating the the cooling system is X l 2 o 3
 Vibrational Control of Chemical Reactors 155

of the system response. The ratio of the maximum

with to minimum flow rate, the ratio of maximum to
minimum concentration, the oscillation frequencies,
and the phase shift between the two oscillations
fl Ut)= 0 -K )A
F CF β' cos^t)-K A
F Cc £ B^cos are among the parameters that can be modified to
improve the behavior of the state K variables. The
k=l 1
effect of these parameters on the average steady
k=l state curves are presented in Figures 3-6. In all
[Κ(ωΐ-ωσΐ)3-Κ Α Αρ £ [a£sin(kujt)+B£cos(tot)] these figures the dimensionless conversion is rep-
ρ(:0 resented with a solid line while the dimensionless
k=l temperature is shown with a broken line. The same
frequency is used for both inputs. The effect of
The coefficient fi' is the same for all cases but
h frequency is shown in the absence (Fig. 3) and in
a"„ and R * vary for each case:
K the presence of phase shift (Fig. 4 ) . As it was
β£ = (2/kïï)sin akïï observed with a single variable oscillations an
s increase in frequency (curve 2-4 correspond to
ajJ (l/kïï){-(l-a)ços(a+2c)kïï+cosakïï-acos(2a-a)kïï} frequencies of 8, 6 and 4, respectively) shifts the
average steady state curve closer to the steady
6J|=(l/kïï){(l-a)sin(a+2a)kïï+(l-2a)sinakïï state curve for fixed inputs (curve 1). The pres-
ence of phase shift (σ=-0.2) causes dramatic
+asin(2a-a)k*rr} changes (Fig. 4) and makes the effect of the fre-
quency variations much more pronounced. In Fig. 5
The resulting averaged equations are the effect of phase shift < Curve 2-7 corresponding
2 2 2 to <r=0,-0. 05,-0.1,-0.15,-0. 2,-0. 25) is presented.
S + Da exp F [(XF^l) + X F 3
Ζ
Λ - - z
1+ Q{1 + γ
2 Of the three types of phase shifts only Type 3
22 (α>-σ>0) gave promising results for this reactor
system. The effect of varying the flow rate ratio
-(ζ,+ζ^Π + λ γ (F* + F )] +
2 C (FJ
? + F )
22 is shown in Fig. 6. As the ratio R is increased
2 2 2 A
(Curve 2-4 corresponding to R = 4, 8, 12) the
ft
average steady state curve moves away from the
+ ζ Π ( ( X F + I ) + X F )1 + C (F + F )}
22 ι + γ
] 22 Y 2£ fixed input curve (Curve 1).
Ζ = - Ζ + S + bp Da exp F {1 + γ [(XF, - I)
2 2 ο α 1 Another achievement with vibrations in two input
2 2 22 2 ) ( 1 0
variables is the intersection of the negative slope
X F 3 F )3 part of the steady state curve with fixed inputs
+
2 + 7 + F 2) - ( z 1z+2) [ l + λ γ
C (F (F +
2 (Curve 1 in Fig. 7) by the average steady state
+ 2 ΖΠ +Ύ ( ( +
λ2 1Ρ +) λ)] 2 ρ+C ( F +)}
F curve (curve 2 ) . Since at the intersection point
1 2 α 1 2 7 Ï 2 the average steady state curve corresponds to
2 2 2 stabilized operation, stabilized operation is
achieved at the same residence time as that of the
z
3 = X{- z
3 Q+ BDa exp F
1 [1 + γ X ((F - l) + ? )
2 fixed operation. Not only the states (concentra-
( z + )2Ί+
( Ύ ( λ 2λ+ ρ( 1 )X )2F ) tion and temperature) but also the residence times
•+ l 2 22 α 2 " 2" 2 1 2 match. Therefore, operation at a closer vicinity of
the desired unstable steady state can be achieved.
+ z (l + y ( X F + (λ-2-XF^ )
?
C X (F + F )
2 l 2 2 2
2 2 Simulation studies using the reactor model indi-
cates two major advantages of vibration of two
+ C (2-X) (F + F ))]}
? 2 input variables. Figure 8 illustrates a case where
large oscillations in concentration and temperature
with 2 A are observed. The conversion achieved in this case
S - (1 - K )A (1/8) α V2k&l - K is substantially larger (0.88) than the conversion
F CF ρ £ FC C
2 2 k=l
corresponding to the average reactor temperature
(0.81). Hence at the expense of an increase in
{(Τ/δ) a β 6 ^ cos k2πσ + (1/4)α 2 «
p£ ρ ] β oscillation amplitude a separation of the conver-
k=l sion curve from the temperature curve is achieved.
+ Figure 9 shows the second contribution. Here the
conversion (denoted by the dashed line) is equal to
(6
1 26 2) 3
cos 2πσ + ( l / 2 ) a F ^ 6 B£ sin k 2πσ}
k that indicated by the corresponding temperature
k=l (solid line) but the amplitude of the temperature
00 00 swings is drastically reduced to about 10°C. Hence
a tradeoff exists between an increase in conversion
- K A A {(a /2) Ε *k Φ
F CF F over and above to that corresponding to the reactor
k-1 k=l temperature and a reduction in the amplitude of the
ind temperature swings in the reactor.
F + 1 n2 n
o - Ε 3 /(-Y) i - V4 4 f β* EXPERIMENTAL RESULTS
n=o
1 Vibrational control experiments support the results n
4 +1 η 2 of the analytical studies and simulations (Fig.
= £ (η+1)ζ^ /(-γ) 6 akïï 10). The amplitude of the oscillations in reactor
k- ( 2 / Λ ) s i n temperature are about 10°C or less. For a given
n=o value of phase shift and forcing frequency (σ=-
4 0.15, ω=8) the average reactor temperature in-
F 2 n(z1 +n + 1 ) a creases with the residence time τ. An increase in σ
2 "Ε 3 /(-Υ)" p - εΑρ from -0.15 to -0.20 results in the reduction of the
n=o average temperature (Fig. 11). A reduction in ω
from 8 to 5 causes an increase in the amplitude of
THEORETICAL RESULTS AND SIMULATIONS the temperature swings and a reduction in the
average temperature (Fig. 11). These results are
Simultaneous vibrations in two input variables much better than the results achieved by oscillat-
provide more freedom in shaping the characteristics ing reactant flow rate or concentration alone.
156 A. Cinar et al.

CONCLUSIONS ACKNOWLEDGEMENT

Forced vibrations in tvo input variables enable
achievement of stabilized behavior with small
swings in the controlled variables or with higher
conversions than the conversions corresponding to
the stabilized reactor temperature. These improve-
ments are due to the contribution of an interaction
term which includes explicit dependence on the
phase relationships between the two inputs (Sincic
and Bailey, 1980). This contribution does not lead
to an improvement for all possible values of phase
shifts and duty fractions. If α and σ are not
selected properly a worsening in the behavior of
the controlled variables may be observed.
Vibrational control creates an opportunity for a
trade-off between reduction of the amplitude in
reactor temperature oscillations and improvement of
the reactant conversion. It is a viable alterna-
tive for stabilizing reactor operation in the
unstable steady-state region. Since an open-loop
controller is used, vibrational control is attract-
ive for cases when measurements are too difficult
or involve considerable delays.
A1r
Supply
This work was performed under the auspices of the
U.S. Department of Energy. A.C. and K. R. are
supported under Grant DE-FG02-84ER13205, S.M.M. and
X.S. are supported under Grant DE-FG02-85ER13315.
REFERENCES
Bellman R., J. Bentsman and S.M. Meerkov (1983).
Nonlinear Systems with Fast Parametric Oscilla-
tions. J. Math. Anal. Appl. 97 572-589.
Bellman R., J. Bentsman and S.M. Meerkov (1984).
Vibrational Control of Nonlinear Systems. Proc
23rd Cntrl and Dec. Conf. 84-89.
Bellman R., J. Bentsman and S.M. Meerkov (1986).
Vibrational Control of Nonlinear Systems. IEEE
Trans. Autom. Cntrl, in print.
Bogoliubov, Ν. N. and Yu. A. Mitropolsky (1961).
Asymptotic Methods in the Theory of Nonlinear
Oscillations, Gordon & Breach, New York, NY.
Cinar Α., J. Deng, S.M. Meerkov and X. Shu (1984).
Vibrational Control of an Exothermic Reaction
in a CSTR: Theory, Simulation, Experiment.
AIChE Annual Meeting, San Francisco, CA. Paper
No. 101b.
Cinar Α., J. Deng, S.M. Meerkov and X. Shu (1985).
Vibrational Control of an Exothermic Reaction
in a CSTR: Theory and Experiments. Submitted to
AIChE J.
Cinar Α., K. Rigopuolos, X. Shu and S.M. Meerkov
(1986). Vibrational Control of an Exothermic
CSTR. 1986 ACC, Seattle, WA.
Meerkov, S.M. (1980). Principle of Vibrational
Control: Theory and Applications. IEEE Trans.
Autom. Cntrl AC-25 755-762.
Meerkov, S.M. (1982). Condition of Vibrational
Stabilizability for a Class of Nonlinear Sys-
tems. IEEE Trans. Autom. Cntrl AC-27 485-487.
Sincic D. and J. E. Bailey (1980). Analytic Opti-
mization and Sensitivity Analysis of Forced
Periodic Chemical Processes. Chem. Eng. Sci.
35 1153-1161.
t

St" S Ί

'Ώ Lsi 11/23

Adiabatic
CSTR

product stream
0 τ (sec) 42

Fig 3 . Effect of frequency when σ=ϋ

Figure 1 . Schematic Diagram of CSTR Network 1

-T-

Ca FO|

Fm
J L 0
(« σ)Τ Τ in Ε , t τ (sec)
+
Fig. 2. Noney mine trie rectangular pulse. Fig 4 . Effect of frequency and σ ( σ = - 0 . 2 )
 Vibrational Control of Chemical Reactors

1.0
CO

>0. 8
χ1- - - B=8.6 5 bp=0.7 5 τ = 17 R. =6
Rc =4 o=0. 4 σ=-0.15
x~O. 6
t/i
Fig 9. Stabilize d operatio n - smal l
amplitud e Τ oscillation s

0 ι (sec ) 42
τ=22 se c
Fig 5. Effec t o f phas e s h i f t σ

τ=18 se c

1 FI τ=17 se c
2 4
3 8
4 12
B=8.S 5 bp=0.7 5 τ=16 se c
R c=10 σ=0 ω=4
Ac=3 . 2 1 α=0.2
0 0
τ (sec ) 42
Fig 6. Effec t o f fee d flov rati o -ί-O
ο 70 M 210 230 0 3 5
TIME C SEC. 3
1 Fig 10 . Stabilize d operatio n - experimen t
1 σ=-0.15, ω=8

B=9. i bp=0.7 5
R*=8 R c=8 ω=4
α=0.2 σ=-0.01

1 FI
2 YC 1=20 sec , ω=8, σ=-0.20

60
0 τ (sec ) 42 u ιοο
Ui
Fig 7 . Intersectio n o f stabilize d state s
σ=-0.15, ω=8 Id

-τ 1 1 1

i=17 sec , ω=5, σ=-0.15

Β=8.65 bp=0.7 5
R.=6 Rc =20 ω=6
A, =2. 26 Ac =2. 0
α=0.3 σ=-0.2
120 180
TIME C S E C 3
t/i
Fig 8. Stabilize d operatio n - hig h conversio n Fig 11 . Stabilize d operatio n - experimen t
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

OPERATION PLANNING AND QUALITY

DESIGN OF A POLYMER PROCESS
Y. Moteki and Y. Arai
Showa Denko K.K., Tokyo, Japan

Abs trac t Moreover, it happened sometimes that some

p r o c e s s e s d e t e r m i n e d by the step C2) could n o t be
This p a p e r p r e s e n t s a report of our technical o p e r a t e d due to safety r e a s o n s .
a c t i v i t i e s to d e s i g n of qulity to p o l y e t h y l e n e
to make matching of m u l t i - u s e r * specification As mentioned above, previous method for
requirements. Two a n a l y t i c a l tools are u s e d : C D quality improvements h a s l i m i t a t i o n s in safety,
Mul t i-var la te statistical a n a l y s i s economical, speediness, and e f f i c i e n c y aspects.
and (2) P o l y m e r k i n e t i c m o d e l i n g a n a l y s i s . To cope with this p r o b l e m , n e w types of a n a l y s i s
were introduced w h i c h are as f o l l o w s :
(Introduc tion)

A polymer p r o d u c t i o n p r o c e s s generally requires One of13 these analysis is " m u 1 1 1 - v a r i a te

the high performances of CI) the operation statistical analysis" which uses statistical
stability, C2) safety , and (3) process models, By using this method, quantitative
e f f ic iency.
r e l a t i o n s h i p s b e t w e e n the p l a n t o p e r a t i o n factors
Fur t h e r m o r e , in this business of J a p a n , users
and the p r o p e r t i e s of resins can be obtained and
strongly demand the v a r i o u s specification of
be used for the i m p r o v e m e n t s of efficiency and
p r o d u c t p r o p e r t i e s for m a t c h i n g their own uses and
s p e e d i n e s s in p r o d u c t i o n trials.
at the same time, from the m a r k e t competitive
atomosphre, the quality improvement is
c o n t i n u o u s l y carried out to desigh the p r o d u c t as Another one is "Reaction engineering
a new "grade". analysis". By this method, the theoretical
We present our s t u d y , taking our low density interpretation to the results from the data
analysis above will be obtained and further
polyehylene p r o c e s s located in O h i t a Factory of
seeking for the location of the best operating
S h o w a Denko Κ. Κ. ,
condition which may exist b e y o n d the limit of
CD By the use of the p r o c e s s daily record as
p r e s e n t data will bo able to be in sight.
input and the m a c r o p h y s i c a l p r o p e r t i e s such as
νiscosities,etc. and the m i c r o p r o p e r t i e s such as
Because of this cooperative character of
m o l e c u l a r d i s t r i b u t i o n s , β t c . as o u t p u t , we applied
these two m e t h o d s , we have n o t an attitude of
the multi varate s t a t i s t i c a l m e t h o d s , a n d d e v e l o p e d
e x c l u s i v e way to this two d i f f e r e n t m e t h o d s . This
the mathematical models which qualitatively and
is a d v a n t a g e o u s . e s p e c a l l y for a p l a c t l c a l field
q u n t i t a t i v e l y e x p r e s s the f u n c t i o n a l r e l a t i o n s of
like ours. F i g u r e C I ) shows the s t r a t e g i c a l map
the process o p e r a t i o n a l c o n d i t i o n s to the d e s i r e d
of the a n a l y s i s of q u a l i t y d e s i g n , which includes
production properties.
a scope to dynamic data t r e a t m e n t s to the p r o c e s s
identi f icatlons.
(2) Meanwhile, we also developed the
polymerization reaction k i n e t i c m o d e l s including
COutline of P o l y m e r Production Process)
elementary reactions such as i n i 1 1 a t i o n , ρ r o p a 1 1 on
terminât ion.etc. to analyze molecular
F i g u r e C 2 , 3 ) shows a s c h e m a t i c figure of the
d i s t r i b u t i o n s and e v a l u a t e the s t a t i s t i c a l models
outline of our p o l y m e r i z a t i o n p r o c e s s , which is a
above through the c o m p u t e r s i m u l a t i o n .
typical high pressure radical polymerization
Reaction stabilities are also analyzed
p r o c e s s of an a u t o c l a v e reactor.
through the locus of m u l t i - s o l u t i o n p o i n t s .
We h a v e , s u c c è s fuily e s t a b l i s h e d the d e s i h n tool of
CTechnical Background)
process condition set-up to a desired product
p r o p e r t y which is p r o p o s e d by the e x p e r i m e n t lab.
Review of P o l y m e r New Grade Design Work

CWork Scheme to P o l y m e r Production Quality Analysis) High p r e s s u r e low density p o l y e t h y l e n e CHP-

LDPE) is p o p u l a r l l y used for film and lamination
The following two k n o l e d g e s are required p u r p o s e s b e c a u s e of its f l e x i b i l i t y , trasparency,
as the m a k e r s of p o l y m e r i c m a t e r i a l s to produce heat-sealingness at low temperatures, and its
polymers with demanded qualities steadily and manufacturing ρrocessabi1ity.
econora 1 c a 11 y.
For f i l m s , the m o s t Important s p e c i f i c a t i o n s
C D The i n t e r - r e l a t i o n s h i p s b e t w e e n o p e r a t i o n are optical properties, such as haze, gloss,
parameters and molecular structures and/or clarity etc., of cource in some purposes,
physical properties. e x t r u d i n g flow rate, p r e s s u r e . t e m p e r a t u r e at the
exit of m a c h i n e , temperature, toughness, and
C2) L i m i t a t i o n s on o p e r a t i n g conditions rigidity are also required. Haze is strongly
correlated to c r y s t a 11 in 1 ty, and it can be
The technological activity of quality improved by d e c r e a s i n g the c r y s t a l 1 inity Cthis can
Improvements of p r o d u c t s had been carried out in be achieved by increasing the n u m b e r of short
the following steps, as d e s c r i b o d later in the chain branchings) and/or by decreasing the
section of " T e c h n i c a l B a c k g r o u n d " in detail. difference in d e n s i t y CI.e. r e f r a c t i v e Index)
b e t w e e n c r y s t a l l i n e and a m o r p h o u s p h a s e . O n the
C D De te rm ina 11 on of the m o l e c u l a r s t r u c t u r e s other hand, the increase in the n u m b e r of high
a c c o r d i n g to the k n o w l e d g e of the r e l a t i o n s h i p s to m o l e c u l a r w e i g h t p o r t i o n s and/or these p o r t i o n s in
the p r o p e r t i e s long chain b r a n c h i n g s cause totally the increase
at p l a s t i c m a n u f a c t u r i n g p r o c e s s .
of haze Ci.o. d e c r e a s e in g l o s s ) , w h i c h came from
C2) E s t i m a t i o n of the m o l e c u l a r structures iter haze increases due to the increase in
determined from r e l a t i o n s h i p s to the factors of e l a s t i c i t y in m o l t e n state. C l a r i t y is b e l i e v e d
the plant process, which have been obtained to be from inhomogenious structures of micro-
q u a l i t a t i v e l y by a n a l i z i n g the a c c u m u l a t e d data of dispersed structure s i m i l a r to ,but much smaller
p r o d u c t i o n trials at the plants. than that of gel and of fish eye in film.
These i n h o m o g e n i o u s s t r u c t u r e s m a y be originated
C3) P r o d u c t i o n trial u n d e r c o n s i d e r a t i o n of from a very small amount of u l t r a high molecular
stability, economical efficiency, and other w e i g h t p o r t i o n Cmlcro g e l ) , which is o b s e r v e d by
p r o b l e m s in the p r o c e s s . drastic Increase in intensity of light s c a t t e r i n g
at small angle region. Molecular weight
Since thore are so many factors of candidates distribution s t r o n g l y influences e x t r u d i n g rate,
taken in p l a n t o p e r a t i o n s i d e , each of ones causes p r e s s u r e and temperature at exit of m a c h i n e .
a wide v a r i e t y of c o n t r i b u t i o n in the molecular
s truc tures.
So far, a n u m b e r of trials had inavoidably to be CData A c q u i s i t i o n s and E v a l u a t i o n )
made by c h a n g i n g the o p e r a t i o n factors by small
amounts at each trials. These p r o c e d u r e s were The information about the i n t e r - r e l a t i o n s h i p s
c o n s e q u e n t l y tedious and e c o n o m i c a l l y inefficient. between reaction conditions and molecular

159
160 Y. Moteki and Y. Arai

structures and p h y s i c a l s t r u c t u r e s are collected

and analized by foiling m a n n e r s . Hystgram of simple c o r r e l a t i o n coefficient
was obtained by using c o r r e l a t i o n matrix. Plus
(1) P o l y m e r i z a t i o n Condition (Production) and minus sign of the coefficients were
distinguished by the p o s i t i o n of * mark in 20
All the sample used for our analysis were columns in each d i r e c t i o n s . In our extended
m e a s u r e d at the p l a n t s i t e , at every hour. BMDP-2R program, the correlation strength was
examined according to a table form,which we
The following p a r a m e t e r s were used as factors dev ised.
of ρ ο l y m o r i z a t i o n c o n d i t i o n s . For example, Figure ( 4 ) shows that, melt
flow index (MI) of named here "Tg grade" has
(1) R e a c t i o n P r e s s u r e negative c o r r e l a t i o n to polymerization pressure
(2) R e a c t i o n T e m p e r a t u r e and monomer temperature .while strong positive
(3) Monomer Feed T e m p e r a t u r e correlation to reaction temperature. MI has
(4) Monomer Feed A m o u n t .further, weakly correlated to monomer feed
(5) Spieces of P r o p a g a t i o n A d d i t i v e s rate,and to présure and temperature at after
(6) Amounts of P r o p a g a t i o n A d d i t i v e s reactor s ep a r a te r (HPS) .
(7) Amounts of Initiators These results are same as the results known
(8) Power of S t i r r i n g in e x p e r i e n c e and can be e x p l a i n e d qualitatively
(9) Pressure in High P r e s s u r e Sep ara tor(HPS) by the k n o w l e d g e of reaction engineering. This
gives encouragement that the multi variate
(10) T e m p e r a t u r e in H P S
analysis can be a p p l i e d to such a technical
sys tern.
(2) M o l e c u l a r Structure and P h y s i c a l Properties of
Polyme rs
(2) Figure of D e n s i t y of C o r r e l a t i o n Matrix
Molecular structures and parameters of
physical properties d i s c u s s e d in this study are The obtained c o r r e l a t i o n coefficients were
devided into several groups and marked after
shown as follows together with m e t h o d s to obtain
arranging in order from the largest number of the
these p a r a m e t e r s .
absolute value of a c o r r e l a t i o n coefficient.
By using this technique, the outline of
M o l e c u lra S t r u c t ue r a nd c o r r e l a t i o n can be o b t a i n e d , where p l u s / m i n u s of
M e a s u r a et n M e t hdo
" P h y s i cl a P r o p e r tsi e coefficients and correlation strength are not
displayed. (Fig. ( 5 ))

Mv, Mn, MV
D
1GPC J(SDK) (3) Correlation between Prarame te rs of
Polymerization and Molecular Structure and/or
S h o rt C h ani B r a n c hgi n C-MNR
, IR Macro P h y s i c a l P r o p e r t i e s
U n s a t u r adt e G r o usp " C - M N,R IR
Among the pairs of the parameters of
L o ng C h ani B r a n c hgi n " C - M N,R [ » 7 ] - GCP molecular s t r u c t u r e and/or p h y s i c a l p r o p e r t y and
L i g ht S c a M e r i n gη- C] polymerization condition, the pairs with large
c o r r e l a t i o n c o e f f i c i e n t s are b l n d e d by solid lines
D i s t r i b u tni o of SCB
, LCB
. Colun F r a c t i o n a nt i o and the ones with small c o r r e l a t i o n coefficients
are blnded by b r o k e n lines according to the result
of the c o r r e l a t i o n m a t r i x to get general v i e w of
V i s c o s iyt I n s t r no R h e o n ert e the c o r r e l a t i o n b e t w e e n the p a r a m e t e r s .
(Fig. S >
Tm, T c, α DSC
ESCR BTL
For example, polymerization pressure has
MI J IS
positive c o r r e l a t i o n to Tm, Tc, , viscosity,
D e n s iyt J IS density, and ESCR, while it has negative
correlations to MI, n u m b e r s of branchings and
vinylidene, Mw, and Mw/Mn. By knowing the
m e n t i o n e d single c o r r e l a t i o n between one p a r a m e t e r
(Multi-Valiate S t a t i s t i c a l M o d e l i n g A n a l y s i s ) and all other p a r a m e t e r s , and by using multi
—Requirement of U n d e r s t a n d i n g to C a u s e and Effect r e g r e s s i o n formula, the e s t i m a t i o n of the quality
R e l a t i o n s between M u l t i f a c t o r s and the design of p r o d u c t s can be made more
easily.
In the case of H P L D P E , m e l t i n g p o i n t (Tm) and
density can be estimated farely accurately as (4) Mluti gresslon Analysis
already m e n t i o n o d , and these results can be used
in some actual cases. However, molecular In this soction, we want to show how the
weight, m o l e c u l a r w e i g h t d i s t r i b u t i o n , long chain p o l y m e r i z a t i o n p a r a m e t e r s affect to s t r u c t u r e s and
branching etc., which influence ρroceesabi1ity, p r o p e r t i e s by using multi r e g r e s s i o n a n a l y s i s .
transparency, and mechanical property, are not
related totally to the p o l y m e r i z a t i o n conditions Here, multi regression formula was obtained
such like as the case of Tm. by stepwise multi regression calculation. The
actual calculation was made by the following
The change in even only one of many steps. First, the input of the single
polymerization conditions, such as pressure, correlation numbers was made in order from the
temperature p r o f i l e , m o n o m e r feed a m o u n t , causes largest one, then the s t r e n g t h of the c o r r e l a t i o n
the changes not only in Tm and d e n s i t y , but also between the n u m b e r s were Inspected a c c o r d i n g to
m o l e c u l a r weight, molecular weight distribution, their d i s p e r s i o n ratio, and invalid v a r i a b l e s were
n u m b e r and d i s t r i b u t i o n of short chain b r a n c h i n g s . neglected. Finally, all the v a r i a b l e s were put
Therefore, we must k n o w h o w much and by which into the calculation process (step back
operating condition's change, the mentioned inspection) again to n e g l e c t v a r i a b l e s with small
molecular structure and p h y s i c a l p a r a m e t e r s are c o r r e l a t i o n s t r e n g t h than the standard.
changed, and finally u n d e r s t a n d the reasons of the
changes. O t h e r w i s e , a p r o d u c t which may be used (5) E s t i m a t i o n from M u l t i R e g r e s s i o n Formula
in a particular purpose, can not be used in some
other p u r p o s e s at all. It is required to find Comparison b e t w e e n the e s t i m a t e d parameters
out optimum values of all the parameters. For of v a r i o u s s t r u c t u r e s and p r o p e r t i e s of products
this purpose, tho relationships between by the m e n t i o n e d muti r e g r e s s i o n formula, and the
p o l y m e r i z a t i o n factor (independent v a r i a b l e ) A and corresponding measured parameters were made
structure factor (dependent v a r i a b l e ) Β and these Table ( 1 ) shows the example of the results.
qantitative r e l a t i o n s h i p s b e t w e e n each variables
m u s t be obtained. Generally, for thermal p r o p e r t i e s (melting
temperature of crystal,, crystallization
Multivariable statistical modelling analysis temperature, crysta 11 inity etc.) and some of
molecular structures (short and long chain
was . t h e r e f o r e , made for high p r e s s u r e p o l y e t h y l e n e
branchings) of p r o d u c t s , the e s t i m a t e d value by
(references: F1A-B1, F1B-46. 107).
the multi r e g r e s s i o n formula and the measured
Original programs used for the calculation values were s a t i s f a c t o r i l y c o i n c i d e n t each other.
were BM-DP4M (prime conpornents analysis) On the other hand, for the p a r a m e t e r s related
including BM-DP2R (multi regression a n a l y s i s ) and to m o l e c u l a r w e i g h t and rheology, the coincidence
a factor analysis p r o g r a m , which is one of the b e t w e o n estimated and m e a s u r e d v a l u e s was less.
BMDP of 1979 versions (extended mutivariate )A Ultimately, good estimated values can be
analysis p r o g r a m of C a l i f o r n i a u n i v e r s i t y ) . These
programs were m o d u l a t e d a c c o r d i n g to purposes. obtained from the data of p r o p e r t y and structure
As a computer, Facom M 1 9 0 of Century Reserch with high accuracy. It must bo pointed out
C e n t e r (CRC) was used. that an extra caution is needed for the
c a l i b l t a i o n of the instruments.
The data used for the c a l c u l a t i o n c o n s i s t of
17 kinds of p o l y m e r i z a t i o n c o n d i t i o n s and 27 kinds Moreover, the same multi r e g r e s s i o n formula
of s t r u c t u r e s and p r o p e r t i e s . All these data should not be used, w h o n e v e r any of process
were used as input. innovation had been carried out to the process in
questioned,such as the reconstruction of the
(Primary Interpretation from Data Correction) Ρ lan ts , e tc .

—Factor C o r r e c t i o n Graph Inspection and Cause

Effect Generation.

(1) P l o t of C o r r e l a t i o n Coefficients
 Operation Planning and Quality Design of a Polymer Process

( PCA (Principle Compornent Analysis) A two demensional distribution map was

Application ) obtained by choosing arbitrary principal
c o m p o n e n t s as h o r i z o n t a l and v e r t i c a l axes and by
It was found that we can estimate the p l o t t i n g the score of the p r i n c i p l e c o m p o n e n t s .
structure and p r o p e r t y of the p o l y m e r farely well Fig. ( g ) shows an e x a m p l e of the map. Since
for p a r t i c u l a r p o l y m e r i z a t i o n c o n d i t i o n s from the this map itself does not e x p r e s s the m e a n i n g s of
multi r e g r e s s i o n formulas. On the c o n t r a l y , by principle compornents, we m u s t analyze the map by
m e a s u r i n g s t r u c u r e and p r o p e r t y of any r e s i n s , for using the factor load and its d i s t r i b u t i o n map.
e x a m p l e , resins of other c o m p a n i e s , it is p o s s i b l e Fig. ( Î ) shows an example of the d i s t r i b u t i o n map
to seeking the o p e r a t i n g c o n d i t i o n s h o w to p r o d u c e of the factor load.
it with such a p a r t i c u l a r quality realized to our
ρ lan t.
1.7.Analysis of P r i n c i p a l Compornent
Even only the o b t a i n e d informations m e n t i o n e d
TableC 3 ) shows the c h a r a t e r i s t i c s of each
above are v a l i d for the improvement of efficiency
principal compornent. The first principal
of a trial production and for education to
compornent Includes variables such as
operators (by telling tho importance of o p e r a t i o n
polymerization temperature and long chain
for the q u a l i t y of p r o d u c t s ) , h o w e v e r , this m e t h o d
b r a n c h i n g s in plus d i r e c t i o n and v a r i a b l e s such as
has some d i f f i c u l t y to apply, because of the
c r y s t a l 1 in 1ty and density in. m i n u s direction.
correlation among operation factors, i.e.
The first principle component includs
correlations among independent v a r i a b l e s and, of
course, the existence of the same independent polymerization and s t r u c t u r e / p r o p e r t y factors in
v a r i a b l e s to each d i f f e r e n t d e p e n d e n t v a r i a b l e s of both plus and minus d i r e c t i o n s , and has 0.52 of
s t r u c t u r e and p r o p e r t y . c o n t r i b u t i o n ratio.

The second p r i n c i p a l c o m p o n e n t has a strong

Therefore, for the p r o d u c t i o n of the resins
characteristics in plus d i r e c t i o n which include
with required quality by controlling this
variables of two structures such as methyl
complicated system well, it is very important to
branchings and v i n y l ends and one polymerization
have systematic view of this phenomena and
factor such a feed a m o u n t of propylene. This
understand their p o t e n t i a l reasons and to expess
means that the feed of propylene causes the
them n u m e r i c a l l y .
increase in methyl branchings due to
c o p l y m e r i z a t i o n and the increase in v i n y l ends due
For this purpose, we applied principle
to the t e r m i n a t i o n of p o l y m e r i z a t i o n .
component analysis for categor1 gica 1 grouping
concept seeking with so called p r l c l p a l axes w h i c h
help us to make sense of p h e n o m e n a in less number (Dominant Reduced Axis and Cate go 1igica 1
of axis. Segrigatlon to Find the D i r e c t i o n of Improvement
to Q u a l i t y )
1. P r i n c i p l e C o m p o n e n t A n a l y s i s of All G r a d e s
1.1 E i g e n Value and A c c u m u l a t e d C o n t r i b u t i o n Ratio The prime p u r p o s e of the muti varlate
analysis is to s i m p l i f y tho c o m p l i c a t i o n of data
Table ( 2 ) shows Eigen values and and to c h a r a c t e r i z e the i n f o r m a t i o n , and not to
get a b s o l u t e a n a l i t i c a l s o l u t i o n and a m p l f i c a t i o n
accumulated contribution ratios of principle
of the imformation.
components obtained from the c o r r e l a t i o n matrix,
which is a data matrix. Since the principle
The mutl v a r i a t e a n a l y s e s were applied to the
m a t r i x whose E i g e n v a l u e is less than 1.0 does not
s y s t e m of high p r o s s u r e p o l y e t h y l e n e , and single
seem to p r e s e n t c h a r a c t e r very w e l l , it is enough
correlations between variables wore obtained.
to consider up to the seventh p r i n c i p l e c o m p o n e n t . During the a n a l y s i s it was found that there are
Accumulated c o n t r i b u t i o n ratio up to the seventh c o m p l i c a t e d c o r r e l a t i o n s among the v a r i a b l e s .
component is 0.88, which m e a n s that 88 % c h a r a c t e r However, we found that, in the principle
of all the samples can be e x p l a i n e d by the seven c o m p o n e n t a n a l y s i s , 88 % of the c h a r a c t e r i s t i c s of
c o m p o n e n ts. complicated phenomena of high pressure
polyethylene and their p o t e n t i a l reasons can be
1.2 Score of P r i n c i p l e Components of Samples e x p l a i n e d by only seven c o m p o n e n t s . Moreover,
even only the first c o m p o n e n t , which c o n s i s t s of
The score of each sample for the seven reaction temperature and long chain b r a n c h i n g s in
principle components were inspected. 1st plus d i r e c t i o n and crysta 11 inity, density, and
principal axis is an axis .drawn to the directon reaction p r e s s u r e in minus d i r e c t i o n , has 52 %
of the m a x i m u m data s c a t t e r i n g in the space of N- contribution ratio. This means that the
th d i m e n s i o n (where Ν is n u m b e r of factor chozen production of high pressure polyethylene is
). 2nd p r i n c i p a l axis m e a n s the axis drawn to 2nd controlled about 50 % by the p r i c i p l e that the
maximum data s c a t t e r i n g d i r e c t i o n and so on. So increase in reaction temperature causes the
each of sample data has the value to each of the increase in the number of branchings, which
P.rlcipal axis as a p r i n c i p a l component (score). reduces crysta 11 inity (density), and that the
By making a hystgram which expresses the decrease in r e a c t i o n p r e s s u r e reduces n u m b e r of
principle score by sample n u m b e r and which uses b r a n c h i n g s and causes the increase in d e n s i t y and
each priciple c o m p o n e n t as a side a x i s , we can crystallinity.
know the d i s t r i b u t i o n of the p r i n c i p a l c o m p o n e n t s .
Fig. ( 7 ) shows the example of the principal
These results suggest that the
component hystgram of the first and the third
principle components. c h a r a c t e r i s t i c s and their p o t e n t i a l reasons of the
c o m p l i c a t e d p h e n o m e n o n seem to be quite simple by
a n a l i z i n g from the m a c r o s c o p i c v i e w p o i n t s .
1.3 Eigen Vector of Each Variables for Seven
Principle Components
By using the p r i n c i p l e c o m p o n e n t analysis,
such as tho score of p r i n c i p l e component, factor
The score of each, p r i n c i p a l c o m p o n e n t s can be
load, the score of factor load, and their
c a l c u l a t e d from tho eigen v e c t o r of each v a r i a b l e ,
d i s t r i b u t i o n m a p s , we found the m e t h o d to estimate
and are shown in T a b l e ( 2 ). The E i g e n vector
the o p e r a t i n g c o n d i t i o n s of p r o d u c t with a r b i t r a r y
can be used for the planning of production
quality (resins with improved q u a l i t y ) and u n k n o w n
condition and for the c o n s t r u c t i o n of a formula
production conditions of products of other
estimating physical property through the
compan les.
distribution figure of the p r i n c i p a l components.

In the d i s t r i b u t i o n map of the score of the

1.4 Single Correlation Coefficient (R) between principle components, each grade, such as A O , T c ,
Used Variables and Tg can be devlded into c l u s t e r s , h o w e v e r , the
overapping p o r t i o n s b e t w e e n the clusters exist.
When tho c o n t r i b u t i o n ratio of a single This result introduces the idea that, when only
correlation coefficient, which is o r i g i n a t e d from one quality is required, we can find a m e t h o d to
self c o r r e l a t i o n , b e c o m e s 1.00, the v a r i a b l e s were improve other q u a l i t i e s also.
neglected from the calculation, because these
variables make the inverse m a t r i x singular.
(Computation of R e s u l t s ans its Validity)
1.5 F a c t o r Load

Factor load is a correlation coefficient Since this is a research in one private

between each v a r i a b l e and a p r i n c i p a l component. company, we can not report in very detail.
Therefore, as in the case of multi regression Here, we will show a few e x a m p l e s in which we
analysis, where the square of multi correlation could estimate the c o n d i t i o n s of the target and
coefficient (R) gives a contribution ratio, made now grades as a result by using the
contribution ratios of 41 variables to each information obtained by the multi variate
principal component can be obtained from the
analys is.
square of the factor load.
(1) L a m i n a t i o n Grade for Low T e m p e r a t u r e Adhesion
Table ( ) shows the factor load of 41
variables for each p r i n c i p l e component. The Since the increase in the ease of thermal
variables which have a large factor load degradation by introducing CI branchings, the
c h a r a c t r i z e the p r i n c i p l e c o m p o n e n t . Table ( ) copolymerlzation with p r o p y l e n e e n h a n c e s a d h e s i o n
shows the c h a r a c t e r i s t i c s and eigen v a l u e s of each to Al type base materials even at low
main c o m p o n e n t . temperatures. B e s i d e , the e x i s t e n c e of p r o p y l e n e
reduces molecular weight because of its
1.6 Distribution Maps of Scores of Principle propagation effect. As a result, high speed
C o m p o r n e n t and F a c t o r Load Ρrocessabi11ty improves, hoever, n e c k - i n becomes
larger. To decrease molecular weight, the
increase in polymerization pressure and/or the
162 Y. Moteki and Y. Arai

decrease in reaction temperature should be made the m e t h o d of A m u n d s e n ' s formulation

usually, however, these operations reduces which m o d e l i n g has
molecular weight d i s t r i b u t i o n and also cause the (a) limiting c o n v e r s i o n of discrete molecular
Increase in neck-in. chain to c o n t i n u o u s connection

Therefore, mutual balance among Cb) L a p l a c e transform t r e a t m e n t under the

p o l y m e r i z a t i o n temperature, pressure, and amount assualysIs.
of propylene is very important for the Improvement
of low temperature a d h e s i o n and to keep small
neck-in and good h i g h speed prοcβ s sab 11ity. MODEL FEATURES

According to the result of muti variate I initiation,propations

analysis of the c h a r a c t e r s of the existing grade, te rmlna t ion
we estimated polymerization temperature, II termination (recombination,
temperature profile, amount of propylene, d l s p r o p o r t l o n a t i o n ) with M O D E L I
III initiatlon, p r o p a g a t i o n and chain
distribution of feed monomer at each level,
trans fer
monomer temperature, and reaction pressure to
IV (MODEL II + M O D E L III)
produce the designed resin which has optimum
with energy b a l a n c e
molecular weight, molecular weight distribution,
V e m p i r i c a l r e a c t i o n cell series
and the number of CI b r a n c h i n g s . n e t w o r k c o n n e c t i o n s of
VI
c o m p o s i t e r e a c t i o n cells of M O D E L IV
After the safety inspection of the obtained
operationg conditions in the actual plant, (model scheme to M O D E L IV and it's simulations)
production trials were m a d e , and we could o b t a i n
the aimed grade at the second trial. This grade T h r o u g h the s e n s t l v i t y checks of the previous
has a slightly larger neck-in than ordinally stage m o d e l , it was p r o v e d the feasibility of the
grades, but possesses good high speed ideas of m o d e l i n g s c h e m e .
processabi1ity and excellent adhesion at low M O D E L IV have b e e n d e v e l o p e d in order to represent
tempe ratures. the more actual reaction which s toichiornetry is
p r o v i d e d as f o l l o w s ;
(2) Injection Grade with High ESCR and High Gloss
kd
I 2Ix degradation
Many users require the injection grade with
high gloss and ESCR. For injection purpose,
high flow property Is required, however, the
improvement of the flow property is limited, Ix XI initiation
because it reduces E S C R due to the decrease in
m o l e c u l a r weight.
For the improvement of gloss, smaller XJ + M
k
P X j+1 propagation
m o l e c u l a r weight and m o l e c u l a r w e i g h t d i s t r i b u t i o n
are required.
XJ + M XI PJ
To solve these contradicting requirements m o n o m e r chain
which are generally believed, we applied the trans fer
result of the muti v a r i a t e analysis. Xj + Ps

Tho data of gloss were not included in the Xs + Pj

multi variate a n a l y s i s , however, it is already p o l y m e r chain
known that gloss is related to ms1t elastisity. trans fer
T h e r e f o r e , amount of high m o l e c u l a r w e i g h t p o r t i o n
must be small and n a r r o w M W D are required. The Xj XI +Pj solvent chain
multi regression formulas of Mw and MWD first tran fe r
require low p o l y m e r i z a t i o n temperature.
At low reaction t e m p e r a t u r e s , Mw and MWD XJ XI +PJ chain transfer
become small and n a r r o w , however, the use of agent transfer
propagation additives increases polymerization
pressure, and as a result, increases density Cat Xj Ρ i+j recombination
the same m o l e c u l a r w e i g h t ) and decrease ESCR. terminât ion

All these results suggest that optimum Xj PJ Disproportionation

operation c o n d i t i o n may exist in c o n d i t i o n s at low te rminat ion
temperature and pressure witho the use of
p r o p a g a t i o n additives.
I :Initlator concentration
According to the result of the muti variate M :monomer c o n c e n t r a t i o n
analysis of various grade produced at low Ρ :produced p o l y m e r c o n c e n t r a t i o n
temperatures and pressures, we first estimated X :ràdical chain c o n c e n t r a t i o n
molecular structure and v a l u e s of basic physical k irate cons tant
properties, which give the aimed quality. And suffix
then, by using the result of the muti variate i.j :degree of p o l y m e r i z a t i o n
analysis (multi r e g r e s s i o n f o r m u l a ) , we decided (number of m o n o m e r u n i t )
optimum operating c o n d i t i o n s which produce the c
aimed structure and p r o p e r t y , and carried out For c o m p u t a t i o n s t r a t e g y ,
trial p r o d u c t i o n s at the plant. After a
trials, the requied resin was obtained and
few
is
(a) v a r i a b l e s incidence matrix(ex. Fig. \
and / 10 )

p r o d u c e d in c o m m e r c i a l b a s e s right now. (b) a function digraph

are developed. These are quite h e l p f u l to avoid
m i s t r e a t i n g of c o n s t r u c t i o n of mass and energy
(Reaction K i n e t i c s Model Study) balance equations and to make the work more
efficient.
--for confidence r e c o g n i t i o n s to statistical Each of rate c o n s t a n t s are used from the data of
analysis work Reference ( ) and our laboratory
m e s u r e m e n t . Fig. Ii shows a s i m u l a t i o n result
H P L D P E is s y n t h e s i z e d at high p r e s s u r e and of molecular weight distribution under various
temperature,and the change of these p o l y m e r i z a t i o n temperature and stage reactions.
conditions cause the changes of molecular We have found a d i f f i c u l t face to c o m p u t a t i o n at
strue tu re s,such as molecular weights, the s i m u l a t a n o u s treatment of energy equation to
distributions, n u m b e r of b r a n c h i n g s , e t c . of the the mass b a l a n c e s . These are of a p r o b l e m of b i -
resulting polymer. A trial p r o d u c t i o n are c a r r i e d furcations of the solution. Fig. 12. shows a
out at the real p l a n t , e v e n though we realized it as triple of reaction temperatures. In this figure,
inefficient one. Yl-curve means energy values of reaction vs.
temperature and Y2-curve means cooling energy
To cope with these p r o b l e m s , we analized the curves. We understand that these phemomena is
relationship between reaction factors (reacti- s u b s t a n t i a l one from the p l a n t o p e r a t i o n experiences.
on rate constant,me an residence time ,e tc.)and So far, we made the various combinations of
m o l e c u l a r s t r u e t u r e s ( m o l o c u l a r w e i g h t .etc.) v a r i a b l e s such as c o o l a n t temperature (Fig. 12 )
The mass balance e q u a t i o n s of s t a t l o n a l state were and monomer feed rate (Fig. 13 )
built under the h y p o t h e s i s of C S T R reactions for for s i m u l a t i o n trial.
each of uni t cell. From the series work, the p r e - e v a l u a t l o n to
The aim of analysis is to u n d e r s t a n d the r e l a t i o n - r e a l i z a t i o n of p l a n t o p e r a t i o n is found improtant
ship b e t w e e n one,even though we have found the conditions of
(a) reaction factor such as rate c o n t a n t s and the s p e c i f i e d p o l y m e r p r o d u c t p r o p e r t i e s from the
molecular distributions results of data analsis and k i n e t i c simulations.
(b) b e h a i v i o r of m o l e c u l a r d i s t r i b u t i o n s and
one from the k n o w l e d g e from the stastical
da ta analys is. (Cone lus i ons)

(Reaction K i n e t i c Model ) A polymer production requires the high

performances of operational stability, safety,
M a n y of the m a t h e m a t i c a l m o d e l s were p r o p o s e d e f f i c i e n c y , and p r o d u c t quality.
from 1950's to I960's on the p r o c e s s of p o l y m e r s ,
which consists, of 'the v a r i o u s sets of e l e m e n t a r y To answer these r e q u i r e m e n t s , it is very Important
r e a c t i o n s . Among them,we Introduced b a s i c a l l y to know the o p e r a t i o n factor with relation of the
target p r o d u c t quality. A f t e r c o n s i d e r i n g a great
 Operation Planning and Quality Design of a Polymer Process 163

number of analytical methods and it's trial Η RtO.IJ = Û-ÏIÏ

applications, we chose two of them; M u 1 t i - v a r i a t β PHI - El
HI0.05)« O.ZÎZ
s t a s t i c a l a n a l y s i s and r e a c t i o n k i n e t i c a n a l y s i s .
- ss -0.5 0.0
-1 0.5" 1.0
As a result of multi~variate analysis, the YAK. NO.
1) • P( 1)
VAR.KAME . I . *
correlations between operation parameters and 2) T120<2) I 1
various m o l e c u l a r s t r u c t u r e s of p l a n t operation 3 ) TL J0(J) I 1
condition were found to produce the target 7L*0{*) 1 1

product resins with required properties, and

IV KSO(S)
FT ( 6 ) 1 ··
1 1
ι
some new grades are successful to come to 7 f~ CC<7 ) 1 · i

realiza t ion. <f) HPS!IV) I1 ι ι

10) ι ι
11i [ NM ΓΙ Ϊ J

Îi i
In tho reaction kinetic modeling analyis 12) PURCI12)
ι 1
starting from Model I which considered very 13 >
n )
"ôFÎUt iJ )
OF(1 * )
......... ι ι
simple e l e m e n t a r y r e a c t i o n s , we finally d e v e l o p e d IS) Of3011S) 1 ..... 1
the Model IV with energy e q u a t i o n s and for more OF«0(It)
1 ι
practical
simulation
purpose,
system
the
of
Model
network
VI which is
calculations
the
to
ΐ!Γ
18 )
TCIAV tiTI
MJ(18)
. 1 I I I
I I
ι
1
c o n n e c t o n of m o d u l a r k i n e t i c u n i t s . r>) Hwimi ! ···· 1
f
21 )
in
Mink
. V1ND12»)
1 1 ..
I1
ι
ιI
Making the s i m u l a t i o n by the Model IV,s e n s i t i v i t y 1.4..·
of various p a r a m e t e r s to m o l e c u l a r weight were
2A )
22
25))
1ERV12*)
TKSV(25> I
J 1
I
obtained, and stability of the reaction was ET10I26) I'""..
Ε 100
(27 >· 1

1
27)
examined. It was found that the system is E1S* 1 25 ) 1
mëtastable at the p r e s e n t o p e r a t i o n temperature, - 2 Ï" El'/SI 29) 1 ··
T«P(30) 1 «·
then more p r e c i s e c o n t r o l such as m u l t i v a r i a b l e 7«E(3))
control is consequently required to maintain j 1) AIP«(32)
1 ··

stable and safe p l a n t o p e r a t i o n u n d e r the target 33 ) TCS(33) 1

1
3* ) TCP(3*)
of p r o d u c t q u a l i t y r e q u i r e m e n t . " 34 ί ALPCI3S) 1 ··
it ) ESCM36) 1 ·· *"!
37 ) MEU137)
Though all the work above st, 3i ) C1I36) I t
r e a l i z a t i o n of so called "new g r a d e " p r o d u c t is
*0) C«(«U) 1 1
successully acomplished. Ci t*1> ' 1
] 1I
42 ) iCb(«2)
ICMO)
We strongly r e c o n g i z e d that the systematic

i.„_
CBUt«* )
work over the span from the u s e r s r e q u i r e m e n t to
product properties and to the plant operation
Fi
technology are e s p e c i a l l y crucial which w i l l be «- 4 Correlation Coefficients between Vôf.Uo. Π and -the Other Variables
essentially coped with the overcome to a newly
coming " q u a l i t y era". 32 ALPMl32 )
19 RHÛC19 > ι
30 T U P ( 3 0 )
2 TL2012)
3 TL3013)
Acknowledgement 35 ALPC135) «ΚΚΚΚΚ
10 C4C40) 8S*X8XH
We thanks for all of tho c o l e a d g u e s 31 THE 131)
in our company who c o n s t a n t l y worked with ,and 34 TCP134)
greatly e n c o u r a g e d u s . 33 TCS133) ïK£lt ί
17 TUAVl17) S I S X Ï S X SXS S
39 C2139) ISSSSX 2 SÏ X SÏ
41 C5141) Χ Ε Χ Κ Η Χ Κ Χ ΗΪ Χ Κ Χ
23 VIND123) KÏKKSKKSKGKliiîi
37 ME.TLl 3 7) Β*ΪΧ3ϊ3ϊ»ΗΚΒ88ίί
22 MWMKM22) SSXXEXXSNNKNSMKS
43 LCB(43). XKKX2ÎKXSKXKIU8X1
8 BClB) BSKNNXXKKtaXXXXÈCXxa
ι pc i ) axaxxxsNENSiSBsaNs + a
4 TL4014) NMXNNKStOîNXiiSXMXN + X X
5 TL5015) XNNXXXSNNHKEXXNKX + KXil
20 MW120) XXXXXXXXXXXXX + - S + +- + XXS
16 0F40116) NXKNNXMXNXKNNNXXX+NXX+î
25 IRSV(25) NKXXXXXKXXXX+NN-XXX++.-X
27 El00t27) XXXXXXKXXXXNXXN.N.XXX.+XK
28 ET5*(28) XXXXXXNXXXXKX!C>Î.K.KXX.+ + aK
29 E1/5C29) XXNXXXNXXXXNXKN.N.NXX.+XSÏS
16 ΜΓl18) XXKXXXNXXXXNXKN-N.NXX
21 MN12.1) xx+xs
36 ESCR136) - .XKXNX2
38 Cl(38)
24 TERV124)
7 CCt7)
10 HPSP(10)
13 QFTL113)
9 HP3T(9) . .+·-£
12 PURC-( 12J -..8
6 FT C 6 ) . .-. -..- . +
14 QF(14) S
15 ÛF30(15) +
- . .. .. +8
11 INR II 11) -· - · -a
/Process Control )
Λ Dcslcn J THE ABSOLUIE VALUES OF
THE MATRIX ENTRIES HAVE BEEN PRINTED ABOVE IN SHADED FORM
ACCORDING TO THE FOLLOWING SCHEME
Fig.l Stratetice Xap of PoLyner Process Quality Design LESS THAN OR EQUAL TO 0 . 125
0.125 TO ANO INCLUDING 0.249
- 0.249 TO AND INCLUDING 0.374
+ 0.374 TO AND INCLUDING 0.499
Χ 0.499 TO ANO INCLUDING 0.623
M 0 . 6 2 3 TD AND INCLUDING 0.748
,00 a 0.748 TO AND INCLUDING 0.873
χ GREATER THAN 0.873

Fig.5. ABSOLUTE VALUES OF CORRELATION'S IN" SORTED ANp SHADED FORM

£ • LtZl ί
(a) (b) (c) (d) (e) (f) (g)

a) Primary Compressor
b) Secondary Conpressor
c) Auto clave Reactor
d) High Pressure Separator
e) Monomer Recycle Line
f) Pelletizer
g) Product
h) urease Pot
T,c ) ( cU ,flc)
Tm
. Pressure
Fig. 2 Outline of a polymer Prosess
-OS

Fig. 3
I: Initiator,
ι

Series Reactor Model

'
H: Monoaer
Fig. 6. Corelation betwen Parameters
{Ci,C±.Cs) ( LCb) ( V,nylid&ne )
164 Y. Moteki and Y. Arai

! 1
1 1
Table 1. TVedîcfad Value Tm, Tc, fron Multi-Reglession
! j
Formula
Sanple τ·. ce) t c. m
:
Mo. Obs. Est. Obs. Est.
-

•.-iJSnP"
•r- rV v
22 112.6 112.5 100.8 98.3
.... ,-it.t u
24 110.3 109.2 98.3 95.8
25 109.5 108.8 95.8 93.7 i \
26 109.9 108.8 90.9 95.8
27 111.7 112.3 98.0 97.9
28 111.1 111.1 97.6 97.9
29 109.3 108.3 96.5 96.0
30 109.2 109.4 96.5 96.0 !
31 108.3 107.3 91.0
j
92.6

j j
32 108.9 108.7 93.4 93.7
33 108.1 106.7 91.7 90.8
34 108.5 109.2 90.8 91.0
35 107.3 109.1 93.2 95.8 C I>
36 108.7 108.4 93.6 93.7
37 108.8 108.7 93.2 93.7 Fig.8 V a r i a b l e f a c t o r load distribution
38 107.9 106.7 90.0 91.0
with respect to principle axises
39 102.8 105.1 91.6 91.1
40 107.7 107.5 90.3 90.9
41 108.1 108.3 91.2 91.0 Fi
.g 9 t I hM
sxiaf
dtoer niV as r i a b l e
42 107.7 107.6 90.8 91.1
43 107.4 107.8 90.3 91.0 •(,) Xii
> PHI Pii)\ p,

Table 2 . Ei gen Value and Accumulated

Contribution
„Acçuipulate
Principle Eigen Value Contribution
1
1 21.3135 0.519
2 4.8608 0.638
3 4.0911 0.738
4 1.9936 0.786
5 1.4646 0.822
6 1.3859 0.855
7 1.1491 0.884

Table 3, Characteristics of each principle Component

Principle Accumulate
Characteristics Eigen Value
Component Contribution
1 C , C » C ». Dens. Crys.
a 4 s
Reactor Temp. Pressure 21.31 0.52
2 Terainal Vinyl
Transter Agent 4.86 0.64
3 Molecular Weight 4.09 0.74
4 Monomer Feed rate, HPS Tenp. 1.99 0.79
5 Mononer Teep. HPS Press. 1.46 0.82

Fit. 10 Digrap
h of computatio
n Schem
e

SIMULATIO
N OF MOLECULA
R HEIGH
T DISTRIBUTION
S FOR HPPF HODFI-
f

h
J/ \
wi 0. J IUil
iUilii n ,j J 1HI
M,j ι nun0.J \\Wa „. 1
D 11 Hill . ι mm
Ι .l UliUl
ΤΤΤΠίΤΤί
0 0 j
MOLECULAR HEIGHT
Fig. 7- Sample Scale d i s t r i b u t i o n w i t h respect

to p r i n c i p l e axises
Fig. J J Simulation of Molecular Weight Distribution
for HP-LDPE Model-(IV)
 Operation Planning and Quality Design of a Polymer Process 165

P a i - C h u a n , E h r l i c h , P. e t a l . ( 1 9 7 2 ) Monte C a r l o S i m u l a t i o n o f
S t r u c t u r e of L o w - d e n s i t y P o l y e t h y l e n e . I n d . Eng. Chem. Prod.
Res. D e v e l o p . , 1 1 , N o . 3 , 3 5 2 .
Th. J . V a n d e r M o l e n ( 1 9 7 3 ) The F r e e R a d i c a l B u l k P o l y m e r i z a t i o n
o f E t h y l e n e a t Low T e m p e r a t u r e . J . P o l v m e r S c i . i S v m P Q S i u m ,
No.42, 29.
R a m k r i s h n a , D. and A m u n d s o n , Ν. R. (1 9 7 4 ) S t i r r e d P o t s , T u b u l a r
R e a c t o r s , and S e l f - A d j o i n t O p e r a t o r s . Chem.Eng. S c i . , 2 9 ,
1 353.
N i c c o , A. ( 1 9 7 6 ) O r i e n t a t i o n e t P r o p r i é t é s I n d u i t e s d e s P o l y m è r e s
A L e t a t S o l i d . U n p u b l i s h e d CdF Chimie R e p o r t .
B o v e y , F. A. e t a l . ( 1 9 7 6 ) S h o r t C h a i n and Long C h a i n B r a n c h i n g
i n Low D e n s i t y P o l y e t h y l e n e . M a c r o m o l e c u l e s , 9 , N o . 1 , 7 6 .
L a p i d u s , L. a n d A m u n d s o n , N. R. (1 9 7 7 ) C h e m i c a l R e a c t o r T h e o r y a
Review. P r e n t i c e - H a l l I n c .
Harmony, S. C. e t a l . ( 1 9 7 7 ) C a l c u l a t i o n o f P h a s e E q u i l i b r i a f o r
Ethylene/Low- D e n s i t y P o l y e t h y l e n e M i x t u r e s . Α.I.Ch.Ε.
Journal, 23, No.5, 758.
B i r n b a u m , I . and L a p i d u s , L. ( 1 9 7 8 ) S t u d i e s i n A p p r o x i m a t i o n s
M e t h o d s - I . S p l i n e and C o n t r o l v i a D i s c r e a t e Dynamic
Programming. Chem. E n g . , S c i . 3 3 , 4 1 5 .
Temperature L e e . L. H. e t a l . ( 1 9 7 9 ) P o l y m e r i z a t i o n R e a c t o r s and P r o c e s s .
Chap. 1 0 , 1 1 , 1 7 .
M o t e k i , Y. ( 1 9 7 9 ) C o u p l i n m g o f GPC and A u t o m a t i c V i s c o m e t r y .
Fig. 12. Reaction Temp, vs Coolant Temp. A p p l i c a t i o n t o t h e S t u d y of High P r e s s u r e P o l y e t h y l e n e .
U n p u b l i s h e d SDK R e p o r t R 8 0 0 1 4 3 .
G u t i n , B. L. ( 1 9 7 9 ) C a l c u l a t i o n and A n a l y s i s o f t h e M o l e c u l a r
W e i g h t D i s t r i b u t i o n (MWD) A l l o w i n g f o r C h a i n T r a n s f e r t o
Polymer i n Homogeneous R a d i c a l P o l y m e r i z a t i o n Process.
Polymer S c i . U . S . S . R . , 2 0 , 6 9 9 .
HPP36 (TR=45-1088) <Λ3Ίο9τ7:Α=12000)_
S h o s t e n k o , A. G. e t a l . ( 1 9 7 9 ) The M e c h a n i s m a n d t h e K i n e t i c s o f
P o l y e t h y l e n e B r a n c h i n g S t u d i e d on a M o d e l S y s t e m . ..Polymer
Sci. U.S.S.R., 20, 1760.
4
_JR^IH/.ENTH,IPΠΝ, Y1 V_2, V3 :_S7...9.
K r e i t s e r , T.V. e t a l . ( 1 9 7 9 ) E f f e c t o f t h e P a r a m e t e r s o f t h e
M a t h e m a t i c a l Model of P o l y m e r i z a t i o n K i n e t i c s , On E s t i m a t i o n
1.4Et07 -j of t h e M o l e c u l a t W e i g h t of P o l y e t h y l e n e . Polymer S c i .
U.S.S.R., 20, 2989.
M o t e k i , Y. ( 1 9 8 0 ) Q u a l i t y D e s i g n o f H i g h P r e s s u r e Low D e n s i t y
P o l y e t h y l e n e I . U n p u b l i s h e d SDK R e p o r t R 8 0 0 0 2 8 .
I.2E + Q7 M o t e k i , Y. ( 1 9 8 0 ) Q u a l i t y D e s i g n o f H i g h P r e s s u r e Low D e n s i t y
P o l y e t h y l e n e I I . U n p u b l i s h e d SDK R e p o r t R 8 0 0 6 5 3 .
R a n d a l l , J . C. ( 1 9 8 0 ) C h a r a c t e r i z a t i o n o f L o n g - C h a i n B r a n c h i n g i n
P o l y e t h y l e n e s U s i n g H i g h - F i e l d C a r b o n - 1 3 NMR. A. C. S.
Symposium S e r i e s 1 4 2 , P . 9 3 .
L u f t , G. e t a l . ( 1 9 8 0 ) The E f f e c t i v e n e s s o f O r g a n i c P e r o x i d e
Initiators in the Polymerization of High Pressure
P o l y e t h y l e n e . J . Macromolecular S c i .
G o t o , S . , Y a m a m o t o , K. e t a l . ( 1 9 8 1 ) Computer Model f o r
C o m m e r c i a l H i g h P r e s s u r e P o l y e t h y l e n e R e a c t o r B a s e d on
E l e m e n t a r y R e a c t i o n R a t e s O b t a i n e d E x p e r i m e n t a l l y . J. Appl.
Polymer S c i . ' . A p p l i e d Polymer Symposium, 3 6 , 2 1 .
6Et06 4- Moteki, Y. e t a l . ( 1 9 8 1 ) Q u a l i t y D e s i g n o f HP-LDPE I I I .
U n p u b l i s h e d SDK R e p o r t R 8 1 0 5 3 0 .
B u d t o v , V. P. e t a l . ( 1 9 8 2 ) D e t e r m i n a t i o n o f The M o l e c u l a r Weight
D i s t r i b u t i o n of High P r e s s u r e P o l y e t h y l e n e and C a l c u l a t i o n

o f t h e Rate C o n s t a n t of t h e Main C h e m i c a l R e a c t o r s . P o l y m e r
S c i . U.S.S.R., 24, No.6, 1367.
Tanaka, N. ( 1 9 8 2 ) Program Package BMD and BMDP. CRC C o m m u n i c a t i o n
No.11 , V o l . 2 1 7 .
A r a i , Υ., M o t e k i , Y. e t a l . (1 9 8 3 ) A p p l i c a t i o n o f Statistial
M u l t i V a r i a t e A n a l y s i s i n C h e m i c a l I n d u s t r y . Kagaku Kogaku,
47, N o . 3 , 150.

•2Etee
600
Temperature

Fig.|3. R e a c t i o n Temp, v s Monomer f e e d rate

REFERENCES

R o d e l , M. J . (1 953 ) . J . Am. Chem. S o c , 7 5 , 6 1 1 0 .

B r y a n t , W.M.D. and V o t e r , R. C. (1 9 5 3 ) The M o l e c u l a r S t r u c t u r e of
P o l y e t h y l e n e I I . D e t e r m i n a t i o n o f S h o r t Chain B r a n c h i n g . _Jj.
Am. Chem. S o c . , 7 5 , 6 1 1 3 .
B i l l m e y e r , F. W. ( 1 9 5 3 ) The M o l e c u l a r S t r u c t u r e o f P o l y e t h y l e n e
I I I . D e t e r n i n a t i o n of Long Chain B r a n c h i n g . J . Am. Chem. S o c .
75, 6118.
B e a s l e y , J . K. ( 1 9 5 3 ) The M o l e c u l a r S t r u c t u r e o f P o l y e t h y l e n e IV.
K i n e t i c C a l c u l a t i o n s of t h e E f f e c t o f B r a n c h i n g on M o l e c u l a r
Weight D i s t r i b u t i o n . J . Am. Chem. S o c , 7 5 , 61 2 3 .
Sperati, C. A. e t a l . ( 1 9 5 3 ) T h e M o l e c u l a r S t r u c t u r e o f
P o l y e t h y l e n e V. The E f f e c t o f Chain B r a n c h i n g and M o l e c u l a r
Weight on P h y s i c a l P r o p e r t i e s . J . Am. Chem. S o c , 7 5 , 6127
B a m f o r d , C. H. and Tompa, H. ( 1 9 5 4 ) The C a l c u l a t i o n o f M o l e c u l a r
Weight D i s t r i b u t i o n s from K i n e t i c S c h e m e s . T r a n s . Faraday
S o c . , 5 0 , 1097
Liu, S h e a n - L i n a n d A m u n d s o n , Ν. R. ( 1 9 6 2 ) C a l c u l a t i o n o f
M o l e c u l a r Weight D i s t r i b u t i o n i n P o l y m e r i z a t i o n . Chem. Eng,
S c i . , 17, 797-802
B i s c h o f f , Κ. B. and L e v e n s p i e l , O c t a v e ( 1 9 6 2 ) F l u i d D i s p e r s i o n -
g e n e r a l i z a t i o n and Comparison o f M a t h e m a t i c a l M o d e l s - I
G e n e r a l i z a t i o n o f M o d e l s . , Chem. Eng. S o c , 1 7 , 245 .
S y m c o x , R. 0 . and E h r l i c h , P. ( 1 9 6 2 ) The F r e e R a d i c a l , High
P r e s s u r e P o l y m e r i z a t i o n o f E t h y l e n e s . The E f f e c t s of
I n i t i a t o r C o n c e n t r a t i o n , Monomer C o n c e n t r a t i o n and P r e s s u r e
on t h e P o l y m e r i z a t i o n r a t e . J . Am. Chem. S o c , 2 0 , F e b , 5 3 1 .
Z e m a n , R. a n d A m u n d s o n , Ν. R. ( 1 9 6 3 ) C o n t i n u o u s M o d e l f o r
Polymerization. A.I.Ch.E. Journal, 9, N o . 3 , 297.
A r i s , R u t h e r f o r d ( 1 9 6 5 ) I n t r o d u c t i o n t o t h e A n a l y s i s of C h e m i c a l
Reactor. Prentice-Hall Inc.
Zeman, R. J . and Amundson, N. R. ( 1 9 6 5 ) C o n t i n u o u s P o l y m e r i z a t i o n
M o d e l s - P a r t I I . B a t c h R e a c t o r P o l y m e r i z a t i o n . Chem. Eng. S c i .
20, 637.
D u e r k s e n , J . H. ( 1 9 6 7 ) P o l y m e r R e a c t o r s and M o l e c u l a r W e i g h t
Distribuyion, Part I. Free Radical P o l y m e r i z a t i o n in a
C o n t i n u o u s S t i r r e d - T a n k R e a c t i o r . A.I.Ch.E. J o u r n a l , 1 3 ,
No.6, 1081.
K i t a , S . and F u k u i , Κ. ( 1 9 6 7 ) Mechanism o f S h o r t Chain B r a n c h i n g
P r o d u c t i o n at R a d i c a l P o l y m e r i z a t i o n of E t h y l e n e . I.
Kobunshi Kagaku, 2 4 , 348
Saito, O. a n d G r a e s s l e y , W. W. ( 1 9 6 9 ) M o l e c u l a r Weight
D i s t r i b u t i o n i n Branched P o l y m e r . J. Polymer S c i - , Λ - 2 , 7,
1937.
E h r l i c h , N. a n d M o r t i m e r , G. A. (1 9 7 0 ) F u n d a m e n t a l s o f The F r e e -
R a d i c a l P o l y m e r i z a t i o n o f E t h y l e n e . Adv. P o l y m e r S c i . , 7 ,
386.
L i n d e m a n , L. P. a n d A d a m s , J . Q. ( 1 9 7 1 ) C a r b o n - 1 3 N u c l e a r
Magnetic Resonance Spctrometry, Chemical S h i f t s for the
P a r a f f i n s t h r o u g h Co. A n a l y t i c a l C h e m i s t r y , 4 3 , N o . 1 0 , 1245
K a m a , T . . e t a l . P972) L o n g - C h a i n B r a n c h i n g and S o l u t i o n
P r o p e r t i e s o f Low D e n s i t y P o l y e t h y l e n e . Die Makromol. Chem.,
155, 2 8 3 .
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

THE DYNAMICS AND CONTROL OF A

DEPROPANIZER
G. A. Carling and R. K. W o o d
Department of Chemical Engineering, University of Alberta, Edmonton,
Alberta, Canada

Abstract. The dynamic and control behavior of a depropanizer that consists of 29 ideal
trays plus condenser and reboiler, with feed introduced at the 12th tray, has been
studied by digital simulation. Open loop dynamic behavior of the column was studied
for step changes of up to 30% in the feed and reflux flow rates and heat input to the
reboiler. For changes in feed flow rate, the disturbance variable used for the control
studies, the column exhibited inverse behavior in the composition response of the light
key component (propane) in the distillate and the heavy key component (isobutane) in
the bottom product. However, no inverse response was observed for the controlled
variables of the concentration of the heavy key in the distillate and the light key in
the bottoms. Control behavior of the five component depropanizer column for conven-
tional single product composition control as well as dual quality control was investi-
gated using the DYCONDIST simulator. Results obtained using indirect control strate-
gies by measuring tray liquid temperature(s) and direct control of the distillate and
bottoms are presented. Performance of the dual quality control schemes is discussed
with respect to predicted behavior based on static information from relative gain array
and relative disturbance gain calculations.

Keywords. Chemical industry; modelling; dynamic response; closed loop systems;

control system analysis.

INTRODUCTION loop control (Ballard, Brosilow and Kahn, 1978;

The continuing interest in the dynamic and control
behavior of distillation columns stems not only
from the central role played by distillation pro-
cesses for separation and purification but also
from demands for improved quality control and
reduced energy consumption, Furthermore the effec-
)
tive use of recently developed process analyzers
in conjunction with the currently available hard-
ware and software of distributed computer control
systems requires more understanding of system
dynamic behavior. Accompanying this trend is the
demand, even for conventional columns, for simul-
taneous control of both product compositions. With
an objective of dual product quality control it is
necessary to understand the inherent interaction
of multivariable systems. Although multivariable
control strategies have been successfully employed
in the aerospace industry for decades, attempts at
direct use for column control have been unsatis-
factory because of nonlinear dynamic behavior. In
tackling these problems, various companies have
developed proprietary simulation programs (Moser,
1982; Cheung and Marlin, 1982; Wei, 1984; Downs
and Vogel, 1985). As a result of the lack of
availability of a general purpose multicomponent
distillation column simulation program, develop-
ment of a simulator for studying column dynamic
and control behavior was initiated at the Univer-
sity of Alberta in 1983.
Gallun, 1979). However, these programs are not de-
signed for general purpose evaluation of different
control strategies with algorithms that may require
use of temperature(s) and/or composition(s) of any
stage. The original development of the simulator
used in this work was conducted by Wong (1985)
after evaluating a distillation column program ob-
tained from Case Western Reserve University (Cook,
1980). The program was used only for studying
column dynamic behavior except for simulation of
the controlled extractive distillation column
studied by Gallun (1979) and described by Holland
and Liapis (1983).
DYNAMIC SIMULATOR
The present dynamic simulator, known as DYCONDIST,
is based on the program developed by Wong (1985)
but has been significantly enhanced by Carling
(1986). Interactive input is now allowed and the
original modular program structure retained to
facilitate easy incorporation of new control strate-
gies. In addition to the input generator which
creates an input data file that can be saved for
future use, DYCONDIST also has a post processor
which allows the user to generate one or more plots
using output data from one or more simulations.
Specific features of the program, written in
FORTRAN 77, are described by Carling (1986).

Simulation of the dynamic behavior of multicom-
ponent columns can be traced to the availability
of increased digital computer capacity in the early
1960's (Rosenbrock, 1962; Peiser and Grover, 1962).
In the last few years further use of multicomponent
dynamic column simulation programs has been re-
ported both for studying open loop dynamic behavior
(Morris and Svrcek, 1981; Prokopakis and Seider,
1983; Kumar, Wright and Taylor, 1984; Stathaki,
Mellichamp and Seborg, 1985) and also for closed
The model used to describe the dynamic behavior
multicomponent plate distillation columns consists
of a collection of ordinary differential and alge-
braic equations. The equations, based on rigorous
heat and material balances, are expressed in a
general form so the dynamic and control behavior of
towers with multiple product streams and/or feeds
can be simulated. The current version of DYCONDIST
is limited to 90 stages and 10 components. The
material and energy balance differential equations

DCCR-L
167
168
are based on the general stage representation
shown in Fig. 1 and may represent a tray with a
feed stream, or a tray with vapor and liquid side-
draws just as well as a partial condenser, total
condenser or reboiler. The column model material
balance differential equations consist of C + 1
equations based on the following assumptions
( i) non-reactive mixture
( ii) perfect mixing in the stage
(iii) negligible material holdup in the vapor
phase and in the liquid in the tray down-
corners .
With the condenser numbered as stage 1 the stage
mass balance equation is
+ )
dt • * + ν V i - W
Λ 1
tn
It then follows that the i*-* component mass balance
for the j stage can be expressed as
- Newton's method. The final equation of the MESH
dx
j+l^j+1 j '
+
3 3
CxL
3
x equation. Adopting the approach of Ballard,
jh - (Vj-^jXyj-Xj)
J-I Î-I
The general form of the energy balance for the j -*
stage, (tray, condenser or reboiler) can be writ-
ten as
dh.
+ Vi'VrV - W ( (
by assuming that there is
( i) negligible energy and material holdup in
the vapor phase
( ii) negligible energy storage in the stage metal
(iii) negligible energy storage in the liquid
contained in tray downcomers
( iv) negligible heat of mixing among species
Thermodynamic properties/data for either ideal or
nonideal species can be easily employed with the
simulator. In the current version of DYCONDIST
the Wilson equation (Smith and Van Ness, 1975) is
used to calculate activity coefficients for non-
ideal systems.
In order to completely specify a column, the simu-
lator provides a set of options from which corre-
lations for stage efficiency, stage heat loss,
pressure drop and liquid holdup may be selected.
Solution of the material and energy balance equa-
tions in conjunction with the defining vapor-
liquid equilibrium relationship requires the solu-
tion of a set of N(2C + 3) equations known as the
"MESH" equations (Henley and Seader, 1981). For
unsteady state operation this involves the solu-
tion of the MESH equations at each time step with
the liquid composition and flow rate determined
first. If the constant molar holdup option is
selected, Equation (1) becomes an algebraic equa-
tion so the vapor flow rate profile is determined
by iteration at the same time as the liquid flow
rate profile. However, for variable stage liquid
holdup, the liquid flow rate profile must be deter-
mined by integration of Equation (1). For either
constant or variable stage holdup, the stage
liquid compositions are established by integration
of Equation (2).
As has been discussed (Wong and Wood, 1985) a
variety of integration methods were considered and
the integration algorithm selected was the adap-
tive semi-implicit Runge-Kutta (ASIRK) method pro-
posed by Prokopakis and Seider (1981). This algo-
rithm features an embedded error estimate and an
G. A. Carling and R. K. Wood
integration step size controlled by an estimate of
the degree of stiffness of the set of equations.
The step size control strategy, adopted from the
work of Ballard, Brosilow and Kahn (1978) , was
selected in place of the computationally complex
strategy suggested by Prokopakis and Seider (Wong,
1985) . The user must specify a minimum and maxi-
mum step size as well as a value of a step size
control parameter which controls the sensitivity of
the method to the estimate of stiffness provided by
the ρseudo-eigenvalue of the stiffest equation in
the set. Judgement must be applied when selecting
the step size control parameter since its magnitude
represents a trade-off between the speed (effi-
ciency) and accuracy of the resulting solution.
( 1
" W The stage liquid compositions are used in conjunc-
tion with the vapor-liquid equilibrium relation-
ships to solve for stage temperatures and vapor
compositions in an iterative procedure based on
A set of equations that must be solved for each stage
of the column is the energy balance differential
Brosilow and Kahn (1978), Equation (3) is not
(2) solved by integration but is transformed into a
1 1 linear algebraic equation of the form
E. = -a.L. 3.V. + ψ . ν . , , Ψ (4)
Λ+
J J J-l 3 3 3 J+l
with the values of the coefficients,E-? ,α^,β-ΐ and ψ*
as defined by Carling (1986). This collection of
equations is solved for the stage vapor flow rates
using the Thomas algorithm since the set of equa-
W (3)
tions forms a band diagonal system matrix.
The solution procedure for DYCONDIST, as described,
is presented in Fig. 2 which also shows the linking
of the control law calculation with the solution of
the column MESH equations. Including models for
local and remote controllers necessitates a slight
alteration of the solution procedure employed for
only studying dynamic behavior. For local control-
lers, calculation of elements of the Jacobian used
in the material balance integration must be modi-
fied and the dimension of the system of equations
increased. In developing the simulator it was
decided that the local controllers associated with
level and flow control would be simulated as con-
tinuous controllers with all other controllers re-
presented in discrete form. Details of the imple-
mentation of proportional (P) level controllers
and proportional plus integral (PI) flow control-
lers are given by Carling (1986). Remote control
algorithms of any degree of complexity can be em-
ployed with the simulator since the control calcu-
lations represent an independent program module.
The current version of the simulator incorporates
a conventional proportional plus integral plus de-
rivative (PID) controller and a self-tuning control
algorithm.
DEPROPANIZER COLUMN
Use of the DYCONDIST simulator for evaluating the
performance of different control strategies for
multicomponent columns and also for studying the
open loop dynamic behavior is illustrated by study-
ing the behavior of a 29 ideal tray depropanizer.
Open loop dynamic behavior was predicted for the
column equipped with a partial condenser. A total
condenser was used for the closed loop simulations.
Operating conditions for the 31 stage column are
shown in Table 1. Liquid enthalpies were corre-
lated as simple first order functions of tempera-
ture and the vapor enthalpies expressed as second
order polynomials in temperature. Vapor-liquid
equilibrium was described by an equilibrium ratio
as a function of temperature for each species
(Wong and Wood, 1985).
 The Dynamics and Control of a Depropanizer 169

TABLE 5 Depropanizer Operating Conditions feed composition and also for feed streams contain-
ing three and four components (Carling, 1986).
Feed:
Temperature 353 Κ
Pressure 2.6 MPa CONTROL BEHAVIOR
Rate 50 kmol/min
Composition (mole fractions) The control behavior of the depropanizer has been
ethane 0.03 studied for a variety of different control strate-
propylene 0.40 gies using DYCONDISΤ (Carling, 1986). Initial
propane (light key) 0.15 simulations involved studying the column responses
isobutane (heavy key) 0.15 for an objective of the control of only a single
cisbutene 0.27 product composition by measurement of tray liquid
temperature and also by direct measurement of pro-
Column Conditions: duct composition. Subsequent simulations of depro-
Pressure 2.6 MPa panizer control behavior were for a dual quality
Reflux Rate 90 kmol/min control objective. A stage for measurement of
Reboiler Duty 500 kJ/min liquid temperature in both the rectification and
Condenser Duty-Total 584 kJ/min stripping sections of the column was selected. The
-Partial 566 kJ/min sensor was located at the stage in each section at
Accumulator Holdup 50 kmol which the temperature exhibits the maximum symmet-
Column Base Holdup 50 kmol rical change in response to changes of equal mag-
Tray Holdup 12-22 kmol nitude and opposite direction in the distillate to
feed ratio (Tolliver and McCune, 1978). Control-
Product Rates and Purities: lers were tuned for a -10% feed rate disturbance.
(_
. i) Partial Condenser : Initial controller settings for the control law of
Vapor Distillate Rate 3.6 kmol/min the form
Liquid Distillate Rate 25.0 kmol/min
Bottoms Rate
Heavy Key in Liquid
21.4 kmol/min OP = κ
£ER + kJ
t
D ^ -dt
^BRdt + K„ 1+ BIAS (5)

Distillate 0.44% were calculated using the Cohen and Coon formulae
Light Key in Bottoms 1.59% (Coughanowr and Koppel, 1965).
(ii) Total Condenser:
Distillate Rate Two indirect control schemes were studied: measur-
28.6 kmol/min
Bottoms Rate ing stage 6 temperature to manipulate reflux rate
21.4 kmol/min
for control of the heavy key concentration in the
Heavy Key in Distillate 0.42%
distillate; and measuring stage 23 temperature to
Light Key in Bottoms 1.63%
manipulate reboiler duty for control of the light
key concentration in the bottoms. Simulation of
OPEN LOOP DYNAMIC BEHAVIOR the schemes indicated that control performance was
improved by adjusting the Cohen-Coon controller
The dynamic behavior of the depropanizer equipped parameters in both cases. Simulation of the cor-
with a partial condenser has been studied for step responding direct schemes in which the product im-
increases and decreases in feed rate, reflux rate purity mole fractions were used as the controlled
and reboiler duty of up to 30% (Wong, 1985; Carling, variables showed that further tuning was required
1986). Due to space limitations, only the response for the case of distillate composition control
of the column to feed rate changes will be con- while Cohen-Coon parameters provided satisfactory
sidered. Furthermore^since composition responses bottoms composition control. As can be seen from
of the light key in the distillate (propane) and of the integral of the absolute error (IAE) values
the heavy key in the bottoms (isobutane) showing summarized in Table 2, use of the direct versus the
inverse response behavior have been presented (Wong indirect control strategy shows the most improve-
and Wood, 1985) the composition responses for all ment for control of isobutane concentration in the
five species are shown in Figs. 3 and 4. As can be distillate. Use of direct measurement of the
seen from the responses for a 30% step decrease in heavy key concentration has reduced the ΙΑΕ value
feed rate, shown in Fig. 3 none of the components from .00988 to .00590 for controlling against a 10%
in the bottoms exhibit inverse response (nonmini- step decrease in feed rate and from .01218 to
mum phase) behavior. However in the distillate the .00268 for a step increase.
light key and the two lighter components, ethane
and propylene, exhibit nonminimum phase behavior. TABLE 2 Single Produce Composition
For the increase in feed rate it can be seen from Control Performance
Fig. 4 that in the distillate only the light key
and propylene exhibit inverse response and in the IAE VALUE
bottoms only the heavy key (isobutane). These FEED FEED
occurences of inverse response behavior involving DECREASE INCREASE
the key component in the streams they dominate is CONTROL OF DISTILLATE: C- 10%) (+ 10%)
consistent with the results reported by Stathaki, INDIRECT: K^ = -10.0 DIS 0.00988 0.01218
Mellichamp and Seborg (1985). The complex nature Kj = 0.10 BTM 0.26867 0.58938
of multicomponent column dynamic behavior is illus- DIRECT: tC = -2000 DIS 0.00590 0.00268
trated by Fig, 5, a three dimensional plot of the Kj = 0.10 BTM 0.29647 0.75263
transient response of the propane molar flow rate CONTROL OF BOTTOMS :
versus stage number for a 30% step decrease in feed INDIRECT: IL, = 0.012 DIS 0.03994 0.06452
flow rate. The inverse propane composition res- Kj = 0.10 BTM 0.04230 0.03081
ponses noted in the enriching section are reflected DIRECT: Κ = - 4.38 DIS 0.01492 0.06943
in the initial response of the molar flow rate of Kj = 0.40 BTM 0.04222 0.01792
propane.
Simultaneous control of the heavy and light key
Additional simulations to study the occurence of specifications for the distillate and bottoms res-
inverse response behavior have been undertaken but pectively was studied using the indirect control
space does not allow for presentation of results. strategy shown schematically in Fig. 6 for ±10%
In addition to the predicted inverse responses for step changes in feed flow rate from the steady
flow rate and reboiler duty changes, nonminimum state value. As can be seen from the responses
phase behavior is also evident for some changes in shown in Fig. 7, obtained using the controller set-
tings for single product control^ the composition
170
control of the distillate (Quality Loop 1) is
satisfactory. However, bottoms composition control
may be considered to be unsatisfactory despite the
IAE value of 0.03526 being lower than the value of
0.04230 for control of only bottoms composition
(cf. Table 2 ) . With the direct control strategy
involving an analyzer for both product streams,
simulations showed that although use of the single
composition controller settings resulted in rela-
tively satisfactory control, bottom composition
control could be improved by further tuning of the
controller. The resulting control performance is
shown in Fig. 8.
Since the simulations showed that there was some
interaction between the two quality loops it was
considered instructive to examine the relative gain
array (RGA) to ascertain if the observed behavior
could have been predicted from a knowledge of the
steady state gains. Relative gains for the 2x2
system were calculated from the expression
r ^ K
12 KK2- 1
j-l
K (6)
L ll 22 J
for direct and indirect dual composition control.
Steady state gains were established from perturba-
tions of + .05%, .1%, 1%, 5% and 10% in reflux
rate and reboiler duty. The resulting relative
gain values are summarized in Table 3, for the
gain array expressed as
Reflux Reboiler Duty
c 1 -
λ 1 - XI
RGA = cl
2 L i- λ
TABLE 3 Relative Gain Values as a Function of
Perturbation Size and Direction
INDIRECT DIRECT
Perturbation Relative Perturbation
( % ) Gain (% )
+ 0.05 10.2 + 0.05
- 0.05 10.6 - 0.05
+ 0.10 86.2 + 0.10
- 0.10 6.25 - 0.10
+ 1.0 - 0.300 + 1.0
- 1.0 1.31 - 1.0
+ 5.0 - 0.07 + 5.0
- 5.0 1.07 - 5.0
+ 10.0 - 0.07 + 10.0
- 10.0 1.07
- 10.0
The relative gain may be used to determine appro-
priate pairings for controlled and manipulated
variables since less interaction between loops may
be expected for pairing with λ - 1. Furthermore,
according to Bristol (1985) and McAvoy (1983), if
λ < 0 only conditional closed loop stability is
expected; for | λ | » 1 the system may be extremely
sensitive; and for |λ|<0.5 interaction between the
loops can be expected with the most severe inter-
action expected for |λ|- 0 . As can be seen from
the relative gain values in Table 3, the values
vary with the direction and size of the perturba-
tion applied to the manipulated variables. When
this situation has arisen (Kim and McAvoy, 1982;
Thurston, 1981; Ralston, 1983) one of two approac-
hes has been followed. Kim and McAvoy (1982)
advocate reducing the magnitude of the perturba-
tion until the calculated relative gains are con-
sistent regardless of the direction of the pertur-
bation while Thurston (1981) states that dual
quality control should not be attempted when this
"degeneracy" is observed. Ralston (1983) found
that the relative gains she calculated were incon-
sistent for perturbation sizes between .05% and 2%
and opted for single point control because of de-
generacy. The inconsistent pattern of the rela-
tive gains shown in Table 3 for very small pertur-
G. A. Carling and R. K. Wood
bations likely results due to the sensitivity of
low steady state gains to small errors in control-
led variables. Although the question of the magni-
tude of perturbations for nonlinear systems has
been considered by Mijares and co-workers (1985)
there was no firm recommendation as to a suitable
perturbation size. For the depropanizer, on the
basis of the values in Table 3, a perturbation size
of 1% of the manipulated variables represents a
reasonable compromise between the desire to obtain
estimates of partial derivatives (small perturba-
tions) and the desire to avoid numerical problems.
The dependence of the relative gain on the direc-
tion of the perturbation is considered to be re-
lated to the strong nonlinear system behavior
rather than to any "degeneracy". Satisfactory dual
quality control performance supports this theory
and suggests that the relative gains based on nega-
tive perturbations are more credible than those
based in positive perturbations since the drastic
interaction suggested by small negative relative
gains is not displayed.
An additional measure of the effect of interaction
is the relative disturbance gain (Stanley, Marino-
tn
Galarraga and McAvoy, 1985). For the steady state
gain of the i controlled variable in response to
a disturbance denoted as K.i3> the relative distur-
bance gains (RDG) for a 2x2 system are
K K
23 12
(7)
13 22
λ(
λ -
(8)
The interpretation by the authors of the expected
control performance based on the value of the gains
was that if |β|>1 larger changes in controller
output would be required to compensate for a dis-
turbance in the case of dual composition control
Relative than for single product control. However, if |β|<1
Gain the interaction is advantageous in that less con*
- 3.63 trol action is required for multiloop control than
2.39 for control of only a single composition. Relative
- 4.60 disturbance gains calculated based on +1% perturba-
2.51 tions in reflux rate reboiler duty and feed rate
- 0.105 are given in Table 4.
1.10
- 0.008 TABLE 4 Relative Disturbance Gain Values as a
1.01 Function of Perturbation Size and Direction
- 0.003
INDIRECT DIRECT
1.00
Relative Relative
Disturbance Disturbance
Perturbation Gains Perturbation Gains
+ 1.0% 3χ = 7.29 + 1.0% 3-L = -0.05
e>
2 = -0.25
$2 = -0.20
1.0% B = -3.48 1.0% βχ = 0.51
x
3 = 1.23 Bo = 0.99
2
For the depropanizer, the RDG values provide little
insight into the control action required or control
performance due to the sensitivity of the measure
to the relative gain values (Carling, 1986).
CONCLUSION
Simulation of the depropanizer dynamic behavior has
shown that the composition responses of the dif-
ferent species exhibit an inverse response behavior
in the product streams and also at intermediate
stages. No general pattern of the specific occuren-
ces of this nonminimum phase behavior could be dis-
cerned. The performance of single product composi-
tion control and dual quality control strategies
investigated using the DYCONDISΤ simulator has
shown that acceptable dual quality control behavior
may be achieved by measurement of tray liquid tem-
perature^) or by direct terminal composition
 The Dynamics and Control of a Depropanizer 171

measurement. Simulation results revealed that regulatory control schemes. Proc. 1982 Sum-
relative gain and relative disturbance gain values, mer Computer Simulation Conference, 462-469.
calculated from steady state information, did not Cook, W.J. (1980). M.Sc. Thesis, Case Western
provide a reliable indication of the control Reserve University, Cleveland, Ohio, USA.
behavior. Coughanowr, D.R., and L.B. Koppel. (1965). Process
Systems Analysis and Control, McGraw-Hill, N.Y.
ACKNOWLEDGEMENT Downs, J.J., and E.F. Vogel. (1985). An interactive
dynamic distillation simulator for analysis
The support of this work by the Natural Sciences and control system development. P r o c 1985
Summer Computer Simulation Conference, 337-342.
and Engineering Research Council under Research
Gallun, S.E. (1979). Ph.D. Thesis, Texas A & M
Grant No. A-1944 is gratefully acknowledged.
University, College Station, Texas, USA.
Henley, E.J., and J.D. Seader. (1981). Equilibrium
NOTATION Stage Separation Operations in Chemical
Engineering, John Wiley & Sons, N.Y.
c
BIAS steady state controller output signal Holland, C D . , and A.I. Liapis. (1983). Computer
l controlled variable, loop " 1 " , temperature Methods for Solving Dynamic Separation Problems.
c or composition McGraw-Hill, N.Y.
2 controlled variable, loop " 2 " , temperature Kim, Y., and T.J. McAvoy. (1982). Computing the
or composition relative gains for pressure and composition
C number of species (components) control of a single distillation tower. Proc.
ER error signal, setpoint minus feedback 1982 ACC, 74-80.
signal
th Kumar, S., J.D. Wright, and P.A. Taylor. (1984).
molar feed rate to the j stage Modelling and dynamics of an extractive distil-
1
enthalpy of the liquid in and leaving the lation column. Can. J. Chem. Eng., 62, 780-789.
j*-* stage
1 McAvoy, T.J. (1983). Interaction Analysis: Prin-
enthalpy of the feed stream to the j ciples and Applications. Instrument Society of
stage 1 America, Research Triangle Park, NC, USA.
enthalpy of the vapor leaving the j*-* stage Mijares, G., C D . Holland, R. McDaniel, C.R. Dollar
"4 proportional action constant
integral action constant
and S.E. Gallun. (1985). Analysis and evalua-
tion of the relative gains for nonlinear
derivative action constant systems. Comput. Chem. Eng., 9_, 61-70.
κ steady state gain between the ,th control- Morris, C.G., and W.Y. Svrcek. (1981). Dynamic simu-
ij 1
led variable in response to a perturbation lation of multicomponent distillation. Can. J.
in the j*-* manipulated variable
tn Chem. Eng. 59, 382-387.
tn
liquid molar flow rate from the j stage Moser, J.H. (1982) DYMODS-A computer program system
liquid molar holdup in the j stage for simulation of process dynamics Proc. 1982
feedback signal Summer Computer Simulation Conference, 514-519.
time instant Peiser, A.M., and S.S. Grover. (1962). Dynamic simu-
Ν total number of stages lation of a distillation tower. Chem. Eng.
OP controller output signal Prog., 58, No. 9, 65-70.
heat loss (gain) from stage " j " Prokopakis, G.J., and W.D. Seider. (1981). Adaptive
tn
liquid molar sidestream flow rate from the semi-implicit Runge-Kutta method for the solu-
j stage tion of stiff ordinary differential equations.
vapor molar sidestream flow rate from the IEC Fund, _20, 255-266.
stage Prokopakis, G.J., and W.D. Seider. (1983). Dynamic
temperature
1
simulation of azeotropic distillation towers.
vapor liquid flow rate from the j*-* stage AIChE J., 29, 1017-1030.
mole fraction of species "i" in the liquid Rosenbrock, H.H. (1962). The control of distillation
in and leaving the columns. Trans. Instn. Chem. Engrs., 40, 35-53.
1 stage
vector of mole fractions in the liquid in Smith, J.M., and H.C. Van Ness. (1975). Introduction
and leaving the j*-* stage to Chemical Engineering Thermodynamics.
11
mole fraction of species "i" in the vapor McGraw-Hill, N.Y.
leaving the j* * stage Stanley, G., M. Marino-Galarraga, and T.J. McAvoy.
1
vector of mole^ fractions in the vapor (1985). Shortcut operability analysis: 1. The
y. relative disturbance gain. IEC Proc. Pes. Dev.,
~3 leaving the j*-* stage
24, 1181-1188.
mole fraction of species i" in the feed Stathaki, Α., D.A. Mellichamp, and D.E. Seborg.
Z stream to the stage
tn (1985). Dynamic simulation of a multicomponent
vector of feed stream mole fractions in distillation column with asymmetric dynamics.
-3
the feed stream to the j stage Can. J. Chem. Eng., 63, 510-518.
Thurston, C.W. (1981). Computer-aided design of
Subscript
distillation column controls. Hydrocarb Proc.,
j denotes the stage 60, No. 7, 125-130.
Tolliver, T.L., and L.C. McCune. (1978). Distilla-
Superscript
1 tion control based on steady state simulation.
ISA Trans., 17, No. 3, 3-10.
i denotes the i*"* species Wei, C-N. (1984). Dynamic simulation of multicom-
ponent distillation columns: a tool for process
REFERENCES analysis, control design and operator training.
Proc. 1984 Summer Computer Simulation Con-
Ballard, D., C. Brosilow, and C. Kahn (1978). ference , 616-623.
Dynamic simulation of multicomponent distilla- Wong, T.T. (1985). M.Sc. Thesis, University of
tion columns. Presented at 71st Ann. Mtg. of Alberta, Edmonton, Alberta, Canada.
AIChE. Miami, Fl., USA. Wong, T.T., and R.K. Wood. (1985). Dynamic behavior
Bristol, E.H. (1985). Book Review: G. Stephanopoulcs of a hydrocarbon separator. Proc. 1985 Summer
Chemical Process Control. Comput. Chem. Eng., Computer Simulation Conference, 331-336.
% 99-101.
Carling, G.A. (1986). M.Sc. Thesis, University of
Alberta, Edmonton, Alberta, Canada.
Cheung, T.F., and T.E. Marlin. (1982). Using dyna-
mic simulation to evaluate distillation tower
172 G. A. Carling and R. K. Wood

A
Liquid • = Ethane
ϋ = Propylene + = Isobutane

j-1
Vapour = Propane X = Cisbutene

ν,,ν,'.Η,
0.15

Vapour Draw 0.10

Stagej
Feed Holdup, M:
F,.z,', h. S XS hDraw
Liquid 0.05

'J- J- i 0.00

-0.05
VapourVj Heat Loss
LLiquid
x h H+ +,1 Ο
^ -0.10
i' l'- i y'j+1' i 1

20 40 60
tn
Fig. 1 Schematic diagram of the j stage
Time (min)

n
S o l u t i o n ® time t

Material
Balances 20 40 60
Time (min)

Fig. 3 Product composition responses for a 30%

step decrease in feed rate
11
Σ K,x,=1
i

Efficiency Control • = Ethane

0 = Propylene + = Isobutane
Correlation Calculation Δ = Propane X = Cisbutene

Enthalpy
Correlation
ah:

dX: Yes

Energy 20 40 60
Balance
Time (min)
0.15
Xj.Vj.Lj.VjTj
0.10

>
Pressure Drop 0.05
Heat Loss Q
Holdup n
0.00

Solution @ time t ^ -0.05

> Ο

CO -0.10
a
-0.15
20 40 60
Fig. 2 DYCONDISΤ solution procedure for variable
stage holdup Time (min)

Fig. 4 Product composition responses for a 30%

step increase in feed rate
 The Dynamics and Control of a Depropanizer 173

Fig. 5 Transient response of propane flow rate Fig. 6 Indirect dual quality control scheme

0. 0.4- 0.4- ....

Ο.ί
-
COMPOSITIOI

tj
....
COMPOSITION

ο ο 0.3-
) E3 Β Ε ] B B •
55 0.2- 0.2-

οι
-
COMPO

COMPO

DISTILLATE BOTTOMS DISTILLATE BOTTOMS

0.1- •-LK •«=LK 0.1- •-LK 0.1- •-LK
O-HK O-HK O-HK
> °~ HQ20 K Ο
40
0
60 20 40
Π
ao «ο oo 9 20 40 60
TIME (min) TIME (min) TIME (min) TIME (min)

QUALITY LOOP 1 QUALITY LOOP 2 QUALITY LOOP 1 QUALITY LOOP 2

CONTR: STAGE 6 TEMP. CONTR: STAGE 23 TEMP. C O N T R : H K IN D I S T I L L A T E C O N T R : L K IN B O T T O M S

MANIP: R E F L U X R A T E MANIP: R E B O I L E R D U T Y MANIP: R E F L U X R A T E MANIP: REBOILER D U T Y
C O N T R O L L E R T Y P E : PI C O N T R O L L E R T Y P E : PI C O N T R O L L E R T Y P E : PI C O N T R O L L E R T Y P E : PI
SETTINGS: SETTINGS: SETTINGS: SETTINGS:
K P = - 1 0 K I = 0.10 K P = 0.012 K I = 0.10 KP= -2000 K I = 0.10 K P = -3.0 K I = 0.30
S P = 331.22K S P = 368.56K S P = 0.0042 S P = 0.0163
P E R F O R M A N C E : IAE= 0.01282 P E R F O R M A N C E : IAE= 0.03526 P E R F O R M A N C E : I A E = 0.01919 P E R F O R M A N C E : I A E = 0.02497
5ITI0

»... .010
- 020
-
.006
-
N

Tl.O-i
LTURE
ITURE

3ITION

.016
60.0
.006
- .016
-
05 zt 0- iOu βτ.ο
5 o .004
- ο ο.014
a.
.012
-
COM]

COMI

.002
u STAGE β W S T A G E 23 iC4 C3
20 40 60 «0 40 60 20 40 60

..^ 10
. On 0.7-

0.·
-
RATE

Η
00.0·
Χ ι-
Of
C

οί at
0.5
eo.o
- >- Χ
DUTÏ

w ΐ3 0.4- H
60.0 0.4-
Q
04
*TIME (min)° *°
20 40
TIME (min)
to
"iME (min*)
" *°
20 40
TIME (min)
«0

Fig. 7 Control performance for indirect control of Fig. 8 Control performance for direct control of
distillate and bottoms composition for a distillate and bottoms composition for a
10% step decrease in feed rate 10% step decrease in feed rate
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

CONTROL OF A DISTILLATION COLUMN

ANALYSIS OF NOMINAL PERFORMANCE
AND ROBUSTNESS FOR MULTIVARIABLE
CONTROL SYSTEMS
M . K u m m e l and H. W . Andersen
Chemical Engineering Department, Technical University of Denmark,
DK-2800 Lyngby, Denmark

Abstract. A general frequency analysis, which for multivariable systems simultaneously

describes nominal performance like set-point tracking and disturbance rejection and
robustness properties like output sensitivity to modelling errors and stability margin
is presented.
As a case study the control of a distillation column is examined. Different control
schemes, such as geometric control, optimal control and PI control, are analysed. These
results are compared to those of transient simulations for both nominal and perturbed
systems. This comparison shows that the analysis yields both qualitative and quantita-
tive insight to the controlled system.
A major result is that geometric control of the distillation column has a sensitivity
equivalent to a similar open-loop configuration. This sensitivity may, however, be re-
duced by applying pole shifting to the geometric controller.

Keywords. Distillation control, Robustness, Control system analysis, Frequency

response, Sensitivity analysis, Stability, Singular values, Process control.

INTRODUCTION c) Sensitivity of outputs to modelling errors.

The last two decades have brought major develop-
ments in advanced control strategies for multi-
variable systems. The design of these controllers
rely on a linear time invariant state space de-
scription. In general these controllers gives im-
proved performance, compared to classical control,
when they are evaluated on the model. Among these
methods are geometric control, which in some cases
can provide total disturbance rejection (Takamatsu,
1979). However, the design of the geometric con-
troller relies heavily on the linear time invari-
ant state space model. The question is how this
controller performs, when it is implemented on a
real plant which deviates from the model.
Singular value methods have proved to be useful in
measuring robustness properties of multivariable
systems (Doyle (1981), Cruz (1981) and Arkun
(1984)). The major results of these works have
been to derive expressions for stability margins
and sensitivity of outputs to modelling errors.
d) Stability margin.
Previous work (Doyle (1981), Cruz (1981) and Arkun
(1984)) has been restricted to cases where the
feedbacks are obtained from the outputs of the
system. In the method presented here, the analysis
might be carried out at the outputs which define
our control objectives, while the feedback is per-
formed from the states of the system.
As an example, the analysis has been used on con-
trol of a distillation column. The use of advanced
control in distillation yields savings in the
energy costs. Thus, geometric control would make
it possible to operate exactly at product specifi-
cations. However, this is only a nominal property.
The main purpose of this work is to examine the
robustness properties of geometric control, com-
pared to other control schemes such as optimal
control and PI control in order to include ro-
bustness properties in the evaluation of proper
control schemes selection.

A reasonable control scheme will in general be a

compromise between nominal performance and robust- ANALYSIS
ness. Thus, designing only for nominal performance
might give systems with poor robustness properties. The nomenclature used in this section is presented
On the other hand, designing only from a robust- in Appendix A.
ness point of view, might yield systems with poor
nominal performance. One aim of this work has
therefore been to derive a method, which simulta- Nominal performance
neously describes nominal performance and robust-
ness for multivariable systems. The first objective is to obtain satisfactory
nominal performance. Here we will focus on how to
evaluate set-point tracking and disturbance rejec-
Palazoglu and Arkun (1985) suggested a method for
tion in the frequency domain.
evaluation of setpoint tracking. In this work a
similar frequency expression for disturbance
rejection in multivariable systems is presented. Set-point tracking. The closed-loop transfer func-
At this point we can simultaneously examine the tion between the outputs y(s) and the set-points
following four properties: Ty(s) can be expressed as: 1 T
a) Nominal set-point tracking. y(s) = C(sI-A + BG(s)H(s))" BG(s)H(s)C r (s)
b) Nominal disturbance rejection. = Tl(s)r (s) (1)

175
176 M. Kummel and H. W. Andersen

According to MacFarlane and Jones (1979) the am- The amplitude ratio between the closed-loop error
plitude ratio between the output vector y(s) and and the open-loop error is determined by:

l()| _
the set-point vector r (s) can be defined as
y
I y (iw) j / I ry (s) J . This vector gain lies in a band
determined by the minimum and maximum singular
ewi

j^ffyj
value of the closed-loop transfer function TI:

VP
m m : a(5(iw)) < < a(S(iw)) (10a)
le (iw)| _
£(Tl(i
W)) <to (2) y(iw) closely
According < ô(TXiw))
tracks the set- (2) m<n: 0
According < _£V good< sensitivity
to (lOa+b)
| e o( ix
a(S(iw)) (10b)
properties
^ ) | obtained if cf(S(iw)) is
points if a(Tl(iw)) and a(Tl(iw)) both are close of the closed-loop are
to unity over a desired bandwidth. close to zero over a desired bandwidth.

Disturbance rejection. The closed-loop transfer Stability margin. The stability margin is calcu-
function between the outputs y(s) and the disturb- lated for right multiplicative errors:
ances q(s) can be expressed as:
1 1 1 P'(s) = P(s)(I+AP(s)) (11)
y(s) = C(I+(sI-A)" BG(s)H(s))" (sI-A)" Dq(s)
The analysis is performed on the feedbacks of the
= T2(s)q(s) (3)
system. Thus, the stability margin is calculated
as :
The vector gain is determined by: _1
ν(iw) lmr(w) < £(I+(G(iw)H(iw)P(iw)) ) (12)
m > d: a(T2(iw)) < Hyr"7 < a(T2(iw)) (4a)
— — — I quw) I — For further details is refered to Doyle (1981) and
Cruz (1981).
m < d: 0 < < a(T2(iw)) (4b)
Iq(iw)I Some remarks on this stability margin are appro-
priate:
According to (4a+b) good disturbance rejection is a) Other types of errors may occur e.g. left mul-
obtained if a(T2(iw) is close to zero over a de- tiplicative errors and additive errors. Mar-
sired bandwidth. gins for these errors may be calculated (Doyle
(1981), Cruz (1981)). However, since these
Robustness errors may occur simultaneously and since the
analysis takes only one error at the time into
The second objective is to obtain satisfactory ro-
account, the results can be too optimistic. An
bustness properties i.e. providing performance in
analysis which calculates stability margin for
spite of model errors. Here we will focus on how
simultaneously occurence of the three error
to evaluate the output sensitivity to modelling types can be derived. However, this margin
errors and the stability margin of the system. will in general be too conservative (Doyle
(1982)).
Sensitivity of outputs to modelling errors. b) This stability margin expresses norm-bounded,
In fig. 1 a general closed-loop configuration is but otherwise unconstrained right multiplica-
shown. A similar open-loop configuration is shown tive errors. Thus, it does not take any know-
in fig. 2. The closed-loop error and the open-loop ledge of the error structure into account.
error is defined as: Therefore the results may be too conservative.
These restrictions in the analysis can be reduced
e (iw) = χ (iw)-x'(iw),
cx c c by using structered uncertainties as proposed by
Doyle (1982).
e (iw) = χ (iw)-x'(iw)
OX 0 0
where ' denotes the perturbed system. If the A CASE STUDY OF DISTILLATION CONTROL
closed-loop and the open-loop give the same nomi- Model
nal response, Cruz et al. (1964) showed that there
exists a linear relation between the closed-loop In this investigation the same mathematical model
error and the open-loop error: _1 for a distillation column as used by Takamatsu
(1979) has been examined. This model describes a
e (iw) = (Ι+Ρ'(iw)G(iw)H(iw)) e (iw) (6) binary distillation. The column consists of 9
cx ox
plates, reboiler and total condenser. The feed,
liquid at its boiling point, is entering at plate
The closed-loop error in the outputs y(iw) may be
5 (state 6 ) . The outputs are the compositions in
expressed as:
_1 the condenser (state 1) and the reboiler (state
11). The manipulated variables are the reflux flow
e (iw) = Ce (iw) = C(I+P'(iw)G(iw)H(iw)) e (iw)
cy cx ox and the vapour flow. Disturbances occur in the
feed flow and the feed composition. The state
= S'(iw)e (iw) (7)
Qx space model is shown in table 1. For further de-
tails, see Takamatsu et al. (1979).
where S' is the sensitivity matrix. Here the sen-
sitivity will be evaluated for the nominal system.
Thus, (7) reduces to: Controllers
-1
e (iw) = Ce (iw) = C(I+P(iw)G(iw)H(iw)) e (iw) Geometric control. The geometric controller was
cy cx ox
designed by Takamatsu et al. (1979). Nominally,
= S(iw)e (iw) (8) this controller yields total disturbance rejection
ox
at the top and bottom composition towards disturb-
where
1 ances in the feed composition. Disturbances in the
feed flow rate are completely rejected at the top
P(s) = (si-Α)" B, _1 composition, but not at the bottom composition.
thus: S(s) = C(I+(sI-A) BG(s)H(s)) (9) According to the control law u = -GHx, the feed-
backs and gains are given by:
 Control of a Distillation Column Analysis 177

0 0 1 0 0 0 0 0 0 0 0 Results and discussion

b
|o ο ο ο ο ο ο ο ο ι o| ( 1 3 Analysis.
) Figures 3-6 show the analysis results
Ρ _ 1330.06 -470.171 for the four control schemes. The analysis is per-
" 1251.47 -632.041 formed at the top and bottom of the column, since
these outputs are those which define the control
This controller is denoted CTRL
objectives.

Geometric control with pole shifting. According to Figure 3 shows the set-point tracking calculated
Kummel and Foldager (1983), state 1, 2 and 11 are by eq. (2). These curves closely resemble those of
available for feedback, which will not alter the the classical Bode analysis (amplitude ratio). The
nominal total disturbance rejection as described geometric controller CTR1 does not include feed-
for geometric control. The gains of the feedback back from the outputs. Thus, CTR1 is not repre-
from state 11 should, however, satisfy the rela- sented with any curves in fig. 3. Geometric con-
tion: trol with poleshifting CTR2 includes feedback from
J (b /b ) the outputs. Thus, this controller is represented
(14) with both an upper and a lower setpoint tracking
2,2 2,l 9 ,ll
i,n 2 gain curve. The offset is between 36-9%, and the
where g is an element in G, and b is an element in bandwidth is between 0.3-2 rad/min. The bandwidth
B. is calculated as the frequency where the gain has
fallen 3 dB compared to the steady-state value.
The following controller was suggested by Kummel The optimal controller CTR3 exhibits improved set-
point tracking compared to CTR2. The offset to-
et al. (1983):
wards step changes in setpoints is between 17-3%,
1 0 0 0 0 0 0 0 0 0 0 and the bandwidth is in the area of 1-5 rad/min.
0 1 0 0 0 0 0 0 0 0 0
0 0 1 0 0 0 0 0 0 0 0 Due to integration in both outputs the PI control-
H =
0 0 0 0 0 0 0 0 0 1 0 ler has no offset towards step changes. The band-
0 0 0 0 0 0 0 0 0 0 1 width is between 0.4-2.0 rad/min. The maximum
curve has a resonance peak at w=l rad/min. The
(15)
size (around 12 dB) suggests an underdamped system
5000 400 330.06 -470.17 -1499.94 with damping ratio around 0.09. Thus the steady
G =
1000 200 251.47 -632.04 -2000.00 state error has been improved at the expense of
the dynamic behaviour.
This controller is denoted CTR2.
Figure 4 shows the disturbance rejection calculat-
Optimal control. An optimal controller was design- ed by eq. (4a). The geometric controller (CTR1) is
ed by Kummel et al. (1983). Since the primary task only represented with one curve. This is due to
the fact that nominally the top and bottom concen-
was to control top and bottom concentration with-
trations are totally decoupled from disturbances
out considering the amplitude in the control sig-
in the feed concentration, thus the transfer func-
nals, the weight matrix is Fu=I. The weight matrix
tions are equal to zero. The top concentration is
Fx was for numerical reasons Fx = I, except for
decoupled from disturbances in the feed flow rate,
the two end elements, which refer to the top and
but the bottom composition is not. Thus, the curve
bottom composition. These two elements were chosen
in fig. 4 represents the transfer function between
to be:
bottom composition and feed flow rate. As seen
8 from fig. 4 there is an offset towards step dis-
f (i,l) = f (ii,ii) 1-10
x x turbances in feed flow rate of around 0.2%.
The resulting feedback matrices are:
Similar to CTR1 geometric control with pole shift-
H ing, (CTR2) is only represented with one curve,
11
(16) i.e. the transfer function between bottom composi-
8286.1 616.4 163.4 40.6 9.0 0.1 tion and disturbances in feed flowrate. Compared
1492.7 252.2 98.1 31.4 8.7 0.2 to CTR1 applying pole shifting has reduced the
offset to 0.06%.
-6.2 -21.1 -55.5 -86.5 2535.8
-6.1 -19.5 -59.4 -220.5 -9250.5
No disturbances are nominally decoupled with opti-
mal control. Thus, two curves represent CTR3. The
This controller is denoted CTR3.
offsets due to step disturbances lie between
0.0004-0.02%. Thus, the offset at the bottom com-
PI control. With PI control the reflux was con- position will be less than for any of the geomet-
trolled from measurement of the top product compo- ric controllers.
sition, and the heat input to the reboiler was
controlled from measurement of the bottom product As expected the PI controller (CTR4) eliminates
composition. The gains and integration times were the steady state error. However, the peaks indi-
determined through dynamic simulation. Dynamic de- cate that Pi-control has increased overshooting
coupling was not implemented. The aspect of ro- compared to both geometric and optimal control.
bustness of different decoupling schemes is dis-
cussed in Arkun et al. (1984). Figure 5 shows the sensitivity of the outputs to-
wards modelling errors as calculated by eq. (10b).
1 0 0 0 0 0 0 0 0 0 0 For pure geometric control (CTR1), this curve
0 0 0 0 0 0 0 0 0 0 1 approximates 1 for all frequencies. Thus, the geo-
(17) metric controller behaves like an open-loop con-
3000(l+l/(3s)) troller. The reason for this is seen by the Graph
G(s) = -1000(l+l/(8s))
0 theoretic approach (Schizas and Evans, (1981).
Figure 13 shows the couplings for the compensated
This controller is denoted CTR4. distillation column. States 1, 2 and 11 are com-
pletely decoupled from the other states.
178 M. Kummel and H. W. Andersen
Thus, deviations in states 1, 2 and 11 will not be
compensated by the geometric controller, since it
does not include feedback to these states.
Geometric control with pole shifting reduces the
lower frequency sensitivity to 45?ό of that of an
open-loop controller. However, at a critical fre-
quency the curve gets larger than 1 (the maximum
is 154%) which is an undesired property of the
closed-loop.
The optimal controller (CTR3) has a lower frequen-
cy sensitivity of 19% and the maximum peak is
119%. Thus, the lower frequency sensitivity is
lower than for both geometric controllers, and the
maximum peak is lower than that of geometric con-
trol with pole shifting.
PI control (CTR4) eliminates the steady state sen-
sitivity, (the curve approaches zero for small
frequencies). However, the curve has an undesired
maximum of 413%. Thus, the closed-loop will be
extremely sensitive to modelling errors with fre-
quencies around 1.1 rad/min.
Figure 6 shows the stability margins towards right
multiplicative errors as calculated by eq. (12).
The geometric controller (CTR1) shows no critical
frequency. It has a draw off point of 1.4 rad/min
(the frequency at which the stability margin is
130% of the steady state value).
Geometric control with pole shifting (CTR2) has a
critical frequency at 1.0 rad/min where the stabi-
lity margin is 82%. Thus, establishing the pole
shifting has reduced the stability margin. The
draw off point is now 2.5 rad/min. This is also a
degradation since lower draw off point indicates a
faster increase of the specific curve.
Optimal control (CTR3) has a greater critical sta-
bility margin than CTR2 (91% at w = 0.6 rad/min),
however, the draw off point is increased to 7.0
rad/min.
PI control (CTR4) has the lowest critical stabili-
ty margin (36% at w = 1.1 rad/min). This is due to
coupling and integration. The stability margin
might be improved by using decoupling as investi-
gated by Arkun et al. (1984).
Simulation. The nine plate distillation column
with reboiler and condenser, and under control of
one of the derived controllers, was simulated with
a 15% step change in the feed flow rate (ALF =
1.5 mol/min). The sample time was 0.2 min and the
disturbance was introduced after 1 min. The nomi-
nal responses were simulated with the linear
mathematical model. Takamatsu (1979) has shown
that there is only minor differences between the
transient from the linear and the non-linear
model. The perturbed responses were also simulated
with the linear mathematical model, but the vapour
flow rate was supposed to be 10% lower than set by
the controller: AV(obtained) = 0.9-AV(set).
Figures 9-12 show the nominal (indicated with an
"N" on the current state) as well as the perturbed
(indicated with a "P" on the current state) re-
sponses for the four controllers. The responses
are presented with 7 of the 11 states, e.g. the
condenser X]_, X 3 , X 5 , the feed plate X 5 , X 7 , X 9
and the reboiler xjj. Further, the two control
signals are shown, e.g. the reflux rate ui and the
vapour rate u . All responses are in deviation
variables.
2
From figs. 9-10 it is seen that the geometric con-
trollers yield total nominal disturbance rejection
in state 1, while state 11 has an offset of 0.2%
(CTR1) and 0.06% (CTR2). This is in excellent
agreement with the results from the analysis
(fig. 4 ) .
Figure 11 (optimal control, CTR3) shows an offset
in state 1 and 11 of respectively -0.017 and
0.014% which is in between the band obtained from
the analysis. As expected the PI control elimi-
nates the steady state error (fig. 12), but also
exhibits an undesired overshooting. It is also
seen from figs. 9-12 that the magnitude of the
control signals is equivalent for geometric and
optimal control.
Figure 7 shows the nominal and perturbed responses
of state 1 for the four controllers. The basic
geometric controller shows the greatest sensiti-
vity towards the error in vapour flow. Applying
pole shifting reduces the sensitivity to 8% of
that of the geometric controller. The optimal con-
troller has an even lower sensitivity (1% of that
of the basic geometric controller). PI control has
a low sensitivity and the steady state sensitivity
is zero. Thus, the simulated results of fig. 7
closely resembles the analysis results on sensiti-
vity shown in fig. 5.
Figure 8 shows the nominal and perturbed responses
of state 11 for the four controllers. Again it is
seen that basic geometric control has the largest
sensitivity that, applying pole shifting, reduces
sensitivity and that optimal control has reduced
sensitivity compared to both geometric controllers
CONCLUSION
An analysis is proposed which simultaneously de-
scribes nominal performance and robustness of a
multivariable closed-loop control system. A case
study of distillation control shows a close re-
semblance between conclusions from the analysis
and those from transient simulations.
Basic geometric control of a distillation column
is shown to have a sensitivity to modelling
errors, which is similar to that of an openloop
controller. The reason for this is seen by the
Graph theoretic approach (Schizas and Evans,
(1981)). Figure 13 shows the couplings for the
compensated distillation column. States 1, 2 and
11 are completely decoupled from the other states.
Thus, deviations in states 1, 2 and 11 will not be
compensated by the geometric controller, since it
does not include feedback to these states. This
sensitivity may, however, be reduced by applying
feedback from the states which does not alter the
nominal performance (total disturbance rejection).
Thus, it seems possible to combine ideal nominal
performance with satisfactory robustness proper-
ties, by using the degrees of freedom in the geo-
metric controller.
The optimal controller exhibits satisfactory ro-
bustness properties. However, it does not provide
total nominal disturbance rejection. Pi-control
eliminates the steady-state sensitivity, but it
has a low stability margin and exhibits undesired
overshooting.
 Control of a Distillation Column Analysis 179

REFERENCES APPENDIX A: NOMENCLATURE

Arkun, Y., B. Manouslouthakis, A. Palazoglu (1984). The control system considered is expressed by the
Robustness analysis of process control sy- following equations:
stems. A case study of decoupling in distilla- T
x(t) = Ax(t) + Bu(t) + Dq(t)
tion. Ind.Eng.Chem.Des.Dev. 23, 93-101.
y(t) = Cx(t); u(t) = G(HC r (t)-z(t)); z(t)=Hx(t)
Cruz, J.B., W.R. Perkins (1964). A new approach to y
the sensitivity problem in multivariable feed- x(t) η-dimensional state vector.
back system design. IEEE Trans Automat. Contr. y(t) m-dimensional output vector.
9, 216-223. u(t) k-dimensional control vector.
Cruz, J.B., J.S. Freudenberg, D.P. Looze (1981). A Ty(t) m-dimensional set-point vector.
relationship between sensitivity and stability
z(t) p-dimensional feedback vector.
of multivariable feedback systems. IEEE Trans.
q(t) d-dimensional external disturbance input
Automat. Contr. 26, 66-74.
vector.
Doyle, J.C., G. Stein (1981). Multivariable feed-
back design: Concepts for a classical/modern A nxn constant matrix.
synthesis. IEEE Trans. Automat. Contr. 26, Β nxk constant matrix. D nxd constant matrix.
4-16. C mxn constant matrix. Η pxn constant matrix.
Doyle, J. (1982). Analysis of feedback systems
G(s) kxp closed-loop control.
with structered uncertainties. IEE Proc. 129,
G (s) kxp open-loop control.
242-250. 0
Kummel, M., L. Foldager (1983). Geometric control P(s) nxk nominal transfer function matrix between
of a distillation column. Symposium on "Appli- and control inputs.
ed control and identification", Lyngby, Den- P'(s) nxk perturbed transfer function matrix be-
mark, Preprints 3, 10-14. tween states and control inputs.
MacFarlane, A.G.J., D.F.A. Scott-Jones (1979).
e(s) error between nominal and perturbed re-
Vector Gain. Int. Journal of Control 29^, 65-
sponses.
91.
Index c: closed-loop.
Palazoglu, Α., Y. Arkun (1985). Robust tuning of
Index o: open-loop.
process control systems using singular values
and their sensitivities. Chem.Eng. Commun. 37, Tl(s) mxm transfer function matrix between the
315-331. outputs y(s) and the set-points r ( s ) .
Schizas, C , F.J. Evans (1981). A graph theoretic
y
T2(s) mxd transfer function matrix between the
approach to multivariable control system de- outputs y(s) and the external disturb-
sign. Automatica Γ7, 371-377. ances q(s).
Takamatsu, T., I. Hashimoto, Y. Nakai (1979). S(s) Sensitivity matrix.
A geometric approach to multivariable control σ(Μ) minimum singular value of matrix M.
system design of a distillation column. Auto-
matica 15, 387-402. σ(Μ) maximum singular value of matrix M.

-0.174 0.105
0.522 -0.943 0.469
0.522 -0.991 0.529
0.522 -1.051 0.569
0.522 -1.118 0.662
A = 0.522 -1.584 0.718
(min)"* 0.922 -1.640 0.799
0.922 -1.721 0.901
0.922 -1.823 1.021
0.922 -1.943 1.142
0.115 -0.171

0 0 0 0
0.00328 -0.00244 0 0
0.00384 -0.00288 0 0
0.00400 -0.00304 0 0
1 0.00376 -0.00280 0 0 1 0 0 0 0 0 0 0 0 0 0
Β =
0.00308 -0.00232 0.4 0 0 0 0 0 0 0 0 0 0 0 1
(mol)" -0.00312
0.00236 0 0.00236
0.00288 -0.00382 0 0.00288
0.00308 -0.00412 0 0.00308
0.00300 -0.00306 0 0.00300
0.00032 -0.00042 0 0.00032

TABLE 1: Linearized Mathematical Model of Binary Distillation Column

Fig.l Closed-loop configuration. Pig.2 Open-loop configuration.

Fig.7 Nominal and perturbated responses of state 1.

Fig.8 Nominal and perturbated responses af state 11.
180 M. Kummel and H. W. Andersen

- CTRl CTR2 CTR3 CTR4

Λ
Λ

Fig.3 Analysis of set-point tracking. Fig.4 Analysis of disturbance rejection. Fig.5 Analysis of output sensitivity.

Fig.6 Analysis of stability margin.

Fig. 13. The digraph of the distillation column
compensated with geometric control.

X IN 0.002
- X 1N
0.002-

o.ooo
• XIP
0.000- p. . ,—· 1
10 20 50
X 5N 0.002
- X 5N
0.002-
^ _- YPI • X3P
o.ooo
1 1
10 20 SO
X5N - X SN
0.002- 0.002-1
-- yqp - X5P
0.000 1 1 so
0.000
10 20
s - * XBN - X 6N
0.002-
— ypp - X6P
10 2Ό ~ -50

r..
X 7N 0.002 - X 7N
0.002-
Y 7P 0.000 - X7P
0.000 -
50

f
10 2Ό
X 9N 0.002
X 9N
0.002-
— X^P X 9P
0.000 - J 1
10
1 1
20 _ 50
0.000
50
XI IN 0.002
X I 1N
0.002-
-- yι i p " - - X I IP
0.000
)
1
10
I
20
t
50
0.000

2-
1 U IN - U 1N

— U IP - IMP
r , 1 1
10 20 50

2-
1 ..„,- U2N - U 2N

î r. 10
Fig.9 Geometric
20
1

control (CTRl).
- - U2P

Fig.10 Geometric control with pole shifting

- U2P

(CTR2).

0.002-
- Χ 1N

0.000-
• X IP

0.002·
- X 5N

0.000
• X5P

0.002
- X 5N

0.000
• X5P

0.002
- X 6N

0.000 • X6P

0.002 - X 7N

0.000· - X7P

0.002
- X 9N

0.000 - X9P

0.002
- XI 1N

0.000 - X HP

- U 1N
- U IP

- U 2N
mN - U2P
SO

Fig.11 Optimal control (CTR3). Fig.12 PI control (CTR4).

Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

EFFECT OF PROCESS AND MECHANICAL

DESIGN ON THE DYNAMICS OF
DISTILLATION COLUMNS
C. McGreavy and G. H. Tan
Department of Chemical Engineering, University of Leeds, Leeds LS2 9JT, UK

Abstract. Conventional design procedures for distillation columns have tended to avoid
consideration of how the interaction of the process and mechanical design affect the
dynamic behaviour and hence influence the control problem. This is because steady
state conditions are used as the criteria for design. Where dynamic changes are a
feature of normal operation, whether arising from feedstock changes, upstream
disturbances or as part of batch processing, the mechanical design features can play a
significant role in determining effective control. This is because it influences not
only the fluid dynamics but heat and mass transfer. In particular, ordered transfer
between operating states, as well as questions of stability of operation can be
strongly influenced by the mechanical details. The strong interaction which arise from
heat and mass transfer and the pressure distribution, together with the thermal
capacitance effects of the structure need to be taken into account. Consequently, to
ensure that the most favourable dynamics are exhibited for control, some care is needed
in the way the physical arrangements are specified. This contribution will present
information on the role played by the dominant process mechanisms and the bearing these
have on the design characteristics. Effective attack on this problem requires an
integrated CAD facility to define a computing environment which allows the complex
inter-relationships to be identified through the user interface. The intention is to
present an analysis of the problem and show how the dynamic behaviour can be shaped by
the mechanical design of the equipment to improve the control.

Keywords. Multicomponent distillation; column dynamics; mechanic design; thermal lags;

pressure profile.

INTRODUCTION the construction details of the column and so does

While it is generally recognized that effective
design and control of a complex chemical process
depends on the identification and characterization
of the principal dynamic features which determine
the time constants of the system, it is not often
the case that this is incorporated into
considerations of the mechanical design. A information is intended to provide a framework for
dramatic example where this can have a significant
effect is in the petroleum industry. The very
large distillation columns have long response times
which are usually not well characterized and it may
take such units hours or even days to arrive at the
desired operational steady state in the initial
start-up or during the transition between steady
states following a change in operating conditions.
Even on plants which operate on a much smaller
scale, the effects are still present. The typical
models on which attempts to carry out a dynamic
simulation are based tend to ignore the thermal
lags arising from the structure and predict much
faster responses than are observed. In addition,
to the slower response, high-order nonlinear
behaviour arises from the phase relations and
produce significant multivariable interactions
which pose difficulties in designing high
performance control systems and devising efficient
start-up procedures.
This work is directed towards an examination of the
factors which dominate these interactions through
the nonl inearities as well as the role that
mechanical design plays in influencing the overall
dynamic behaviour. To illustrate the general
principles, a case study based on a ternary mixture
comprising methyl-ethyl ketone (MEK), benzene and
toluene is used, with emphasis being placed on how
the constructional details influences the response.
This part of the analysis is concerned only with
not include the heat transfer processes in the
reboiler and condenser or the condenser-accumulator
holdup. Attention is specially directed towards
the general nature of the dynamics of startup under
various conditions, with particular reference as to
how the nonlinearities and interaction between the
variables determine the overall behaviour. This
further studies related to possible designs which
might give better response characteristics,
especially in achieving faster changes in operating
states. Two specific areas are considered:
1. The thermal lag associated with the metal
structure comprising the tray flow and column
walls.
2. The concentration lag arising from the liquid
holdup in the downcomer section of each stage.
MODEL OF A PLATE DISTILLATION COLUMN
Essentially a plate distillation column is designed
to be a sequence of staged equilibrium processes.
The relationships between the variables
representing the dynamic model of a distillation
column can be described by a set of differential
and algebraic equations. In principle, the dynamic
model should comprise of a full set of differential
equations describing how all the variables change
with time. However, because certain phenomena
occur relatively faster than others, these elements
can be regarded as instantaneously adjusting to new
states, so that the transient response is
controlled by the rate of change of the slowest
components. The complete set of equations is
essentially made up of:

181
182 C. McGreavy and G. H. Tan

mass balances The heat balance over the metal structure is:

heat balance
C dT
vapour-liquid equilibrium relationships w
12 W
U . (T. - Τ ) + U (T. . - T„.)
dt wlj j wj' j -0l J" J
momentum balances describing the tray
hydraulics. Associated with these are
equations for the physical properties such as
U
liquid molar density. w3 ^

The heat balance needs to include thermal lags

which are likely to dominate the overall response,
especially those which give rise to the
u
S( Tj w. - τ )
β (3)
characteristic slow response of operational units.
In particular these are associated with the heat
capacitance of the structure rather than the liquid
and vapour. The associated additional algebraic equation
relating the liquid molar enthalpy to the tray
Since this study is concerned with the column it is compositions and temperature is given by:
only necessary to consider a single generic
element, namely the general stage, i. (When the nc
whole separation system is included i.e. column, 0 = h. I x. . f{T.} (4)
reboiler and condenser, additional models need to
be added for the latter two units. Details of
these are given in (Tan, 1983). Vapour-Liquid Equilibrium

The basic equations describing the vapour-liquid

DYNAMIC EQUATIONS FOR A GENERAL DISTILLATION STAGE equilibrium on each tray is expressed as

A conceptual representation of a general

distillation stage is illustrated in Fig. 1. It
K. . x. . (5)
assumes :

1. Vapour holdup on and between stage is

negligible.
nc
2. The liquid on each stage is completely - I 1J
= 1 .0 (6)
mixed.
P. .γ. .
3. Local vapour-lquid equilibrium is
K. . 1,2,... (7)
instantaneously reached on each tray. φ. .P.
Failure to achieve this can be accounted
for by using an efficiency factor. Liquid and vapour fugacities (or activity
coefficients) which determine the equilibrium
Implicit in this model is the inclusion of the relationships are complex nonlinear thermodynamic
various heat transfer effects relating to the functions and , of course, are specific to the
liquid holdup on the plate and the column particular components being separated.
structural capacitances. Usually, this effect is
not included despite the fact that it can have a
significant effect, since the thermal capacity of Tray Hydraulics
the tray can be comparable in magnitude to that of
liquid. Associated with this is the influence of The liquid and vapour flow rates within the column
the various parameters controlling the rates of can be derived by taking momentum balances across
heat transfer between the capacitances. Figure 2 each stage. Since these are relatively rapid
illustrates the various heat flows between the tray hydraulic phenomena they can be assumed to reach an
metal on adjacent stages, as well as the liquid instantaneous steady state determined by the state
holdups. Given these assumptions, a mass balance variables. As a result, the relations describing
taken over the tray contents becomes: the flow can be expressed algebraically. In what
follows, the effects of weeping and entrainment are
neglected and the hydraulic gradient is ignored,
d ( XM although, including them does not affect the
Li i,i> = L. . x. . . + V. , + F . χ.. . + general conclusions. The total liquid holding on
dt T any plate can be expressed as:

F_ . y.. . - (L. - P .)x. . - (V. + P„.)y. .

T T (8)

with the vapour flow and pressure drop across the

i = 1,2,...,nc (1)
tray correlation being given by:

and the corresponding heat balance is given by:

0 (9)
dh
7 v
\i .i V. . (H. . - h.) + L. ,(h. . - h.) +
dt j-l j-l h j+l j+1 j and the height of liquid in the downcomer section

F .(h . - h.) + F„.(H.. - h.) - (V. + P„. ) (H. - h.) 0 Z Z

V r
Tv /v
y y3 y
U T (10)
« D . " FR, PR- l'j-1
~ Qj fj
LJ v j -fj ^ . ( Τ
- w l / j - wj> - Tj ) vj j (2)
2 Wj+ i
 Effect of Process and Mechanical Design
The liquid head necessary to generate flow and
overcome frictional losses of the liquid through
the downcomer apron is given by:
(11)
which results from a height of liquid on the tray
comprising the sum of the weir height and the
hydraulic head for liquid flowing over the weir
ζ Z +f } { L )
τ. - o j
Wj
While it is possible to refine this description to
provide more information regarding the downcomer
design, this would only be required for the very
detailed examination of a system: it does not
affect the characterization of the dynamics.
A consequence of this spread of intrinsic time
constants for the elements and the interactions
which occur is that the resulting set of equations
is stiff so that this places a considerable
computational burden on the numerical solution
(Tan, 1983). The consequence of this is that some
care is needed in ensuring that the solutions used
to predict the responses are reliable.
STARTUP OF A DISTILLATION COLUMN
The start-up of a column can be examined using the
above equations and specifying any two of the
variables Q , D, L. and R so that the overhead
vapour flow can be found. Using this model, it is
possible to show that the procedure adopted can
have an important effect on the efficient and
stable operation of the column during start-up, as
well as influencing the time to reach steady state.
The cases to be considered are taken from the data
given in Table I.
Start-up Procedure
A possible policy for start-up is:
1. Initially charge the column with liquid
at its bubble point at two positions:
(a) the feed plate
(b) the top plate
It is assumed that there is liquid at the feed
composition and temperature on all the trajs
and downcomers. The effects of weeping are
neglected and heat is introduced through the
reboiler.
2. Vaporization of the liquid increases the
pressure in the reboiler until it is large
enough to overcome the pressure drop due to
the orifice frictional losses and the tray
liquid head on the bottom tray, so that the
vapour passes up to the next. The hot vapour
heats up and vapourizes some of the liquid
contents on this tray and the whole sequence
is repeated until vapour reaches the top plate
of the column.
3. When the vapour flow up the column is
established, that leaving the column as the
overhead is condensed and returned as reflux.
The feed flow on to the top plate of the
column is stopped and the column is operated
at total reflux until the overhead composition
reaches prescribed conditions. Distillate
product is then withdrawn and the column
brought to a steady state.
DCCR-M
183
DYNAMIC BEHAVIOUR DURING STARTUP
Plots of the hydraulic variables during startup are
shown in Fig. 3. The time constants associated
with this process are of the order of 1 to 2
minutes, the larger value being that associated
with the longer holdup in the reboiler. It is
apparent that there is a considerable variation,
particularly during the first two minutes. The
column does not exhibit equimolar overflow
response. Instead the vapour flows vary
considerably from tray to tray. The vapour flow
response in the column is very fast: it takes less ( ,
than 1 second after startup for the vapour flow
changes to reach the top of the column. Hadju et^
al. (1975) found that significant vapour lags,
ranging from 10 seconds to 10 minutes, can arise
when the dynamics of the heat transfer process in
the reboiler are included, but the concern here is
with the intrinsic response of the column.
It should be noted that despite the very fast
vapour response, it is coupled to the slower
thermal and composition transients and so can still
take more than ten minutes before settling to a
steady state value. A comparison of the vapour
transients with the liquid composition transients
(Fig. 4) reveals that the shape of the vapour
response is similar to that of the light component
transients (MEK) . This is a consequence of the
fractionation which takes place so that the heavier
components accumulate near the bottom end of the
column and raise the boiling point of the mixture.
The light components rise up the column with a
corresponding drop in the temperatures on the upper
trays. For a fixed heat input, Q, , and cooling
rate, Q , the amount of vapour which is generated
increases progressively as the concentration of
light materials increase up the column. After
about 10 minutes, when the bulk of the
concentration and temperature changes (see Fig. 5)
have taken place, the vapour flows effectively
reach steady state values.
This response illustrates how the vapour flowrate
is coupled to the tray temperature (or pressure)
and composition through the
condensation-evaporation process. Thus, even under
normal unsteady state operation, whenever there are
feed significant changes the tray temperature or
compositions, as might occur with a change-over of
a feedstock, significant changes in the vapour
flows will arise.
The transients of the liquid flows and holdups are
closely linked, and the relationship between them
is dependent on the liquid head which generates the
flow. During startup, there is initially a very
sharp drop in the tray liquid flowrates as the
vapour boilup creates a pressure profile over the
column. This pressure build up on each stage has
the effect of increasing the downcomer liquid head
and holdup, but effectively reduces the liquid
flowrate on to the tray through the downcomer
apron. The liquid buildup in the downcomer will
continue until it is sufficiently large to overcome
the tray pressure drops arising from the pressure
profile across the tray. If the pressure
difference is unduly large relative to the
downcomer area, then there is the likely
possibility of the occurrence of downcomer
flooding. Essentially it is an operational problem
which can be attributed to poor structural design
of the column, and emphasizes the need to take this
into consideration at the design stage. The
example quoted by Peiser and Grover (1962) of a
field column that was experiencing operational
difficulties due to a tendency for the lower trays
to flood and was identified as arising from poor
tray hydraulics confirms this.
184 C. McGreavy and G. H. Tan

As can be seen from Fig. 5, the column pressure the lower stripping stages and decreasing in the
profile is established in less than 1 second, and rectifying stages (Fig. 6 ) . However, an imbalance
this is approximately the time taken for the vapour between the heat input and removal rate, such as a
flow to be propagated to the top tray. The lag in decrease in the reboiler heat input, can cause
vapour flow is essentially negligible and, since these composition-induced temperature changes to be
the top pressure is not controlled, the resultant masked by the larger changes in the heat flows, as
column pressures are determined primarily by the can be seen in Fig. 6.
pressure in the reboiler, i.e. any changes in the
reboiler pressure appear very rapidly in the There is evidence of considerable interaction
indicated pressures on the trays above. It can be between the mass and heat flows changes within the
seen that the general response of pressure on the column. The composition transients for three
trays is identical to the reboiler pressure different strategies using varying rates of reboil
variations, despite the differences in the can be seen in Fig. 7. It is apparent that, as the
temperature and composition transients. Thus, reboil rate is increased, the immediate effect
although the column pressure profile is established is to increase the vapour flow up the column,
in a very short time and the pressure drop per bringing more of the lighter components towards the
plate remains relatively constant, the individual upper trays. However, after about 100 seconds,
fluctuation pressure on the tray is considerable when distillate product is taken off at the
and may take more than 40 minutes to reach steady specified rate, more of the lighter components will
state. These long pressure transients arise leave the column. The net effect is a fall in
because the tray pressure is also coupled to the concentration of the lighter components throughout
tray temperatures and compositions through the the column leading to higher pressures and
vapour-liquid equilibrium process and so is linked temperatures.
to the slower concentration and temperature
changes. Because temperature has such a strong effect, any
drift in the column temperature profile which
Consequently, there are always two separate arises for example from the column structure acting
processes simultaneously influencing the tray as a temporary heat source or sink during the
pressures: the tray hydraulic mechanism which, is transient period can be expected to be significant
very fast and the thermodynamic equilibrium and needs to be examined in more detail.
constraints which cause the pressure to follow the
behaviour of the slower temperature and composition
transients. The latter can produce some unusual INFLUENCE OF TRAY METAL THERMAL CAPACITANCE
pressure transients. Generally, the tray pressure
will increase when there is either an increase in The considerable interaction which can arise
the tray temperature or the concentration of the between the heat and mass flow changes within the
lighter, more volatile components which have higher column has already been noted. Consequently since
vapour pressures. Conversely, the tray pressure the thermal capacitance of the metal structure is
will decrease with a fall in tray temperature or an large, the thermal lags within the system will be
increase in the concentration of heavier important in affecting the behaviour of the column
components. during transient conditions. Apart from the column
liquid holdup, there is considerable thermal
However, the open-loop pressure response is capacity associated with the metal mass of the
determined primarily by the pressure in the tray structure, C . This behaves as a large heat
reboiler. As heat is introduced into the reboiler, reservior during the transient period, which slows
the temperature rise produces a corresponding down the thermal responses of the column.
increase in pressure. As fractionation takes
place, with more of the heavier components As C increases, the time taken for the column to
accumulating in the reboiler, the w . . . .
reach steady state is significantly longer due to
temperature-induced increase is masked, resulting the larger thermal lag of the system. Typically,
in a slight fall in the column pressures. After the total time taken to reach steady state is about
about 2 minutes, when the bulk of the concentration 1,000 seconds when ^C^ is neglected and from 3,000
change has taken place, the continual addition of seconds (C = 5*10 ) to more than 25,000 seconds
the heat causes the pressure transients to be (C = 2.5*Υθ ) when it is included. In this case
dominated by the temperature effects, resulting in (see Table I ) , although the difference in the top
a sharp increase in the tray pressures until steady tray temperatures predicted for each of the
state is reached. It would be expected that if constants above are not very large, about 2 C at
there is a gross imbalance in the heat flows into most, this represents about 30% of the total
the column, e.g. either an exceptionally large temperature change of 7 C. Thus, although the
increase in the heat supply or the cooling rate, temperature response is slow, it does not have a
then the column pressure transients will be significant influence on the composition.
dominated by the thermal effects, and the
concentration induced effects cannot be clearly Generally, a major source of large variations in
identified. tray temperatures arises from heat input to the
reboiler. This causes significantly different and
The column composition responses are largely slower transient responses because of the thermal
determined by the mass flows within the column as effects arising from C . There is associated with
well as the external product flow distributions. w
this a relatively large damping of the temperature
Thus, any change in the product distribution, e.g.
responses which produces noticeably different tray
an increase or decrease in distillate flow, will
pressure and composition transients.
cause the column composition to shift so as to
satisfy both the internal and overall material The compositions can also exhibit rather unusual
balances. Prior to startup, the compositions on transient behaviour as a result of the effects of
all the trays should be the same as the feed the thermal lag. The resultant slower temperature
composition. After startup, the vapour boil-up changes acts as an additional constraint on the
will bring the lighter, more volatile components variation in tray composition, which in turn
towards the upper part of the column, whilst the influences the resulting vapour-liquid equilibrium.
returning reflux carries the heavier components A consequence of the lower tray temperature and
towards the lower section. This fractionation pressure transients arising from the slower rate of
process is accompanied by corresponding temperature temperature change is to drive more of the lighter
changes, with the tray temperatures increasing in components (MEK and benzene) up the column. This
 Effect of Process and Mechanical Design
can be seen in Fig. 8 where the concentration of
toluene in the distillate product is ,0.1313 (for C
7 composition, e.g. by temperature.
= 0.0) and 0.1258 (for C = 2.5*10 ) after 1 ,ΟΟυ
w
seconds, i.e. a difference of 0.0045. As a result
of the large thermal capacitance, the steady
composition of 0.1313 is only reached after a
further 8 hrs.
It is evident that the effect of C on the column
dynamics is significant and consistent with the
long response times observed in the large
distillation columns, where the value of C can^be
expected to be significantly larger than ^ t ^ c ) .
These long transients are obviously undesirEble
because they make control of the column difficult.
Although, improvement in the response might be
possible by increasing the heat transfer
coefficient between the liquid and the metal
structure but this proves not to be the case (Tan,
1983).
EFFECTS OF DOWNCOMER DESIGN
During startup, the effect of a large fraction of
downcomer area with respect to the plate area, FDA,
is to increase the time taken for the liquid
buildup in the downcomer to reach a level
sufficiently large to overcome the pressure drop
acting in the opposite direction. This results in
lower tray liquid flows, reducing the liquid
inventory in the reboiler and hence the bottoms
liquid product flow since a flow level controller
is used. The effect of this reduced holdup and
liquid product flow is to increase the separation,
as shown in Fig. 9.
On the other hand, the total time taken to reach
steady state is longer for the case where FDA is
large. This arises from an increase in the total
column holdup due to the larger downcomer holdups.
The resulting variations in the total liquid holdup
might amount to as much as 20% on a particular
stage. This difference in the downcomer
capacitances results in significantly different
composition transients (see Fig. 9). In
particular, for large columns, with a large number
of stages, the effect of neglecting the downcomer
hydraulics results in significantly different
transients and response times being predicted.
This highlights the need to pay attention to those
aspects of mechanical design which will have an
influence on the dynamic characteristics,
particularly where they would be of assistance in
start-up and control.
It is evident that simple models which neglect the
downcomer hydrodynamics would be limited by their
inability to predict the occurrence of operational
problems such as downcomer flooding which are
likely to occur during the transient period, e.g.
as a consequence to a small value of the downcomer
clearance area. The effect of decreasing the
downcomer clearance area, AGAP, results in higher
frictional losses and lower downcomer liquid flows
during the transient. This suggests that the
mechanical design of the trays should take account
of the dynamic characteristics which would be
desirable to assist in start-up and control.
GENERAL OBSERVATIONS
The following characteristics of the dynamics of a
column during start-up have been highlighted:
(a) The direct response to disturbances which
affect the vapour flows and pressure is very
rapid, i.e. of the order of seconds.
3.
(b) The variations in column pressures can make it
185
difficult to obtain indirect measurements of
This has
implications in designing control systems.
(c) The assumption of equimolar overflow is not
valid because significant variations in flow
rates arise. Knowledge of this is required to
avoid sudden surges in vapour boil-up.
Failure to recognize this tends to cause a
temporary reduction in the liquid flow through
the downcomer clearance and a consequent
liquid build-up leading to flooding during the
transient period.
(d) Considerable interaction occurs between the
heat and mass transfer processes within the
column. An important consequence of this is
that the column dynamics is strongly
influenced by the thermal lags arising from
the thermal capacitance of the structure.
While these observations arise from consideration
of start-up, they are also relevant to factors in
the design of a control scheme. Since the
objective is usually to maintain a minimum product
purity at all times, it focuses attention on the
need for rapid measurement of compositions if high
purity products are required. The scope for using
temperature to infer the composition is very
limited unless the effect of pressure variation is
accounted for. This means that the model necessary
for minimum realization is correspondingly more
complex. Moreover, given the tendency for the
pressure transients to exhibit inverse response,
columns which operate under temperature control
should include some form of pressure compensation
if good composition control is required. In any
case, for multicomponent systems, temperature is
not strictly a unique indicator product
composition, even if the pressure is held constant.
Where high purity products are required the
temperature changes will tend to be so small that
the controller will be essentially insensitive to
the small but significant variations in
compositions that occur. Distortion which arise
from pressure transients will obviously exacerbate
this. Without direct measurement there will be a
need for an efficient state estimator if high
performance is to be achieved, which will mean a
considerable increase in the complexity of the
control algorithm.
All this points to involving the pressure as an
integral part of an overall control strategy. If
such an approach is adopted, it is possible to take
advantage of the speed of propagation of the
pressure response, which has been shown to be very
fast. The pressure can then be used as a control
variable to manipulate the other column variables.
To do this, it is necessary to use a dynamic model
which is able to account for the effects of
pressure variations along the lines demonstrated
above.
Consideration of the thermal capacitances and the
heat transfer properties of the tray metal on the
column dynamics has revealed that:
1. The effect of the tray metal capacity, C , on
the column dynamics can be considerable,
particularly when it is largej: than the liquid
holdup thermal capacity, M^c^ .
2. The consequential influence of the large C on
W
the column composition transients is
significant only when there is a considerable
change in the column temperatures, e.g. more
than 10 C.
When C^ is large, smaller values of U ^ (the
product of the overall heat transfer
186 C. McGreavy and G. H. Tan

coefficient between the tray liquid holdup and REFERENCES

metal capacitance and the total heat transfer
surface area) will cause an increase in the Tan, G-H., (1983). Influence of capacitances on
thermal lag. the dynamics of multicomponent distillation
columns. Ph.D. Thesis, University of Leeds.
4. The effects of the interstage heat transfer Hajdu, H., Foldes, P., and Mizsey, P. (1979). New
parameters on the time for the column to reach Interaction effects in distillation column
steady stage are negligible. dynamics. I.Chem.Ε. Symp. Ser., 56,67-87.
Grover, S.S., and Peiser, A.M. (1962). Dynamic
The magnitudes of the time constants are consistent simulation of a distillation tower, Chem. Eng.
with those observed in industrial operations and Progr ., 58.» 65-70.
hence points to the factors which dominate the
response of such columns. The implications this
has for control system design are that reduced
models should not be based only on concentration,
but should include temperature as an explicit state
variable and also allow for the effects of
pressure.

NOMENCLATURE

Fy Vapour feed flowrate (mol/s)

h Liquid molar enthalpy (j/mol)
hf Liquid feed molar enthalpy (J/mol)
H Vapour molar enthalpy (j/mol)
Hf Vapour feed molar enthalpy (J/mol)
Κ Equilibrium coefficient
L Liquid flowrate (mol/s)
L
r Reflux flowrate (mol/s) H
M Stage liquid holdup (mol)
L Saturated vapour pressure of
component i (mmHg)
Liquid sidestream flowrate (mol/s)
Ρ^ Vapour sidestream flowrate (mol/s)
Q Heat loss from stage (j/s)
t Time (s)
Τ Stage temperature (C)
T Metal wall temperature (C)
U
w Overal heat transfer coefficients between Fig. 1 Conceptual representation of distillation
w l ' liquid column

hold-up and metal wall on and between adjacent

w3 stages (J/(s C))
Overall htc between metal wall and
surroundings (J/s C)
Vapour flowrate (mol/s) κ \
\

*Di
Mole fraction of component i in liquid stream
Mole fraction of component i in downcomer \ N \
Mole fraction of component i in liquid feed \
*fi \ T \ \
Mole fraction of component i in vapour stream
\t
J Height of liquid on tray (cm) J

k \
Total tray pressure drop (cm)
PR
Dry tray pressure drop (cm)
Pressure drop due to frictional and flow
^FR
resistance in downcomer clearance (cm)

\\_
M,
GREEK LETTERS \ \ Lj
^
Ύΐ
φ.
Liquid activity coefficient of component i
Vapour fugacity coefficient of component i
\ \
\ y » K \ \

\
\
SUBSCRIPTS
\

i Component number
j Stage number
L Liquid phase
Ν Condenser stage
0 Reboiler stage
V Vapour phase
Fig. 2 Conceptual representation of the column
ABBREVIATIONS structure

htc Heat transfer coefficients

MEK Methyl-ethyl ketone
 Effect of Process and Mechanical Design 187

MO -
810
135 0 - 102.5
140 - / Λ -
132.5 - 100 Û

79(^ 9 7 .5
/
/ .
. 30 0 -
<î iuo • 780 95.0

i 80 '
127.5-
ô 92.5
/ /
•7 .' 125 0 ^ 770
J \ \
S 90.0
* 60 122 5 - \ 750

\ 8 7 .5
iu - 120.0- 750
\
20 · 7 3 117.5- J 1 4 7*0
J 4
85.0
4
82 5
10 10 10* 10° 10 10 10 10 ' 10 10 ' 10 1 0' io' io
Π Μ Ε (sec) T I ME ( s e c ) T I ME (sec) T I ME (sec)

1 0? s.

VJQ υ

97 5 - —reboiler

υ // ---tray \

//y
95 0
en f e e d I ray

Τ39 2 5 — tray 3
- t op tray

Ml
90 0

B7 5

7 1 4 36 0 rt J 4
10 ΙΟ ΙΟ
10'
10 10* 10 10
T I ME ( S e c )
time ( s ejc

- Reboiler Fig. 5 Temperature and pressure transients during

-tray 1
the startup of a 5-staged column
-teed tray
-tray 3
- t op tray

Fig. 3 Hydraulic transients during the startup

of a 5-staged column

0.225
0.50 |
0.200|
0.45
„ 0.175,0 5 0.4 0
/

{/
ο °! 104
c S 0.3 5
830
~ 0 125
102
Ζ ο.ιοο
g 0.3 0
s 820
1 0 0|
x 0.075 \* 810
I
0.2 0 \ 96
Σ
//
0.050-
800
-
0 025i 0.15
:
10 10 10
T I ME
10 10' » » 10
790
// 4
94
/ 1
//
(sec ) TIME (sec )
780
92UL
10° „y ι • 10 10° 1 0* 10 10"
io' io'
T I ME (sec )
T I ME (sec )

UO 0 .12 \

102
0.10 -
100

u 98 £ 0.08
σι Ε
*» 96
TJ

e 94

I 0' 10* 10* 10 W 10 ' W* »'

92
/
0 .04
V 4
90 1—L-
T I ME (sec ) T I ME (sec )
10° 10' 10* 10' 10' 10* 10' 10

T I ME (sec ) T I ME (sec )
reboile r
D is î3 0 0
tray 1
0is= 2 5 . 0
feed tra y
tray 3
top tra y Fig. 6 Effects of varying distillate product
flow on the startup of a 5-staged column
Stage No. 0 (Reboiler)
Fig. 4 Liquid and vapour compositions transients
during the startup of a 5-staged column
188 C. McGreavy and G. H. Tan

7S0
:
700

ON 6S0
X
Σ
Σ 500 -

C
550-

£00 e 1 1 4
10 10 ' 10 10 10' 10' 10' 10* 10
TIME ( s e c ) T I ME (sec)

r e b o i lre
tr*y 1 0.160
0 032f
f e «d t r ay
t r ay 3 _ 0.155J
0.030
t op t r ay £ 0.tSo |
ω 0.028
u s
-
χ
0.025
I °· iλ
0.02 Ν o.uo
0 022|
0.135 i . /
10 15 20 0.130 10 15 20
Π Μ Ε (mins) T I ME ( m i n)s
Fig. 7 Temperature and pressure transients during
the startup (when the reboiler heat flow 0 . 8 S0
0.08 5
is decreased) 0 845

0 8i0|
Ο.Οβ
Ο
~ 0 835
I F
Υ 0.830 0.07 5
J 0825
0.070 |
^ 0.820ff
χ
0.815 0.06 5
0.8101
0.06 0
10 15 20 25 10 15 20 25
T I ME (mins) T I ME ( m l h) i

F O A « 00.
FDA » 0.05
F DA . 0.12

Fig. 9 Effects of varying downcomer area (FDA)

on startup response. (AGAP = 0.030)
Tray composition transients
Stage No. 0 (Reboiler)

10 ic 10*
TfMÉ (sccj
TIME (iecj
Cw = 0 . 0
C w « 2 . 540
- C w . 5 . 050
C w *2 5 0 6

Fig. 8 Effects of varying (when is increased)

on startup response. Liquip and vapour
composition transients. Stage No: 4
top tray
Qb = 4829070 J3s = 1300 J/ (s deg C)
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

PARAMETRIC TRANSFER FUNCTIONS FOR

BINARY TRAY-DISTILLATION COLUMNS
J. B. Edwards* and M . H. N. Tabrizi**
* Department of Control Engineering, University of Sheffield, UK
**Control Systems Centre, University of Manchester, Institute of
Science £s? Technology, UK

Abstract

Starting from familiar assumptions for dynamic analysis, a transfer-function-matrix

(T.F.M) model for long, symmetrical, binary distillation columns is derived
completely analytically. The model relates twin output composition changes to
perturbations in the internal liquid and vapour flow rates. The model is intended for
use in controller design and, being expressed in terms of plant parameters and operating
conditions should provide useful common ground between control engineer and plant
designer. Though derived precisely for a symmetrical plant, it is expected that the
model should apply more generally as a good approximation. Experimental data is given
supporting this claim. Inverse Nyquist loci are generated in precise and approximate
form and compare favourably with computed step responses. They accord broadly with
previous, part-empirical, part-numerical research results. A second companion paper
is devoted to packed-column analysis.

Keywords. Chemical variables control, Distillation, Fluid composition control, Freq·resp-

onse, Modelling, Multivariable systems,Nyquist criterion, Partial differential equations.

INTRODUCTION design (since the parameter- formulae include real

f
The accurate calculation of transfer-function mat-
rices (TFM s) f or the composition dynamics of col-
umns, by completely analytical methods, has not
been achieved hitherto despite several decades of
column research worldwide. TFM's have been either
experimentally or part-analytically derived.
Experimental data has been used to overcome diffi-
cult steps in the analysis so that the range of
model application to varying plants and operating
conditions is limited. It is common experience
that different columns differ enormously, yet
inexplicably, in their behaviour, especially if
trials are allowed to run for the full settling-
time. The difficulties arise from the complexity
of the expressions which develops early in the ana-
lysis of even the simplest hypothetical system
because of nonlinearity, spatial distribution and
from the two-stage construction of columns and the
associated boundary conditions. To simplify the
equations, apparently small terms are often discar-
ded but this can be disasterous because their effe-
cts cannot be properly judged à priori. Column dyn-
amics are driven by small differences in large
terms. Other, late, approximations may involve
lumping of spatial variables and relaxation of
boundary conditions. Reasonable agreement with
experiment has been achieved on trials on one-off
plants but it can be difficult to tell whether
success is due to the quality of the analysis or
the incorporation of empirical data often derived
from the same, or similar plant to that used for
final model validation.
Despite the difficulties of derivation, the need
for parametric TFM's remain for four main reasons:
(a) to test detailed column simulations (which are
fraught with numerical pitfalls).
(b) as a base for control system design.
(c) to check coarse TFM models obtained theoreti-
cally or by experiment, and
(d) to allow control design to influence plant
niant parameters and operating conditions).
In the present papers we model the primary driving
forces only, i.e. those thermodynamic drives which
make the column distil rather than secondary phen-
ena such as heat loss and hydrodynamics, We there-
foremodel ideal columns but, having stated our
initial idealing assumptions, we proceed without
further approximation to generate what we believe
are important benchmark models for small-signal
(i.e. linear) behaviour. Space restrictions allow
us to include only a few intermediate results of
the straightforward but lengthy manipulations.
(Edwards and Tabrizi, 1986).
2. SYSTEM EQUATIONS AND IDEALISATIONS
2.1 Equilibrium assumptions
1
We approximate the mixture s equilibrium curve by
the two straight lines:
(1f- X) = α(1 1 - Υe ) , rectifier (1)
and Y = α X , stripping section (2)
e
where Y is the equilibrium vapour mol-fraction of
e
the light component associated with a liquid of
mol-fraction
X whilst X ' is the equilibrium liquid mol-
e
fraction associated with a vapour of mol-fraction
Y' . .Slope parameter α is chosen such that 1<α<β =
the relative volatility and the lines cross on the
-45° line. Such piecewise linearisation has often
been used in earlier dynamic analyses, and notably
by Wilkinson and Armstrong (1957) et al. In tray
columns it is usually assumed that theoretical
trays operate in equilibrium so that
Y(h',t) = Y (h',t) (3) and X\h',t) = X'(h',t) (4)
f e e
where h is the distance from the top and bottom of
the tower respectively. As in introductory texts
(Judson King, 1971) on steady-state design, each

189
190 J. Β. Edwards and M. H. N. Tabrizi

tray is assumed ideal. The non-attainment of equi- are governed by

2 2 2 2
librium can have important dynamic results however,
as shown in the companion paper. 3 Y/Oh') =9 X'/Oh') = ο (11)
so that the quiescent composition gradients are
2.2 Heat-and mass-balances
constant, giving uniform loading throughout each
The plant variables and parameters are shown in section; a good operating condition. Furthermore,
Fig.l. For the heat balance, we neglect sensible for
heat changes, vapour compression and assume adiabatic
- 9Y/9h' = 9X'/9h' = G = constant > 0 (12)
operation as in introductory steady-state analyses
designed to yield insight rather than to generate then, nominally, V =L (=V) , V =L (13)
r s s r
detail. With equal latent heats per mol of the two
components therefore,rates of evaporation and con- and vapour and liquid feeds must equate, i.e.
densation on any given tray will balance so that the
F = F £(â F)
F„(ï (14)
vapour and liquid flow rates V , V , L and L will be
r g r s
independent of h': The well known constant-molar- yielding equal output rates, V - L and L - V =
overflow used by many investigators including
r r g g
F at top and bottom. Finally to avoid pinch effe-
Jawson (1965) Voetter (1957) Rosenbrock (1966), cts, the feed mixture should be in equilibrium as
Friedly (1974), Edwards (1977). In common with regards both sections, i.e. satisfy Eq. (1) and (2)
others, we neglect vapour capacitance in comparison so that vapour and liquid feed compositions ζ and
to the molar liquid capacitances Η^δη and Η^δη of Ζ should be:
each rectifier and stripping section tray. The mass
balance for the lighter component-on a typical tray ζ = a/(l+a)and Ζ = 1/(1+α) (15)
therefore yields a simple differential equation in- Solution is thus confined to a range of special
volving finite differences w.r.t H\ to which Taylor's cases but this is not unduly restrictive as it is
theorem may be applied assuming a large number of a range of ideals for which practical design should
! This, and eliminating
trays at small spacing oh. strive. Even tray-loading is but one attribute of
X(h',t) in favour of Y(h ,t) using Eq. 1 gives the this special-case range and is a condition which
partial differential equation (pde) has often been assumed implicitly in other column
studies, both dynamic and steady-state.
3(HY)
α 3* δη= -CaL -V ) Φ - δη+i (aL^)
9
r r 2.5 Normalisation
(5) Inserting the foregoing operating conditions in
>(<5h)
pde's (7) and (8) and normalising by setting
Oh')'
whilst Tfor the stripping section we deduce: h = h'/δη , τ = t/T (16)
3(H'X ) jvy! where base time Τ is the time for liquid to travel
— | — ôh = (L -aV ) Sh+iU+aV ) bâse-distance δη ?.e.

dit s s AH s s Τ = αΗ δη/V = Η δη/L (17)

Δ ι , η £ £ r
(δηΓ (6)
Oh')"
2
we obtain the greatly2 simplified forms
9y/9x = 92
y/9h 2
- (G/V) (v - ai) (18)
Criteria for the implied long column assumption (i.e.
trunkating at Sh^ are investigated in Section 4. and 9x'/9î - 9 ' / 9 h + (G/V) (£- αν) (19)
x
2.3 Linearisation for small perturbations Their solution requires knowledge of inputs ν(τ)
and £(τ) and also steady-state composition gradient
Linearisation of nonlinear pde's (5) and (6) is
G (see Section 2.7) and the process boundary
essential at the outset for T.F.M's derivation.
conditions considered next.
Linearisation about the steady-state, for constant
1
H^ and H^' produces the following p.d.e's relating
2.6 Boundary conditions (large-signal)
small variations y and x in Y and X' to small per-
turbations ν and I in vapour and retlux rates viz: 2.6.1 Accumulator and reboiler. Denoting accumul-
ator and reboiler mol-fractions by Y(o), and X(o)
JUL 2

νν f^* w
(α h
6
+
- (a and Y'(ο), X'(o) respectively mass balances on
Oh') 2, these vessels may be written:
(δη)
9y δη(α£-ν) + {
2 (δη) (a£+v) (7)
9{H Y(o)}
a 9Y δη
V {ε-εΥ(ο) + } (20)
9h' Oh') at 9h'
h'=o
9x' (δίιΓ and 3iH,X'(o)} , oh}
On ; L (21)
2 Tt s {
-*'^ w +
h'=o
dh
+ ^ i - δηα-αν) + J ϋ - ν (6h-T (A+av)(8)
if ε = α-1 (22)
Oh')'
normalising to: 9iT Y(o)}
and if ν and I are kept sufficiently small, steady- (23)
=ε{1-Υ(ο)}+
state values may be substituted for all the capital 3τ
variables in (7) and (8) thus yielding linear rel- h=o
ationships between y,x' and v,£. 9{T X'(o)}
b = -εΧ'(ο) +
9X'
(24)
9h
2,4 Operating conditions and plant parameter h=o
constraints
where T
Solution of pde's (7) and (8) is practical only if
a = H a/ H ^ h and T
b = H b/Hôh
H and H^ being the molar capacitances of the acc-
they and their boundary conditions are rendered a
umulator and reboiler.
symmetrical by appropriate choice of plant para-
meters and quiescent operating conditions. We ther-
2.6.2 Feed-trays. A mass balance on the rectifier
efore make L' and oh the same for the rectifier and
stripping section and set feed tray yields
9H X(L)
ctH =
£ H^ (=H) (9)
F * V
x^ δη (26)
Now if, nominally, aL = V and aV =L„ (10)
+ eY'(D-Vr Y ( D - L
r h'=L'
' r r s s
then, from (5) and (6), steady state compositions on applying a first-order Taylor expansion and from
 Binary Tray-Distillation Columns 191

Eq. (10), (13) and (14) we deduce that Eqs. (32) - (36). Their symmetry allows the ready
diagonalisation of the system if we adopt the out-
F - e V (27)
put and input vectors:
so that substituting for F and ζ in (26) gives,
y(h,r) - χ'(ΐι,τ)
after normalisation q(h,T) (37) and
y(h,x) + χ'Οι,τ).,
9Y(L)
^_ X<(L) U-Y(L)}-f
(28)
3τ + +
h=L ν(τ) + 1(τ)
u(x)= £ (38)
and, similar treatment of the stripping section's
ν(τ) - £(τ)_
feed tray yields
Laplaoe transforming the pde's in s w.r.t. h and in
3X'(L)_
^-x'a)-{i-Y(L)} fl (29) ρ w.r.t. τ gives
3x + h=L
-ε 0
2 s * -1
2.7 Large-signal steady-state solution: (s - p) £ - s q(o) - q(o)=s (39)
0 a+1
justifying symmetry.
where q_(s,p) is the double transform of £(h,x),
Subject to boundary Eqs. (23), (24), (28) and (29)
,c[(o) is the transform of qXo,x) w.r.t. τ, u that of
and setting 3/3τ = 0, the large signal steady state
U ( T ) and £(o) that of 3q/3h at h = o. The end
Eqs. (11) may be readily solved to produce the sol-
vessel conditions (32) and (33) yield, if Τ =T =T,:
ution for Y(h) and X'(h) from which we note the a b
constant value of slope; Τ ρ q(o) = - eq(o) + q(o) (40)
G = 2 ε/{(α+1) (2cL+a+l)} (30) 2 ~ -1
so: (s -p)£ - (s+ c-Tp)q(o)+ s u=0(41)
and the fact that X'(h) = 1 - Y(h) (31) ο -(α+1)
The results are necessary for substitution in the
and inversion back to the h,p domain, and setting
solution to our linearised small-signal model (18)
h = L therefore yields
and (19) but have a strong practical appeal in them-
selves. The constancy of G, leading to even tray q(L) - Q ( )q(o) - Q (L)u = 0 (42)
loading, has already been anticipated and discussed
iL 2
where Q^(L) = {/pcosh/p L + (c+Tp)sinh /pL>I_//p(43)
but equation (31) represents another important att-
ribute of the special case we have adopted. It
cosh/p L-l -ε 0
indicates that, from a 50/50 mixture of components and Q (L) = (44)
(ζ + Ζ = 1.0), a symmetrical plant and operating
2 [0 α+Ι^
regime produces top and bottom products that are
Now the feed Eqs. (36) and (37) may be expressed
equally pure, nominally. In cases where the feed +
composition were richer (weaker) then, to avoid 3
Q £(D ^ 1^ • ν (45)
pinch effects, this would be entered nearer to the 3 h=L
accumulator (reboiler) so approaching a single-
stage situation, capable of easier analysis (e.g. p+2 , 0
where = 0.5ε 0 (46)
and Q.=
Armstrong and Wood (1961)). —4
0 0 , ρ
-0.5(3α+,1)
2.8 Small signal boundary conditions. allowing unknown q(L) tobe eliminated between (42)
and (45) giving
These are derived from the large signal boundary
equations (20), (21) and (26) (and a similar equa- q(o) = G(o,p)u (47)
tion for the stripper feed tray) and the assumed
G(o,p) relating and compositions to flow rates
(symmetrical) operating conditions. For the acc-
being given by:
umulator for instance we get, on implicit different-
iation :
G(o,p) = Q Q + 3h (48)
Q -Q Q ah
ctH M2l ^
ν{-Υ(ο)ε+ε·
3Y
6h} V {- y(o) ^ '<Sh}
— L7"3—1
h=L
4 32-
a 3 t 3h'
t + rf i£ +ηf h'=o
3 The matrices of Eq. (49) are all diagonal so that
the first-term of the R.H.S. being zero in steady ,P) 0" )
state so that normalising we get simply G(o,p)
g (o,p)J
22
h=o
(32)
for the transfer F Jg(o,
and knowing Q .,.Q^, we quickly obtain expressions
n functions g^(o,p) and g ( o , p ) ,
22
these being, respectively:
whilst for the stripping section we obtain
dx'(o) 3χ' ε{(ρ+2) (cosh/pL-l)/p+(sinh/pL)/^ρ+Ο.δ}
εχ'(o)+ (33) (50)
b 3τ 3h {(l+T)p+2+e}cosh/pL+{(p+2) (e+Tp)+pKsinh/pL)//p
h=o
(a+1) (cosh/p L-l) + (a+l) (sinh/pL)//p+0.5(3ot+l)
From feed tray Eq. (26) and noting, from the steady (51)
state solution, that: {p(l+T) + c}cosh/p L + /p(l + ε + Tp)sinh/p L
Y'(L) - Y(L) = -G(a+l)/2 (34) 2 for ρ = ο are:
the limiting values
we deduce g (o,o) = e(L + L + 0.5)/(2 L+ct+ 1) (52)
11 E
9y(L) (a+1)
= V
ν
+a.}- f +x'(L)-y(L) (35) and g (o,o) = -{(a+l)L + 0.5(3a +1)}/ε
22 (53)
h=L
4. PREDICTION OF COLUMN BEHAVIOUR
Similarly, the stripper feed equation is found to
be:
A completely analytic T.F.M. model relating the
1 S behaviour of y(o,x), χ'(ο,τ) to ν(τ), £(τ) has thus
+y(L)-x'(L) (36) been derived in terms of normalised complex fre-
3τ V 2 3h
h=L 1
quency ρ and normalised column length L, both read-
T
ily converted to real frequency p and real length
SOLVING FOR THE T.F.M. L by the formulae
The small signal behaviour of the column given the
flow disturbances v(t), &(t) is now completely spe- p L /(H oh) (54) and L' = L6h (54)
cified by p.d.e's (18) and (19) and boundary
192 J. Β. Edwards and M. H. N. Tabrizi
f
The real time constant T of the end vessels is ob-
tainable from its normalised value thus:
τ' = Τ H ôh/L = H /L = H. /L (55)
£ r a r b r
?
In terms of y,x,v and I, the model may be expressed:
y(o,p)-x (o,p) g Co,P)
=G V 11
y(o,p)+x'(o,p) g (o,p)
22
(56)
where composition gradient G readily obtained from
α and L via Eq. (30). Some aspects of dynamic be-
haviour are now obvious whilst others require an
examination of the formulae for g-Q and g22· 0^
the former, we note that, over the entire frequency
range, composition-tilt, y(h,x) - x'(h,T) is driven
noninteractively by average circulating flow {v.(τ)
+ £(τ)}/2 whilst composition-total y ( h , O + χ'(η,τ)
is driven purely by take-off rate ν(τ) - £(τ). We
note also that the tilt gain g^(o ,o) is positive
whilst g22(o,o)is negative and that for low-relative
volatility mixtures (ε << 1.0) requiring long col-
umns (L >>1.0-see steady state solution), then
I g-Q(°>°) I « | g (o,o) I · These findings accord well
22
with simulation experience (Rosenbrock, 1966).
Rademaker's deductions (1975) and Shinskey's empi-
rical data (1963). They, reported that separation
is influenced predominantly by ν(τ) but our conclu-
sions are in close accordance for low-relative
volatiles where large reflux ratios are needed, so
allowing |(v + i)\ » |v - i\.
4.1 Inverse Nyquist Loci
Typical loci of g;Q(o,ju)) and g (o,joo) computed
22
from (50) and (51) are shown in Figs. 2 and 3 for
α= 1.1 (ε = 0.1), L = 10, Τ = 1.0 (corresponding
to a separation 2 GL = 0.46). Their near-vertical
nature over a wide range of ω, predicts a nearly
exponential step response as confirmed by the sim-
ulation results of Fig.4. The observed static gains
(=2.6 and -225.0) and time constants (= 25 and
140) accord closely with those predicted by the
loci (i.e. gains of1 0.37~1 and -0.0048~1 and time
constants of 0.04" and 0.0077~1). The broadly
first-order lag nature of the responses, may be de-
duced analytically from Eqs. (50) and (52) however,
without recourse to precise computation by examin-
ation of the expressions for g-Q and g22 at large
p: Considering 5the band
1.0 » | p ° * | » 1/L
which is (a wide band for long columns), only the
lowest powers of ρ outside the hyperbolic functions
need be retained so that
&
(ο,ρ) +ε {2/p + l//p}/{2 + ε + 2 ε / / ρ } (58)
11
and g (ο,ρ)->-( +1)(1+1//ρ)/Ιρ(1+Τ)+ε+/ρ(1+ε+Τρ) }
α (59)
Now for mixtures of low 3 (ε << 1.0 and requiri-
ng L >>1.0) therefore:
gï (o,p) = (1+α)ρ/2ε
1 , 5
1 L «|p° |«1.0
and g (o,p) -αρ/(1+α)
22
The system therefore tends to integrate over a wide
range of frequency as observed, though its behav-
iour does become more complicated at higher frequ-
_ 1 05
encies. The band considered is wide since in the
region, L < < ω · << 1.0
|g * (o,juj)| - ω/ε and |g (o,jcu)|
22
and these values are large compared to the static
inverse gains
g ( o) , oL2 d/ a n ethanol/water curve). Now for ν = 3, I = 2 we
g ( o , o ) | - / 2 L (62)
l n I " 22 e
The near first-order lag nature of the systems
responses is thus predicted quantitively, for any
long column without computation. The prediction
accords qualitatively with previous experience.
4.2 Effect of terminal capacitance
Since Τ drops out of the high-frequency analysis
above it follows that only the final portion of the
step-response is influenced by changes in T. This
is indeed confirmed by the simulation result of
v(p)+£(p) Fig.5 for ε = 0.1, L = 10 with Τ = 1,20 and 50.
v(p)-£(p) 5 . APPROXIMATE MULTIVARIABLE FIRST-ORDER LAG MODEL
One of the motivations for deriving G(o,p) has been
to validate simpler models. Because the system is
class-0 and tends to an integrating process the
multivariable first-order lag approximant, Ga(O,p)
of Owens (1975) would seem to be applicable where
0^(0,p) A + A
—1 —ο (63)
-1
where A^ = Lim G (ο,ρ) and A = Lim50{p ^G \o,p)}
—ο ι ι —
p-»o
IpI-*
Here, we should strictly replace |p|-*°° by L <<
|p0.5| << -j^q approximant should still app-
bt u
e
t b
ly for controller design provided excessive gains
are avoided. Now A^ and ^ are readily derived
analytically. A-^ merely involves solution of the
small-signal spatial differential equations obtai-
ned by setting 8/8τ = 0 in (18) and (19) yielding
static gain formulae (52) and (53) directly. ^ can
be estimated even more simply from (18) and (19) by
ignoring after transformation all but the p-depend-
ent coefficients of the dependent variables giving
ol
ρ q -
The vertical d
a+1
a thus
loci
A -
—ο
n obtained
ω f o1 for
-(ot+1)
G Uthe
J aelements
r
•'J (64)
gÂntojJu) SÂ22 ^ »J ^ Â ^°»J ) f shown
alongside the corresponding elements of G l(o,ju))
in Figs, 2 and 3. The agreement is clearly ade-
quate over the very wide frequency range considered.
(This is not so with packed columns!). It is inter-
esting that A unlike A-^ has not required use of
0
the boundary conditions for its approximate deter-
mination above. This indicates that boundary cond-
itions affect only the low frequency behaviour of
the system, in accordance with our observations and
deductions concerning the effect of terminal capac-
itance in Section 4.2.
Controllers designed on the basis of the m.v. 1st-
(57) order lag derived from simulated asymmetric, tray-
column step responses have been successfully tested
in simulation by Edwards and Owens (1977) for gains
which increase the speed of response by up to five
or six times.
6. CORRELATION WITH PILOT PLANT DATA
Although the paper is intended primarily as a theo-
retical contribution, some comparison with earlier
experimental data has been possible. Step-tests
carried out on the 7m, 15 kw, 12-tray, water-etha-
(60) nol pilot column at Sheffield produced the compo-
sition changes, shown in Fig.6, measured on trays 1
and 10 with a liquid feed F = 8.8 mole/min @ Ζ =
0.16 into tray 6. The vapour rate changes from 5.0
to 8.0 mole/min at a constant reflux ratio of 2:1
making ν = 3.0 and i = 2.0. Steady-state output
compositions of 0.52 and 0.08 yield an average for
2GL of 0.44 and hence G = 0.0367. from which an
1
effective value of ε = 1.88 (α=2.88) is obtained
ω/2 (61)
via Eq. (30). (This suggests an effective relative
volatility of about 4.0 which produces an ideal
equilibriumcurve coarsely approximating the true
deduce from our linearised model that
1
= {-|g l+5|g |}G/2V and x'= {-| g ^ | -5 | \ }G/2V
2 2 ' ~u1 1 '
'"22' ' ~ Ζ Ζ ' "il'
and calculating the steady-state gains from (52)
and (53), therefore predicts y = zero (the and
 Binary Tray—Distillation Columns 193

g-^-j contributions nearly cancelling one another) No, 2, p. 177.

and χ = -0.069: in remarkably close accordance with McMorran, P.D. (1971) 'Application of the inverse
the measured change. Nyquist method to a distillation column model',
Proc. of 4th U.K.A.C. Control Convention p.122.
The pilot column has large terminal capacitances of
Owens, D.H. (1975) First and second-order-like
some 330 moles yielding a non-normalised value for
structures in linear multivariable control
Τ = 62 mins. Now where end-capacitances dominate
system design, Proc. I.E.E., 122, (9),
it is readily shown from (50) and (51) that
pp. 935-941.
cL Rademaker, 0., Rijnsdorp, J.E. and Maarleveld, A.
(1975) 'Dynamics and control of continuous
1
Lim {p g x
11(o,p)} = and 1Lim {p g 1(o,p)}
22 distillation units', Elsevier, Amsterdam.
L »p»T L »p»T Rosenbrock, H.H. and Storey, C. (1966) 'Computat-
(ot+l)L ional techniques for chemical engineers,
Τ Pergammon, London.
Shinskey, F.G. (1967) Process control systems (New
and from this latter formula and the static gain
York: McGraw-Hill).
g ( o , o ) , a time constant of 42 min is predicted
22 f
for the x' response. The predicted near-exponen-
Stainthorp, F.P. and Searson, H.M. (1973) Trans.
Inst. Chem. Engrs. 51, 42.
tial response for x shown in Fig.6 is clearly most
Voetter, H. (1957) 'Plant and process dynamic
encouraging as is the very small deviation of the
characteristics', Butterworths, London.
y-trace, in accord with prediction. The transient
Wilkinson, W.L. and Armstrong, W.D. (1957) Chem.
initial reversal of this trace is not unreasonable
t s
since, in the near counter-balance situation bet-
Eng., Sci., 7(1/2).
ween the ffects of g-Q and g 2 J ^ ^ quite poss-
2
ible, with slight parameter-change, for the high-
Fig.l. The complete system
frequency response to differ in sign to that of
the ultimate response.

7 CONCLUSIONS

A TFM model relating output composition changes y

and x' to changes ν and I in vapour and liquid
flow-rate has been found analytically. Its para-
meters are derived-functions of plant parameters
and operating conditions. Whilst its purpose is
controller design, it should allow dialogue between
control engineer and plant designer at an early
stage in a project. Our assumptions are those fre-
quently made in dynamics studies, (usually implici-
tly). These are: a piecewise-linear equilibrium
curve, spatially independent flow rates, constant
tray loading and holdup, zero vapour capacitance,
tray equilibrium and symmetry. The model should
be sufficiently robust to apply, as a good
1
approximation, to well-designed and non-symmetrical
plants. The TFM between Qr-x , y+x'] and [v+£,
v-z] is diagonal and the individual transfer-funct-
ions predict near-first-order lag behaviour and a
rapid method for determining such approximants has
been demonstrated. These findings accord with
earlier, part empirical, research. More complex
behaviour at very high frequency is, however, pre-
dicted and a companion paper on packed columns
shows that first-order lag behaviour is not always
a safe assumption. Terminal capacitance is proved
to affect only the low-frequency end of the system's
frequency response. Many more predictions could be
made and experiments planned using the model as a
basis.

8. REFERENCES

Armstrong, W.D. and Wood, R.M. (1961) 'An intro-

duction to the theoretical evaluation of the
frequency response of a distillation column to
a change in reflux flow rate', Trans.Instn.
Chem.Engrs., Vol.39 pp. 80-85.
Edwards, J.B. and Jassim, Η .J. (1977) 'An analyti-
cal study of the dynamics of binary distilla-
tion columns', Trans. Inst. Chem, Engrs., 1977,
55, pp. 17-29.
Edwards, J.B. and Owens D.H. (1977) First-order
type models for multivariable process control,
Proc. I.E.E., 124, (11), pp. 1083-1088.
Jawson, M.A. and Smith. W. (1965) Counter current
transfer processes in non-steady state, Proc.
Royal Soc., (London), A225, p. 226.
Judson King, C. (1971) 'Separation Processes'
McGraw Hill, New York.
Kim, C. and Friedly, J.C. (1976) 'Approximate
dynamic modelling of large staged systems', Ind.
Eng., Chem., Process Pes. Develop,, Vol. 13,
194 J. B. Edwards and M. H. N. Tabrizi

Fig.2. 1 and lst-order lag loci for

Accurate Fig.4 Unit step responses
gi i" (o>Jo)

-I U
300 600 900 1200 τ

ω = 0.91

lOj Accurate

— 1st Order lag ε =·1, L = 10, Τ = 1

Δω = 0.2

8j
0.72
m I

0.4j , 0-.039
6j
0.54

0.2j - ιω - 0 . 0 2
1
Ο 60 120 180 240 τ
' 1 1 1__
.36

1 Fig.5 Effect of Terminal Capacitance

2j ,τ - 1

I 1

Fig.3. 1 (
Accurate ω) ( lst-order lag loci for
and
£23.~ Μ

-0.01
1 0
400 800 1200 1600

Fig.6 Pilot plant response to V-change

Y(tray 10)
u; = 0.01& -O.Olj

II 0.525

ll
0.032/1 -0.02j 0.520

0.515

0.049 J 1 -0.03j measured

predicted
X'(tray 1) , step applied

0.066 1 I -O.04J

0.085 1 1 JD.05J

IO 0 t 02105

(ω-values same on both loci)

Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

PARAMETRIC TRANSFER FUNCTIONS FOR

PACKED BINARY DISTILLATION COLUMNS
J. B. Edwards* and M . Guilandoust**
^Department of Control Engineering, University of Sheffield, UK
^Department of Chemical Engineering, University of Newcastle upon Tyne, UK

Abstract. Using similar assumptions to those adopted in the companion paper on tray
columns, a parametric transfer-function matrix is derived packed columns. For simpli-
city, calculations are illustrated for a column that is symmetrical statically and
dynamically i.e. having a vapour/liquid capacitance ratio c = 1.0 producing a diagonal
T.F.M. between the sum and difference of output composition changes and the circulating
and product take-off flow rates. The non diagonal T.F.M. is also presented for
c φ 1.0 and subsequent analysis and simulation show that the column behaviour is not
greatly affected by changes in this parameter. Long packed columns are shown to produce
novel nonminimum phase effects when twin product control is attempted, whereas serious
travelling-wave phenomena can limit controller performance in short columns. The poss-
ibility of zero separation-gain is also revealed. Computed inverse Nyquist loci are
confirmed by analysis and by numerical simulation. The chief cause of discrepancy
between tray and packed column behaviour stems from the continuous equilibrium assump-
tion for theoretical trays.

Keywords. Chemical variables control, Distillation, Fluid composition control, Freq.resp-

onse, Modelling, Multivariable systems, Nyquist criterion, Partial differential equations.

1. INTRODUCTION 2. EQUILIBRIUM AND MATERIAL BALANCE FORMULATIONS

In the companion paper, Edwards and Nassehzadeh Vapour and liquid streams are imagined separated
Tabrizi derived a parametric transfer-function mat- vertically by a conceptual inter-phase barrier and
rix (T.F.M.) for long, binary columns of the tray the column split horizontally into a number of con-
type. Simplifying assumptions were made at the out- ceptual cells for which dynamic mass balances are
set and, thereafter, the derivation proceeded com- derived. The equilibrium curve is again linearised
pletely analytically and without subsequent approx- thus :
imation. Their assumptions involved the piecewise
linearisation of the equilibrium curve, constant (1-X) = a(l-Y ) (1) Υ' = αΧ' (2)
e e
molar overflow, zero vapour capacitance and equal
tray-loading. The additional assumption made for Taking a material balance for the flow of lighter
ease of analysis, was that of static symmetry, in- component through and within an arbitrary cell of
volving equal lengths of rectifier and stripping the rectifier and stipping section, applying a
section, feed-composition coordinates located at Taylor expansion and letting cell length δη 0,
the knee of the linearised equilibrium curve, equal
T
yields the following partial differential equations
vapour and liquid feed rates and nominally equal (pde's), having eliminated X and Y using (1) and
take-off rates of top and bottom product. Such (2): f
operating conditions were shown to produce top and 3(H Y)/3t - V 3Y/3h = k (Υ -Y) (3)
bottom products of nominally equal purity, in terms ν r r e
of lighter and heavier component respectively: a -ot3 ( H Y )/3t - L a3Y /3h' = k (Υ -Y) (4)
condition judged by the authors to represent ideal £ ef r e T r e
plant design for the composition control of twin - 3(H'X )3t + L 3X'/3h = k (Χ'-Χ') (5)
products and therefore a practically useful special 1
case. Although precisely derived for this special Z s s e
a3(H'X')/3t + aV 3X'/3t = k (Χ' - X ) (6)
case, it is anticipated that the model should still ν e 1 s e s e
stand as a useful approximation for asymmetric sit-
where Y and X are the actual vapour and liquid mol-
uations, hopefully up to the point where a single-
fractions in rectifier and stripping section respec-
rather than a twin-stage analysis might become more tively, V (V ) , L ^ L ^ are the vapour and liquid
appropriate. Supporting experimental data was given molar flow rates in che rectifier (stripper), Η (Η')
In the present paper we apply similar techniques to H£(H£) the vapour and liquid capacitances p.u.length
packed columns which yield rather more complex equ- for the rectifier (stripper). As before, h' denotes
ations and important behavioural differences. Cal- the distance measured from the end-vessel (accumu-
culations are outlined for the simpler care of unity lator or reboiler).
vapour/liquid capacitance ratio, c, because of
space limitations but the results for the asymmetric 3. NORMALISATION
case of c φ 1 are given. For derivation details see
Edwards and Guilandoust (1981 and 82). All symbols
As previously, we shall choose a plant built and
used have the same significance as in the companion
operated symmetrically thus
paper.
V
r= aL
r = L
s= aV
s = V (7)

195
196 J. Β. Edwards and M. Guilandoust

with a mixed feed of vapour, at rate F , composit- perturbation model) may be calculated by setting
ion ζ and liquid at rat
v
rate F , composition Ζ where 3/3τ=0 and solving the resulting spatial de's sub-
ject to their now static boundary conditions. The
F = F (8) ζ = αΖ α/(1+α) (9) ea r
solutions, given in Fig.l, for X'(h){=1-Y(h)} and
£ Ν
Η = αΗ' = Η. (10) Η' = αΗ = Η Xé(h)i=l-Y (h)1 linear in h for our chosen
ν ν 1 £ £ 20 (11) e
symmetrical conditions. In particular:
and, by suitable choice of packing density and
dY , dX'
section geometry: dY
dh
e
=dX'vf= e
= (28)
k = k = k (12) dh dh dh
s r
where, as for the tray-golumn, G is given by
Eqs. (7) and (8) producing nominally equal product
flows (=F) at top and bottom. It should be empha- G=2c/i(a+l)(2eL + a +1)} (29)
sized however, that Eq.(7) describes only the nom- Also, the equilibrium and actual composition prof-
inal (quiescent) operating condition of the plant iles are separated by unit normalised distance i.e.
and does not prevent the application of small, by base distance V/k in real distance h'.
independent changes of v,£ in, say V and L .
g r 6. SMALL PERTURBATION EQUATIONS
Under these conditions the system pde's may be nor-
malised to
χ x νd a nr
Implicit differentiation of pde's (3)to(6), writing
ΎY Ye» '» è> ^ f ° small changes in Y,Y ,
3(CY)/3T - 3Y/3h = Y "Y e
X', Xg, V and L respectively and substituting the
e s r
calculated steady-state values for the remaining
- 3Y /3τ ~ 3Υ /§h = Υ - Υ
upper-case symbols yields the following normalised
e e e 1 small-signal pde's.
- 3Χ'/3τ + 3X'/3h = X' - Χ
c 3y/3T-3y/3h + G v/V = Y ^ - Y
3(cX')/3x + 3X'/3h = Χ' -eΧ'
e e e
where normalised distance h and time τ are n b - 3y /3x-3y /3h + aG£ V = y -y
give y
(30)
h = h'k/V (14) and τ = tk/Hζ (15) - 3χ'/3τ+3χ'/3η + G£/V = x'-x '
the base time H /k being the time for liquid (in
2
the stripping section) to travel base distance V/k c 3x'/3T+3x'/3h+aG v/V = x'-x'
which may be regarded as the tray spacing of the e e e
equivalent tray-column. It has a further physical After double Laplace transformation in s w.r.t. h
significance demonstrated in Section 5. The vapour and in ρ w.r.t. τ using superscript ~ to denote
/liquid capacitance ratio c is defined as variable transforms w.r.t. h and τ and ~ to repre-
c = H!/H (16). sent transforms w.r.t. τ only, these equations may
2 be written thus
4. LARGE-SIGNAL BOUNDARY CONDITIONS
1 y y(o) 1 0
4.1 At the feedpoint. -1
Q (s) + (31)
If L is the normalised length of either section of 0 α
y <°>
the columns then L = L' k/V (17) e
and the feed boundary conditions are simply and
rx^(o) α 0
x'
V Y'(L) + F ζ
s v r
T
V Y(L) (18) Q ~ V s ) ~e -x'(o). 1
=0
(32)
x'_ 0 1 L- J
and L X(L) + F Z
r £
L X (L)
s
(19) χ x 2
1
Substituting operating conditions (7) (8) and (9) y(°)>y
e
(°)> '(°)> '(°) being the transformed vari-
e
and eliminating X(L) and Y (L) in favour of Y (L) ables It h=o and
and X'(L) using Eqs. (1) and (2) yield the norm-
1 + cp - s , -1
alisea" forms (33)
Q \s) =
+1 = V V, -(l+p+2)
X'(L) + {1 - Y(L)} = 2/(a+l) (20)
e (34)
and [ z z l (G/ )[ ^J
and {1 - Y (L)} + X'(L) = 2/(a+l)
e
(21) 12>
Now some of the unknowns, say y (o) and x^(o), may
e
be eliminated at this stage using the small pertur-
4.2 At the end vessels.
bation version of boundary Eqs. (25) and (26) which
Material balances on the accumulator and reboiler may be written:
(of constant capacitance H and H moles respecti-
a b
vely) , yields the differential equations (de's) y (o)
e f 1
y(o)
(35)
H aadY (o) x'(o)
i (o)J
-~- = V [ail-Y (o)}-{l-Y(o)}] (22)
r e
X C)6 f
where h (ρ) = 1/(1+Tp) (36)

V e = L {X (o)-aX'(o)}
Inverting back to the h,p domain, then gives
(23) setting L = h, :
dt s e
if the end vessel runs in equilibrium. y(L) ' Γι olRI
For symmetry, we set H a= = H (24)
+ y(o) Q(l) -1 «a, L o 4 j £ j - o .0(37)
a e
so that dY (o) and

dT
=α{1-Υ (o)}-{l-Y(o)}
e
(25) x'(L)'
x -HP α 0
e '(o)Q (L) -R (L) (38)
and dX '(o) i ' ( L) 0 1
X'(o) - X'(o) (26)
dx where Q/h), Q*(h) are the inverse Laplace transf-
-1
orms (w.r.t.^h) of Q(s^, Q(-s) and R(h), R*(h) are
where Τ = Η /V (27)
e those of s Q(s), - s Q ( - s ) . The feedpoint bound-
ary conditions (derived by implicit differentiation
5. STEADY-STATE SOLUTION of (18) and (19) and substitution of the now known
steady-state operating conditions) may be expressed
The system is now completely specified and steady- thus
state conditions (needed as parameters for the small
y(L)=x^(L)-(c/2)
Z1 (39)and x'(L)=y (L)+(c/2)z
e 2 (40)
 Packed Binary Distillation Columns 197

g ( o) , p
where ε = α-1 (> 0) (41)
n
so that (39) and (40) may be used to eliminate the (51)
feedpoint variables (i.e. at h = L) from (37) and
(38) giving
-1 g11= _î 1
(ε/Ρ)(coshqL-l)-(l+q)(sinhqL)/q- ,/2 (52)
1 y(o) (l-a h )(q/p)sinhqL+(l+a~ h )coshqL
QUO - ι , e e
,Q*U) X'(P) (cp/q2) (coshqL-l)-(l+a) (sinhqL/q^/2
(53)
§2?" -1 -1
(42) (p/q)(l+a h )sinhqL+(l~a h )coshqL
1 0 α θ"| [L 6 e e
RCL) +R(L) and, for the static gains, taking limits as p-*o we
0 a.
[oij " I [olJ
obtain
g n( o , o ) =a{cL 2
-(a+l)L- e
/2}/{2 £
L+a+r} (54)
Only outputs y(o), x'(o) inputs ΖΓ and z now rem-
2
ain and Q(L), Q*(L), R(L) and R*(L) are readily and g (o,o) = -a{(a+l)L+ /2}/ (55)
determined from Eq.(33). Straightforward matrix 22 e £
T
algebra can then obtain the TFM between 8. BEHAVIOURAL PREDICTIONS
£y*(o) ,χ' (o)J and [v,£jT. The work required is
tedious and painstaking to avoid disastrous errors. Much can be learned by inspection- and straightfor
Space constraints here demand that we illustrate ward simplification of Eq. (52) thro (55) as is now
the method with the simple special case of c=1.0 demonstrated.
but results for c φ 1.0 are given in Section 9 and
their detailed derivation is described elsewhere
8.1 Static gains
(Guilandoust and Edwards, 1982).
We compare firstly the expressions for g-Q(o,o),
8. T.F.M. DERIVATION FOR c = 1 g^2(o,o) deduced for packed columns (Eqs. 54 and 55)
with those derived in the companion paper for tray-
From (33) we deduce that columns, viz: 2
-(1+p+s) 1 g (o,o) = e(L +L+0.5)/(2cL+a+l) (56)
11
Q(s) = 2 2 (43) and g (i,o) = -{(a+l)L+0.5(3a+l}}/c (57)
j -q -1 1+p-s 22
2 ? The results are very similar for long columns
where q -p^+2p (44) (L»1.0) but it is important to note that g-Q(o,o)
can go negative for shorter packed columns. Such
thus giving Q(L),{Q*(L)} columns can still yield practical separations 2GL
(see Eq.(29)). It is at high frequency major diff-
-(1+p)(sinhqL)/q+coshqL, (sinhqL) /p (45) erences between the two columns appear, however.
-(sinhqL)/q ,(1+p)(sinhqL)/^coshqL
8.2 High-frequency (H.F.) behaviour
{top sign for Q(L) and bottom sign for Q (L)} from Noting that , if ρ = jco,
which we obtain the L.H.S of (42) as:
q ·+ jo) + 1 , L 0 « ω (58)
then it is readily show, from (52) and (53), that
Q(L) Y(o)+Q (L) x'(o)
-1, Lim {ja>g (o,ja)) }+ -J , Τ » 1.0 (59)
1;L
T 1 << ω << 4/ε _1
a sinhqL-coshqL , bsinhqL-c coshqL y(o) or + - J/{l+exp(-2qL)a }, Τ = 0 (60)
(46) where J = 1 -aexp(-2qL)+ε exp(-qL) (61)1
-(bsinhqL-c'coshqL) ,-(asinhqL-coshqL)J x' (o).
where a={a ^ h • ( l + p ) > / q , b={l-a h ( l + p ) } / q and c' g therefore approaches pure integration at H J ,
0 e
= -a -41 · Exploiting the obvious symmetry of (46) but here the H F., gain is negative, unlike the tray
we can therefore write _ column where H.F. gain g-j^ is positive.
e
(a "*~h -1) (q/p) sinhqL-(l+a *"h ) coshqL , 0
e
For long columns separating difficult mixtures (i.e.
-1 -1 ε-^l.O requiring L>>1.0 for a reasonable separation
0 , -(1+a h )(ρ/q)sinhqL-(l-a h )coshqL 2GL: see Eq.(29) the exponential terms in (60) and
" (47) (61) are clearly negligible but for larger values of
y(o)-x'(o) 1 l" ε , necessitating a shorter column for the same
separation, the term exp(-qL) acquires importance.
_y(o)+x' (o)_ 1 -i_ R.H.S. of equation(42) It represents a composition wave reflected from the
Turning to the R.H.S of (42), R(s) is just s ^ ( s ) feed boundary and yields the approximate H.F. tran-
and hence R(L),{R*(L)}= sfer function
1
Lim {jo)g (o,joi)+ {l+ exp (-L) exp (-j L ) (62)
±(l+p) (l-coshqL)-qsinhq 1 ,+ (1-coshqL) ~|
(48) 1« «4/έ
e W
|_±(l-coshqL) ,+(1+p) (1-coshqL)-q sinhqlj
ω (JJ A
An identical expression may also be derived for the
and after careful manipulation we obtain H.F. limit of g22(°>J )* Diverging loops on the
inverse Nyquist loci of gx|(o,ju3) and g22^°>J^) may
Γι -il therefore be anticipated to become increasingly not-
R.H.S (42)
1 -1 iceable as L is reduced, each loop occupying a fre-
quency increment Δω~"2πί-1
(ε/ρ(1-coshqL)+(1+α)(sinqhL)/q+c/2,0 ζ„ +ζ
J(49) L Λ 8.3 Inverse Fyquist Loci 0
, , 2. ,„ , , Χ · , ,Χ /
Λ 1 2 Loci of g{j;(o,jaj) and g22( »J(A)) from Eqs (52) and
(47) and (49) therefore yield the TFM relationship
(53) are given in Figs 2 and 3, the 'short' column
parameters being ε=0.75, L=2.8 whilst for the
fy(o) - x'(o)" V+£*
'longer'column, ε=1.0 and L=5.0. T=5.0 in both
=G(o,p) (G/V) (50)
+ x' (o) cases. The stronger wave effects on g~l in the
e short column case are obviously present} whilst in
where G(o,p) is diagonal, viz G(o,p) = the longer column (which has a positive
198 J. Β. Edwards and M. Guilandoust
g^(o,o)) nonminimum-phase behaviour of g-^ is pre-
dicted by the locus orbiting the origin. The pred-
1
ictions are borne out by unit-step-responses compu-
ted from the pde s (30) and shown in Figs. 4 and 5.
The nonminimum phase behaviour of g-^ for the large
column is obvious. The steady-states and basic
time-constants of the step-responses accord with
locus predictions. The inverse Nyquist loci also
confirm the analysis of 8.2 and by step tests.
Being open-loop do not bring out the predicted wave
motions which become significant only on closed loop
control and it is shown from (63) that, in the sho-
rter column, taking feedback temperature measure-
_1
ments at, say, h=0.5L (=1.0) would generate an add-
itional H.F.phase-lag φ, [=sin {eexp (-L+h)}], of
22°. Twin product control of packed columns is
therefore made difficult by additional phase-lag
due to travelling waves in shorter towers and by
the non-minimum-phase behaviour of g-Q in taller
plants. Neither problem was revealed by analysis
of tray columns and, in practice, the continuous
equilibrium assumed for the tray column may not
occur leading to behaviour between the two idealised
cases.
8.4 Non-minimum-phase behaviour of g^^*
A physical explanation
The initial negative response of y(o,τ)-χ'(ο,τ) is
readily explained in physical terms. When V and
g
L are simultaneously increased, (i.e. v+£>1.0),
r
weak vapour from the bottom of the column will be
f
moved initially upwards, whilst rich liquid will be
moved initially downwards so reducing Y(o)-X (o).
Only then, will evaporation and condensation begin
the change towards the new operating condition.
The final response may be positive, with a suffic-
ient value of L, then the predicted non-minimum-
phase response is explained. The argument does not
rely on the assumption that c=1.0, or indeed that
c= any specific value.
8.5 Low-frequency (L.F) behaviour: Effect
of Terminal Capacitance
Additional features of the system frequency resp-
onse can be deduced directly from expressions (52)
and (53) without computation: The L.F. behaviour
of the system may be examined by retaining the first
power of ρ in approximating these expressions as
ρ->Ό. Such a procedure applied to (52) leads to the
result
g^(o,p)^(o,o)(l+T p)/(H-T ),L|p|»1.0
2 2 2 lP
Τ = -L {2L(a+l)-3e}/6{cL -(a+l)L-e/2j
χ )
T = L{(a+l)L+e}/(2eL+a+l), (T=0) }
2 their similarity we deduce that control is little
2
Τ =T, T =aT(2L+l)/(2cL+a+l) , Τ » 1
1 0
so that, for small |p|
g ^ ( o , p ) - g^(o,o) { l + i ^ - T ^ p }
As with tray-columns therefore, end-vessel capac-
itance, affects only the L.F. response and hence
the final tail of the step-response, this predict-
ion being confirmed by the step-responses of Fig.6
for L=5, ε=1.0. In addition provided Τ >T ,
then the initial direction of locus departure is
+ 90° for gn(o,o)>0,0 (i.e. for long columns) and
-90° for g-Q(o,o)<0,0 (i.e. for short columns), in
accordance with the computed loci of Figs. 2b and
2a respectively. T clearly exceeds T^ as required
2
(a) in the presence of T, provided L>0,5, and (b) with
very large and very small L when T=o. Intermediate
cases need further study.
9. UNEQUAL VAPOUR AND LIQUID CAPACITANCE (c φ 1.0)
The condition c=l can be approached practically
(Strigle and Perry (1981)) but generally, c « 1 . 0 .
Fortunately, the system behaviour is not greatly
sensitive to wide variations in c, as is now demon-
strated.
Once we set c^l.O, matrices are no longer diagonal
and expressions become much more cumbersome. Great
care and persistance however will produce G(o,p)
given in matrix Eq .(6.9)friext page). It is readily seen
that the off-diagonal terms become zero at p=o and
gll(°>p) and g 2(o,p) are identical to the values
2
given for the c=l case(Eqs. 54 and 55) as would be
expected. At higher-frequencies, however, the
T.M.F ceases to be diagonal since the plant is now
asymmetric.
As an extreme example, setting c=o, then from (69)
we deduce:
1 , 1
LimG(o,j )^,exp(-2L)-l-1.5eexp(-L)
ω-χ» .1 > 1
+ eexp (-L) exp (-j OJL) Γ (70)
4
-ι
and the Nyquist loci converge to small orbits of
radius eexp(-L)/4 about the real point -{1+1.5 ε
f
exp(-L)-aexp(-2L)}/2 . This checks with the original
pde's (30) noting that y and x will remain zero
e
initially after a step in v+& because of liquid
capacitance. Furthermore terminal capacitance will
keep y (o) and xl(o) at zero initially. Thus from
e
Eqs. (31) and (32) :
(l-s)y= -y(o)-i /s (71) and (l+s)x^= -z /s (72)
1
whichjwith boundary Eqs. (30) and (40) and, knowing
that χ'(ο) = ο initially gives:
1
y ( o ) - x ' (o)
y(o) +X (o)
u; - si
= "£ 1.5cexp(-L)+l-ctexp(-2L) } ' " I V i,l (73)
2
Apart from the small orbital terms therefore, the
initial unit step-response Eq.(73) is identical to
LimG(o,jou) as ω-*».
This compares most favourably with the initial neg-
ative step-response of Fig.7 computed for L=5,
ε=1.0, T=5, and 0<c<l. The persistence of the ini-
tial negative dip clearly is not a feature of the
c=l case alone.
Fig.8 compares the loci of g^Co, joo) for the extr-
emes c = 0.0, 1.0 aiidthe parameters of Fig.l. Their
1
similarity confirms the transient response. Fur-
thermore, if G*(o , ju)) =G~- -(q, ju)) then loci of
g^[J(o,ju)) and g*^(o,jmy contrast the performance of
(63) the control loop linking y(ο)-χ'(ο)and ν+£ with the
other loop (between y(o)+x'(o) and v-£) open and
tightly closed respectively. Fig.8. also compares
( 6 4
these two loci for the extreme case of c=o and from
affected by c.
(65)
10. RECONCILING TRAY AND PACKED COLUMN BEHAVIOUR
Parameter c (ignored in tray columns) has little
(66)
effect as we have seen. Parameter L causes the
major variation however. Furthermore, as L is in-
creased the behaviour of packed and tray columns
approach equality. This is demonstrated by Fig.9.
f
This suggests a reduction in the nonminimum phase
dip of y(o)-x (o) as L increases as confirmed by
Fig.10 for L=20, for packed and tray columns of
identical static gain. The effect of c - change is
f
also shown. Recalling that L = L'k/V for packed
columns then, for a given L and V, an increase of
L implies an increase in evaporation coefficient k
and hence a closer approach to equilibrium through-
out the tower, so explaining the convergence of
behaviour at large L.
11. CONCLUSIONS
A parametric T.F.M. model for packed columns has
been derived completely analytically. Apart from
symmetry, all other assumptions are those made in
the part-analytic studies of other investigators.
 Packed Binary Distillation Columns 199

1 - c
Defining ρ (67)

cosh PL sinh PL
and q = /(l+c)p(4+p+cp)/4 (68) then 6 ( o , p ) -
T^ f sinh PL cosh PL

P L P L P L VIr
ί ( l + c ) c o s h q L [ ( + 2γ- p ) e +(c- ^ = | ^ p ) e " ]- q s i n h q L [ ( a + c ) e + (1+ca)e~ \ - ε (1+c) [2+ ( i + + ^ | ) p ]
C (l+c)p(l+a h )c'osh q L + 2q(l-a h )sinh q L
c
e PL e P L P L P L 2
(1+c) c o s h q L [ ( l + a + ρ) e -(1+af ^ | 2 p ) e ~ ] - q sinhqL[(a-c) e + ( l - c a ) e " J - ε (l-c )p/2

( l + c î p d + c f ' H ^ î c o s h q L + 2q(l - a'S i ) sinh q L

P L P L P L P I (69)
q c o s h q L [ ( a + c ) e - ( 1 + c a ) e " ] - ( 1+c)sinh q L [ ( + ^ p ) e
e_ 1 + ( - +
e ^ψ- ρ)e" ^-
e(1-c)q
(1+ΰ)ρ(1+αΛe )sinh P - Ia
P qLL + 2q(l h
e
)cosh q L
P L P
q cosh qL[(a-c)e -(l-ca)e" 3 -(l+c)sinh qL[(l+a+ )e +(l+cc+^^p)e~ ^-2e d+c+^)
p q
d+cïpd+a^h )sinh q L + 2q(l - a~\ )cosh q L

The model should apply as a good approximation to

Fig.l. Steady-state composition profiles
towers operated with a degree of asymmetry. If the
chosen outputs and inputs are Qr(o)-x'(ο),
y(o) , + x' (o)] and [v+£, v-jQ then the T.F.M. is
diagonal if vapour/liquid capacitance ratio c-l.
When c^l, the T.F.M. is diagonal only at zero
frequency. Variation of c does not greatly affect
system behaviour however.
An essential difference between packed and tray
f
columns is the nonminimum phase response of y(o)
- x (o) to v+£. The negative dip in the step-
response is found to persist when c is reduced in
the range 0.0<c<1.0. The static separation-gain
g-Q can vary in sign depending on parameter values
and, for short columns, becomes negative. Short
columns have been shown to produce travelling wave
phenomena which can contribute additional phase-
shift of some 22° so affecting controller design.
Thepossibility of zero separation-gain arises, so
explaining the difficulties sometimes encountered
in controlling this quantity in practice. As might
be expected, increase of normalised length L (e.g.
by increase of evaporation constant k in a given
column) brings the responses of packed and tray
columns into close proximity.

Finally we should stress that the control problems

raised by this analysis apply to two-product Fig.2. Inverse Nyquist Loci for g^o.ju)
control and not necessarily to single-product
control.

REFERENCES " J2.0

Edwards, J., and Guilandoust, M. Nov. 1981. 'The -4.0 -2.0 0 -2.0 y^l.O Q Λ Re
• ι
influence of vapour capacitance on the compo- Re
sition dynamics of packed distillation columnsJ
Qy ω=1 -0 / to-1.0 1.0
• -jl.O
University of Sheffield, Dept. of Control Eng., \ 2.0 • 2.0
Research Report No. 166. -J2.0 . -J2.0
Guilandoust, Μ., Edwards, J.B. Feb. 1982. Ά 1 3.0 i 3.0
parametric transfer-function matrix for packed - -J3.0
binary distillation columns having unequal vapour
vapour and liquid capacitance', ibid, P.esearch
4.0 % > / 5.0 .-J4.0 Κ 4.0
- -j4.0
Report No. 171. / 5.0
• -J5.0
Strigle, R.F. and Perry, D.A. 1981, 'Packed towers • 6.0
reduce cost', Hydrocarbon Processing, Vol.60, "-J8.0 - -J6.0
No.2, p. 103. 7.0 X 7.0

- -J7.0
10.0 j --j 10.0
8.0 · ^
. -j8.0
J9.0
° > 10.0

(a) Shorter Column (b) Longer Column

DCCR-N
200 J. Β. Edwards and M. Guilandoust

Fig*3. Inverse Nyquist loci for 8??(ο^ω) Fig.6. .Effect of terminal capacitance on g-^(o,p)

Im J0.5
- J0.5

e R
0 Re o
1 * I
-0.5 0.5 -0.5 0.5

/
ω=0.3 ω=0.3
• -J0.5 - -J0.5
o.e
0.6
0.9"
- -jl.O
0.9 / - -jl.O
1.2

1.2 /
1.5
- -jl.5 - -j.15
1.5 /
1.8
21
1.8·

Λ 2.1 - " -J2.0

- -J2.0
2.4
"T
ΙΟ 0 0 21 0- 13 0 Γ4 0 5 0 6 0 7 0 8 9
2.7 \ - -J2.5
\-j2.5
-\
3.°7
* 3.0

(a) Shorter Column (b) Longer Column Fig. 7. Effect of c on g^(o,p)

Fig.4. Unit step responses of g (o,p)

11

(b) longer column

- 2 1.
- 01 . /
- 80 . /
- 60 . /
- 40 . /
- 0.2/

0 / τ

—r— 1
1 • ·
240 30 5
ι
/ 0 01 0

(d) shorter column

F i8
-1 *
§ « * Loci of g-^CojJo) and g^o.jaj)

Fig.5. Unit step responses of g^Ccsp)

Ι ς,(ο,τ) Im
L = 5
Τ = 5 , .
ε = 1 *-s
ω=0.1 > y \

* /
4 JO/ 0.5j \λ
* //

s<p 1 1,5 Re

9 J' -1.0 -0.5 0.5 1.0

 Packed Binary Distillation Columns 201

Fig. 2. An alternative
Fig. 2. An scheme for scheme for
alternative Fig. 2. Fig.
An alternative
2. An alternative
scheme for
scheme for
Fig. 2. An alternative scheme for Fig. 2. An alternati
Fig. 2. A
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

ESTIMATION AND CONTROL OF

DISTILLATION PRODUCT COMPOSITION
USING TRAY TEMPERATURE
MEASUREMENTS
M . T. Guilandoust, A. J. Morris and M . T. T h a m
Department of Chemical and Process Engineering, University of
Newcastle upon Tyne, UK

Abstract. Two adaptive methods for inferring process outputs which are subject to long
measurement delays, from other more rapidly sample secondary outputs, are presented.
Both e s t i m a t o r s have quite general structures. W h i l e one is based upon state-space
considerations the other adopts an input-output representation of the plant a s the
starting point for synthesis. In contrast to other published techniques, the proposed
methods require m i n i m a l design effort; do not result in static e s t i m a t i o n error a n d
help to reduce the problem of secondary measurement selection. Since their parameters
can be continuously estimated and updated on-line, the estimators are able to cope with
slow variations in process dynamics. Evaluation of the proposed schemes on a non-linear
binary distillation column model demonstrates that a significant improvement in perfor-
mance is achieved using the inferred values of plant output rather than the infrequent
measurement of the controlled output as feedback for process control.

Keywords. Inferential control; adaptive systems; distillation column control; adaptive

control; estimation.

INTRODUCTION
A common process engineering problem is the control
of outputs w h i c h can only be m e a s u r e d by slowly
operating, possibily expensive, analysing e q u i p -
ment. In many cases, such measurements are also too
infrequent to be used (directly) for the control of
the plant. E x a m p l e s of such situations are in the
product composition control of distillation columns
and chemical reactors. Large m e a s u r e m e n t delays
often result in poor load disturbance rejection
properties. A popular approach to o v e r c o m e this
problem is to infer values of the controlled output
from other easily measurable variables (secondary
outputs). Here, the l i m i t i n g rate of control is
determined by t h e s a m p l i n g i n t e r v a l o f t h e
secondary outputs, eg. Joseph and Brosilow( 1978);
Morari a n d Stephanopoulos (1980a,b); Patke and
coworkers(1982) and Wright and coworkers (1977).
From the published work, it appears that a m a j o r
problem in developing an inferential control scheme
is that of e s t i m a t o r design. This is because the
relationships b e t w e e n controlled a n d secondary
outputs can be quite complex. In this paper, simple
SISO estimator models, the parameters of which can
be identified on-line, are derived using two diffe-
rent approaches. The first is a state-space based
technique while the second is derived directly from
an input-ouput model. The adaptive techniques pro-
posed here avoid the complexities of non-adaptive
estimator design and are able to cope w i t h slowly
time varying plant parameters.
PROCESS DISTURBANCES
Process disturbances have been observed to be char-
acterised by sustained random changes lasting for
periods of a t l e a s t the process response time, eg.
Joseph and Brosilow (1978), Morari and Stephanopou-
los (1980a,b) a n d Tuffs a n d Clarke (1985). To
describe this behaviour, the noise (disturbance)
has been modelled as Brownian motion. This process
can be pictured a s a sequence of random steps of
independent a m p l i t u d e s , occurring at times d e s -
cribed by a Poisson distribution. Although inter-
pretations of this model for load disturbances may
well describe the physical situation, its use poses
an additional problem - the Brownian motion process
has a pole on the limit of stability (non-station-
ary) and, hence, is unobservable from the output.
For effective control of the plant, the controller
must also possess an equivalent pole w h i c h can
result in adverse closed loop behaviour (Wittenmark
and A s t r o m , 1984). The random step-like behaviour
of disturbances can alternatively be m o d e l l e d as
exponentially correlated noise (ECN) - a stationary
process, therefore avoiding observability problems.
In this paper, disturbances affecting the plant are
all assumed to be stationary processes so that each
may, according to the representation a n d spectral
factorisation theorems (Astrom,1970), be represen-
ted as the response of a stable filter to a w h i t e
noise sequence.
STATE-SPACE APPROACH

The state-space estimator assumes that the plant is

203
204 M. T. Guilandoust, A. J. Morris and M. T. Tham
deesribed by the following model:
x(tf1)=Ax(t)+Bu(t-m)+L0Kt) (1)
v(t)=Hx(t)+l^(t) (2)
y(t)=Dx(t-d)+i§(t) (3)
x(t) is the state vector and <*Kt) is a vector of
independent zero mean sequences. v(t) (available at
each time step) a n d y(t) a r e m e a s u r e m e n t s of the
secondary and controlled outputs respectively.
^j"(t) and y?(t) a r e the corresponding m e a s u r e m e n t
noises, assumed to be zero mean independent random
sequences, d is the analyser delay associated w i t h
y(t). m is the s m a l l e s t of the t i m e delays in the
responses of controlled a n d secondary outputs to
changes in the m a n i p u l a t e d input u(t). Any dif-
ference b e t w e e n the t w o delays can be included by
extending x. If the m a t r i c e s A, B, L, H, D, a n d
the noise covariances are known, Kalman filtering
can be used to compute the state e s t i m a t e s χ from
v(t). The controlled output is then estimated as:
y(t.-d)=Dx(t) (4)
The K a l m a n filter for Eqs. (1) a n d (2) are:
χ ( t> 1 ) =Ax ( t )+Bu ( t-m)+Ke( t)
€(t)=v(t)-v(t) (6)
v(t)=Hx\t) (7)
x(t) is the optimal estimate of x(t) w h i l e v(t) is
the filtered value of v(t). 6(t) is the innovations
sequence w h i c h is w h i t e a n d Κ is the l i m i t i n g
Kalman gain. The technique, h o w e v e r , does n o t
assume a k n o w n plant model. x(t) m u s t therefore be
computed by c o m b i n e d parameter a n d state e s t i m a -
tion. As Eqs. (5) and (6) a r e equivalent to Eqs.
(1) and (2) (Anderson a n d Moore, 1979), either
representation m a y be employed. It is, however,
more convenient to use the innovations representa-
tion. Since the state e s t i m a t e s can be c o m p u t e d
from Eqs. (5) and (6), filtering is not necessary.
Further, it is reasonable that the parameterisation
be p e r f o r m e d such that the identified p a r a m e t e r s
would directly correspond to an input-output model
(Ljung and Soderstrom, 1983). This is best achieved
by formulating Eqs. (5) to (7) in an o b s e r v e r -
canonical form. In such case, using v(t), the p a r a -
meters of Eq. (5) can be e s t i m a t e d recursively
(Shieh a n d c o w o r k e r s , 1983)· The state e s t i m a t e s
are subsequently obtained from Eqs. (5) to (7) a t
each time step w i t h o u t recourse to filtering.
Knowing D, y(t+d) c a n then be c o m p u t e d from Eqn.
(4). If the state e s t i m a t e s are not correlated, D
could be e s t i m a t e d using values of y(t) a n d Eq.
(3). However, this is not the case as m a y be noted
from Eq. (5). This problem can be o v e r c o m e by
noting that in the observer-canonical structure,
Eqs. (5) and a (7) a r e of the f o r m :
" 1
•
x(t-1)= b u(t-m)+ At) (8)
- r,
v(t) (9)
Elimination of the state e s t i m a t e s x«|(t) to x ( t )
n 1
from the RHS of Eqn. (8) yields:
x.,(t*1)=v(t*1)
x (t+1 )=-a v(t)-. .-a v(t-n+2)+b u(t-m)+..
2 2 n g
! .+b u(t-m-n+2)+k f(t)+..+k £(t-n+2)
n 2 n 2
x ( t> 1 ) =-a v ( t )+b u ( t-m)+k e( t ) (10 )
n n n n
Substitution of the state estimates into Eq. (3),
results i n :
y ( t ) =y&| u ( t-m-d-1 )+. .+^ _ ·, u ( t-m-n-d+1 )+
n
y v ( t-d )+..+tf _f t-n-d+1 )+
o n
<T,e(t-d-1 )+. . + J 6 ( t - n - d + 1 )+1£(t) (11)
n-1
Since Eqn. (11) does n o t dependent on the state
estimates,T its parameters could be estimated from:
e [θf(t-d)+e(t)
W hy(t)=0 Τ = (12)
$f τ l^1'"'Αι-1 A
'*o»"»*n-1 >">4-1 ]
r(t-d) = [u(t-m-d-1 ),.., v(t-d),.. ,£(t-d-1 ),.. ]
T7
and e(t) is t h e equation error. After obtaining θ
at time t, y(t)=9 f (t-d) is computed to update <P(t-
d). Finally, the updated θ is used to estimate the
controlled outputT from:
y(t*d)=e f(t)
=^<|U(t-m-1 )+. .+/5 -|U(t-m-n+1 )+ Î^v(t)+..
+1ίη-1 ν<t-n+1 )+ ci, £( t-1 )+..+ <£_ ·,£(t-n+1 ) (13)
The p a r a m e t e r s of Eqs. (11) and (13) are again
updated at time tfd when y becomes available. Equa-
tion (13) thus enables the controlled output to be
estimated a t the faster secondary output sample
(5)
rate. T h e number of p a r a m e t e r s to be e s t i m a t e d in
Eq. (11) is 3n-2 (our experience indicates that n=1
or n=2 is normally sufficient). After updating the
parameters of Eq. (8), the values of £(t) in Eqs.
(11) a n d (13) a r e calculated from Eqs. (6), (8) and
(9) (Tsay a n d Shieh, 1981).
Often, v(t) c a n be obtained w i t h l o w noise levels
and load disturbances only occur infrequently.
Under these c i r c u m s t a n c e s , the t e r m s tend to be
negligibly small. There is therefore no need to
identify a secondary output model. The result is a
significant reduction in the n u m b e r of unknown
parameters in Eq. (11), and Eq. (13) simplifies to:
y ( U d ) =y&,u ( t-m-1 )+. .+/3 _ 1 u ( t-m-n+1 )+
n
X v(t)+..+Î _ v(t-n+1) (14)
0 n1
So far, the discussion h a s a s s u m e d the complete
observability of the plant from v(t). T h e question
now arises as to what happens if this assumption is
violated. It is evident from Eqs. (1) to (3), that
y(t) w i l l then share only some of its poles w i t h
v(t). The zeros of y(t) a n d v(t) are, in general,
different from each other. Loosely speaking, the
substitution of χ into Eq. (3) may be regarded as
assigning the poles of v(t) to y(t). The zeros
assigned to y(t) are then determined by the parame-
ter e s t i m a t i o n m e c h a n i s m w h i c h places the zeros
such that the effect of the difference between the
poles of v(t) a n d y(t) is m i n i m a l . It is also
possible that some of the poles of v(t) are cancel-
led. Simulations performed on systems with signifi-
cantly different controlled a n d secondary output
dynamics have indicated satisfactory estimation.
INPUT-OUTPUT BASED ESTIMATOR
This estimator assumes a system of the form:
v(t)=G u(t-m )+L w(t)+1^(t) (15)
1 l
y(t)=G u(t-d-m )+L2w(t-d)+^(t) (16)
2 2
m-| and m are the time delays i n t h e responses of
2
secondary a n d controlled o u t p u t s to changes in
u(t). G-j and G are proper transfer functions while
L. a n d L are r o w vectors of proper transfer
2
functions. A s s u m i n g that w is a vector of statio-
 Estimation and Control of Distillation Product Composition
nary random sequences, Eq. (15) may be written as:
v(t)=G u(t-m )+H o>(t)
1 1 1
where c*)(t) is a zero m e a n w h i t e sequence and
a stable proper transfer function. By defining:
£(t)â^(t)-^(t-d)
Eq. 0 6 ) m a y be w r i t t e n as:
y ( t ) = G u ( t-d - m )+H o)( t-d )+£( t )
2 2 2
where again H is a stable proper transfer function
2
and £(t) is zero m e a n a n d white. E l i m i n a t i o n of
m + m(17) and (19) results i n :
a)(t) b e t w e e n Eqs.
y ( t ) = [ G z " 2 1 - G H / H Mt-d-m., )
2 12 1
+[H /H ]v(t-d)+£(t)
2 1
If G p G , H-j and H are k n o w n , u(t) a n d v(t) c a n
2 2
be used in Eq. (20) to compute:
y(t+d)^y(t+d)-£(t+d)
which is the e s t i m a t e of the controlled output.
Equations (18) a n d (21) imply that deviations of
y(t+d) from its true value are l i m i t e d to the
magnitude of secondary output m e a s u r e m e n t noise,
1^j(t), which, by suitable choice of secondary var-
iable, is normally s m a l l e r than that associated
with y(t). Thus, the e s t i m a t e y(t) is, in theory,
closer to the controlled output a t t i m e t-d than
its measurement y(t). In practice, the accuracy of
y will depend on the accuracy of the plant model
used. Since y can be computed a t the rate a t w h i c h
ν is obtained, the p r o b l e m s of controlled output
measurement delays can be avoided. In many instan-
ces, due to the large number of disturbance inputs
and complex plant d y n a m i c s , ,G , a n d H can
2 2
only be approximately obtained using off-line sys-
tem identification. In this work, h o w e v e r , these
transfer functions a r e n o t a s s u m e d to be k n o w n .
Instead, Eq. (20) is recursively a p p r o x i m a t e d o n -
line. From the definitions of G-j, G , H-j and H ,
2 2
it seems reasonable to approximate Eq. (20) by:
Ay (t)=Bu(t-d-m)+Cv(t-d)+£(t)
A, B, and C are p o l y n o m i a l s in z~^ of order η and
m=Min(m-j ,m ). A is m o n i c a n d the leading c o e f -
2
ficient of Β is assumed to be non-zero. The order η
is chosen to provide the required accuracy of
approximation. The p a r a m e t e r s of Eq. (22) cannot be
estimated using available recursive techniques as
y(t) is n o t available a t each time step. This is,
however, overcome by writing Eq. (22) for t i m e s t
to t-n(d-1)+1 and eliminating intersample values of
y from the d resultingn recursions
d 1 yield:
to 2 (NOP), 260 (-10%) and 380 (NOP) m i n s . The open loop
( 1 + o z - . . ^ z "z )ny (d t ) =)
dm
( Su )
z - 4 .
( ++ s
tiz z
( .1i
- . +
< d + +n d /1 y2 +
n d
Vnd " - - o 1 " "
T +i nz -d )v(t-d)
For parameter estimation, Eq. (23) is written as:
y(t)=0 ^(t-d)+e(t) (24)
where 1^ [ ^ , . . , ^ , ^ , . . , ^ , ^ , . , , ί ^ ]
n d
f (t-d)=[-y(t-d),..,u(t-d-m-1),..,v(t-d),.. ] dered: using GC outputs as feedback and using stage
and e(t) is the equation error, θ is recursively
estimated a t every d time step. y(t)=0^^(t-d) is
computed, after estimating θ a t time t, to update
^ t - d ) . The controlled ouput is then e s t i m a t e d
during the time T interval t to t+d using:
y(tfd)=6 f(t) (25)
The p a r a m e t e r s of Eq. (23) a r e again updated a t
time t+d when a new y becomes available. Estimates
of the controlled output can now be computed at the
faster secondary output sampling rate using Eq.
(25). E q u a t i o n (23) has n(2d+1)+1 u n k n o w n p a r a -
205
meters and d may typically lie in the range 2 to 6.
(17) A first order e s t i m a t o r (n=1) c a n therefore have
is between 6 to 14 u n k n o w n p a r a m e t e r s a n d a second
order e s t i m a t o r , b e t w e e n 11 to 27. These figures
(18) may suggest that the tuning in period for larger
values of d will be inconveniently long. This is
(19) not, however, the case. Although the number of
unknown parameters increases with d, the data being
supplied to the e s t i m a t o r at each sample instance
of the controlled output also increases [Eq. (23)].
Simulation studies, investigating t h e effects of
(20) analyser delays up to 6 sample intervals, have
indeed c o n f i r m e d t h a t t h e t u n i n g i n p e r i o d
increases only slightly for larger values of d.
(21)
The controlled output can often be estimated with
sufficient accuracy using a first order estimator.
The p a r a m e t e r of the first order e s t i m a t o r can
d
be s h o w n to be related to a 1 as o^ ——3«j . It m a y be
d
noted from Eq. (20) that la-j | is normally less then
unity (unless H has a dominant unstable zero). For
2
larger values of d, it therefore f o l l o w s that
<?^«1, especially w h e n | a ^ | « 1 . A s a result, c*
d
becomes negligible whereupon Eq. (25) reduces to:
y ( t+ d ) = ^ u ( t-m-1 )+. .+/# u ( t-m-d )
d
+ if v(t)+..+ï v(t-d) (26)
0 d
It is noted that Eq.(26) h a s a structure w h i c h is
similar to Eq. (14). This agreement between the two
design approaches provides additional confidence in
the arguments presented in this section.
SIMULATION STUDIES AND DISCUSSION
The control performances of the state-space estima-
tor ( E s t D , Eq. (14), a n d the input-output e s t i m a -
tor (Est2), Eq. (26), w e r e evaluated on a d y n a m i c
(22) model of a 10 stage binary distillation column. The
feed of 50wt% methanol-water, with a flow rate of
18 g / s w h e n the c o l u m n is a t its normal operating
point (NOP), enters the c o l u m n a t stage 5. Bottom
product composition, analysed by a gas c h r o m a t o -
graph (GC) w i t h a cycle t i m e of 3 mins., is to be
maintained at 5wt% methanol by manipulating steam
flow rate to the reboiler (stage 1). The disturban-
ces to the c o l u m n are step changes in feed flow
rate, w h i c h are introduced a t t i m e s 20 (+10%), 140
responses of bottom product composition and stage 4
temperature to +_10% step changes in feed and steam
flow rates (Fig. 1) illustrates the non-linear
(23)
behaviour of the system to the respective inputs.
Firstly, two common control schemes will be consi-
1 t e m p e r a t u r e feedback. Figures 2 and 3 show the
responses of the bottom product c o m p o s i t i o n a n d
steam flow rate, respectively, when PI control was
applied with G C output feedback. Compare this with
the responses obtained w h e n stage 1 temperature
feedback w a s used (Figs. 4 and 5). T h e i m p r o v e m e n t
in p e r f o r m a n c e is clear. W h e n G C outputs are used
for feedback control, the rate of control is l i -
mited by the analysis t i m e , ie. 3 mins. Stage 1
temperature, on the other hand, can be sampled and
controlled a t a faster rate, in this case 0.5 min.
206 M. T. Guilandoust, A. J. Morris and M. T. Tham
Disturbances can be detected sooner and corrective
action taken. The result is a significant improve-
ment in control performance. As this is a binary
system, stage 1 temperature bears a direct relatio-
nship to bottom product composition, and hence can
be employed successfully as the feedback variable
for bottom product composition control. For a
multicomponent system, however, reboiler liquid
temperature is not an accurate indication of bottom
product composition. This s c h e m e may, therefore,
not m e e t w i t h s i m i l a r success, unless t h e system
can be closely a p p r o x i m a t e d as a psuedo binary
system. This can be demonstrated on the same binary
column by using stage 4 temperature as feedback to
the PI controller ( Figs. 6 and 7). Large offsets
were observed for feed deviations away from its
NOP. The importance of secondary variable selection
for inferential control is therefore highlighted.
Figures 8 and 9 show the responses when Est1, using
stage 4 temperature, w a s used to provide inferred
composition values to a PI controller. T h e full
line plot in Fig. 8 s h o w s the response of b o t t o m
product composition while the dashed line shows its
estimates. It can be seen that the e s t i m a t e s w e r e
fairly accurate. More importantly, no offsets were
observed a n d the deviations from set-point w e r e
noticeably smaller than those obtained using analy-
ser feedback (cf. Fig. 2). The use of Est2 resulted
in almost identical composition responses as shown
in Figs. 10 and 11. The estimates obtained by Est2,
however, exhibited larger transients (Fig. 10) than
those observed using Est1 (Fig. 8). Since Est2 h a s
a larger number of parameters to identify, a change
in process dynamcis (as a result of a disturbance
step - see Fig.1) requires a correspondingly longer
period for adaptation. Nevertheless, both e s t i m a -
tors could provide for better regulation than that
achieved by direct composition feedback control.
Both Est1 and Est2 c a n also be used in adaptive
control schemes. The responses of bottom product
composition and steam flow rate using a self-tuning
controller (STC) w i t h composition feedback, to
regulate the loop against the same disturbances are
shown by the dashed line plots in Figs. 12 and 13
respectively. A l t h o u g h t h e S T C s c h e m e l e d t o
an improved performance over that achieved by PI
control w i t h composition feedback (Fig. 2 ) , the
results of adaptive inferential control were better
(Figs. 8 and 10). However, w h e n the STC w a s used in
conjunction with Est1, ie. an adaptive inferential
adaptive control scheme, a vast i m p r o v e m e n t in
performance w a s observed (Figs. 12 and 13 - full
line plots). Indeed, the performance of this adap-
tive inferential adaptive control scheme approached
that achieved by the PI controller using stage 1
temperature feedback (Fig. 4). This is r e m a r k a b l e
considering that the estimator w a s using the tem-
perature of the stage immediately below the feed
tray as the secondary output v a r i a b l e for infer-
ring bottom product composition. In practice, such
a choice of secondary output w o u l d be considered
naive as the d y n a m i c s of stage 4 temperature are
very different from those of bottom product compo-
sition (Fig. 1). The above examples, however, demo-
nstrate the robustness of the proposed estimators
to secondary variable selection and their applica-
bility to multicomponent systems.
CONCLUSIONS
Two new approaches to inferential control have been
outlined. Application of the adaptive estimators to
regulate the b o t t o m product composition of a non-
linear binary distillation column model has demon-
strated their usefulness in providing for improved
control of processes with long measurement delays.
In particular, it has been s h o w n that the e s t i m a -
tors are insensitive to secondary variable selec-
tion; can cope with non-linear dynamics and can be
used in conjunction w i t h either non-adaptive or
adaptive feedback controllers.
ACKNOWLEDGEMENTS
The authors gratefully acknowledge the UK SERC for
supporting this work.
REFERENCES
Anderson, B.D.O. and J.B. Moore (1979). Optimal
Filtering. Prentice-Hall.
Astrom, K.J. (1970). Introduction to Stochastic
Control Theory. Academic Press.
Joseph, B. and C.B. Brosilow (1978). Inferential
control o f p r o c e s s e s P a r t s I & I I I . A I C h E
Journal, 24. No.3, 485-491 & 500-509.
Ljung, L. a n d T. Soderstrom (198 3). Theory and
Practice of Recursive Identification. MIT Press.
Morari, M. a n d G. Stephanopoulos (1980a). Optimal
selection of secondary measurements within the
framework of state estimation in the presence of
unknown disturbances. AIChE Journal, 26, No.2,
247-260.
Morari, M. and G. Stephanopoulos (1980b). M i n i m i -
zing unobservability in inferential control
schemes. Int. J. Control, 31, No.2, 367-377.
Patke, N.G., P.B. Deshpande a n d A.C. Chou (1982).
Evaluation of inferential and parallel cascade
schemes for distillation control. Ind. Eng.
Chem. Process Pes. Dev., 21, 266-272.
Shieh, L.S., C T . W a n g and Y.T. Tsay (1983). Fast
suboptimal state-space self-tuner for linear
stochastic m u l t i v a r i a b l e systems. Proc. IEE,
Pt.D, J30, 143-154.
Tsay, Y.T. a n d L.S. S h i e h ( 1 9 8 1 ) . S t a t e - s p a c e
approach for self-tuning feedback control w i t h
pole-assignment. Proc. IEE, 128, Pt.D, No.3, 9 3 -
101.
Tuffs, P.S. and D.W. Clarke (198 5). Self-tuning
control of offsets: a unified approach. Proc.
IEE. 132. Pt.D, No.3, 100-110.
Wittenmark, B. a n d K,J. A s t r o m (1984). Practical
issues in t h e i m p l e m e n t a t i o n of self-tuning
control. A u t o m a t i c a , 20, No.5, 595-605.
Wright, J.D., J.F. M a c G r e g o r , A. J u t a n , J.P.
Tremblay and A. Wong (1977). Inferential control
of an e x o t h e r m i c packed b e d tubular reactor.
Proc. JACC, 1516-1522.
 1
500
500
(MIN)
(MIN)

DEVIATIONS O F B O T T O M P R O D U C T COMPOSITION (FEED F L O W CHANGES).

Fig. 2. PI C O N T R O L O F B O T T O M P R O D U C T COMPOSITION.

3
o'

1 η
500 ο
(MIN) 500
(MIN) 3

DEVIATIONS O F S T A G E 4 T E M P E R A T U R E (FEED F L O W CHANGES).

Fig. 3. S T E A M F L O W R A T E .CONTROL VARIATIONS U N D E R PI COMPOSITION C O N T R O L .

Ο
CL
500 C
(MIN) i 500 η
DEVIATIONS O F B U T T O M P R O D U C T COMPOSITION (STEAM F L O W CHANGES). (MIN) ο
Β
Ο
Fig. 4. PI C O N T R O L USING S T A G E 1 T E M P E R A T U R E .

500
(MIN) t Ί ' '—Κ 500
(MIN)
DEVIATIONS O F S T A G E 4 T E M P E R A T U R E (STEAM F L O W CHANGES).

Fig. 1. O P E N L O O P RESPONSES (- 10% C H A N G E S F R O M S T E A D Y STATE). ί '

Fig. 5. S T E A M F L O W R A T E VARIATIONS U N D E R PI T E M P E R A T U R E C O N T R O L .

Ο
-α
 ÏN0
Ο
00

500
(MIN) 500
(MIN)

Fig. 6. PI C O N T R O L USING S T A G E 4 T E M P E R A T U R E .
Fig. 10. DEVIATION RESPONSES IN B O T T O M P R O D U C T COMPOSITION A N D ITS ESTIMATE (Est. 2).
(ADAPTIVE INFERENTIAL C O N T R O L )

Η
Ο
500
c
(MIN) 500
(MIN) 3
σ-
ο
Fig. 7 . S T E A M F L O W VARIATIONS U N D E R PI T E M P E R A T U R E C O N T R O L .
Fig. 11. S T E A M F L O W R A T E C U N T R U L VARIATIONS
UNDER ADAPTIVE INFERENTIAL C O N T R O L .
Γ
3
α.
500
(MIN)

500 Η
(MIN)
Η
Fig. 8. DEVIATION RESPONSES IN I3UTTUM P R O D U C T COMPOSITION A N D ITS ESTIMATE (Est. 1). Fig. 12. COMPARISON O F ADAPTIVE INFERENTIAL ADAPTIVE C O N T R O L (STAGE 4 T E M P E R A T U R E )
(ADAPTIVE INFERENTIAL C O N T R O L ) A N D SELF TUNING C O N T R O L O F B O T T O M COMPOSITION.

15.8-j

In 1
bp I
500
(MIN)
3 13.8- Κ-
500
IT (MIN)

STEAM
F i9.g . F L O W R A T E C O N T R O L VARIATIONS
UNDER ADAPTIVE INFERENTIAL C O N T R O L .
FS - i· 1S T E3A M F L O W R A T E VARIATIONS CORRESPONDING T O FIGURE 12.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

SIMULTANEOUS ADAPTIVE REGULATION OF

TERMINAL COMPOSITIONS IN BINARY
DISTILLATION
M . T. Tham*, F. Vagi**, A. J. Morris** and R. K. Wood**
^Department of Chemical and Process Engineering, University of
Newcastle upon Tyne, UK
^Department of Chemical Engineering, University of Alberta, Canada
1
Abstract. A ' p o s i t i o n a l m u l t i v a r i a b l e1 s e l f - t u n i n g c o n t r o l l e r i s d e r i v e d
and extended to admit an 'incremental control policy. Both multivariab-
le s e l f - t u n i n g c o n t r o l l e r s a r e capable of m u l t i v a r i a b l e - m u l t i r a t e s e l f -
tuning c o n t r o l : different loops making up the multivariable control
s y s t e m c a n be a s s i g n e d d i f f e r e n t control a c t u a t i o n r a t e s . The 'positio-
nal' a n d 'incremental' a l g o r i t h m s w e r e i m p l e m e n t e d in the d u a l - c o m p o s i -
tion control of a pilot scale distillation column. The results show that
both a l g o r i t h m s provided good s i m u l t a n e o u s r e g u l a t i o n of terminal product
compositions. T h e ability of m u l t i r a t e s e l f - t u n i n g c o n t r o l to a c h i e v e
improved overall p e r f o r m a n c e s are also demonstrated.

Keywords. Adaptive control; Digital control; D i s t i l l a t i o n c o l u m n c o n -

trol; M u l t i v a r i a b l e s y s t e m s ; M u l t i v a r i a b l e - m u l t i r a t e c o n t r o l ; P r o c e s s
control.

INTRODUCTION position control of a p i l o t scale binary

Distillation is t h e m o s t c o m m o n o p e r a t i o n
used i n t h e p r o c e s s a n d p e t r o c h e m i c a l
i n d u s t r y to a c h i e v e p r o d u c t s e p a r a t i o n a n d
purification. Unfortunately, the distilla-
tion process is very energy i n t e n s i v e w i t h
the e n e r g y r e q u i r e m e n t s c l o s e l y l i n k e d t o
the desired s e p a r a t i o n . Energy usage can,
h o w e v e r , b e m i n i m i s e d if s a t i s f a c t o r y s i m -
u l t a n e o u s control of t e r m i n a l c o m p o s i t i o n s
can be a t t a i n e d d e s p i t e d i s t u r b a n c e s to the
column. However, simultaneous terminal
c o m p o s i t i o n control is n o t easy d u e to n o n -
linear dynamics a n d the interaction that is
inherent in a m u l t i v a r i a b l e system. Hence,
the p e r f o r m a n c e s a c h i e v e d using single loop
conventional P I / P I D c o n t r o l l e r s w i l l , in
general, n o t be s a t i s f a c t o r y . Much effort
has t h e r e f o r e been d e v o t e d to the d e v e l o p -
ment and evaluation of new control schemes
and s t r a t e g i e s as d o c u m e n t e d in a r e c e n t
review (Seborg a n d c o w o r k e r s , 1 9 8 3 ) . D e s p i t e
the n u m e r o u s a p p l i c a t i o n s o f a d a p t i v e c o n -
trol s t r a t e g i e s ( A s t r o m , 1 9 8 3 ) , n o i n d u s -
trial d i s t i l l a t i o n c o l u m n c o n t r o l s t u d i e s
are m e n t i o n e d . F u r t h e r m o r e , in a r e c e n t
paper addressing s e l f - t u n i n g i m p l e m e n t a -
tion i s s u e s ( W i t t e n m a r k a n d A s t r o m , 1 9 8 4 ) ,
only s i n g l e - i n p u t s i n g l e - o u t p u t systems
were c o n s i d e r e d . M o r r i s a n d c o w o r k e r s
(1981,1982) have n e v e r t h e l e s s s h o w n that
very a c c e p t a b l e o p e r a t i o n c a n b e o b t a i n e d
using b o t h single a n d m u l t i v a r i a b l e s e l f -
t u n i n g c o n t r o l l e r s (STCs) f o r p r o d u c t c o m -
column. These experimental and simulation
studies u s e d d e t u n e d m i n i m u m variance
'positional' c o n t r o l s t r a t e g i e s b a s e d on
extensions of the S T C of Clarke a n d
Gawthrop( 1 979). E v a l u a t i o n of a m u l t i r a t e
derivation of the m u l t i v a r i a b l e a l g o r i t h m
w a s , h o w e v e r , o n l y p e r f o r m e d by s i m u l a t i o n .
Here, t h e d i f f e r e n t l o o p s m a k i n g up t h e
multivariable s y s t e m a r e c o n t r o l l e d a t
rates chosen to satisfy the d y n a m i c requi-
rements of each control loop. Chien and
c o w o r k e r s ( 1 983) a l s o p r e s e n t e d s i m u l a t i o n
results, using a m u l t i r a t e S T C b a s e d on
modifications to t h e m u l t i v a r i a b l e S T C
g i v e n by K o i v o ( 1 9 8 0 ) , This paper pre-
sents results of the application of m u l t i -
variable a n d m u l t i r a t e STCs to the s i m -
ultaneous r e g u l a t i o n of the top a n d b o t t o m
compositions of a pilot plant binary dis-
tillation column. The algorithms used were
the m u l t i v a r i a b l e a n d m u l t i r a t e ' p o s i t i o -
nal' S T C s o r i g i n a l l y s u g g e s t e d by M o r r i s
a n d c o w o r k e r s ( 1 98 1 ,1 9 8 2 ) , a s w e l l a s t h e
corresponding 'incremental' STCs based upon
a recently p r o p o s e d i n c r e m e n t a l p r e d i c t o r
(Hodgson, 1 982).
SYSTEM REPRESENTATION AND SELF-
TUNING CONTROL
Although virtually all c h e m i c a l p r o c e s s e s
are n a t u r a l l y m u l t i v a r i a b l e s y s t e m s w i t h
many i n t e r a c t i n g p a t h s , in p r a c t i c e , the
control of m o s t large dimensioned multiva-

209
210 M. T. Tham et al.

riable s y s t e m s c a n be r e d u c e d to t h e c o n - i n t o 2 ( M I SkO ) s u b s y s t e m s_: k

trol o f t h e f e w m o s t i n t e r a c t i v e loops. A y ( t ) = z - di i B u ( t ) + z i j B j U j ( t )
i i i i i
This o f t e n r e s u l t s in a t w o - i n p u t t w o - +z- iD v (t)+C 5 (t), i i j (4)
output (2x2) s y s t e m w h i c h c a n be d e s c r i b e d
ii i i
The m i n i m i s a t i o n o f Eq.(2) i s n o t r e a l i s -
by t h e f o l l o w i n g d i s c r e t e e x p r e s s i o n : able a s i t r e q u i r e s f u t u r e o u t p u t v a l u e s .
Ay(t)=Bu(t)+Dv(t)+C«t) (1 ) This p r o b l e m i s o v e r c o m e by r e p l a c i n g t h e
T h e p o l y n o m i a l m a t r i c e s A, B, C a n d D a r e unknown variables w i t h their k ^ - s t e p - a h e a d
defined as follows: predictions. Using the identity:
A = d i a g ( A ,d d
1 A 2) ; C = d i a g C C j , C 2) ; C=EA+F (5)
D = d i ak g ( z ~ 1 D , z_" * k2 D ) w h e r e Ε = d i a g ( E ,kE ) , d e gk( E ) = k ^ - 1
B r z " k1 1 B 1z k1 2 B 2 ] and
1 2 i
F = d i akg ( z " 1 1 F , z " 2 2 F ) , ie.:
n l 2 1 2
= Lz~ 21B 21 z " 2 2 B 2 J2 C =E A +z'" iiF (6)
i ii i
A^, B j j , and a r e p o l y n o m i a l s i n z'\ The k ^ - s t e p - a h e a d p r e d i c t i o n of y^ c a n be
A^ a n d are assumed monic, with a further o b t a i n e d by m a n i p u l a t i n g Eqs.(4) a n d (6) to
assumption that t h e roots of C a r e s t r i c - yield :
i
tly s t a b l e . The subscripts i and j denote, yJ(t+k /t)=F y (t)+G u (t)
respectively, the loop number and the
i i ii i i
i n t e r a c t i o n f r o m t h e j - t h l o o p , y, u a n d ν
+G i u ,j ( t +k i - ki i ) +jL iv i( t +k i - di i)
+H y*(t+k -1/t-1) i i j (7)
i i i
are v e c t o r s o f s y s t e m o u t p u t s , control wGi t h E yBJ ( t + kd / at ) =n y ( t + k ) - E ^ ( t + k ) ;
inputs a n d m e a s u r a b l e d i s t u r b a n c e i n p u t s
l i i i i i i i i
ij" i ij H = ( 1 - C ) . S u b s t i t u t i o n of
i i
w h i l e S i s a vector of uncorrelated random Eq.(7) into Eq.(3)1 a n d minimising results
sequences w i t h zero m e a n s a n d f i n i t e v a r - in t h e p o s i t i o n a l control signal for each
iances. T h e t i m e d e l a y s , k^j a n d d ^ , a r e loop a s :
expressed as integer multiples of the sam- u (t)=[w (t)-F y (t)-G u (t k -k )
pling i n t e r v a l of t h e i - t h l o o p . E q u a t i o n
1 1 ii i j + 1 1i j
-H y (t+k -1/t-1)]/CQ +G ] (8)
(1) i s n o t t h e m o s t g e n e r a l r e p r e s e n t a t i o n
ii i i i i i
where Q = q {(0)Q|/g (0). This control stra-
of a (2x2) s y s t e m . Another input-output
i ; ii
tegy w i l l be r e f e r r e d to a s the P M V S T C .
model u s e d in the d e s i g n of m u l t i v a r i a b l e k
STCs h a v e A a n d C a s f u l l r a t h e r than Operating o n E q . ( 7 ) by = 1( 1 - z ~ i i )
diagonal m a t r i c e s , eg. K o i v o ( 1 9 8 0 ) . How- results in t h e ^ - i n c r e m e n t a l predictor
ever, in s e l f - t u n i n g c o n t r o l , t h e a d v a n - ( H o d g s o n , 1 9 8 2 ) : (i i j)
tages to be g a i n e d f r o m u s i n g t h e m o r e yJ(t+k /t)=yJ(t/t-k )F Aiy (t) (9)
general r e p r e s e n t a t i o n a r e only m a r g i n a l . l i 1 1i i
+G A u (t)+G jû Uj(t+k -k j)
Further, the s i m p l i f i c a t i o n a r i s i n g f r o m i i i i i i ii
+L A v (t+k -d )+H A y (t+k -1/t-1)
the u s e o f d i a g o n a l A a n d C p o l y n o m i a l iii i ii iii i i 1
Substitution o f Eq.(9) into Eq.(3) and
matrices provide certain a d v a n t a g e s , f o r minimising yields the ' k - i n c r e m e n t a l con-
example : trol signal for t h e i - t h loop a s :
a) T h e p r o c e s s m o d e l f o l l o w s d i r e c t l y f r o m u ( tG) = [ w (ut (
) - yt* ( +
tk/ tk- )
k )-F A y (t)
i i i i iii
the P - c a n o n i c a l form and allows the
" ii^i i ii~ ii
simple n o n - p a r a m e t r i c i d e n t i f i c a t i o n of
- . H i y J ( t + k - 1 / t - 1 ) ] / [ Q + A G ] (10)
process dynamics for initial controller il i i l ii i
The i n c r e m e n t a l c o n t r o l s t r a t e g y w i l l be
parameterisation.
referred to as t h e I M V S T C .
b) T h e s y s t e m s t r u c t u r e c a n b e e a s i l y d e -
composed into m u l t i - i n p u t single-output It i s a p p a r e n t t h a t t h e s y n t h e s e s o f t h e
(MISO) s u b s y s t e m s , t h e r e b y p r o v i d i n g a PMVSTC a n d I M V S T C h a v e b e e n s i m p l i f i e d to
direct link to e s t a b l i s h e d s i n g l e v a r i a t e the s y n t h e s e s o f i n t e r c o n n e c t e d M I S O S T C s .
self-tuning theory and practice. k
To m a i n t aki n c adu s a
a l i td
yn, kit i sorr e q uei rte d e
h
t h a tc a s
c) L e s s p a r a m e t e r s a r e r e q u i r e d t o a d e q u a - ki j = ^k at is ^ ei i * ^
ii»
tely describe t h e s y s t e m .
ii ij» °^ s i m u l t a n e o u s e q u a t i o n s
must be s o l v e d s i n c e b o t h i n p u t s w i l l be
Multivariable Self-Tuning Control u n k n o w n a t c u r r e n t t i m e t. A l t h o u g h f e e d -
forward load com pensa stion h a s been i n -
The S T C s c o n s i d e r e d h e r e , a r e r e q u i r e d to cluded, t h e r e m a y be i n s t a n c e s w h e r e t h e
m i n i m i s e , w i t h r e s p e c t to u(t), t h e d e t u n e d load d i s t u r b a n c e , v^, c a n n o t be m e a s u r e d .
m i n i m u m v a r i a n c e c o s t f2u n c t i o n : 2
Under s u c h c i r c u m s t a n c e s , t h e p r e d i c t i o n
J = E 2{ I| y ( t + k ) - w ( t ) I| + | | Q ' u ( t ) I| l t} (2) error of t h e i - t h l o o p c a n be used as an
||x|| denotes t h e norm of vector χ w h i l e estimate for v
E{.j l is the e x p e c t a t i o n o p e r a t o r c o n d i -
i (Morris and coworkers,
t 1982). In the control of u n k n o w n processes,
tional upon i n f o r m a t i o n a v a i l a b l e up to a n d a r e c u r s i v e p a r a m e t e r is u s e d to obtain
i n c l u d i n g c u1r r e n t t i m e t . The control estimates of the coefficients of t h e
weighting Q is d e f i n e d as a diagonal c o n t r o l l e r p o l y n o m i a l s F ^ , G ^ , G ^ j , Lj. a n d
t r a n s f e r m a t r i x . E q u a t i o n (2) c a n b e s e p a - H^. T h e r e g r e s s i o n e q u a t i o n s f o r p a r a m e t e r
rated as J = [J-|,J ]^, 2 where: 2 estimation a r e o b t a i n e d f r o m Eqs.(7) a n d
2
J =E{[y (t+k )-w (t)] +[Q' u (t)] | } (3) (9) f o r t h e p o s i t i o n a l and incremental
i i i i i ii t
E q u a t i o n (1) c a n s i m i l a r l y be s e p a r a t e d predictors r e s p e c t i v e l y . W i t h the latter,
 Terminal Compositions in Binary Distillation
p a s t p r e d i c t i o n s o f o u t p u t a r e r e p l a c e d by
y (t/t-k )=y (t)+ e (t), w h e r e e.^ i s a n
i i i i i
a s s o c i a t e d z e r o - m e a n error (Hodgson,1 982).
It s h o u l d a l s o b e n o t e d t h a t t h e a l g o r i t h m s
described a l l o w for i n d e p e n d e n t tuning of
the c o n t r o l w e i g h t i n g , Q ^ , f o r e a c h l o o p .
T h i s w e i g h t i n g c a n be e x p r e s s e d i n t h e f o r m
of a n i n v e r s e P I / P I D ( M o r r i s a n d c o w o r k e r s ,
1 981,1 98 2) u s i n g s e t t i n g s s p e c i f i e d i n
terms of p r o p o r t i o n a l b a n d , i n t e g r a l t i m e
and d e r i v a t i v e t i m e . c a n t h e r e f o r e be
adjusted using familiar tuning rules.
Multirate Self-tuning Control
The a b i l i t y to d e c o m p o s e a m u l t i v a r i a b l e
s y s t e m to M I S O s u b s y s t e m s a l s o i m p l i e s t h a t
d i f f e r e n t s a m p l i n g / c o n t r o l i n t e r v a l s can be
c h o s e n to suit t h e d y n a m i c c h a r a c t e r i s t i c s
of d i f f e r e n t l o o p s . T h e p e r f o r m a n c e u s i n g a
multirate control strategy should therefore
be b e t t e r t h a n t h a t w h i c h m a k e s u s e o f a
single s a m p l i n g rate for all l o o p s . T h e
only r e s t r i c t i o n o n t h e m u l t i r a t e f o r m of
both the PMVSTC a n d I M V S T C is that s a m p l i n g
intervals of e a c h loop a r e integral m u l t i p -
les o f t h e f a s t e s t s a m p l i n g t i m e . T h e
structures of t h e m u l t i r a t e algorithms
remain i d e n t i c a l to t h e i r multivariable
counterparts.
EXPERIMENTAL IMPLEMENTATION AND RESULTS
The highly non-linear m e t h a n o l - w a t e r dis-
tillation c o l u m n p i l o t p l a n t used in this
s t u d y w a s t h e s a m e 22.5 c m d i a m e t e r b u b b l e
cap c o l u m n w i t h 8 t r a y s a t 30.5 c m s p a c i n g
used in a n u m b e r of other s t u d i e s (Sastry
and c o w o r k e r s , 1 9 7 3 ) . I t is c o n t r o l l e d by
an L S I 11/03 m i c r o c o m p u t e r w i t h a f o r e g r o u -
nd/background operating system, communica-
ting w i t h an HP1000 c o m p u t e r used for chro-
matograph control and c h r o m a t o g r a m analysis
(Vagi,1985). T o p p r o d u c t c o m p o s i t i o n (XD)
is m e a s u r e d u s i n g a n i n - l i n e c a p a c i t a n c e
cell w h i l e the b o t t o m p r o d u c t c o m p o s i t i o n
is a n a l y s e d , o n a 3 m i n . c y c l e , b y a n H P -
5720A g a s c h r o m a t o g r a p h . Normal steady
s t a t e o p e r a t i n g c o n d i t i o n s a r e : F e e d (FE) -
18 g / s a t 5 0 m a s s % , T o p p r o d u c t - 8.6 g / s
a t 9 5 m a s s % , B o t t o m p r o d u c t - 9.4 g / s a t 5
m a s s ? , R e f l u x ( R E ) - 8.4 g / s a n d S t e a m ( S T )
- 11 g / s . F o r t h i s c o l u m n , s t e p c h a n g e s i n
feed f l o w rate a r e r e g a r d e d as b e i n g the
most s e v e r e d i s t u r b a n c e s to a f f e c t i t s
operation. T h e e x p e r i m e n t s thus c e n t r e s on
the r e g u l a t o r y p r o p e r t i e s o f t h e S T C s
a g a i n s t f e e d r a t e c h a n g e s o f ±25% a b o u t i t s
normal o p e r a t i n g p o i n t . F o r a l l r u n s , t h e
feed c h a n g e w a s i n t r o d u c e d a t t h e 3 0 t h
minute a n d b o t h c a s e s o f m e a s u r e d a n d
unmeasured disturbances were considered.
When the singlerate form of the a l g o r i t h m s
211
were i m p l e m e n t e d , both t h e XD a n d XB loops
were s a m p l e d a t 3 m i n u t e i n t e r v a l s . The
sampling i n t e r v a l for t h e XD loop w a s r e d -
uced to 1 m i n u t e w h e n m u l t i r a t e self-tuning
control w a s a p p l i e d . An UD f a c t o r e d
recursive least squares e s t i m a t o r (Bier-
m a n , 1976) w a s u s e d i n e a c h l o o p t o i d e n t i f y
the c o e f f i c i e n t s o f t h e p o l y n o m i a l s F
if
G , G j , Hj^ a n d L . T h e o r d e r s o f t h e
i J L± ±±
polynomials w e r e chosen based upon an unde-
rstanding o f t h e n a t u r e of p l a n t d y n a m i: cs
a n d w e r e s p e c i f i e d a s ; F-^:2, 6 - ^ : 3 , G ^ 2 ^ »
Ι :2 for the XD loop a n d F : 2 , G : 5 , G : 4
λ 2 2 1 2 2
and L : 2 f o r the X B loop. T h e c o e f f i c i e n t s
2
of were not estimated. When evaluating
the p e r f o r m a n c e s o f t h e S T C s a g a i n s t u n m e a -
surable d i s t u r b a n c e s , no L ^ p a r a m e t e r s
w e r e s p e c i f i e d . T h e c o n t r o l w e i g h t i n g s , Q^,
were c h o s e n to h a v e t h e f o r m of i n v e r s e PI
compensators. The estimator covariance
matrix o f e a c h l o o p w a s i n i t i a l i s e d a t
10001 w i t h the data forgetting factor fixed
at 0.998. All the runs were started with
the c o n t r o l l e r p a r a m e t e r s s e t t o z e r o
except for the l e a d i n g c o e f f i c i e n t s of G ^
w h i c h w e r e s e t t o 0.5. T h i s v a l u e i s o f t h e
same order as that e s t i m a t e d from open loop
tests. The ST1 C s w e r e i n i t i a l l y t u n e d in
the b a c k g r o u n d while the column w a s under
conventional PI c o n t r o l . S e l f - t u n i n g c o n -
trol w a s s w i t c h e d on only w h e n good output
predictions were observed.
The cases w h e n feed d i s t u r b a n c e s w e r e r e -
g a r d e d as being u n m e a s u r a b l e w i l l first be
considered. Figure 1 shows the responses of
XD a n d X B a n d t h e i r r e s p e c t i v e m a n i p u l a t i v e
i n p u t s , R E a n d S T , t o a +25% s t e p i n c r e a s e
in F E u n d e r s i n g l e r a t e P M V S T C control.
Although the transient responses were good,
there w a s a n o f f s e t in t h e r e s p o n s e of X D .
System r e s p o n s e s to a - 2 5 % step d e c r e a s e
in F E , u s i n g the same algorithm, are shown
in F i g . 2 . I n t h i s c a s e , n o o f f s e t s w e r e
observed in e i t h e r loop. It is also i n t e r -
esting to note that a c c e p t a b l e control w a s
still achieved even though the reflux flow
' b o t t o m e d out.' N e x t , t h e s i n g l e r a t e I M V S T C
w a s applied to regulate the system. Figure
3a s h o w s t h e c o m p o s i t i o n r e s p o n s e s t o a
+ 25% c h a n g e i n F E w h i l e F i g . 3b s h o w s t h e
responses to a d i s t u r b a n c e of - 2 5 % F E . As
expected, n o o f f s e t s w e r e o b s e r v e d in
either loop. Whereas positional self-tuning
algorithms d e p e n d on p a r a m e t e r a d a p t a t i o n
for o f f s e t r e j e c t i o n , t h e i n c r e m e n t a l a p -
proach g u a r a n t e e s offset free responses.
In g e n e r a l , d i s t u r b a n c e r e j e c t i o n w a s
f a s t e r t h a n t h a t a c h i e v e d by t h e s i n g l e r a t e
PMVSTC as is q u i t e a p p a r e n t f r o m t h e r e s p o -
nses o f X B . T h i s h a s , h o w e v e r , been
achieved a t the expense of more vigorous
transient behaviour. A p p l i c a t i o n of the
multirate P M V S T C to regulate the c o m p o s i -
tion loops r e s u l t e d in the responses s h o w n
212 M. T. Tham et al
in F i g . 4 a a n d F i g . 4 b . C o m p a r i s o n w i t h t h e
previous r e s u l t s w i l l r e v e a l that the
multirate PMVSTC provided the best control-
led X D r e s p o n s e s , t h u s f a r , a g a i n s t t h e
+ 25% change in F E . There w a s , h o w e v e r , a
prolonged p e r i o d of o f f s e t in the r e s p o n s e
of X B . R e g u l a t i o n o f t h e c o m p o s i t i o n s a -
g a i n s t t h e -25% c h a n g e i n F E r e s u l t e d i n
very o s c i l l a t o r y r e s p o n s e s i n b o t h l o o p s
(Fig. 4 b ) . F o r t h i s p a r t i c u l a r f e e d c h a n g e ,
XD e x h i b i t s a n i n v e r s e r e s p o n s e b e h a v i o u r
(Tham,1 9 8 51) . T h e c o n t r o l l e r t h e r e f o r e h a s
to ' l e a r n about the new conditions. A l -
though control of b o t h l o o p s w a s m a i n -
tained, transient behaviour w a s poor. Since
feedforward t e r m s have n o t been s p e c i f i e d ,
the p o s i t i o n a l a l g o r i t h m c o u l d n o t d i s t i n -
guish b e t w e e n d i s t u r b a n c e a n d i n t e r a c t i o n
effects. T h e p a r a m e t e r s of t h e c o n t r o l l e r
thus b e c a m e b i a s e d r e s u l t i n g i n t h e l o s s o f
dynamic decoupling capability. This problem
was n o t a s s e v e r e w h e n t h e s i n g l e r a t e
PMVSTC w a s used because the higher freque-
ncy d y n a m i c s h a v e b e e n i g n o r e d d u e t o t h e
use o f a l o n g e r s a m p l i n g r a t e f o r t h e X D
loop. T h e r e s p o n s e s o b t a i n e d u n d e r m u l t i -
rate I M V S T C r e g u l a t i o n a r e s h o w n i n F i g s .
5a a n d 5 b . F o r t h e +25% F E c h a n g e , n o
offsets w e r e o b s e r v e d in t h e r e s p o n s e of
XB. A s i n t h e p r e v i o u s r u n s , h o w e v e r , u s e
of t h e i n c r e m e n t a l s t r a t e g y r e s u l t e d i n
larger initial d e v i a t i o n s from s e t - p o i n t .
This is c l e a r l y e v i d e n t f r o m t h e r e s p o n s e
of X D . W h e n t h e - 2 5 % F E c h a n g e w a s i n t r o d -
uced (Fig. 5 b ) , t h e r e s p o n s e o f t h e XD l o o p
deteriorated. In this r u n , h o w e v e r , r e g u l a -
tion of X B w a s c l e a r l y b e t t e r t h a n t h a t
a c h i e v e d by t h e m u l t i r a t e P M V S T C . S i n c e t h e
incremental a l g o r i t h m u s e s differenced
data, the p r o b l e m of p a r a m e t e r b i a s h a s
been avoided. P a r a m e t e r r e a d j u s t m e n t in the
XD l o o p to a c c o u n t for the inverse
response, therefore, h a s very little effect
on t h e X B l o o p . T h i s w o u l d s e e m t o s u g g e s t
that t h e u s e of an i n c r e m e n t a l control
strategy imparts better decoupling proper-
t i e s to m u l t i v a r i a b l e S T C s a s h a s r e c e n t l y
b e e n c o n f i r m e d by P e e l a n d c o w o r k e r s ( 1 9 8 5 ) .
The c a s e s w h e n t h e f e e d d i s t u r b a n c e s w e r e
measured a n d used to p r o v i d e f e e d f o r w a r d
c o m p e n s a t i o n w i l l n o w be d i s c u s s e d . F i g u r e s
6a a n d 6 b s h o w s t h e c o m p o s i t i o n r e s p o n s e s
to a +25% a n d a - 2 5 % c h a n g e i n F E , r e s p e c -
tively, u s i n g t h e s i n g l e r a t e P M V S T C . T h e
improvement in p e r f o r m a n c e s o v e r those
shown in F i g s . 1 a n d 2 is o b v i o u s . D i s -
turbance rejection w a s achieved with better
transient b e h a v i o u r a n d a t a m u c h faster
rate. T h e i n i t i a l o f f s e t in t h e XB l o o p to
a + 2 5 % F E d i s t u r b a n c e c a n b e o b s e r v e d to be
slowly r e m o v e d as t h e f e e d f o r w a r d p a r a m e -
ters t u n e d - i n . A m a r k e d i m p r o v e m e n t in
overall p e r f o r m a n c e w a s o b s e r v e d w h e n t h e
multirate PMVSTC with feedforward compensa-
tion w a s i m p l e m e n t e d . Very t i g h t c o m p o s i -
tion control w a s achieved against both the
+ 2 5 % a n d t h e - 2 5 % F E d i s t u r b a n c e s (Figs. 7 a
and 7 b r e s p e c t i v e l y ) . T h e i n c l u s i o n o f
feedforward t e r m s e n a b l e d changes in load
dynamics to be s p e c i f i c a l l y m o d e l l e d a n d
hence p a r a m e t e r bias d i d n o t occur. The
performances of the m u l t i r a t e IMVSTC w i t h
feedforward c o m p e n s a t i o n against the feed
disturbances (Figs. 8 a a n d 8b) were better
than those achieved w h e n feedforward compe-
nsation w a s n o t u s e d (cf. F i g s . 5a a n d 5 b ) .
W h i l e X D r e s p o n s e s w e r e c o m p a r a b l e to t h o s e
obtained under m u l t i r a t e PMVSTC control, XB
responses to t h e - 2 5 % FE step w e r e poorer.
Larger s e t - p o i n t d e v i a t i o n s were again
observed when the incremental strategy w a s
implemented.
CONCLUSIONS
This paper has shown that the multivariable
STCs considered here a r e c a p a b l e of
providing g o o d s i m u l t a n e o u s r e g u l a t i o n of
the t o p a n d b o t t o m c o m p o s i t i o n s o f t h e
p i l o t s c a l e b i n a r y d i s t i l l a t i o n c o l u m n . It
is a c k n o w l e d g e d t h a t t h e c o m p a r i s o n o f
control algorithms through on-line applica-
t i o n s is d i f f i c u l t d u e to c h a n g e s in c o n d i -
t i o n s b e y o n d t h e c o n t r o l of t h e e x p e r i m e n -
ter. T h e f o l l o w i n g r e m a r k s a r e , h o w e v e r ,
based upon t r e n d s o b s e r v e d d u r i n g t h e i m -
p l e m e n t a t i o n of the m u l t i v a r i a b l e STCs over
almost a year of continuous study.
E x p e r i m e n t s h a v e s h o w n t h a t if d i s t u r b a n c e s
c a n b e m e a s u r e d (or e s t i m a t e d ) a n d t h e i r
effects i n c l u d e d in the control l a w , the
performance of t h e r e s u l t i n g S T C is s i g n i -
ficantly i m p r o v e d . W h i l s t t h e tuning of
c o n v e n t i o n a l f e e d f o r w a r d e l e m e n t s is k n o w n
to be d i f f i c u l t , t h e a d v a n t a g e h e r e l i e s in
t h e a u t o m a t i c t u n i n g o f t h e s e e l e m e n t s . In
this p a r t i c u l a r c a s e , i n c r e m e n t a l algo-
rithms do n o t a p p e a r to o f f e r s i g n i f i c a n t
improvements over p o s i t i o n a l algorithms.
For u n m e a s u r a b l e d i s t u r b a n c e s , the quality
of d i s t u r b a n c e r e j e c t i o n u s i n g t h e l a t t e r
is d e p e n d e n t on parameter adaptation
whereas the former guarantees offset remo-
val. F u r t h e r , the incremental strategy
i m p a r t s b e t t e r d e c o u p l i n g p r o p e r t i e s to t h e
multivariable S T C s . H o w e v e r , t h e r e m a y be
instances where use of incremental STCs m a y
l e a d t o d e g r a d a t i o n o f p e r f o r m a n c e s d u e to
increased controller sensitivity. This sen-
sitivity c a n , n e v e r t h e l e s s , be r e d u c e d
through a n a p p r o p r i a t e c h o i c e of control
weighting. The multirate self-tuning stra-
tegy p r o v i d e d a m e a n s f o r a c h i e v i n g i m -
proved o v e r a l l p e r f o r m a n c e . The faster
sampling rate a l l o w s the early detection of
disturbances a n d hence faster disturbance
rejection can be achieved.
 Terminal Compositions in Binary Distillation 213

ACKNOWLEDGEMENTS M o r r i s , A . J . , Y. N a z e r a n d R.K. W o o d ( 1 9 8 D .
In C . J . H a r r i s a n d S . A . B i l l i n g s
The a u t h o u r s g r a t e f u l l y a c k n o w l e d g e t h e (Eds. ) Self-Tuning a n d Adaptive
financial s u p p o r t f r o m t h e N S E R C o f C a n a d a Control: Theory and Applications.
under grant A-1944; the Robert W o o d Fellow- Peter P e r e g r i n u s , New York, 2 4 9 -
ship a n d t h e O R S S c h e m e A w a r d o f t h e U K . 281 .
Thanks are also d u e to the Depts. of C h e m i - M o r r i s , A . J . , Y. N a z e r a n d R.K. W o o d ( 1 9 8 2 ) .
cal E n g i n e e r i n g a t b o t h t h e U n i . o f N e w c a s - Multivariate self-tuning process con-
tle u p o n T y n e , U K , a n d t h e U n i . o f A l b e r t a , trol. O p t . C o n t r o l A p p . M e t h . , 3, 36 3-
Canada, for the u s e of their facilities. 387.
Peel, D., A.J. M o r r i s a n d M.T. T h a m (1985).
REFERENCES Single variate self-tuning control in
a d i!s t r i b u t e d e n v i r o n m e n t . P r o c .
Astrom, K.J. (1983). Theory a n d a p p l i c a - Control 8 5 . C a m b r i d g e .
tions o f a d a p t i v e c o n t r o l - a s u r v e y . Sastry, V.A., D . E . S e b o r g a n d R.K. W o o d
A u t o m a t i c a . 12., 4 7 1 - 4 8 6 . (1977). Self-tuning regulator applied
Bierman, G.J. (1976). M e a s u r e m e n t u p d a t i n g to a b i n a r y distillation column.
using the UD factorization. Automati- Automatica. 13 , 4 1 7 - 4 2 4 .
c a , 1_2, 3 7 5 - 3 8 2 . Seborg, D . E . , S . L . S h a h a n d T . F . E d g a r
Chien, I.L., D.A. M e l l i c h a m p and D.E. (198 3 ) . A d a p t i v e c o n t r o l strategies
Seborg (1983). A d a p t i v e c o n t r o l s t r a t - for process control-a survey. Paper
gies f o r d i s t i l l a t i o n c o l u m n s . P r e - presented a t A I C h E D i a m o n d Jubilee
prints A C C . Meeting, Washington, D.C., Nov.
Clarke, D.W. a n d P.J. G a w t h r o p ( 1 97 9). S e l f - Tham, M . T . (1985). S o m e A s p e c t s o f M u l t i -
tuning c o n t r o l . P r o c . I E E
t 12 6 1 No.6. variable S e l f - t u n i n g Control, PhD
H o d g s o n , A . J . F . (1 9 8 2 ) . P r o b l e m s o f i n t e g - Thesis, U n i , o f N e w c a s t l e upon Tyne,
rity i n a p p l i c a t i o n s o f a d a p t i v e c o n - U.K.
trollers. O.U.E.L. R e p o r t N o 1 4 3 6 . V a g i , F. ( 1 9 8 5 ) . M u l t i v a r i a b l e S e l f - t u n i n g
K o i v o , H . N . (1 9 8 0 ) . A m u l t i v a r i a b l e self- Control of a Binary Distillation
tuning controller. Automatica, 16, 3 5 1 - Column. M S c . T h e s i s , Uni.of A l b e r t a ,
366 . Edmonton, C a n a d a .
Wittenmark, B. a n d K . J . A s t r o m ( 1 98 4 ).
Practical issues in the implementation
of s e l f - t u n i n g c o n t r o l . A u t o m a t i c a ,
20, 5 9 5 - 6 0 5 .

S 95.25
cc
g 95.00 fa
S
Χ 94.75

94.50
I I .60-!
9.0500-
~ 10.90
£ Β.6125-
ΰ 10.20

Κ
I£ Β . 1 7 5 0 -
LU
k 9.50·,

7.7375 a.
βϋ I2U
BU 120
7.3000
TlflE(MINS) TIME(MINS)
Fig. 1. Positional STC - Unmeasured Disturbance (+2$% FE)

95.50-1

> î 95.2s4

94.50

8.2500

«5.8125

091 5 . 3 7 5 0
UJ

4.9375·

4.5000'
80 120 80 120
TIME(niNS) TlflE(MINS)
Fig. 2. Positional STC - Unmeasured Disturbance (-2$% FE)
 214 M. T. Tham et al.

95.60 95.50 (-2596 F E )

95.50
(+2$% FE)
95.50
(-25% FE) (+259<i F E )
; 95.25 $95.25
CO l\ 05.26 g 95.00
g 95.00 co
Q 94.75
g 94 . 7 6
94 . 5 0
g 95.00

94.60 U § 94.75. 120

94.50·
12·

A 8
Π p.
<o = U)
3 O)
1d 6 3
!5 "
CD 3-1 X
X X
o- 0 o
U 60 12U
60 120 0 60 120
TIME(flINS) TIMEiniNS) TIME(ÎIINS)
TIMEIMINS)
(a) (b) (a)
(b)
Fig. 3. K-Incremental STC Unmeasured Disturbances F i g . Ij.. P o s i t i o n a l M u l t i r a t e STC
Unmeasured Disturbances

95.60 (+259Î F E ) (-25?é F E ) (+25# F E ) (-25% FE)

$ 96.25
CO ft 95.25
O)
g 95.00
g 95.00
g 94.75
g 94.75
94.60·
12- 94 .50-1
Γι p.
CO
g β.
A 3
DÛ q.
X

0 60 120 60 120 0 60 120

0 60 120
TINE(MINS) TIME(NINS) TiriEiniNS) TIÎIE(MINS)
(a) (b) (a) (b)

F i g . 5. K - I n c r e m e n t a l Multirate STC Unmeasured Disturbances F i g . 6. Positional STC Measured Disturbances

95.50 (+25%
(+25% FE) (-25% FE) 95.50 FE) (-25% FE)
$ 96.25 $95.26
CO ι
CO
g 95.00-|
g 95.00 g 95.00-j
g 94.76 11
g 94.75· Ω 94.75
9 4 . 5 0 "1 1 1 1 Γ —1
0 60 120 94.50·

CO
- Λ
g 6-1 CU 6 -

1 r 1 1 1
-i 1
60 120 60 120 0 60 120 0 60 120

TIME(fllNS) TINE(MINS) TIME(MINS) TIME(MINS)

(a) (a) (b)
(b)
F i g . 7. Positional Multirate STC - Measured Disturbances Fig. 8, K - I n c r e m e n t a l Multirate STC Measured Disturbances
 Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

ANALYSIS AND CONTROL OF DISTILLATION

COLUMNS: A QUASI CLASSICAL APPROACH
H . Rotstein, A. Desages, J. A. Romagnoli and N. Karin
Planta Piloto de Ingenieria Quimica (PLAPIQUI), UNS-CONICET,
8000 Bahia Blanca, Argentina

Abstract, In this paper the application of Quasi Classical Approach to the analysis of
distillation columns with dual composition control is presented. The method relies
heavily on a computer graphic package because it retains the basic fundaments of classi-
cal control theory such as Nyquist and Bode plots. Firstly the main measurements and
properties in which the methodology is based are briefly described. Then the advantages
of using reserved frame controllers is discussed and through the example it is shown
that they can handle the principal aspects of a control system.

The experimental model of the column, TOFA, has been chosen to show the applicability
of the methodology.
Keywords. Multivariable control systems. Process Control. Frequency response.
1. INTRODUCTION
Distillation processes have been the subject of
numerous control and optimization studies due to the
large benefit obtained through better understanding
and control of these units. They constitute typical
examples of nonlinear multivariable systems and pro^
vide a permanent challenge to the control system
designer. Lubyen (1967) has shown the benefits a-
chieved when both products compositions are con-
trolled; such a control scheme has been called
'"dual composition control" or "two point control"
However, the interaction between the two compo-
sition feedback loops can be harmful. As a means to
decrease interaction so called decouplers are used,
also industrially but additional problems of sensi-
tivity and robustness with respect to model uncer-
tainties arise, for controllers implemented on a
real plant are derived from a mathematical model
which may describe the process only approximately.
If this fact is not taken into account the closed
loop performance of the plant deteriorates from the
predicted and may compromise stability.
The last several years have been something of a
revolution in multivariable control theory. Central
to this revolution is a renewed focus on the funda-
mental issue in feedback design: providing perfor-
mance in the face of uncertainty, which supported
the interest in the frequency domain. Several re-
sults of both practical and theoretical importance
come out of this research and it is now well known
that multivariable sensitivity and robustness can
not be reliable evaluated one loop at a time (Le-
thomaki, 1981). These new results have seldomly been
applied to study chemical processes, mainly because
well established techniques were prefered (for an
example of INA applied to distillation columns, see
Waller, 1984).
In this paper mathematical analysis and design
tools are presented which can be systematically
used to study the properties of the multivariable
system resulting in the development of dual compo-
sition control, within the mathematical framework
the degree of interaction, sensitivity and robust-
ness can be quantified and controllers of different
degrees of sophistication can be designed. In order
to demonstrate the utility of the approach we will
dwell on the case study of two typical distillation
columns currently found in the literature. The
study is developed in three stages: first the plant
DCCR-0
without compensation is investigated, then simple
controllers are designed to study the factibility
of controlling the system with a compensator of
given order and finally some simulation results are
shown.
2.SYSTEM ANALYSIS VIA SPECTRAL FACTORIZATION
In order to understand the advantages of n asi -
u
Classical approach in the analysis and control of
multivariable systems, a short description of the
conceptual framework is necessary. The main goal of
the whole method is to present the overall aspects
of the system under study in terms of pain and
phase parameters, associated with graphical displays
An appropriate interpretation of these graphics
gives a good knowledge of the system and its be-
haviour.
Let G(s) be the open loop nominal transfer function
matrix of the system under study, then two standard
descompositions are used:_1
G(s) = M(s)A(s)M(s) (1)
the eigenvalue decomposition and
G(s) = Y(s)Z(s)U(s)* (2)
the singular valued decomposition (SVD). Provided
some care is taken it can be assumed that if G(s)
is continuous in s, then the continuity properties
carry over both to the eigenvalues and eigenvectors
and to the singular values and vectors. Let Dnyq de
note the standard Nyquist D contour, then as s
traverses Dnyq in the Clockwise direction, the set
of eigenvalues,,
X(G(s)) = i g ^ s ) , g (s) g^s)}
2
traces out a set of eigenlocies in the complex
plane that is called the "characteristic gain loci"
(CGL) which is used to assess stability by means of
the following;
Theorem 1 : Generalized Nyquist Stability Criterion
(Mac Farlane, 1976),
mxn
Let G ( s ) e C ( s ) . then the closed-loop nominal
system (I+G(s))^^-G(s) is stable if and only if
-Wg.Dnyq and #(g.Dnyq.-l) = # SKP(G(s) ,C) )
where the left side of the equality is the number
of encirclement of point -1 by the CGL and the
right side is the number of Smith MacMillan poles

215
216
of G(s) in the open right half plane.
Following a straight generalization of SISO stabili
lity criterion, it would be nice to be able to
measure stability margins from the CGL but unfortu-
nately that is not a good approach for it is well
known (Lethomaki,1981) that the loci may be very
sensible to small plant perturbations. This justi-
fies the use of SVD whichhas
E(s) = diag (a,(s), a ( s ) , . . . , σ (s)}
ι 0z m
where
1 I η with G e C
and the a^(s) are real valued, frequency dependent
singular value functions that in the present
contextare called the principal gains because they
can be used to give a frequency domain characteri-
zation for the limits to some appropriately defined
gains and also to provide a condition number for
inversion (Wilkinson, 1965). Spectral sensitivity
is very important for Quasi Classical approach and
is addressed next.
Normality and spectral sensitivity
A matrix Q is said to be normal if and only if it
conmutes with its conjugate transpose. The reason
why normal matrices are important in our context
is brought out by the following theorem.
Theorem 2 (Wilkinson, 1965): Let A be a matrix
having linear elementary divisors and Η be a matrix
such that
Η *AH diag {\Λ
if A + εΒ then
λ is an eigenvalue of
I λ—λ. I < ε k(H)IIBll
-1
where k(H)=IH^UH' is the spectral condition num-
ber.
The theorem shows that the overall sensitivity of
the eigenvalues of A is dependent on the size of
k(H) may be regarded as a condition number of Awith
respect to its eigenvalue problem. Furthermore if A
is normal then
k(H) = 1
and the eigenvalue problem is always well con-
ditioned. This nice spectral properties may seem to
have little practical importance, for normal matri-
ces are a small set of all matrices and very rarely
a matrix transfer function will be normal in the
whole frequency range. However in theorem 2 we can
replace "normal" by "approximately normal" and the
eigenvalue problem will still be well conditioned.
It is necessary to formalize what is called approxi
mately normal" and to give a measure of the
distance to normality in the present scope.
Misalignment
A measure of departure from normality will be found
in a seemingly elyptical way using the following
definitions and properties (Hung, 19δ2) .
11
Definition: two unitary matrices U, Υεϋ™" are
called aligned if there exists
Θ = diag {exp(0i)} such that
U*Y = θ
Definition: let GeC . If all possible SVD's of G
G = ΥΣυ*
are such that Y, U are aligned, then G
is said to be aligned.
Property 1: G is aligned G is normal.
Property 2: If G has distinct singular values then
the converse of property 1 holds.
The singular vector frame misalignment of G(s) is
defined to be
H. Rotstein et al.
m(G) = sup min β U*Y - diag {exp(je.)}H*2
U,Y Θ.
which in case of distinct singular values reduces
to
m(G) = min I U*Y - diag {exp(JO ) } l
2
Moreover it can be proved that
near misalignment near normality
and so misalignment may be used to measure the
departure mfrom
nx inormality. In the general case,
we have
0 < m(G) < 2
and for m = 2
0 < m (G) < /2
Spectral sensitivity may also be measured by means
of the sensitivity indices (Postlehwaite, 1983).
Let x± and y^ be right and left normalized eigen-
vectors associated with eigenvalue X±. Then the
sensitivity indices are given by
with 0 < s. < 1.
ι
Note that for normal matrices s = 1 and the eigen-
values are least sensitive. If s -> 0 sensitivity
of X-j[ increases because the eigenvalue problem is
worst conditioned. Normality is an adequate
structural measure for the conditioning of the
eigenvalue problem; however it is only a sufficient
condition for small spectral sensitivity and there-
fore the indices can be used as a more detailed ve-
rification.
Ouasi-Nyquistmxm
Loci
Let GeC have an SVD
G = YZU*
and let Θ = diag{exp(jG^)} such that
m(G) = ||U*Y - Θ.|
2
If we now define
Δ = ΘΣ
V = ΥΘ*
we can write
G = (Υ Θ*) (ΘΣ) U* = VAU*
which is called the Quasi Nyquist decomposition of
G. Now assuming that G(s) has no poles on the Ny-
quist D contour, the OND can ba applied for each s
on the contour and set
G(s) = V(s)A(s)U*(s), s ε Dnyq
The diagonal elements of A(s) trace a set of m loci
in the complex plane called the "Ouasi-Nyquist loci"
of G(s). Note that the modules of the curves for
each frequency are the principal gains and the
phase is related with a normal approximation and
that comparison of the CGL and the ONL gives a
great deal of information about the system.
Robustness and interaction
No set of measurements would be complete without
measures for robustness and interaction. It has
been already stated that the distance of the CGL
to the critical point -l.+jO is an optimistic mea-
sure for robustness due to the sensitivity of the
CGL to perturbations. However, when a plant is
normal (near normal) this distance becomes a good
measure for robustness and for instance the M
circles or the reciprocal M circles may be used.As
taesystem departures from normality another way of
measuring the multivariable distance must be found
 Analysis and Control of Distillation Columns 217

and a natural canditate is to study dual composition control of distillation

M(G) =0(1 + G)
which gives an upper bound for the norm of multi-
plicative ir.odel errors. A more general measurement,
which applies to arbitrarilly define model errors
is
M(G) =Ô(h(G))
where h(.) is a bilinear matrix function of G, de-
pending on the defined type of model error.
Regarding to interaction, the different approaches
give rise to one or more measurements with graphical
display. As singular value decomposition is central
to the present theoretical frame, interaction is
measured by means of the total interaction angle
(Lau, 1985) with the observations made by Rotstein
(1986).
CONTROLLER DESIGN
It has been shown the importance of having normal
(or almost normal) plants: therefore a good target
for design is normalizing a plant while simultane-
ously handling the three aspects of closed loop
behaviour: stability, performance and robustness.
This may be achieved using a reversed frame con-
troller (Hung, 1983). Let
G(s) = Y U ^ ( s ) U * ( s )
G vity indices. The interaction angle shows that the
and a diagonal matrix Y (s) = diag {y^(s)}; then
K s
the compensator
K(s) = U(s)Y (s)U*(s)
K the system has two mayor disadvantages: a great
will be called a Reverse Frame Controller (EFC).The
reuiting compensated system is
Q(s) = G(s)K(s) = Y(s)Z (s)Y (s)Y*(s)
G K
It can be seen that the system is now normal, with
a CGL given by
(S) = X (S) (S)
Y Q q Y K
In general it is not possible to realize the con-
troller exactly, and an approximation shall be found
using an optimization program. However in two par-
ticular cases the design may be achieved with
simple and fast procedures: 00
columns has been investigated on the following
systems.
1. TOFA, a 10 plate column modeled from first prin
ciples previously studied by Toijala and Fager-
vik (1972).
2. WAWO, a 30 plate industrial column whose model
has been obtained experimentally by Wardle
(1969).
The first column studied was TOFA which has only
one relatively small time delay in element G]_i(s).
The CGL has a shape that is usual at -least in the
low frequency range- for distillation columns: one
branch in the first and the other in the third
quadrant and the absence of resonant poles is
clearly denounced. TOFA is open loop stable and
from the figure it can be seen that it is also
closed loop stable, due to the small module of the
eigenvalues. The principal gains are disbalanced
in a wide range, and for s = 0:
α = 1.05
χ
σ
2= 0.076
The superposition of CGL and QNL shows that TOFA is
far from normality, a fact that is quantisized by
the misalignment plot. The bad conditioning of
the eigenvalue problem is confirmed by the sensiti
system i tightly coupled, a common characteristic
to all distillation columns with dual composition
control. From the set of graphics it is clever that
departure from normality and a high interaction
angle.
Controllers Design
The first approach on designing a controller for
the distillation column was to design a multivariji
ble PI which normalizes the plant in 0 and °°, how-
ever the controller should not only reduce misalign^
ment but also widen the bandwith and lower interac-
tion. The election of the singular values of the con
troller is a problem of trial and failure; the final
value were

a) Normalization at 0 and : in this two , .375 α° - 5.

frequency the model structure are real matrices and
a PI can be designed with t which balanced up the gain in the critical frequen-
cy range, and
K = K(o) = U ( o ) Z Y ( o )
κQ = K o)t ( 0 0
Κ =Z U Y 00
σ, = .01
οο (°°> (°°) K c o
and . 0 0
Κ = -- Κ + Κ
TI s o
to assure stability.
which normalizes the system at low and high frequen-
cies. The results of the compensated plant are shown in
figure 2, Note that the bandwith is not wide be-
b) Linear design using linear leasts squares:
cause the CGL would get to near to the -1.0,0.0
in this case the set of poles for the controller
point thus compromising robustness; this would
and a desired CGL have to be chosen. Given the OND
mean a slow response and small disturbance re-
for the plant and a set of frequencies Ω = iw.},
jection. However two goals have been partially
i=l,...,n
achieved: i) Interaction has been reduced and ii)
G(jw) = Y(jw,)E (jw.)U*(jw.) Misalignment is substantially lower on a wide
p frequency range.
the desired plant is found in this frequencies:
Q(jw.) = Y(jw.)y (jw )Y*(jw ) Simulation
Q i i
Then, after some algebraic manipulation and an ade-
quate selection of the poles for the controller a Figure 3 shows the behaviour of the precompensated
linear least squares problem can be stated which column TOFA for changes in the set points and the
can be solved analytically be means of the pseudo disturbance rejection when the outputs are pertur-
inverse. bed. Note that the control objectives (stability,
performance and robustness) were partially obtained.
No doubt, better results shall be expected if the
3. TOFA COLUMN
controller is designed by a more sophisticated
System Analysis algorithm wich optimizes over some of the parame-
ters j however the proposed controller shows that a
The feasibility of using Quasi-Classical approach PI should be enough to give good control over TOFA,
218
4.WAW0 COLWiN
System analysis
The following column studied (figure 4) was UAUO.
An interesting fact about VJAWO is that it has a
pair of complex conjugate zeros close to the imagi-
naty axis which reflects both in the principal gains
(observe the low frequency gain increment) and the
CGL (which has a phase lead)· then the column pre-
sents a ''bandpass" behaviour, Comparison of the CGL
and the QNL shows that the input and output struc-
tures are severely misaligned,· this misalignment is
of critical importance because the smaller zeros
and poles are close together and then a small per-
turbation could drastically change their behaviour.
Controller Design
Due to the low frequency phase lead, problems were
expected when attempting to control VAUO with a PI.
One of the branches of the CGL got deep into the
second quadrant and only at the crossing frequency
both branches behave alike: this made the principal
gains equating at critical frequencies quite diffi-
cult. The final PI obtained had:
k = 19.5k
n 12 = .540 k
21 = 13.8 k £2 = -12.4
T = 91 7 =,T5 9 T= 21 6 =,T 9
ll 12 " ° 21 22
In previous applications of Quasi Classical Approach
misalignment variâtes in a smooth fashion at high
frequencies due to the asymthotic trends of the
different parameters of the model. Instead the
presence of time delays introduce a typical be-
haviour in the CGL which in turn severely affects
00
the misalignment. In spite of this it is still pro
fitable to use normalization because high frequen
cy gains can be handled and generally misalignment
is partially reduced. Note that the bandpass' ef-
fect could not be surpressed -in fact not even
compensated- by this simple controller and that
interaction angle is quite high. In order to reduce
interaction a higher order controller should be
used. A natural candidate for denominator matrix
was:
D(s) = s(s+.013).I
which then trends to compensate the troublemaker
zero. The final design met most of the desired
specifications (figure 6 ) ; misalignment is low,
bandwith has augmented interaction is lower and
robustness has not been too compromised.
Simulation
Figure 7 showsthe simulation of the compensated
column. Note that the harmfull effect of coupling
between loops is still present and this can
doubtly be reduced with a second order controller,
suggesting that very likely a higher order con-
troller should be needed if interaction reducing is
critical.
5.SUMMARY AND CONCLUSIONS CONCERNING MODERN
FREQUENCY DOMAIN APPROACH FOR ANALYSIS OF
DISTILLATION COLUMNS
By using two models it has been shown that Quasi-
Classical approach is useful for the analysis of
distillation columns and serve as a reliable guide
in the design of multivariable feedback controllers.
The study of the uncompensated plant has shown to
beappropiate to investigate the system and to detect
the problems that will arise when controlling with
a feedback scheme. The simple controllers serve as
a clue over how much could be expected from a com-
pensator of a given order while helping in providing
the necessary specifications.
Time delays introduce serious difficulties in the
H. Rotstein et al
application of quasi-classical approach, for the
typical behaviour of the CGL is transmited to the
rest of the measurements including misalignment
(all present and 'oscillatory" behaviour at high
frequencies). Nevertheless the analysis is still
enlighting and a proper PI compensation could some-
times be found.
Robustness, which we have established as a key
property of control systems, could seem to be only
slightly attacked by this approach. A second sight
shows, however, that it is really present in the
whole analysis philosophy which is based in singu-
lar value decomposition. The close relationship
between robustness and this decomposition has been
thoroughly established in the literature (Lethomaki,
1981).
As a final conclusion, Quasi-Classical approach is
a powerfulltechniques of interactive analysis for
the user has at hand a set of measurements which
are necessary to specify his needs and understand
his results.
REFERENCES
Hung, Y. S. and Mac Farlane, A. G. J. (1983). Multi
variable Feedback: A Quasi-Classical Approach",
1 0 , Springer-Verlag, Berlin Heidelberg New York.
Lau Η., Alvarez J. and Jensen K. (1985). "Sunthesis
of Control Structures by Singular Value Analy-
sis; Dynamic Measure of Sensitivity and Inter-
action", AIChE Journal, 31. 427-453.
Lehtomaki Ν. Α., Sandell," N. R. and Athans M. (1981)
"Robustness Results in Linear-Quadratic Gaussian
Based Multivariable Control Design", IEEE Trans,
on Automat. Contr,, 26, 75-92.
Lubyen, W. L. and Vinante, C. D. (1972). "Experi-
mental studies of distillation decoupling' ,Kem.
Teollisuus 29, 499-514.
Mac Farlane, A. G. J. and Postlethwaite, I. (1972).
"The generalized Nyquist stability criterion
and multivariable rood loci". Int. J. Control,
25, 81-127,
Postlethwaite, I (1983). "Sensitivity of the Cha-
racteristic Gain Loci" Proc. of the second
IFAC symposium 163-174.
t
Toijala, Κ, V. and Fagervik, K. C. (1972). A digi-
tal simulation study of two point feedback
control of distillation columns", Kem. Teolli-
ssus, 29, 5-16.
Ualler, Κ. V,, Wikman, Κ. E, and Gustafsson, S. E.
(1985), "Decoupler Design and Control System
Tuning by ΙΝΑ for Distillation Composition Con^-
trol", Chem. Eng. Commun , 35, 149-174.
L
Wardle, A, P. and Wood, R. M, (1969). "Problems of
Application of Theoretical Feedforward Control
Models to Industrial-scale Factionating Plant' ,
I. Chem. E, Symp, Series No, 32, 6, 68-81.
Wilkinson J, H. (1965). "The Algebraic Eigenvalue
Problem", Oxford University Press.
 Analysis and Control of Distillation Columns 219

CHARACTERISTIC GAIN LOCI CHARACTERISTIC GAIN LOCI

-.3063
QUASI NYQUIST LOCI AND CHARACTERISTIC GAIN LOCI PRINCIPAL GAINS

.10000 .10000 .mm .mm

E--M E+?« E+01 £+02
ROBUSTNESS MEASURE - MULTIPLICATIVE PERTURBATIONS

FIG. 2? TOFA WITH PI COMPENSATOR

FIG 1Î UNCOMPENSATED COLUMN TOFA
SET POINT CHANGE RESPONSE SET POINT CHANGES

: :
-Π:.. H : : : . ) - ^ I · I
.2 12.18 24.16 36.14 48.12 60.1 72.08 84.06 96.04 108.02 120. v. 16. 32. 48. €4. 8*. 56. 112. 128. 144. 160.

UNITS OF TIME UNITS OF TIME

DISTURBANCE REJECTION

.02
f- \\ \ 1
\
-.e2

: ! ! 5 ! ! ! ! ! :: ! ! !! !!
.6 12.54 24.48 36.42 48.36 60.3 72.24 84.18 96.12 108.06 120. η 5 ΐ ί 1 i 51 i i ; i hi i 1 · ·· 1 ' i 1 i ' 1 •· = 1 ·· 1
16 . 32 . 48 . 64 . 80 . 96. 112. 128. 144. 160.
UNITS OF TIME

FIG. 3: SIMULATION RESULTS FIG 7: COLUMN WAUO - SIMULATION RESULTS

220 H. Rotstein et al.

CHARACTERISTIC GAIN LÛCI

characteristic gain loci

κ . )
\
s
J
^^^^ ^y'
/ ^
-.«67
a5. 5
QUASI NYQUIST LOCI RINClPAL GAINS

(
\

...... .mw
E-%2 E-Gi ROBUSTNESS MEASURE - MULTIPLICATIVE PERTURBATIONS
GLOBAL INTERACTION ANGLE
I 1 1
-

- / -
—
1 1 1
•&S? -k%v
FTR 4i UNCOMPENSATED COLUMN UAWO FTG 5: COLUMN UAUQ - PI COMPENSATION

CHARACTERISTIC GAIN AND QUASI NYQUIST LOCI

Λννψ Λνΰνν .lew ,iww

E-02 È-èi E+M E+01
PRINCIPAL GAINS kubuS'inESS MEASURE - MULTIPLICATIVE PERTURBAT [UNS

I 1 1

38.

26,

ι*.
- / -
/
t. /

-ie. i i 1
m? -m
ETG 65 COLUMN UAUO - 2nd ORDER COMPENSATOR
-WT -gR* I W
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

IDENTIFICATION OF PERIODIC CYCLED

DISTILLATION FOR CONTROL PURPOSES
B. Toftegârd and S. B. j0rgensen
Instituttet for Kemiteknik, Technical University of Denmark, Building 229,
DK-2800 Lyngby, Denmark

ABSTRACT. Identification of periodic cycled distillation is carried out in

three steps. First a qualitative description of the periodic process dynamics
is obtained. Secondly a discrete time model structure is formulated, and thir-
dly the model parameters are estimated. The formulated model explains about 99%
of the output variances. This model may directly be used in an adaptive contro-
ller design. As an example a simulation with a minimum variance controller is
presented and demonstrated to have good output variance reduction capabilities.

KEYWORDS. Periodic distillation, Identification, Dynamic quantity, Interaction

diagram, Model structure, Process control.

1. INTRODUCTION. Matsubara et. al. (1982) introduced 'Relay

Periodic cycled distillation was first intro-
duced by Cannon (1961). A cycle period con-
sists of two parts: The vapor flow part (VFP)
and the liquid flow part (LFP). In the VPF the
liquid stays on the trays and the vapor flows
up through the column and is collected in a
condenser. In the LFP bottom product is taken
out of the reboiler, the liquid is shifted one
tray down without being mixed, and the conden-
ser holdup is partly filled onto the top tray
and partly taken out as distillate. This ope-
ration form gives ideally a much higher column
efficiency than conventional distillation. With
an infinite reboiler holdup the performance of
periodic cycled distillation is exactly the
same as for a Lewis case II distillation (Le-
wis, 1936) if the Murphree point efficiency is
the same in the two cases. The principle in a
Lewis case II separation is that the liquid is
flowing in plug flow across the tray in the
same direction on every tray and the vapor is
flowing in plug flow up through the column.
Periodic cycled distillation is one method to
perform a Lewis case II distillation, where
length is substituted with time.
In this paper 'control' means feed back con-
trol. Open loop control during the VFP and LFP
is not considered here. The control strategy
for a periodic cycled distillation column is
different from control of most other chemical
processes. From a control point of view the
process operates directly in discrete time with
the sample time equal the cycling period, be-
cause products are withdrawn during the LFP
only. A few control strategies for this pro-
cess has been proposed in the littérature.
Dale and Furzer (1978) proposed a discrete time
PI (proporsional-integral) controller. They
used the duration of the VFP as manipulated
variable, controlling an appropriate output.
feedback periodic control'. This control
strategy gives perfect control if the control-
led variable can be measured without delay and
without noise, and uses also the duration of
the VFP as manipulated variable. The control
strategy is simple. If e.g. a top product
concentration is to be controlled it is mea-
sured continuously. The concentration will be
decreasing and when it reaches the desired
value the top product is withdrawn. Then the
VFP is continued until enough liquid is collec-
ted in the condenser to be returned to the
column. The control strategy proposed by Dale
& Furzer is SISO (Single Input Single Output).
This gives some limit in the control perfor-
mance, e.g. fluctuation in the controlled con-
centration, e.g. the distillate concentration,
on changes in the feed concentration. The
control strategy proposed by Matsubara et. al.
is also SISO, but the control is perfect if the
measurement is undelayed and noise free. How-
ever this is not always the case. Α ΜΙΜΟ
(Multi Input Multi Output) controller will be
able to take interactions between all measure-
ments and all manipulated variables into ac-
count. Furthermore a ΜΙΜΟ controller designed
from a properly formulated model automaticaly
may include a feed forward part, such that some
fluctuations are reduced if not eliminated.
However to design a ΜΙΜΟ controller it is es-
sential to have a reasonable model describing
the process dynamics.
The purpose of this paper is to describe the
identification of a periodic cycled distilla-
tion column. The method used requires only
qualitative knowledge of the physical interac-
tions in the column. The first step is to make
a DYQUID (DYnamic QUaninty Interaction Dia-
gram) . From this interaction diagram a model
structure may be formulated. This model struc-
ture may then directly be fed into a proper
estimation algorithm for parameter estimation.
Using the resulting parameters a control design

221
222 B. Toftegârd and S. B. J0rgensen
may be carried out. This estimation and con-
trol may also be combined into adaptive con-
trol. This paper is also an example of how a
general model formulation for an adaptive con-
troller may be carried out. The structure of
the paper is as follows. Section 2 contains a
description of the qualitative dynamics of a
periodic cycled distillation column and a model
structure is formulated. To illustrate the use
of the model structure in section 3 an estima-
tion of the model parameter for an example is
described, and in section 4 the control strate-
gy is formulated and illustrated by a simu-
lation example.
2. DYNAMICS.
The process dynamics are described qualitative-
ly with a DYQUID. From this DYQUID a ΜΙΜΟ
model structure may be formulated. A nine tray
column with feed on tray five is used as an
example for illustration of the DYQUID and the
model.
The procedure for developing a DYQUID for con-
tinuous time systems has been introduced by
Jorgensen et. al. (1985). From this DYQUID
they formulated a discrete time model. Due to
the nature of the periodic cycled distillation
the DYQUID in this paper is extended to handle
discrete time events. All continuous time
interactions are carried out over a specified
time interval. Discrete time events result in
momental interaction at a specified time.
The DYQUID for a periodic cycled distillation
column is shown in-Fig. 1. A circle designate
a quantity, the thick arrows mass flow quantity
transport, the thin arrows other interactions,
and the cut symbol (dubble slash) a discrete
time event. If the discrete time events occur
at different times the sequencing can be noted
next to the cut symbol. For simultaneous dis-
crete time events interaction between the quan-
tities occur only if they are directly connec-
ted by an arrow. The significant quantities
for periodic cycled distillation are the compo-
nent mass holdups on the trays, in the reboiler
and in the condenser . The quantities interact
with themselves, since the flows from the trays
depends upon the quantity. This is illustrated
in Fig. 1 by the arrows from the interaction to
themselves. The up going arrows on the left
symbolize component mass flow in the vapor flow
up through the column. This component mass
flow interacts immediately with all quantities
above if the tray efficiency is less than one
and is active during the whole VFP. The influ-
ence is however greatest on the nearest trays.
The arrows on the right symbolize the component
mass flow in the liquid flow during the LFP.
In the LFP top and bottom product are withdrawn
and feed enters the column as illustrated on
Fig. 1. The LFP may for modelling purposes
just as well be assumed to have no duration.
Contrary to the interaction through the vapor
flow the interaction through the liquid is an
immediate event happening once in a period.
All discrete time interactions for periodic
cycled distillation is carried out simultaneous
in the LFP, that is at the end of the cycling
period. The concentration on the trays dec-
rease during the VFP as illustrated on Fig. 2.
Varying the cycling time the concentrations,
and thus the quantities, at the end of a
period vary. This interaction from the cyc-
ling time is illustrated by the thin arrows on
the rigth hand side of Fig. 1.
The next step is to express the interactions
between variables in some algebraic way which
may be used by an identification and control
algorithm. The model is discrete and the samp-
le time is equal the period time. It should be
pointed out that this description also is qua-
litative .
Assuming the tray holdup is constant a quan-
tity, i.e. component mass holdup, may be repre-
sented by the mole fraction. This will be done
here as concentration may be meassured. In a
column operating at constant pressure tempera-
ture may be measured instead.
Generally only a few concentrations are mea-
sured. The concentrations to be controled are
usually the top and bottom product concentra-
tions . The concentration on tray one is used
here instead of the bottom product concentra-
tion which is equal to the reboiler concentra-
tion since the reboiler holdup is quite large
and therfore changes slowly. In addition to
these concentrations the concentration on the
feed tray is used, as this is where disturban-
ces in feed concentration enter.
Interaction matrices for these three measure-
ments are shown in Fig. 3. The elements may be
identified from the interaction shown in the
DYQUID on Fig. 1. First the quantities inter-
acts with themselves. This causes the diagonal
elements in A^. The elements just above the
diagonal in A^, A^, and A^ are due to vapor
flow from measurepoints below. Actually there
is also an interaction from x^ to x^, due to
successive continuous interactions over the
cycling period. This interaction is however so
weak that it is neglegted. For a first simple
model the vapor interaction elements in A^ are
omitted as the interaction after just one cycle
period is rather weak. The main interaction
occur after two cycling periods giving the
elements in A .
? After a preliminary identifi-
cation the elements in A^ were omitted as they
did not give a significant improvement of the
degree of explanation. The interaction due to
the liquid flow occurs exactly four cycling
times delayed. This delay may be counted di-
rectly from the DYQUID, noting that the concen-
tration on tray nine and the concentration of
the distillate are the same at the sample time.
This gives the elements in A^. The elements in
B ^ are due to the interaction from the cycling
period, Τ . As the feed enters at the end of
the cycliRg periods the feed concentration
only influences the feed tray giving the ele-
ment in Άγΐ' Finaly to be able to handle
coloured noise and some model error a C^ matrix
is included.
This model structure is a simple description of
the dynamic behaviour directly made from pro-
cess physical knowledge. However as shown
later this simple model is able to explains
about 99 % of the output variance.
3. IDENTIFICATION.
To verify the model structure a parameter esti-
mation is performed using data from a nonlinear
simulation of a periodic cycled distillation.
The model structure is not unique and it may be
neseccay to extend it with e.g. additional Β
matrices. Therefore a degree of explanation of
the output variance is calculated. The exita-
tion of the input to the process is moderate as
the identification is a test of the model stru-
 Identification of Periodic Cycled Distillation
cture and not of the estimation algorithm.
The example column is for benzene toluene sepa-
ration with nine trays. Feed enters on tray
five, reflux ration equals 1.57, feed concen-
tration equals 0.50 mole fraction benzene,
equal top and bottom product amounts, and the
reboiler holdup is 20 times the holdup of the
upper most tray.
A recursive extended least squares parameter
estimation routine with forgetting factors
equal 0.98 is used for the identification.
This algorithm is a part of the program package
MIMOSC (ΜΙΜΟ Selftuning Control) described by
Hallager et. al. (1984). The model structure
as shown on Fig. 3 is an input to the program
package.
First the process is simulated for 200 cycling
periods with a PRBS (Pseudo Random Binary Se-
quence) signal on the cycling time and the feed
concentration in order to obtain initial values
of the parameters. The pertubation amplitudes
are about 3% on the cycling time and 4% on the
feed concentration. The parameters after this
simulation are shown on Fig. 4. The degree of
explanation, here defined as one minus the
ratio between the variance of the prediction
error and the variance of the measurements, are
for the three measurements about 98.8%, 98.5%,
and 99,3%. Thus the model structure provides a
reasonable description of the process dynamics.
4. CONTROL.
After some initial parameter values are esti-
mated the model may be used for design of a
controller. To demonstrate this a simple mini-
mum variance controller is chosen. The perfor-
mance of this controller based upon the model
formulated in section 2 is shown by a simula-
tion example. As control object it is chosen
to maintain the distillate concentration using
the cycling period as manipulated variable.
The tray holdup is held constant, hence the
product split is varied. The vapor flow is
assumed constant, this flow could however be
used as manipulated variable instead of the
cycling time. The feed concentration is varied
with gaussian distributed white noise with
standard deviation equal 0.01 and a reduction
in mean value of 0.02 in the interval from 350
to 370 cycling period as process input destur-
bance. The controller is started after 301
cycling periods.
The controller is a simple minimum variance
controller based on the model formulated in
section 2 where the parameters in C^ is not
taken into account. Other controllers, e.g.
for controlling both top and bottom product
concentrations where also the reflux ratio
(tray holdup) is used as manipulated variable,
may by implemented in a similar way.
From the estimated model parameter it is simple
to evaluate the control law shown on Fig. 5.
Using this specific controller it is only
necessary to estimate the parameters in the
upper row in the model shown on Fig. 3. One
have to measure x^ and x<-. The term with x,_
coresponds to a feed forward term.
Fig. 6 shows simulation results as deviations
fxoTft the stationary values, which are
(χ ,χ χ ) =(0.1846, 0.5109, 0.9558), Τ -
1.6377 and x = 0.50
f
223
5. DISCUSSION.
From Fig. 6 it can be seen that the control
performance is good. There is almost no influ-
ence of the steps in x^. on the distillate
concentration in the controlled case. That is
the model structure formulated from the DYQUID
is indeed suitable for control design. Other
controllers than minimum variance may be used,
if desired.
The procedure for obtaining the DYQUID and the
model structure presented here can also be used
for processes as periodic cycled stripping and
absorption.
6. CONCLUSION.
The main conclusion is that it is possible to
obtain a discrete time model structure from a
DYQUID for a periodic cycled distillation col-
umn; where the DYQUID is obtained from quali-
tative process knowledge. The model structure
is demonstrated, by parameter estimation on a
nonlinear simulation, to give a low prediction
error variance, thus it provides a high degree
of explanation of the process output. Further-
more it is demonstrated that the identified
model directly may be applied for control de-
sign. This control may be adaptive if necessa-
ry.
REFERENCES.
Cannon,M.K.: "Controlled Cycling Improves Va-
rious Processes". Ind.Eng.Chem,1961,53,629
Dale.E.B. and I.A.Furzer: "Periodic cycling of
plate columns. Control, simulation and experi-
ments". Chem.Eng.Sci.,1978,33,905-911
Hallager,L., L.Goldschmidt and S.B.Jorgensen:
"Multivariable adaptive identification and con-
trol of a distributed chemical reactor". Large
Scale Systems,1984,6,323-336.
Jorgensen,S.Β., L.Goldschmidt, L.Hallager:
"Sparse process modelling for robust multiva-
riable adaptive chemical process control". IFAC
Identification and System Parameter Estimation,
1985, York UK.
Lewis,W.K.: "Rectification of binary mixtures.
Plate efficiency of bubble cap columns". Ind.-
Eng.Sci.,1936,28,399-402.
Matsubara,Μ., Y.Nishimmura, N.Watanabe and K.-
Ongi: "Relay feedback periodic control of plate
columns", Chem.Eng.Sci.,1982,37,753-758.
SYMBOLS.
M^, Matrices in the model (see Fig. 3 ) .
e. nois for i'th measurement,
t Cycling period number.
Τ
P Cycling period duration.
x Distillate concentration,
x^ Feed concentration,
x. Concentration on tray i.
χ Reboiler concentration, p
y. i'th measurement.
224 B. Toftegârd and S. B. j0rgensen

0.270
0.946 /
y
t

-0.017
-0.013 )
-0.077
u
f • ( 0.558

Fig. 4. Estimated model parameter in the model

structure on Fig. 3.

u
==(p + yp ) / p
Control law:

Yt t - i>i,t 2 2.t-i 3

Fig. 1. DYQUID for periodic cycled binary dis- where p^, p?and p3are thu
e estimate d paramete r for the measuremen t y^-
p y + p +P
tillation column. Continuous interactions oc-
cur during a cycling period. Dicrete interac- >l,t+1 " l l,t 2,t-l 3 t
tions occur at the end of a cycling period. The
quantities are the component mass holdups . Fig. 5. Control law for a minimum variance
controller. Τ is used as manipulated variable
and x, and x Ire measured,
d 5c

j 0.004
-CONTROL
' 0.002 NO CONT.

0.000
-0.002
-0.004
-0.006
-0.008 , 1 1 1
350 400 450 500
NPER
Fig. 2. Principal behavior of the concentration
during the VFP. Variation in Τ gives varia-
tion in the concentrations. ^

450 500
NPER
+ + 0 12 η
0 10-
t +
y i =M
+ c
+
t ' h * t - l * *3 h-2 * U h-3 luiH
t 8 v l It 0 08-
0 06 -
-t+1 =i i t 0 04-
+ U) 0 02-
(x) (x) 0 00-
(x) y
t it-1 I ) h- -0 02-
-0 04 -
+ +
XU X V+ + -0 06- —ι— — I —
400
1
500
350
h-3 t t -t+1 NPER

Fig. 3. Model structure for a peridic cycled Fig. 6. Simulation results using minimum vari-
distillation column with measurements on tray ance control. Control is started after 301
1, tray 5, and distillate. T and x are input cycling period. Values are plotted as devia-
variables.
p f tion from stationarity.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

CONTROL OF AN EXTRACTIVE
DISTILLATION PLANT
1
B. Retzbach
HOECHST AG, Abt. TLS/Geb. C584, POB 800320,
D-6230 Frankfurt/Main 80, FRG

Abstract. A new separation scheme for extractive distillation which uses

a vapor sidestream is presented together with an appropriate control sys-
tem. Starting with a simplified model of the dynamic behaviour, the in-
fluence of the level control loops is investigated and a multivariable
control system is developed using state space methods. Simulation results
and experimental investigations at a pilot plant show the high robustness
and reliability of the control during startup and operation.

Keywords. Extractive distillation; distillation columns with side stream;

startup of distillation columns; profile position control; application of
observers; multivariable PI control.

INTRODUCTION with the problems appearing during the im-

To save energy and raw materials separation
schemes are designed showing close coupling
between the columns either by energy inte-
gration or by recycling of materials. Side-
streams are often used to remove intermedi-
•ate boiling or tailing components. This may
cause a complex dynamic behaviour and unex-
pected transients may lead to serious con-
trol problems. This is especially true if
the columns operate at low reflux ratios or
if strongly nonideal mixtures are separat-
ed. In both cases, pinch points occur with-
in the columns. The profiles are sharp and
their positions are very sensitive to dis-
turbances. The conventional single loop
control system is inadequate for these col-
umns. As an example, a new scheme to per-
form extractive distillation is presented
in this paper.
The high number of physical state variables
of a distillation column together with the
strong nonlinearity of the balance equa-
tions and the thermodynamic relations are
the main obstacles preventing the use of
state space techniques to design the con-
trol system. To overcome this difficulty,
we use a simplified model which is develop-
ed from the overall balance equations of a
column section. The model is first employed
in the investigation of the liquid level
control loops. These loops are assumed to
be in a quasi steady state later. Neverthe-
less, they strongly influence the remaining
model.
Based on the simplified model a state ob-
server and a state feedback are designed to
control the separation. The control system
is tested by simulation with a detailed mo-
del ( i.e. the material- and energy balance
equations of the trays ) and by experimen-
tal investigations at a pilot plant. De-
scribing these experiments, the paper deals
plementation of the multivariable control
system and the startup of the plant.
EXTRACTIVE DISTILLATION
Extractive distillation is a frequently
used separation process in the chemical
industry. A solvent is fed into the distil-
lation column in addition to the mixture to
be separated. This so called extractant is
selected to increase the relative volatil-
ity of the components. By that it is pos-
sible to separate homogenous azeotropes
like propanol and water. The extractant is
a high boiling component ( we use glycol in
our example ) , which is removed from the
column as a part of the bottom product.
Usually a second column is used to recover
the solvent.
A large scale industrial extractive distil-
lation plant was investigated by Silberber-
ger (1979), who developed a control system
to overcome operational problems. This con-
trol system proved to be of a high reliabi-
lity and has been in continous operation up
to now ( Gilles, Retzbach, Silberberger,
1980). During these investigations exten-
sive simulation studies were carried out
using a detailed stage-to-stage model of
the plant. This simulations yield mole
fraction profiles, which make a new ar-
rangement of the columns self suggesting
(Fig. 1 ) .
The vapor in the lower part of column K01
contains approximately 98 mole percent wa-
ter. Therefore this component is easily re-
moved by a vapor side stream S. The steady
state and dynamic composition and tempera-
ture profiles of the new scheme were calcu-
lated using a full set of balance equations
and including the nonideal thermodynamic
behaviour of the mixture. A typical result

The results presented in this paper were obtained by the author during his time with
Prof. Dr.Ing. E.D. Gilles at the Institut fur Systemdynamik und Regelungstechnik der
Universitât Stuttgart, Pfaffenwaldring 9, D-7000 Stuttgart 80

225
226 Β. Retzbach

late D E ~ is w a t e r .
If t h e f l o w r a t e o f t h e v a p o r rising in the
c o l u m n is i n c r e a s e d , t h e regions of h i g h
mass transfer rates and the temperature
fronts are forced t o m o v e upward in t h e
c o l u m n . T h e s t a t e v a r i a b l e s of a l l o t h e r
r e g i o n s a r e only s l i g h t l y altered. A f t e r
some t i m e , t h e c o m p o s i t i o n of t h e side
s t r e a m c h a n g e s a n d h e n c e K02 is affected
too.

SIMPLIFIED DYNAMIC MODEL

T h e e s s e n t i a l d y n a m i c q u a l i t y of the extrac-
tive d i s t i l l a t i o n c o l u m n is the m o v e m e n t of
the r e g i o n s of h i g h m a s s t r a n s f e r r a t e s .
F u r t h e r m o r e , t h e r e is a strong d e p e n d e n c e
of t h e p r o d u c t q u a l i t y on the loci of these
r e g i o n s . A c o n t r o l s y s t e m w h i c h e n s u r e s the
s e p a r a t i o n m u s t t h e r e f o r e s t a b i l i z e the
Fig. 1. Extractive distillation plant m o v e m e n t s of t h e f r o n t s in case a d i s t u r b -
ance o c c u r s .
is shown in f i g . 2 : S i l b e r b e r g e r (1979) i n t r o d u c e d the locus of
There are two regions of high mass transfer a t e m p e r a t u r e f r o n t as a state v a r i a b l e of
rates in the lower p a r t of K 0 1 . P r o p a n o l is a s i m p l i f i e d d y n a m i c m o d e l . T h e same p r o c e -
removed from t h e liquid p h a s e a b o v e t h e d u r e w a s u s e d by G i l l e s and R e t z b a c h (1983)
s i d e s t r e a m S. W a t e r and g l y c o l a r e s e p a r a t - to d e v e l o p a s i m p l i f i e d m o d e l of an e x t r a c -
ed b e l o w . A t t h e s e l o c a t i o n s , t h e c h a n g i n g tive d i s t i l l a t i o n c o l u m n w i t h a v a p o r s i d e -
compositions cause sharp temperature p r o - stream. It h a s b e e n g e n e r a l i z e d b y Retzbach
f i l e s , t h e so c a l l e d t e m p e r a t u r e f r o n t s . (1986) t o a l l o w t h e d e r i v a t i o n o f s i m p l i -
The b o t t o m p r o d u c t SU-j c o n s i t s of t h e p u r e fied m o d e l s for a l l k i n d s of d i s t i l l a t i o n
extractant, which can be recycled after plants where sharp separations occur.
c o o l i n g . A small a m o u n t of t h e e x t r a c t a n t The c o l u m n s a r e s e p a r a t e d into s e c t i o n s a t
the feed a n d t h e s i d e s t r e a m l o c a t i o n s .
W r i t i n g d o w n t h e o v e r a l l m a t e r i a l - and e n -
ergy b a l a n c e e q u a t i o n s of t h e s e c t i o n s ,
d i f f e r e n t i a l e q u a t i o n s of the loci o f the
r e g i o n s o f h i g h m a s s t r a n s f e r r a t e s are
o b t a i n e d . If the p r o f i l e s are flat at the
b o u n d a r i e s o f t h e s e c t i o n s , t h e linearized
b a l a n c e s y i e l d i n t e g r a t i n g e l e m e n t s . The
d e t a i l e d d i s c u s s i o n o f the m o d e l i n g p r o c e -
d u r e is b e y o n d t h e scope o f this p a p e r . B u t
it is i m p o r t a n t t o k n o w t h a t t h e m o d e l s are
o b t a i n e d in t h e c o m m o n state s p a c e m a t r i x
notation.
In f i g . 3 the b l o c k d i a g r a m o f t h e s i m p l i -
fied d y n a m i c m o d e l of t h e e x t r a c t i v e d i s t i l -
lation p l a n t is g i v e n . In this e x a m p l e m o v e -
m e n t s of t h e r e g i o n s of m a s s t r a n s f e r are
p o s s i b l e a b o v e and b e l o w t h e side stream
(m=4, m = 5 r e s p e c t i v e l y ) and in t h e side
water K02 c o l u m n ( m = 8 ) . T h e liquid h o l d u p s in the r e -
1 b o i l e r (m=6) and in the c o n d e n s e r s (m=1,
"ζ-
j^T—- -A — — m = 7 ) are state v a r i a b l e s of t h e simplified
SU, liq. mole fractions m o d e l as w e l l .

3 5 0 4 Θ 0
' temperature l Κ] ·

F i g . 2. P r o f i l e s o f t h e e x t r a c t i v e d i s t i l -
lation p l a n t .
a) mmm Steady s t a t e
b) === M o v e m e n t after a s t e p
c h a n g e of t h e v a p o r flow

m u s t b e stored in a tank only to p r o v i d e

the g l y c o l d u r i n g the s t a r t u p of t h e p l a n t . ASU-
T h e t o p p r o d u c t DE-| is t h e p u r e p r o p a n o l .
T h e r e m a i n i n g g l y c o l in t h e side s t r e a m is
r e m o v e d from t h e w a t e r in t h e r e c t i f y i n g
s e c t i o n K 0 2 and fed b a c k t o K 0 1 . T h e d i s t i l - F i g . 3. S i m p l i f i e d d y n a m i c model
 Control of an Extractive Distillation Plant
A l l p a r a m e t e r s of the s i m p l i f i e d m o d e l c a n
be deduced from the steady state plant pro-
p r o f i l e s . H e n c e s t a r t i n g from t h e p l a n t
d e s i g n c a l c u l a t i o n s the c o n t r o l s y s t e m can
be developed.
LEVEL CONTROL LOOPS
The s e l e c t i o n of b a s i c c o n t r o l loops like
level c o n t r o l and flow rate r a t i o c o n t r o l
is the f i r s t step t o w a r d s a c o n t r o l s y s t e m
for the p l a n t w e c o n s i d e r . A s t h e s e c o n t r o l
loops show a r a t h e r fast d y n a m i c b e h a v i o u r ,
they m a y later b e t r e a t e d as q u a s i s t a t i c .
H o w e v e r , it is of g r e a t i m p o r t a n c e to r e -
c o g n i z e t h a t t h e s e b a s i c c o n t r o l loops
strongly i n f l u e n c e the d y n a m i c b e h a v i o u r
of the r e g i o n s of m a s s t r a n s f e r .
A s an e x a m p l e , the r e s p o n s e of the p r o f i l e s
in the lower p a r t o f K01 is s h o w n in f i g . 4
( the i n s t r u m e n t a t i o n a t the c o n d e n s e r s is
given in f i g . 1 ) . If a step d i s t u r b a n c e
d e c r e a s e s t h e f l o w r a t e of the feed F ,
A
the p r o f i l e s a r e m o v i n g u p w a r d o r d o w n w a r d
d e p e n d i n g o n t h e i n p u t v a r i a b l e u s e d for
level c o n t r o l .
liq. mole fractions
Fig. 4. I n f l u e n c e of the level c o n t r o l
loop o n t h e d y n a m i c b e h a v i o u r .
T h e i n v e r s i o n of the g a i n d u e to the s t r u c -
tural variation of the basic control s y s -
tem can b e i n v e s t i g a t e d u s i n g the s i m p l i -
fied m o d e l ( f i g . 3 ) . If the level is c o n -
t r o l l e d by c h a n g i n g t h e f l o w rate o f the
b o t t o m p r o d u c t , the r e g i o n s of m a s s t r a n s -
fer a r e a f f e c t e d o n l y b y the liquid flow
in t h e c o l u m n . O n the o t h e r h a n d , c o n t r o l -
ling the liquid level w i t h the h e a t i n p u t
c a u s e s changes in the flow r a t e of the v a -
p o r w h i c h o v e r c o m p e n s a t e the i n f l u e n c e of
t h e r e d u c e d liquid f l o w .
T h e r e are of c o u r s e a l a r g e n u m b e r of p o s -
s i b l e c o n t r o l s t r u c t u r e s in a c o m p l e x d i s -
t i l l a t i o n p l a n t . S o m e of t h e m are u n f e a -
s i b l e b e c a u s e they for e x a m p l e act a g a i n s t
the o v e r a l l m a s s b a l a n c e of the p l a n t .
T h e s e s c h e m e s are d e t e c t e d a u t o m a t i c a l l y
d u r i n g the d e v e l o p m e n t of the s i m p l i f i e d
m o d e l . T h i s is b e c a u s e a m a t r i x w h i c h h a s
to b e i n v e r t e d b e c o m e s s i n g u l a r . A f e a s i b l e
b a s i c c o n t r o l s y s t e m for the e x t r a c t i v e
d i s t i l l a t i o n p l a n t is s h o w n in fig. 1 ) . T h e
c o n t r o l loops are s e l e c t e d m a i n l y to a l l o w
easy e x p e r i m e n t a l i m p l e m e n t a t i o n at a s m a l l
s c a l e p l a n t , w h i c h w a s b u i l t to t e s t t h e
c o n t r o l system p r e s e n t e d in the s u b s e q u e n t
chapter.
227
PROFILE POSITION CONTROL
C l o s i n g the l e v e l c o n t r o l loops and t r e a t -
ing t h e m q u a s i s t a t i c a l l o w s a f u r t h e r sim-
p l i f i c a t i o n o f the b l o c k d i a g r a m in fig. 3.
F u r t h e r m o r e , the p o s i t i o n a s 8 of the p r o -
file in K 0 2 is e a s i l y s t a b i l i z e d b y a t e m -
p e r a t u r e c o n t r o l that acts on the r e f l u x
r a t i o ( A R 2 / 4 D E 2 ) . T h e b l o c k d i a g r a m of t h e
r e m a i n i n g d y n a m i c s y s t e m , i.e. the m o v e m e n t
of the r e g i o n s of m a s s t r a n s f e r p r o p a n o l /
w a t e r A S ^ and w a t e r / g l y c o l A S 5 , is g i v e n in
fig. 5.
AS
reflux ratio .
control propanol/water
mass transfer
reboiler water/glycol
δΗ
Fig. 5. S i m p l i f i e d m o d e l for the profile
position control.
A f t e r i n t r o d u c i n g t h e s t a t e - , i n p u t - and
disturbance vectors T
AX = US ,AS )T
àu =
4 5 T
( S ,IH ) ( ! )
Δζ =
A
UZ ,Ax )
A Z A
the s i m p l i f i e d m o d e l m a y b e r e p r e s e n t e d in
the w e l l k n o w n s t a t e s p a c e n o t a t i o n . T h e
d y n a m i c b e h a v i o u r of t h e r e b o i l e r c a n e a s i -
ly b e i n c l u d e d in the s i m p l i f i e d m o d e l by
a u g m e n t i n g the s t a t e v e c t o r .
If t h e m o v e m e n t s of t h e p r o f i l e s are m e a s -
ured by t h e r m o c o u p l e s i n s t a l l e d at the d e -
s i r e d loci of t h e p r o f i l e s , t h e e f f e c t i v e
r a n g e is l i m i t e d d u e to s a t u r a t i o n ( see
fig. 2 ) . W i t h i n the l i m i t s , t h e m e a s u r e m e n t
v a r i a b l e s c h a n g e n e a r l y l i n e a r l y w i t h the
p o s i t i o n s of t h e f r o n t s . If t h e limits are
e x e e d e d , the o u t p u t of t h e m e a s u r i n g d e v i c e
c a n b e c o n s i d e r e d to b e c o n s t a n t .
This s i m p l i f i e d m o d e l w a s u s e d b y G i l l e s
and R e t z b a c h (1983) to d e s i g n a s t a t e o b -
s e r v e r and a state f e e d b a c k . By s i m u l a t i o n
s t u d i e s in w h i c h the c o n t r o l s y s t e m w a s a p -
p l i e d to t h e s t a g e - t o - s t a g e m o d e l it h a s
b e e n s h o w n , t h a t the c l o s e d loop is stable
and t h a t u n k n o w n feed c o m p o s i t i o n d i s t u r b -
ances may be estimated.
H o w e v e r , in p r a c t i c e t h e r e are o t h e r d i s -
t u r b a n c e s in a d d i t i o n to t h o s e in eq. ( 1 ) .
For example unknown heat losses, unexpected
e r r o r s in the c h a r a c t e r i s t i c s of the c o n -
t r o l v a l v e s e t c . . On the o t h e r h a n d , it is
c l e a r from t h e rank c o n d i t i o n s d e v e l o p e d in
c o n t r o l t h e o r y , t h a t it is i m p o s s i b l e to
compensate more than two independent dis-
t u r b a n c e s by t w o c o n t r o l i n p u t s . B e c a u s e
m o s t of the d i s t u r b a n c e s i n f l u e n c e the
steady s t a t e of t h e p l a n t , it is c o n v e n -
ient to introduce a new disturbance vector
&z, w h i c h s t a n d s for the o f f s e t s of the i n -
p u t v a r i a b l e s . C o n s t a n t v a l u e s of the e l e -
m e n t s of a r e a s s u m e d . The d y n a m i c m o d e l s
of t h e d i s t u r b a n c e s are t h e r e f o r e i n t e g r a -
t i n g e l e m e n t s . F i n a l l y , w e h a v e the s i m p l i -
fied m o d e l
228 Β. Retzbach

It w a s found d u r i n g e a r l y e x p e r i m e n t s t h a t
cUx/dt Β ( IU + 4Z ) it w a s n e a r l y i m p o s s i b l e to set the e x t r a c -

άγ
d^^/dt Ο
C ΔΧ
( 2 ) t i v e d i s t i l l a t i o n p l a n t into o p e r a t i o n
w i t h o u t p r o d u c i n g a lot of w a s t e m a t e r i a l .
T o r u n the e x p e r i m e n t s m o r e e f f i c i e n t an
A c o n t r o l s y s t e m h a s b e e n d e v e l o p e d to sta-
appropriate startup procedure was developed.
b i l i z e the p l a n t a n d t o c o m p e n s a t e t h e i n -
A d d i t i o n a l l y w e t r i e d to m e e t the p r o d u c t
f l u e n c e s of the d i s t u r b a n c e s . A c c o r d i n g to
s p e c i f i c a t i o n s d u r i n g s t a r t u p . This r e q u i r e -
the d e s i g n m e t h o d g i v e n b y B a l c h e n and c o -
m e n t is o f i m p o r t a n c e in a large scale
w o r k e r s (1973) an o b s e r v e r is u s e d to e s t i -
p l a n t i f , for e x a m p l e , the p r o d u c t s are r e -
m a t e the s t a t e v a r i a b l e s and the d i s t u r b a n -
cycled.
ces
S t a r t i n g a p l a n t is a d y n a m i c p r o c e s s . H e n -
d4x/dt = Β (Au+Δζ) + ce d y n a m i c s i m u l a t i o n is a v a l u a b l e tool to
dâï/dt z e ( 3 ) i n v e s t i g a t e the b e h a v i o u r of the columns
Ai = C Δχ , = Ay-ay during this period. The step-by-step pro-
c e d u r e g i v e n b e l o w w a s f i r s t d e v e l o p e d by
F r o m these e s t i m a t e s , the i n p u t variables
s i m u l a t i o n s t u d i e s and then i m p l e m e n t e d at
are c a l c u l a t e d
the s m a l l s c a l e p l a n t .
Au = -Κ - Δζ ( 4 ) B e c a u s e of the h i g h s e n s i t i v i t y of the p o -
χ ΑΧ s i t i o n s of the r e g i o n s of h i g h m a s s t r a n s -
T o g e t h e r w i t h the o b s e r v e r , eq. (4) m a y b e
fer r a t e s e a c h of t h e s e r e g i o n s m u s t b e
r e g a r d e d as a p r o p o r t i o n a l / i n t e g r a l s t a t e
s t a b i l i z e d b y a c o n t r o l loop i m m e d i a t e l y
feedback controller.
a f t e r it h a s b e e n d e t e c t e d . Due to the
T o tune the o b s e r v e r and t h e f e e d b a c k , the
s e l e c t i o n o f the m a t r i c e s Ε
matrices E , E
x z and K x are c o m p u t e d b y p o l e χ and E z, the
o b s e r v e r and the s t a t e feedback c a n b e
a s s i g n m e n t . M o r e e l e m e n t s are a v a i l a b l e
s p l i t i n t o t w o p a r t s w h i c h are c o n n e c t e d
than poles. The matrices E , E
x z are select-
ed d i a g o n a l l y t o s i m p l i f y t h e s t r u c t u r e of
only b y a few c o u p l i n g e l e m e n t s . It is
t h e r e f o r e easy to p u t the c o n t r o l system
the o b s e r v e r . T h e o f f - d i a g o n a l e l e m e n t s of into o p e r a t i o n s t e p w i s e d u r i n g the startup
Κ
χ are d e t e r m i n e d w i t h r e s p e c t to the p h y s -
ical s i g n i f i c a n c e of t h e i n p u t v a r i a b l e s :
procedure.
S i x steps are n e c e s s a r y to force the p l a n t
In the c l o s e d l o o p , the i n p u t A H is u s e d to
to its steady s t a t e :
c o n t r o l the o v e r a l l b a l a n c e of g l y c o l in
K01 ( A S 5 ) . W e e m p l o y the s i d e s t r e a m A S to Step I : P u r e p r o p a n o l is p u m p e d into the
c o n t r o l the b a l a n c e of w a t e r ( a w e i g h t e d r e b o i l e r of K01 and v a p o r i z e d . T h e c o l u m n
sum of A S 4 and A S 5 ) . is o p e r a t e d w i t h t o t a l r e f l u x to fill the
trays.
Step I I : G l y c o l is i n t r o d u c e d a t Z
IMPLEMENTATION
E and
p r o p a n o l is r e m o v e d at the top u s i n g a r a -
t h e r low r e f l u x r a t i o . A s G l y c o l is a h i g h
In o r d e r to i n v e s t i g a t e the n e w c o n c e p t of b o i l i n g c o m p o n e n t , it r e m a i n s in the liquid
e x t r a c t i v e d i s t i l l a t i o n and the c o n t r o l s y - p h a s e and is t r a n s p o r t e d d o w n w a r d (fig. 6 ) .
stem in m o r e d e t a i l , a p i l o t p l a n t w a s c o n - W i t h the m o l e f r a c t i o n of g l y c o l in the
s t r u c t e d . The t e c h n i c a l d a t a of it c a n b e b o t t o m the t e m p e r a t u r e r i s e s . If the b o t t o m
found in T a b l e 1 at the end of the p a p e r . c o n s i s t s of p u r e g l y c o l , a t e m p e r a t u r e
T h e two c o l u m n s are e q u i p p e d w i t h t h e r m o - f r o n t o c c u r s and m o v e s u p w a r d . T h e b o t t o m
c o u p l e s o n e a c h tray t o a l l o w t h e d e t e c t i o n product may now be recycled to Z .
of the e n t i r e t e m p e r a t u r e p r o f i l e s . H o w -
E
e v e r , only two tray t e m p e r a t u r e s are n e c e s -
sary for c o n t r o l .
The o b s e r v e r and the s t a t e f e e d b a c k (eqs. 3
and 4) w e r e i m p l e m e n t e d as p a r t of a r e a l
time computer program on a PDP11/34 process
c o m p u t e r . Only the d e v i a t i o n s from t h e
o p e r a t i n g p o i n t of the p l a n t are g i v e n b y
the f e e d b a c k ( 4 ) . T h e r e f o r e a s i m p l i f i e d
c a l c u l a t i o n of the s t e a d y s t a t e of t h e
p l a n t w a s i m p l e m e n t e d as w e l l . A l l q u a n t i -
t i e s , w h i c h can e a s i l y b e m e a s u r e d a t the
p l a n t ( i.e. flow r a t e s and t e m p e r a t u r e s )
are u s e d to d e t e r m i n e the o p e r a t i n g p o i n t s
of H,S from the o v e r a l l m a s s - and e n e r g y
balances. This steady state calculation 1. 350.
l i q . mole fractions temperature [K]
acts as a f e e d f o r w a r d c o n t r o l .
C o n v e n t i o n a l PID c o n t r o l l e r s in i n d u s t r y
m a y b e t u n e d o r o p t i m i z e d d u r i n g the o p e r a -
t i o n of t h e p l a n t b e c a u s e the s i g n i f i c a n c e
of the c o n t r o l l e r p a r a m e t e r s is e v i d e n t . At = 600 I s ] At = 60 (s)
T h i s is n o t true for the e l e m e n t s of the
matrices E , E
X Z and K x. T o p r o v i d e a m e t h o d
for the on line t u n i n g o f the o b s e r v e r and MW Δρ= .005 (bar]

the s t a t e f e e d b a c k , the p o l e - a s s i g n m e n t
p r o c e d u r e w a s i n c l u d e d in the r e a l t i m e sr-U
program. temperature I K ] temperature I K ]

S T A R T U P A N D O P E R A T I O N OF THE
PILOT PLANT F i g . 6. S t a r t u p step I I ; a) simulation,
b) e x p e r i m e n t .
T h e s t a r t u p of c o m p l e x p l a n t s is a d i f f i -
c u l t t a s k , e s p e c i a l l y if t h e u n i t o p e r a - S t e p I I I : B o i l u p and r e f l u x are switched to
t i o n s are very s e n s i t i v e t o d i s t u r b a n c e s . t h e i r o p e r a t i n g p o i n t s . T h i s forces the
 Control of an Extractive Distillation Plant 229

profiles in the upper part of K01 to their
steady states (fig. 7 ) . The temperature
front is still moving upward at this time.
As soon as the temperature at tray 37 is
near the desired set point, the observer
and the control loop for s 5 is turned on to
stabilize the movement by adjusting the
heat input Q v# the electric equivalent of H
Step IV: The feed Z A is now the propanol/
water mixture. The intermediate boiling
component water accumulates in the lower
part of K01 and a second temperature front
occurs (fig. 8 ) .
Step V: If the temperature at the desired
locus of this front (tray 27) starts to ri-
se, the side stream valve is opened by a
ramp function. Now the second part of the
observer/controller is switched on to regu-
late the overall balance of the water
(fig. 9 ) .

step

TT
[Kl

< 37»w

350.
40. - control A s
AUTOMATIC
S[%l control As#
j AUTOMATIC

time [ s ]

temperature [ Κ ]

Fig. 7. Startup step III; temperatures

profils and transients (experim.).
Fig. 9. Startup step V; experiment

X /I
glycol
ZA
^-swwater

480.

°- liq. mole fractions

T [Kl
Τ
® < 37»*-

350. . . . „ . 480.
temperature [ Κ ]
T
< 27>w

350.

80
time I s ]
temperature [ Κ ] *> ·

Fig. 8. Startup Step IV; a) simulation, Fig. 10. Closed loop experiment; step
b) experiment. distubance: increase of Z A by
10 % .
230 Β. Retzbach

Step VI: K 0 2 is operated with total reflux

until its trays are working properly. Then
REFERENCES
the bottom product SU2 can be recycled to
K01 .
Balchen, J. G., T. Endresen, M. Fjeld,
After step V I , the plant is in its usual
T. 0. Olsen (1973). Multivariable PID
steady state operating point. The control
estimation and control of systems
system presented makes the tight control
with biased disturbances. AUTOMATICA
of the positions of the temperature fronts
9, pp. 295-307.
possible. The efficiency of the control is
Gilles, E. D., Β. Retzbach, F. Silberberger
demonstrated in fig. 10. The close loop
(1980). Modeling, simulation and con-
response after a step change of the flow
trol of an extractive distillation
rate Z is shown. There was now feedfor-
A
ward control used during this experiment.
column.
In: Squires, R.G., G.V. Reklaitis
(Eds.). Computer applications to che-
mical engineering. ACS Symp. Series
CONCLUSION
no. 1 2 4 , pp. 481-492.
Gilles, E. D., Β. Retzbach (1983). Reduced
B a s e d on a s i m p l i f i e d m o d e l of t h e d y n a m i c
models and control of distillation
behaviour, linear state space methods have
columns with sharp temperature profi-
b e e n e m p l o y e d in t h e d e s i g n of a c o n t r o l
les. IEEE Trans. Autom. Control, AC-
s y s t e m for an e x t r a c t i v e d i s t i l l a t i o n
28, pp. 629-630.
p l a n t . The s i m p l i f i e d m o d e l is d e d u c e d f r o m
the b a l a n c e e q u a t i o n s of c o l u m n s e c t i o n s .
Retzbach, B.(1986). Mathematische Modelle
T h e r e f o r e o t h e r d i s t i l l a t i o n s c h e m e s in
von Destillationskolonnen zur Syn-
w h i c h sharp s e p a r a t i o n s o c c u r e c a n b e
thèse von Regelungskonzepten. Ph.D.
t r e a t e d in t h e s a m e m a n n e r . F r o m t h e p r a c -
dissertation. University of Stuttgart
t i c a l p o i n t of v i e w , n o t o n l y t h e d e s i g n o f Silberberger, F. (1979). Modellbildung. Si-
the c o n t r o l l e r is of i m p o r t a n c e . F u r t h e r - mulation und Regelung einer Extrak-
m o r e , t h e c a l c u l a t i o n of the o p e r a t i n g tiv-Destillationskolonne. Ph.D.
p o i n t s c a n b e u s e d as a f e e d f o r w a r d b r a n c h dissertation. University of Stuttgart
in t h e c o n t r o l l o o p . T o a l l o w t h e t u n i n g of
the c o n t r o l s y s t e m d u r i n g o p e r a t i o n o f t h e
p l a n t , a d e s i g n m e t h o d (for e x a m p l e p o l e - ACKNOWLEDGEMENT
assignment) should be implemented into the
r e a l t i m e c o m p u t e r p r o g r a m as w e l l . D u r i n g The project was sponsored by the Deutsche
s t a r t u p of t h e p l a n t , s e v e r a l s t r u c t u r a l Forschungsgemeinschaft (DFG). The chemi-
c h a n g e s of t h e c o n t r o l s y s t e m are n e c e s s a - cals used in the experiments were freely
ry. M a k i n g u s e of t h e d e s i g n f r e e d o m in t h e given by the HOECHST AG.
controller matrices can simplify these
s w i t c h i n g s to a g r e a t e x t e n t .

TABLE 1 Technical data of the pilot plant

material glass
trays bubble cap trays, one cap each,
efficiency greater than 70 %
K01: 40 trays, 10 cm diameter
Feed: glycol tray 12 (from top)
propanol/water tray 22
Sidestream: tray 32
K02: 10 trays, 5cm diameter
hight K01: 7 m, K02: 2 m
reboiler electrically heated 0 to 12 kW
level control by overflow pipe
condensers water cooled, reflux ratio ad-
justed by switching funnel
pressure K 0 1 : atmospheric, K02: 0.8 bar
feed electrically heated 0 to 3 kW
preheater with temperature control
pumps adjustable piston pumps
heatjacket 12 sectors independently con-
trolled
computers Z80 μ-processor for operation,
secondary control loops, super-
vision and data aquisition.
PDP11/34 process computer for
control, display of profiles
and data storage.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

COMPARISON OF THREE DIFFERENT

APPROACHES TO THE SEMIBATCH
DISTILLATION COLUMN CONTROL DESIGN
M . Atanasijevic, R. Karba, F. Bremsak, T. Recelj*, J. Golob**
and L. Fele***
University E. Kardelj, Faculty of Electrical Engineering, 61000 Ljubljana,
Trzaska 25, Yugoslavia
*KRKA Pharmaceutical, Novo mesto, 68000 Novo mesto, Cesta herojev 45
^University E. Kardelj, Department of Chemistry and Chemical Technology,
61000 Ljubljana, Hajdrihova 19, Yugoslavia
***Chemical institute Boris Kidric, 61000 Ljubljana, Hajdrihova 19, Yugoslavia
ABSTRACT
The w o r k d e a l s w i t h the c o m p a r i s o n o f t h r e e d i f f e r e n t a p p r o a c h e s t o the s e m i b a t c h d i s t i l -
lation column c o n t r o l d e s i g n . F i r s t s h o r t d e s c r i p t i o n of the p r o c e s s is g i v e n . F o r c o n t r o l
p u r p o s e s the r e d u c e d o r d e r m u l t i v a r i a b l e m o d e l w a s d e v e l o p e d . I n v e r s e N y q u i s t array techni'
q u e , p o l e a s s i g n m e n t a n d d e c o u p l i n g m e t h o d are m o s t f r e q u e n t l y u s e d for the d e s i g n o f m u l -
t i v a r i a b l e c o n t r o l . S o the s t a t e feedback d e c o u p l i n g c o n t r o l l e r in the c o m b i n a t i o n w i t h
o b s e r v e r and u n i v a r i a b l e P I c o n t r o l l e r s , o u t p u t feedback p o l e a s s i g n m e n t m u l t i v a r i a b l e PI
c o n t r o l l e r and r e g u l a t o r for a c h i e v i n g d i a g o n a l d o m i n a n c e (using INA t e c h n i q u e ) t o g e t h e r
w i t h u n i v a r i a b l e P I c o n t r o l l e r s w e r e d e s i g n e d . F o r e a c h o f this a p p r o a c h e s only the m a i n
d e s i g n i d e a s are b r i e f l y d i s c u s s e d . E f f i c a c y of d e v e l o p e d c o n t r o l l e r s for the r e a l p i l o t
p l a n t is s h o w n t h r o u g h h y b r i d s i m u l a t i o n w h e r e the s y s t e m w a s r e a l i z e d on a n a l o g - h y b r i d
c o m p u t e r E A I - 5 8 0 and c o n t r o l l e r s on the d i g i t a l c o m p u t e r P D P 1 1 / 3 4 . In the p r e s e n t p h a s e
of the study a l l d i s c u s s e d a p p r o a c h e s give a c c e p t a b l e r e s u l t s . F o r the f i n a l d e c i s i o n
about the m o s t s u i t a b l e c o n t r o l s t r a t e g y some m o r e m e a s u r e m e n t s on the p l a n t is n e e d e d .

1. INTRODUCTION
D i s t i l l a t i o n is a t y p i c a l e n e r g y - c o n s u m i n g
p r o c e s s . E v e n for a b i n a r y d i s t i l l a t i o n c o -
l u m n , the - u c c e s s f u l c o n t r o l of the com-
p o s i t i o n s of b o t h t o p and b o t t o m p r o d u c t
can y i e l d s u b s t a n t i a l p r o f i t r e s u l t i n g from
the p o t e n t i a l s a v i n g in u t i l i t y cost and of
course n o m a n s u p e r v i s i n g is n e e d e d in s u c h
cases.
F o r the s e p a r a t i o n o f c o m p o n e n t s on the b a -
sis of v o l a t i l i t y d i f f e r e n t k i n d s of d i s t i l -
lation / 1 , 2 , 3 / are u s e d in c h e m i c a l i n d u s -
t r y . In the cases of s m a l l p r o d u c t i o n and
o f t e n c h a n g i n g m i x t u r e the use o f b a t c h d i -
s t i l l a t i o n is very s u i t a b l e . In some cases
t h i s type of d i s t i l l a t i o n c a n b e i m p r o v e d
b y the use of s e m i b a t c h i n s t e a d of b a t c h r e -
c t i f i c a t i o n / 3 / . In d i s t i n c t i o n from b a t c h
d i s t i l l a t i o n h e r e the r e b o i l e r is c o n t i n u o -
usly fed w i t h the s a m e flow as the a c c u m u -
lator. The a d v a n t a g e o f such d i s t i l l a t i o n
is n o t only the e n l a r g e d m a s s of final p r o -
d u c t b u t a l s o c o n s t a n t l e v e l of m i x t u r e in
the r e b o i l e r w h i c h as a c o n s e q u e n c e e n s u -
res a p p r o x i m a t e l y c o n s t a n t v a p o r flow r a t e .
T h e w o r k deals w i t h m o d e l l i n g and c o n t r o l
of m e t h y l a l c o h o l - w a t e r s e m i b a t c h d i s t i l -
l a t i o n . First t h e o r e t i c a l n o n l i n e a r m o d e l
of the p r o c e s s w a s b u i l t / 3 / . It d e s c r i b e s
the w h o l e d y n a m i c a l b e h a v i o u r of the s y s t e m .
Equilibrium equations were derived directly
from the p r o c e s s . O f c o u r s e some a p p r o x i m a -
t i o n s m u s t b e t a k e n i n t o a c c o u n t . In s p i t e
of t h e m the a c c u r a c y of the m o d e l r e m a i n e d
s u f f i c i e n t . B e c a u s e of the r e l a t i v e l y s l o w
d y n a m i c s the l i n e a r i z a t i o n is j u s t i f i e d , of
course only in the n e a r n e i g h b o u r h o o d of the
chosen w o r k i n g p o i n t s . F o r c o n t r o l p u r p o s e s
the r e d u c e d o r d e r linear m o d e l w a s d e v e l o -
p e d b y the aid o f d i s c r e t e i n p u t - o u t p u t
identification using p r o g r a m p a c k a g e A N A
/4/.
M u l t i v a r i a b l e control d e s i g n r e p r e s e n t s a
w i d e l y e x t e n d e d set o f d i f f e r e n t m a t h e m a t i -
DCCR-P
23
c a l m e t h o d s a n d a p p r o a c h e s / 5 / . T h e y can b e
c l a s s i f i e d in m a n y d i f f e r e n t w a y s . In the
g e n e r a l f r a m e w o r k of the m u l t i v a r i a b l e s y s -
tems control, considerable attention has
b e e n given to the i d e a o f n o n - i n t e r a c t i o n .
F r o m this p o i n t o f v i e w , w e u s u a l l y t e n d
to compensate a system so that each input
affects only one o u t p u t , a n d p e r t u r b a t i o n
o f one o u t p u t a f f e c t s n o o t h e r o u t p u t . The
property of n o n - i n t e r a c t i o n is u s u a l l y
d e s i r e d due to the s i m p l i f i c a t i o n o f the
c o n t r o l l o o p . W h e n n o n - i n t e r a c t i o n is o b t a -
i n e d , the m u l t i v a r i a b l e s y s t e m is t r a n s f o r -
m e d i n t o a s e t of i n d e p e n d e n t s i n g l e - v a r i a -
b l e s u b s y s t e m s , w h i c h can t h e n b e t r e a t e d
by classical methods.
For the discussed plant we tried to design
the m o s t s i m p l e i . e . c o n s t a n t c o m p e n s a t o r
for d e c o u p l i n g / 6 / . B u t , for this type of
c o m p e n s a t o r i n f o r m a t i o n of a l l p r o c e s s s t a -
t e s w a s n e e d e d . O f c o u r s e , m e a s u r i n g of c o m -
p o s i t i o n s on a l l s t a g e s w o u l d n ' t be a good
s o l u t i o n . T h i s p r o b l e m w a s s o l v e d b y the
use of m i n i m a l o r d e r s t a t e o b s e r v e r / 5 / .
The d e s i r e d c l o s e d - l o o p b e h a v i o u r w a s t h e n
a c h i e v e d b y t w o u n i v a r i a b l e PI c o n t r o l l e r s .
F r o m s o c a l l e d interactive d e s i g n p r o c e d u r e s
p o l e a s s i g n m e n t m e t h o d ,as one o f the m o s t
i m p o r t a n t , w a s u s e d / 5 / . H e r e the c o m p l e t e
s y s t e m w i t h a l l its i n t e r a c t i o n s is t r e a t e d .
T h e use of t h i s a p p r o a c h o f t e n r e s u l t s in
more simple closed-loop systems. Similar
as in the p r e v i o u s e x a m p l e P I m u l t i v a r i a b l e
c o n t r o l l e r p r o v e d t o b e very s u i t a b l e / 7 / .
R o s e n b r o c k ' s i n v e r s e N y q u i s t array (INA)
m e t h o d is w e l l k n o w n a n d w i d e l y u s e d t e c h n i -
q u e for m u l t i v a r i a b l e c o n t r o l s y s t e m s d e -
sign / 8 / . T h e s i m p l i c i t y of d e s i g n e d c o n -
t r o l s c h e m e s , the p o s s i b i l i t y of e n g i n e -
e r i n g c o n s t r a i n t s i n c o r p o r a t i o n in the d e -
sign p r o c e d u r e a n d h i g h d e g r e e o f c o n t r o l -
lers r e l i a b i l i t y cause the a p p r o a c h to be
u s e d in m a n y p r a c t i c a l i n d u s t r i a l a p p l i c a -
t i o n s . The I N A m e t h o d use the concept o f
diagonal dominant transfer function m a t r i -
232 M. Atanasijevic et al.

ces and theorems o f G e r s h g o r i n a n d O s t r o w - o b t a i n e d n i n t h o r d e r linearized m o d e l can

ski t o e n s u r e the c o n d i t i o n s for the design b e s u c c e s s f u l l y r e d u c e d to fourth o r d e r
of i n d e p e n d e n t u n i v a r i a b l e s u b s y s t e m s c o n - u s i n g discrete i d e n t i f i c a t i o n / 4 / . The r e -
t r o l . Many a p p r o a c h e s e x i s t for a c h i e v i n g s u l t i n g s y s t e m can b e d e s c r i b e d in w e l l
d i a g o n a l d o m i n a n c e ( D D ) . In o u r w o r k w e s t u - k n o w n state s p a c e f o r m , w h e r e m a t r i c e s and
dy the p o s s i b i l i t y of u s i n g the data o b t a i - v e c t o r s in o u r case are as follows :
ned from o t h e r m u l t i v a r i a b l e s y s t e m s design
t e c h n i q u e s , i n our case the m e t h o d of d e c o u - -0.435 0.438 0.017 -0,019
p l i n g , to s y n t h e s i z e c o n t r o l l e r for a c h i e - -0.123 0.121 -0.009 0.010
v i n g D D not l o o s i n g any a d v a n t a g e of the -0.198 0.193 - 0 . 3 4 3 0.354
original Rosenbrock's approach. -0.102 - 0 . 0 9 7 -0.170 0.161
Efficacy o f d e v e l o p e d c o n t r o l l e r s for the 0.126 -0.097
real p i l o t p l a n t is shown t h r o u g h h y b r i d s i - 0.118 -0.080
Β = (1)
m u l a t i o n w h e r e the s y s t e m w a s r e a l i z e d on 0.292 -0.117
a n a l o g - h y b r i d c o m p u t e r E A I - 5 8 0 and c o n t r o l - 0.220 -0.093
lers on the d i g i t a l c o m p u t e r P D P 1 1 / 3 4 .
Γ 1 0 0 01
Q= Au
2. D E S C R I P T I O N OF T H E PROCESS Lo ο ι oj Uyi (t)J

The p r o c e s s u n d e r c o n s i d e r a t i o n as s h o w n in CONTROL DESIGN

F i g u r e 1. consists of r e b o i l e r , w h e r e a p -
p r o x i m a t e l y c o n s t a n t v a p o r flow rate is e n - 3.1. Design of c o m p e n s a t o r for d e c o u p l i n g ,
sured. A v a p o r from the r e b o i l e r goes up state o b s e r v e r and two univariable PI
through the c o l u m n , gives up p a r t of its controllers
e n e r g y at e a c h p l a t e and h e l p s the v a p o r i -
In o u r case the n u m b e r o f input and output
zation of m o r e v o l a t i l e c o m p o n e n t . The role
v a r i a b l e s of the s y s t e m is e q u a l , s o w e de-
of d i s t i l l a t i o n c o l u m n is to s e p a r a t e the
c i d e d t o use the d e s i g n p r o c e d u r e for con-
m i x t u r e s o that d i s t i l l a t e flow rate h a s
s t a n t d e c o u p l i n g c o m p e n s a t o r in time d o m a -
the p r e s c r i b e d c o n c e n t r a t i o n . In the c o n d e n -
in / 6 / .
ser such h e a t e x c h a n g i n g m u s t b e e n s u r e d that
all v a p o r w i l l c o n d e n s e , b u t there is n o n e - In this case the c o n t r o l law for the state
ed of c o o l i n g the d i s t i l l a t e , b e c a u s e p a r t feedback d e c o u p l i n g h a s the form:
of liquid is r e t u r n e d to the top of the c o -
lumn as the r e f l u x flow w i t h a t e m p e r a t u r e Au(t) = £ zlx(t) + g Aœ(t) (2)
w h i c h is n e a r t o the b o i l i n g p o i n t . The plant O b s e r v e r is d e f i n e d w i t h / 5 / :
is a r r a n g e d for r e c t i f i c a t i o n a t c o n s t a n t
q u a l i t y of d i s t i l l a t e , s o there is a l s o a ζ
r e f l u x d i s t r i b u t o r w h i c h r e t u r n s one p a r t (3)
_ e z + 5 Y _ + s u
of liquid from the c o n d e n s e r t o the column 0 0
x*= g z + g y_
and the o t h e r p a r t is drawn as a top product. 2 1
w h e r e £ , g , D , g , § , g , and g
o o 2 are c o n s t a n t m a -
t r i c e s of a p p r o p r i a t e d i m e n s i o n s , w h i c h in
condenser o u r case h a v e the following v a l u e s :
.949 2.348 -1.671 - 2 . 6 6 4 ^
.184 7.533 -1.982 - 3 . 6 4 2 .
(4)
,008 0.0061 Γ 0 -0 .0625 ]
,019 0.008J= L-l - 0 . 5 J

4L· r U
co n/r o1ler|
listillate
So" Κ 00228
01168
0 0 1 5 7 0 .00189 "
00645 0 . 0 1 0 3 3 .
1 0 0 0
1.074 0 4 . 388 1.562|
= [Si * S ]
2 0 1 0 0
column ,0 .095 0. 886 99.075 - 1 6 . 2 6 3
ί 3,
The p a r a m e t e r s of t w o uni v a r i a b l e PI control-
lers are a l l e q u a l ( 3 3 . 3 3 ) . The closed-loop
s y s t e m is s h o w n in Figure 2 .

univariable
PI c o n t r o l l e r

υ
decoupling
SYSTE M γ
compensator " fef.

Figure 1. S c h e m a t i c r e p r e s e n t a t i o n of an
η-stage d i s t i l l a t i o n column
IL observe r

The g o v e r n i n g e q u a t i o n s for the d y n a m i c m o -
d e l can be d e r i v e d b y a p p l i c a t i o n of m a t e -
r i a l b a l a n c e and e q u i l i b r i u m e q u a t i o n s u n -
der some u s u a l a s s u m p t i o n s . The s i m u l a t i o n
o f n o n l i n e a r m o d e l s h o w e d t h a t in spite o f
this a s s u m p t i o n s the a c c u r a c y of the m o d e l
remains sufficient.
B e c a u s e of relatively s l o w d y n a m i c s o f the
p r o c e s s the l i n e a r i z a t i o n is j u s t i f i e d , of
c o u r s e only in the n e a r n e i g h b o u r h o o d o f
the chosen w o r k i n g p o i n t s . T h r o u g h the s t u -
dy of the plant i t w a s e s t a b l i s h e d t h a t
univariable
PI c o n t r o l l e r
e
2
Figure 2. C l o s e d loop s y s t e m w i t h state d e -
c o u p l i n g c o m p e n s a t o r , observer and
t w o u n i v a r i a b l e PI controllers
C l o s e d - l o o p b e h a v i o u r o b t a i n e d through h y -
brid simulation using discretized decoup-
ling c o m p e n s a t o r , o b s e r v e r and two univa-
raible PI c o n t r o l l e r s is s h o w n in Figure
3.
 The Semibatch Distillation Column Control Design 233

neering in Ljubljana (YU) ) which serves for

— Ai
the computer aided design of seven diffe-
rent types of pole and zero assignment mul-
1.4
'N
.;.ι !Τ
ta tivariable controllers,the PI regulator for
α ο : ::

- —
b) discussed system was designed.

— <M
il-.

C.
(
3 *

• :1 :
ιω.
] :; : f

r ........ / r -

:.! ,
Figure 4. Output feedback pole assignment
•
controller

Multivariable PI regulator designed by pole

: assignment method can shift 2JL-m-l poles to
the specified locations. Shifting a larger
number of poles requires a calculation of
- — nearest possible locations which can be at-
• tained. The order of closed-loop system is
... because of I-part in direct path enlarged
to n+JL, while P-part of regulator is in the
feedback loop (Figure 5 ) . In this case the
1
control law has the form:
— j .—

:
\ < ----- - - — •
u = g Z + Q j (v-y_)dt = - [ K : Q ] j (6)

tea I
α ί> " !
7=>
-Γ
d) —

α : : ί Usui

1j
Figure 5. Block diagram of the PI pole as-
signment control
1
- — The matrices of the designed controller we-
i re designed to be

I 1
Γ 1.381 5.874] "2.297 1.5901{ ],
I ! j •· L-1.609 5.842J - [ ;.0.384 1.471J
Figure 3. Closed-loop responses of the sys- where closed-loop poles have the following
tem excited by initial conditions values: -0.365, -0.066, -0.370, -0.300 +
and step disturbance on first in- + j 0.200.
put at t=50s. Sampling time is ls
a) first output signal, b) first Closed-loop behaviour obtained through hy-
control signal, c) second output brid simulation using discretized multiva-
signal, d) second control signal riable pole assignment PI controllers is
shown in Figure 6.

3.2. Pole assignment control design 3.3. Control design by INA method
For the model given in eq.(1) such control- The multivariable INA method was developed
ler was desired that the resulting closed- by Rosenbrock / 8 / . It extend the classical
loop system has a certain number of its po- univariable Nyquist control design methods
les on the previously prescribed positions. to multivariable systems using the concept
The approach bases on the equivalence of of diagonal dominance (DD) together with
closed-loop characteristic polynomial of Gershgorin and Ostrowski theorems to esta-
multivariable system and equivalent system blish the conditions for the control desigi
with one input and multiple outputs. This of independent univariable subsystems. The
results in compensator matrice which is dya- criterion of diagonal dominance is much lo-
dic i.e. has rang equal to one /!/ (Figure oser than the request for strict diagonali-
4 ) . If g is a product of column and row ve- zation of transfer function matrix which is
ctor: too extreme for general use. It is to say
Κ that we should only like to restrict the
2 k (5) influence of other loops to a given loop
where cj is (m χ 1) vector and k is (1 χ a) to be appropriately small.
vector, multivariable system (A,B,g) is re-
duced to equivalent system with one input A large amount of work was devoted to the
(£,i2,g). problem of diagonal dominance achievement
since it is prerequisite for any succes-
By the aid of interactive program package sful INA design effort. Original Rosenbrock's
POLASS / 3 / (which is installed on PDP 11/34 approach to the problem allows the complete
minicomputer in Faculty of electrical engi- freedom in designing the DD compensator and
234 M. Atanasijevic et al.

dels in the d e s i g n as w e l l as its e x p e r i e n -

ces and i n t u i t i o n . In a d d i t i o n , he h a s good

a) i n s i g h t i n t o the c l o s e d - l o o p s y s t e m b e h a v i -
our t h r o u g h o u t the design process w h i l e the
d e s i g n e d c o n t r o l l e r s are easily r e a l i z a b l e .
In o u r a p p r o a c h w e w a n t e d to combine the
a d v a n t a g e s of R o s e n b r o c k ' s m e t h o d and i n f o r -
m a t i o n a b o u t the s h a p e of control s i g n a l v e -
c t o r u for the d e c o u p l e d system o b t a i n e d by
the a i d of the c o r r e s p o n d i n g computer aided
p r o c e d u r e for d e c o u p l i n g (chapter 3 . 1 ) . The

b)
Tït c o m b i n a t i o n o f b o t h m e t h o d s e n l a r g e s their
a d v a n t a g e s and r e d u c e s t h e i r d e f i c i e n c i e s .
The d a t a a b o u t the s i g n a l s w h i c h act in the
sense o f i n t e r a c t i o n s c o m p e n s a t i o n facilita-
tes the d e v e l o p m e n t o f DD c o n t r o l l e r (struc-
ture a n d p a r a m e t e r s ) w h i c h i s , using R o s e n -
b r o c k ' s a p p r o a c h , o b t a i n e d p r a c t i c a l l y with
"cut and t r y " m e t h o d . O f course w e don't
w a n t to design c o n t r o l l e r w h i c h w o u l d p r o -
duce s t r i c t l y the s a m e s i g n a l s as obtained
w i t h d e c o u p l i n g a l g o r i t h m , b u t w e take i n t o
a c c o u n t only t h e m a i n d y n a m i c p r o p e r t i e s
and steady s t a t e . The p r o c e d u r e of DD c o n -
troller development considering mentioned
d a t a can b e r e a l i z e d e i t h e r b y the aid of
s i m u l a t i o n or u s i n g a p p r o p r i a t e i d e n t i f i c a -
tion algorithm.
The s t r u c t u r e o f the complete c o n t r o l s c h e -
me for the d i s c u s s e d p l a n t is shown in F i -
gure 7.

dl Figure 7. C o n t r o l s c h e m e for d i s c u s s e d p l a n t
u s i n g INA m e t h o d , ( w = w = 0 , s y s -
1 2
t e m w a s e x c i t e d t h r o u g h the d i s -
t u r b a n c e inputs u - a n d / o r u
and i n i t i a l c o n d i t i o n s )
m2

The D D c o n t r o l l e r w a s found t o b e

Γ-0.9867 1.266
U (8)
L-2.4000 1 706
T h e r e s p o n s e s of s o c o n t r o l l e d s y s t e m are
s h o w n in F i g u r e 8 t o g e t h e r w i t h correspon-
ding Gershgorin bands.

11
F i g u r e 6. C l o s e d - l o o p r e s p o n s e s of the s y s -
tem e x c i t e d by i n i t i a l conditions
and s t e p d i s t u r b a n c e on f i r s t i n -
p u t at t = 5 0 s . S a m p l i n g time is
0.5s. a) first o u t p u t s i g n a l , b )
first c o n t r o l s i g n a l , c) s e c o n d
o u t p u t s i g n a l , d) s e c o n d c o n t r o l
signal
gives only some s u g g e s t i o n s for its s u c c e s -
sive g e n e r a t i o n . The a i m of the a p p r o a c h is
to d e v e l o p DD c o n t r o l l e r w h i c h is as s i m p l e
as p o s s i b l e . The I N A m e t h o d allows the d e -
s i g n e r to use s i m p l e t r a n s f e r function m o -
 The Semibatch Distillation Column Control Design 235

Figure 8. a) Closed loop system time res-

ponses w = 1 , W 2 = 0 (1-y-L and
2 - y ) , bf INA curve for the ele-
2
ment (1,1) with Gershgorin bands,
c) closed loop system time res-
ponses w = 0 , w = l (1-yjL and 2 -
1 2
- y ) , d) INA curve for the ele-
0
ment (2,2) with Gershgorin bands

From Figure 8 it can be seen that in spite

of nonnegligible interactions (a-curve 2
and c-curve l)the DD is achieved (Gershgo-
rin bands exclude the origin). The univari-
able PI controllers for each independent
loop were then designed with the constants:
k , = 10, k = l , k and k^=l. Closed-loop
i:L
behaviour obtained through Hybrid simulati-
on using discretized DD controller and two
univariable PI controllers is shown in Fi-
gure 9. . . Figure 9. Closed loop responses of the sys-
: tem excited by initial conditi-
i-J. ' , .._.ia) J -1, ons and step disturbance on first
input at st
t=50s. Sampling time st
nd b) 1 is
Is. a) 1 output signal,
nd
control signal, c) 2 output si-
UsJ
gnal, d) 2 control signal
236 M. Atanasijevic et al.
The Ostrowski bands for the case in Figure
9 is shown in Figure 10.
Figure 10. Ostrowski bands for the case in
Figure 9, a) element (1,1), b)
element (2,2)
5. CONCLUSIONS
In the work three most frequently used mul-
tivariable controller design methods were
applicated on the pilot semibatch distil-
lation column. All results of hybrid simu-
lation indicate that the proposed approa-
ches can be successfully used for on-line
control of discussed plant. Besides the
given results also some other situations
were simulated, like: presence of system
parameter changes, presence of measurement
noise, time-delayed input signals and limi-
tation of control signals to appropriate
operative area. In spite of this additional
tests the final decision about the most su-
itable control strategy can not be made in
this phase of investigations. Some additi-
onal experiments on the plant are therefo-
re initiated.
6. REFERENCES
1. S.Domenech et M. Enjalbaert: Modèle ma-
thématique d'une colonne de rectifica-
tion discontinue - I. établissement du
modèle, Chemical Engineering Science,
1974, Vol. 29, pp. 1519-1528
2. Claude Foulard, Guy Bornard: Identifica-
tion and System Parameter Estimation of
Distillation Processes, Identification
rd Estimation, Part 1,
and System Parameter
Preprints of the 3 IFAC Symposium,
the Hague (Delft, the Netherlands , 12-
15 June, 1973
3. M.Atanasijevid, R.Karba, F.Bremsak: Se-
mibatch Distillation Modelling and Con-
trol Design, Proceedings, 3. Symposium
Simulationstechnik, pp. 46 4-46 8, Bad
Munster am St.-Ebernburg, September 1985
4. M.Sega, S.Strmônik, R.Karba, D.Matko:
Control System Treatment by t h
Program
Package ANA, Proceedings, 1 1 IMACS
World Congress on System Simulation and
Scientific Computation, Vol. 4, pp. 95-
9 8, Oslo, August 19 85
5. Rajnikant V. Pateland Neil Munro: Multi-
variable System Theory and Design, Per-
gamon Press, Oxford, 19 82
6. Peter L. Falb and William A. Wolovich:
Decoupling in the Design and Synthesis
of Multivariable Control Systems, IEEE
Transactions on Automatic Control, Vol.
IAC-12, No. 6, Dec. 1967
7. H.Seraji: Design of Proportional-puls-
Integral Controllers for Multivariable
Systems, Int. J. Control, Vol. 29, No.l,
1979
8. Rosenbrock H.H., Design of multivariable
control systems using the inverse Nyqu-
ist array, Proc. IEE, Vol. 116, No. 11,
1969, 1929-1936
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

NON-INTERACTING CONTROL SYSTEM

DESIGN OF A DISTILLATION COLUMN BY
THE PARTIAL MODEL MATCHING
T. Takamatsu*, I. Hashimoto*, Y. Togari** and Y. Hashimoto**
^Department of Chemical Engineering, Kyoto University, Kyoto, Japan
**Department of Systems Engineering, Nagoya Institute of Technology,
Nagoya, Japan

Abstract. A non-interacting dual composition control scheme is developed for a dis-

tillation column by using the partial model matching technique and it is realized with
four conventional PI controllers. The partial model matching technique developed by
Kitamori enables to design decoupler and main controller at once» with partial
knowledge of the process to be controlled. Also, the features of the closed loop
response are easily adjustable by choosing a suitable reference model. It is shown
through simulation calculations that I-P non-interacting controller system designed by
Kitamori's method is effective as a dual composition control of a distillation column.

Keywords. Process control; control system synthesis; PID control; partial model
matching; decoupling.

INTRODUCTION
where
Distillation is a typical energy consuming G (s) is a transfer function of Plant, Whole
process. The successful control of the composi- control system or Reference model.
tions of both top and bottom products can yield
substantial profit resulting from the potential
saving in utility cost. Some serious problems are Secondly, the control system parameters are deter-
often experienced when a feedback control of both mined so as to coincide the coefficients of power
product compositions is attempted. One of the series of s in the closed loop transfer function
principal difficulties is control loop with those of a desired reference model, from 0-th
interaction. In order to overcome such problems order to the higher order of s.
an application of a multivariable control scheme
seems indispensable. In this paper, the partial It is not an easy task to readjust the controllers
model matching technique developed by Kitamori which are obtained by applying a state space con-
(1980.a) is applied to dual composition control of trol theory. However, with partial model
a distillation column. matching, controller can be readjusted easily by
changing the reference model and number of
parameters of the controller.
THE MERIT OF PARTIAL MODEL MATCHING

Many control theories based on state space Partial model matching methods have further
model [e.g. LQ( Linear Quadratic) regulator, pole advantages, e.g., performance of controller can be
assignment, exact model matching, decoupling con- preestimated more visually than in LQ regulator or
trol by Falb & Wolovich, etc.] have been proposed. pole assignment. Also, for the selection of a
However the implementation of such technics in in- desired reference model, a guide has been
dustrial field has not always been so successful. developed.

One of the main reasons is that identification of A design formula for I-PD type decoupled systems
the state equation is a hard task especially in is proposed by Kitamori in 1980 (1980.a) and PID
the field of process control. Therefore, a design type decoupled systems are also proposed ( 1980.b).
method will be more advantageous if it does not In other non-interacting control methods such as
need complete knowledge about the process INA (Inverse Nyquist Array method), GGBM
dynamics. A partial model matching method (Generalized Gershgorin pseudo Band Method) and
proposed by Kitamori (1979a) is based mainly on SRD (Sequential Return Difference method), it is
the low frequency characteristics of the process, necessary to design decouplers first and then to
which can be identified more accurately than the design main controllers. With Kitamori's method,
high frequency characteristics. the decoupler and the main controller can be ob-
tained simultaneously. As a result, design proce-
Firstly, in this method, the dynamic characteris- dure becomes much simpler and the response curves
tics of the three elements: (i) the plant to be of the controlled system can also be perceived
controlled, (ii) the closed loop system and (iii) very easily from the reference model.
the desired reference model, are all to be
expressed in the following form. In this paper, a dual composition controller,
2 i.e., a non-interacting control system which has
Λ ,sv=b0+bis+...+b.s»
G — —
.
m
exp(-T
.
Ls =C o + c i S + c a s + " (1) the structure shown in Fig. 1, is designed by ap-
a +ais+...+a s» * plying the partial model matching method.
0 B
237
 238 T. Takamatsu et al.
i

t
diagonal matrices M j · Σ .
XDS
I-P
Design Algorithm of Control Systems
I-P
The structure of I-PD type control system is shown
•H O
in Fig, 2.
I-P -ρ ϋ
CO
R(s)+ E(s) Κ U (s) Y (s)
« I-P K? » Process
XB S e s

Fe+Fis+F s2+—
Fig. 1. I-P non-interacting control system 2
for distillation control

DESIGN PROCEDURE OF I-PD TYPE NON-INTERACTING CON- Fig. 2. I-PD control system
TROLLER

I-PD has the same property as PID with respect to

Expression of the Purpose of Control the control against disturbances. I-PD controller
is as familiar to operators as PID controller is.
The specifications for a control system are given Therefore, operators can easily cope with abnormal
as follows: situations. I-PD controller, when compared with
(CI) Each decoupled subsystem has zero PID, has the advantage of being able to avoid
steady-state error. pole-zero cancellation. With regard to set point
(C2) Each decoupled subsystem has adequate change, I-PD controller can avoid excessive fluc-
damping characteristics. tuations in manipulating variables, because it
<C3) Each decoupled subsystem has the shortest does not utilize the derivatives of reference
rise-time on satisfying (CI) and (C2). inputs. Furthermore, the structure of this con-
<C4) The remnant coupling as a result of imperfect troller can be equivalent to that of the state
decoupling among the subsystems has to be as feed back controller, if higher order derivatives
small as possible. of the outputs, up to the order of the plant to be
In a model matching method, these specifications controlled, can be fed back.
are represented by using a reference model. In
Kitamori's method, the reference model for m in-
puts m outputs system is expressed in following Y <s)=H (s)-i-U (s) (9)
form. U (s)=V ( s ) - F (s)-Y (s) (10)
V (s)=K ·Ε ( s ) / s (H)
E (s)=R ( s ) - Y (s) (12)
M(e).Y(s)=R(8) (2)
where 2 where
H (s)=H o+H is+H 2 S * + . -
M (β)=Μ ol + M ιΣ s+M 2 Σ s*-h.. (3) F (s)=F o+F îs+F 2S*+...
M o=M i=I (4)
M i=diag[cr u α κ ·» α n ] (5) The closed loop system can be expressed as follows:
Σ =diag[o- 1 (X 2 ··· <7 M ] (6) 1
{ I +s-K - ( H (s)+F (s)) } -Y (s)=R (s) (13)
Y (s) is the output m dim. vector
R (s) is the reference input m dim. vector
where Κ is assumed to be nonsingular.
M ( s ) is a polynomial matrix
A set of a i 3 (0=0,1,···) expresses damping of Controllers are designed in such a way that each
controlled response between the i-th input and coefficient matrix in Eq. (13) coincides with the
the j-th output corresponding coefficient matrix of reference
ο* ι is a measure of rise-time of controlled model from the 0-th order and upward as follows:
response between the i-th input and the
i-th output since it is the first moment of the
0-th order I = Μ 0 (14)
impulse response.
1-th order K'MHo+Fo) = Μ ιΣ 2 (15)
2-th order K'MHi+Fi) = Μ 2Σ (16)
The closed loop system is also expressed as
follows:
When the controller consists of a pre-compensator
Κ and feedback compensators F o,F 1,··· , F p , Σ
W ( s ) - Y <s)=R (s) (7) can be utilized so that diagonal elements of the
where
(p+3)-th coefficient matrix of the closed loop
W (s)=W ol + W is+W 2S2+-. (8) p t î
system W ρ + 3 can coincide with those of the
reference model Μ Ρ + Ϊ Σ . Σ is, therefore,
The controller is designed in such a way that each
calculated by the following equation.
coefficient matrix of the closed loop system, A 1
( W o , W i , W 2 > - ) i coincides with the corresponding
2 matrix of the reference model(Μ ο,Μ ιΣ = M p 3 - M p 2-diag ( H P + I" !! ) (17)
coefficient + + P+
2
,Μ 2 Σ >··· ) from the 0-th order and upward. When the rise-time of each output i.e., Σ , is
chosen by a designer, coefficient matrices, W t,
Specification (CI) is satisfied by setting W o = M 0 = are matched to M i l ' from the 0-th to the (p+1)-
I · th order. In the both cases, controller
Specification (C2) is satisfied by choosing a parameters can be calculated by the following
suitable set of M j( j=2,3,... ). equations.
Specification (C3) is expressed as the problem of + 1
f inding the smallest positive o" i. Κ = Η + ιΣ -<P
+2 >1Mp+ 2 -
Ρ (18)
Kitamori's method differs from other partial model F j = K M ά.ιΣ J - H j (j=0,l,2,- (19)
,P)
matching methods in that the parameter G \ is
automatically calculated. p=0 means I-P controller.
Specification (C4) is satisfied by matching coef- p=l means I-PD controller.
ficient matrices of closed loop system, W j to
 Non-interacting Control System Design 239

Process Modeling Table 1.1 Parameters of Kitamori models

It is necessary to express process dynamics in the order (Χ ο (Χ ι a 2 a 3 a « a s

form as shown in Eq. (9). There are many ap- arbi. 1 1 0.5 0.15 0.03 0.003
proaches to get this representation. If the
process dynamics is being expressed by a transfer
function matrix, let the all elements have a com- Table 1.2 Parameters of Binomial Models
mon denominator as follows.
order a ο α ι a 2 a a a 4 a s
Y ( S) = 2i£Lu<s) (20) 2 1 1 0.25
- - -
a(s) 3 1 1 0.333 0.037
- -
where B (s) is a polynomial matrix
a(s) is a polynomial
4
5
1
1
1
1
0.375
0.400
0.063
0.080
0.0039
0.0080
-
0.00032
6 1 1 0.417 0.093 0.0116 0.00077
Even if the transfer function matrix includes 0.00123
7 1 1 0.429 0.102 0.0146
dead-time lag, transformation to Eq. (20) can be
realized by using Tayler expansion of exponential
function. The representation of Eq. (9) is ob-
tained from Eq.(20) by using Maclaurin expansion.

When the process dynamics is expressed by a dis-

crete model, z-transformation of the model is
first performed. Then, by utilizing the following
relationship,
2
z=exp(r s)=l+-jys + - | ^ s +-~ps3 + ... (21)

where τ is the sampling period

the transfer function matrix in s-domain can be

obtained.

Selection of Reference Model Fig. 3. Unit step response of the closed loop
systems
Not only the closed loop response but also the
robust stability is determined by the reference
model to be utilized. Therefore, it is important SIMULATED DISTILLATION COLUMN
to choose a suitable reference model in order to
design a controller which can satisfy the given The effectiveness of the non-interacting control
specifications. For the sake of simplicity, let system shown in Fig. 1 is demonstrated by
us consider a single input, single output system. simulation. The column used in the simulation
A reference model used in Kitamori's method has study is fairly small. The design values are
parameters β and a j(.3=0,1,··· ) to express rise- given in Table 2.1. The dynamic model is based on
time and damping of a closed loop response. the standard assumptions: vapor holdup is negli-
gible in comparison with liquid holdup; each plate
22 acts as an ideal stage; perfect mixing on each
Y(s) = •R (s) (22)
a o+CR I < R s + Α < R s +..« plate and in the reflux accumulator and reboiler;
2 adiabatic column operation; perfect level control
As the values of α j i those shown in Table 1.1 are
in the accumulator and reboiler; perfect pressure
recommended. The α j give about 10% damping, ir-
control and zero pressure drop across the plates;
respective of the number of system parameters
and energy dynamics are much faster than material
matched to reference model's. If overshoot is
balance dynamics. Dynamic material balances and a
permissible, these values in Table 1.1 is
steady state energy balance are written for each
recommendable. Response curves without overshoot
stage in the column and the resulting equations
can be always be obtained, if it is possible to
are solved numerically.
coincide the transfer function of the closed loop
system with a binomial mode. As a reference model
where both the magnitude of overshoot and the Table 2.1 Specification of the distillation column
rise-time can be adjustable by a single parameter,
the following model is being proposed (Shigemasa System Methanol--Water
et. al 1983). Feed Flow Rate 1000.0 (mol/hr)
Feed Composition 50.0 (mol%)
α )·α B J + λ -a K J q-value of Feed 1.0 <-) [Liquid]
=j(1-λ (j=0,l,2,..- ) (26)
Distillate Composition 95.0 (mol%)
where a B J are parameters of binomial model. Bottom Product Composition 5.0 (mol%)
a Kj are parameters of Kitamori's model, Operating Pressure 1.0 (atm abs)
λ is response shape factor. Number of Trays 10 (-)
Feed Tray Location 7 (from top)
In Fig. 3, the closed loop response of a 4-th or- Tray Efficiency 100 (%)
der plant is shown, where I-PD controller is Diameter of Column 12.0 (cm)
Height of Weir 2.0 (cm)
designed by using the reference model given Eq.
Depth of Column's Bottom 5.0 (cm)
3
(26).
Accumulator Holdup 1000.0
3
(cm )
From this figure, it is easily perceived that the
rise-time of response can be decreased by making Reboiler Holdup 2000.0 (cm )
response shape factor, λ , larger, especially for Reflux Ratio 0.710 (-)
λ >0.4, overshoot emerges in the response curve.
Even when the plant has a large dead time, a
desirable response curve can be realized by using MODELING FOR CONTROLLER DESIGN
a reference model in which a proper dead time is
included. In order to obtain the process dynamics in the
240 T. Takamatsu et al.

form of Eq. (9), we calculated the step responses Table 3.1 Parameters of the Model of Ea. (9)
of the compositions of top and bottom products to
changes in the two inputs (reflux flow rate, (l,l)element (l,2)element (2,l)element (2,2)element
boilup rate). The results are shown in Fig. 4. Η ο 3.3476E+3 -8.9473E+2 2.7931E+3 -1.4515E+3
Η ι 1.1006E+3 -1.7041E+2 7.1052E+2 -4.5714E+2
Η 2 2.6013E+1 -1.3194E+1 1.5904E+1 -1.3407E+1
Η 3 1.1495E-1 -1.7751E-2 7.4013E-2 -4.7619E-2
R e f l u x Flo w Rat e chang e + 5 ( m o l / h r ) Η 4 1.3548E-3 -6.8721E-4 8.2835E-4 -6.9828E-4
Η 5 3.5827E-6 -5.5472E-7 2.3129E-6 -1.4881E-6
unit of H j (j=0^5) are (mol-hr-i-%-i)

The design parameters of I-P non-interacting con-

trol system were determined by using Eqs. ( 1 7 ) ^
(19) and Kitamori model given in Table 1.1. The
obtained parameter values are shown in Table 3.2.
B o i l u p Rat e chang e + 5 ( m o l ' h r )

Table 3.2 Parameters of I-P Non-Interacting

Controller
Fig. 4(a). Step response of distillate composition (1,1)element (l,2)element (2,l)element (2,2)element
Σ 0.0801 - -0.0469
Κ 3.4304E+5 -1.5463E+5 2.2146E+5 4.1481E+5
F ο 2.4132E+4 -6.3649E+3 1.4947E+4 1
-1.8023E+4
R e f l u x Flo w Rat e + 5 ( m o l / h r ) unit of Σ is hr,unit of Κ & F ο are (rool'hr- ·%-*)

In order to evaluate the effectiveness of the I-P

non-interacting control system, the performance of
this control system was compared with those of two
other control scheme:
Bo i lup Rat e + 5 ( m o M " i r ) (i) 2-loops I-P controller system shown in Fig. 5
and
(ii) conventional decoupling control system shown
50 10 0 15 0 20 0 25 0 in Fig. 6.
T i me (m i n)

Fig. 4(b). Step response of bottom composition

Xd
I-P
From the data sampled from the response curves
shown in Fig. 4, we obtained the following auto-
regressive model by using the least square method.
•Η Ο
-Ρ ϋ
03
Y ( k + l ) = Z A j - Y ( k - j ) + B o-U(k) (27)
j=0 I-P
XBse t Xb

where Y (k) and U (k) mean the k-th sampled data

of outputs and inputs
Fig. 5 2-loops I-P controller system
Table 2.2 Parameters of AR model

(1 ,1)element (1,2)element (2,1)element (2,2)element

Ao 1.2981 -0.2545 Xose t + Lt
0.1104 1.0249 I-P
Α ι -0.3329 0.2461 -0.0643 -0.0833

c
Β ο 3.140E-5 -2.516E-5 4.070E-5 -6.530E-5
Α ο,Α ι are dimension less,
unit of Β ο is (mol%»mol-i «hr)
sampling period is 1.5 min.

I-P
Then, the following equation was derived from Eq. XBse t Vb
(27) by using z-transformation.

q+i 1 Fig. 6 conventional decoupling control system

Σ G j«zi" Y (z) = U (z) (28)
j=0 The control parameters of the 2-loops I-P control-
ler system were tuned by Kitamori's method. The
By substituting Eq. (21) into Eq. (28), Eq. (9) is resultant parameter values are shown in Table 4.
obtained. The resultant parameters are shown in
Table 3.1.
Table 4 Parameters of 2 loops I-P Controller
I-P configuration is chosen for a non-interacting
control system of a distillation column. The 1st loop 2nd loop
structure of the controlled system is illustrated Σ 0.05974 0.1169
in Figs. 1 and 2. Y(s) ,U(s) ,R(s) and E(s) are all Κ 4.8505E+5 -5.5215E+4
2 dimensional vectors, Fo.Fi,··· and Κ are 2x2 F ο 2.7345E+4 -5.7454E+3
matrices, and F i , F 2 , » » * are all zero matrices. unit of Σ is hr.unit of Κ & F ο are (mol-hr-i·%-ι)
 Non-interacting Control System Design 241

As a cross controller in the conventional decou-

pling control system, a constant matrix shown in
Table 5.1 is utilized. All the elements of this
matrix are calculated from the steady state gain
of the process. Two I-P controllers in the main
control loops were designed independently of each
other by using Kitamori*s design procedure. The
resultant parameter values are shown in Table 5.2.

Table 5.1 Parameters of Cross Controller I-P nonintercting control

95.2
2 loops controller
(l,l)element (l,2)element (2 l)element (2,2)element ; cross controller + m a i n controller
C ιj| 3.3476E+4
y
1 1 2.7931E+3 1 1 1
-8.9473E+2 -1.4515E+3
30
unit of C υ are mol'hr- **-
Τ I ME (M Ί Η)

Table 5.2 Parameters of I-P Controller for Fig 7.1(a) Response of X D to set point change
Decoupled System XD=0.95->0.96(mol%) & X =0.05->0.06
B
1st loop 2nd loop 6.5

Σ 0.0725 0.2630
Κ 1.4036E+2 1.4252E+1
F ο 9.1702E+0 2.7411E+0 6.Θ

unit of Σ is hr, Κ & F ο are dimension less / /

f /

Simulation calculations are performed in order to

compare the performances of the three different
control systems; /
i) I-P non-interacting control system X :I-P nonintercting control
+ '2 loops controller
ii) 2 loops I-P controller system Δ :cross controller + m a i n controller
iii) conventional decoupling control system.
10 20 30
Fig. 7.1 shows the step responses to set points
change ;X =95%-* 96% and X =5%-* 6%. The rise-time
D B
of the controlled response of X D is largest when
Fig 7.1(b) Response of X B to set point change
XD=0.95->0.96(mol%) & XB=0.05->0.06
I-P non-interacting control system is used.
However, the differences in response of X D are in-
significant among the three control systems. On Table 6.1 Parameters of I-P Non-Interacting
the other hand, with respect to the response curve Controller
of X B , I-P non-interact ing control system results
in the smallest rise-time and the differences (1,1)element (1,2)element (2,1)element (2,2)element
among the three controlled responses are sig- Σ
Κ
0.08
3.4394E+5 -5.3254E+5
- 0.08 -
nificantly large. Judging from the above, I-P 2.2204E+5 1.4286E+5
non-interacting control system seems to have the F ο 2.4167E+4 -3.3656E+3 1.4970E+4 -9.9770E+3
most preferable characteristics for the set point unit of Σ is hr,unit of Κ & F ο are (mol«hr-*·%-*)
change.

In the above simulation, the calculated values of Table 6.2 Parameters of 2 loops I-P Controller
Σ are different among the three control systems.
1st loop 2nd loop
In order to analyze how effectively the two con-
trol loops are decoupled by these three control Σ 0.08 0.08
systems, a simulation calculation was carried out Κ 2.7046E+5 -1.1784E+4
by using the same Σ =diag(0.08,0.08) for all three F ο 2.0006E+4 -8.7195E+3
unit of Σ is hr.unit of Κ & F ο are (mol-hr-»·%-»)
control systems. The controlled responses of X D
and X B for the three control systems are shown in
Figs. 8.1—8.2. Fig. 8.1 shows the response
curves of X D and X B * when the set point of X D
alone is changed. I-P non-interacting control Table 6.3 Parameters of I-P Controller for
system can bring X D up to the newly assigned set Decoupled System
point in the shortest time, and, at the same time
can regulate X B at the set point with the smallest 1st loop 2nd loop
deviation. These simulation results show that I-P Σ 0.08 0.08
non-interacting control system seems to be most Κ 1.1530E+2 1.5441E+2
effective in reducing interaction between the two F o l 8.2171E+0 1.1318E+1
control loops. unit of Σ is hr, Κ & F ο are dimension less

Fig. 8.2 shows the response curves of X D and X B

with respect to the change of feed flow rate,
F=1000-* 1100mol/hr. The deviations of X and X
D
from the assigned set points are restrained most
B CONCLUSIONS
effectively, when I-P non-interacting control sys- In this paper, the partial model matching tech-
tem is employed. From the foregoing discussion, nique developed by Kitamori is applied to the dual
it can be said that I-P non-interacting control
composition control of a distillation column.
system is the most preferable among the three con-
This method is advantageous in that the decoupler
trol systems with respect to transient charac-
and the main controller can be developed
teristics to set point change as well as distur-
simultaneously. Based on this design method, I-P
bance rejection property.
non-interacting control system shown in Fig. 1 is
designed for a dual composition control. Through
simulation calculations, it is analyzed that this
 T. Takamatsu et al.

I-P non-interacting control system is the most

96.0
advantageous, when compared with a conventional
decoupling control system and two independent
loops controller system, with respect to both
properties of attaining desired responses to
changes in set point and achieving sufficient
95.6
feedback properties.

Nomenclature
X :I-P nonintercting control
+ :2 loops controller F :Liquid feed flow rate
Δ :cross controller + main controller
F j.Feed back matrix of the i-th order derivative
10
of controlled variables
H j'matrix in Eq. (2.1) which expresses process
dynamics
Fig 8.1(a) Response of X to set point change Κ : Multiplying matrix for integration of errors
D
X =0.95->0.96(mol%) Lt :Reflux flow rate
D η : order of denominator polynomial of reference
models
U :Vector of manipulated variable;
X : I - P nonintercting control
Vector of reflux flow rate and burlap rate
5.4 + :2 loops controller
Δ :cross controller + m a i n controller in distillation control
Vb :Boilup rate
5.3
X D .Composition of Distillate
X B :Composition of Bottom Product
5.2
Y :Vector of controlled variable;
5.1
Vector of distillate composition and
bottom product composition
5.0i
in distillation control
a j'.Parameters of reference model expressing
4.9 damping of controlled response
o* :Parameters of reference model expressing
4.8 rise-time of controlled response
Σ : Diagonal matrix consisting of o"

Fig 8.1(b) Response of X to set point change

B
X =0.95->0.96(mol%)
REFERENCES
D
Graham, D. & R.C.Lathrop (1953). The Synthesis of
"Optimum" Transient Response:Criteria and
Standard Forms. AIEE, Nov.. 273-288
Kitamori, T. (1979a). A method of control system
design based upon partial knowledge about
controlled processes. Trans. Soc. Instrum. &
Control Eng. (Japan), 15, 549-555
Kitamori, T. (1980a). A design method for I-PD
type decoupled control systems based upon
partial knowledge about controlled
processes. Trans. Soc. Instrum. & Control
s Eng. (Japan), 16, 112-117
94.9 ''',
+ — #
+
X :I-P nonintercting control
:2 loops controller
Kitamori, T. (1980b). A design method for PID
Δ :cross controller + m a i n controller type decoupled control systems based upon
-1 I J J L i partial knowledge about controlled
15 processes. Trans. Soc. Instrum. & Control
Time(min ) Eng. (Japan), 16, 139-140
Fig 8.2(a) Response of X to disturbance Shigemasa,T., Y.Takagi, Y.Ichikawa and T.Kitamori
D
F=1000->1100(mol/hr) (1983). A Practical Reference Model for
Control System Design. Trans. Soc. Instrum.
& Control Eng. (Japan), 19, 7, 74-76
Takamatsu,Τ., I.Hashimoto, M. Iwasaki and M. Nakaiwa
X : I - P nonintercting control (1981). An Experimental Study of the Multi-
\ \ + :2 loops controller variable Control of a Distillation Column,
: cross controller + main controller
Proceedings of the 8th IFAC world congress,
Kyoto, 107.3
\\ \

Fig 8.2(b) Response of X B to disturbance

F=1000->1100(mol/hr)
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

RELATIONS BETWEEN STEADY-STATE

PROPERTIES OF DISTILLATION CONTROL
SYSTEM STRUCTURES
Kurt-Erik Haggblom and Κ. V. Waller
Department of Chemical Engineering, Abo Akademi, 20500 Abo, Finland

Abstract. Recent research in dual composition control of distillation has, to a large

extent, treated various control system structures. Special attention has been paid to
configurations where the manipulative variables have been chosen so as to minimize
interaction between the control loops.
In the present paper consistency relationships between the process gains of single
control structures are derived. It is also shown how the process gains of arbitrary
control structures can be calculated from the process gains and steady-state data of
just one suitable base structure. The equations obtained are exact relations based
on fundamental first principles. By means of these relations it is possible easily
to investigate (e.g. calculate relative gains (RGA) for) a large number of control
system structures.
A structure where the reflux flow and the boilup, or functions of them, are the compo-
sition manipulators is a suitable base structure. The reason for this is that a model
for such a scheme contains more information than does e.g. a model, such as it usually
is given, for a material balance scheme. The latter one cannot therefore be utilized
to the same extent for variable transformation purposes as the former.
The consistency relationships also provide useful constraints for parameter estimation.

Keywords. Distillation control; modelling; system analysis; control system synthesis;

multivariable control systems; parameter estimation.

INTRODUCTION a binary (or one that can be treated as binary)

Recent work in distillation control has largely
been concerned with control system structures.
Emphasis has been put on finding structures that
reduce, minimize, or eliminate the interaction
between the composition control loops. Then the
manipulators have normally been various flow
ratios.
In this paper consistency relationships between
the process gains of the models for some commonly
studied distillation control structures are de-
rived. It is also shown how transformations to
other sets of manipulators can be performed by
means of these relationships and steady-state
data. Thus, it is possible to obtain the process
gains related to other control structures if the
process gains and some steady-state data related
to just one base structure are known. However, it
is also shown that only control structures involv-
ing the reflux flow and boilup (or variables di-
rectly affecting them) as composition manipulators
can serve as base structures for arbitrary vari-
able transformations.
two-product distillation process. It is, however,
fairly straightforward to generalize the treatment
to include more complex distillation processes
with an arbitrary number of components, feeds, and
sidestreams (Haggblom, 1986).
CONSISTENCY RELATIONS
BETWEEN PROCESS GAINS
In this section consistency relationships between
the steady-state gains of models for some commonly
studied distillation control structures are de-
rived. As it is sufficient to consider linear
(or linearized) steady-state models to obtain
relations between steady-state gains (which, of
course, can be part of a dynamic model), only such
models are used in the sequel.
Consider the distillation process in Fig. 1. Con-
tinuous operation requires control of the column
inventory. This is usually accomplished by con-
trolling the holdup in the reboiler and the reflux
drum. Hence, the material balances

Obviously these derived relations can be utilized F = D + Β (1a)

to, e.g.,
Fz = Dy + Bx (1b)
ο test consistency and improve accuracy in
parameter estimation must hold in steady-state. Linearization of
Eq. (1) when F and ζ are constant (only relations
ο investigate a large number of control system
between the product compositions and the product
structures, e.g. by means of relative gain
flows are considered here) gives
(RGA) analysis, with a minimum of experimen-
tal or simulation work
Δϋ + ΔΒ = 0
ο remove unnecessary or incorrect assumptions
y Δϋ + D Ay + χ ΔΒ + Β Δχ = 0

For simplicity, the relations are here derived for or

243
244 Kurt-Erik Haggblom and Κ. V. Waller

ventional control scheme (L,V). As shown later in

the paper, these relations are needed when making
transformations with the (L,V) structure as base
structure.

Actually, the distillation model for the (L,V)

structure consists of both Eqs (3) and (4), where
the gains are related by the consistency relations
(5). It is, thus, sufficient to give the four
gains in Eq. (3) together with steady-state data
completely to define the steady-state model when
product compositions are the controlled variables.

Where steady-state models for the conventional

scheme are given in the literature, the four gains
in Eq. (3) are reported, but steady-state data are
often missing. Even if the steady-state data are
available, the usual approach in the literature is
to use
K =K = 1 r (ΔΔ0
oΐ=
Β DL
K
" BL
K =
" " - ' V=const.)
= R = 1 Δ( Δ0
O LΝ
(6a)
=6 C
B O N S
DV " RBV
V ' - ) ( )

These relations are, however, valid under very

Fig. 1. A two-product distillation process special circumstances only and probably never in
real columns. Numerical illustrations are given
later in the paper.

When the four gains in Eq. (3) are estimated, e.g.

AD = _ - JL-ty from experimental or simulated data, they should
ΔΒ = - -Η-Δχ (2)
be determined in such a way that also Eq. (4),
where Δ denotes a (small) steady-state deviation with the gains calculated by Eq. (5), fits well
from the nominal steady-state defined by D, B, with the data in question. Or more precisely, the
y, χ (and Eq. 1.). task is to determine the eight gains in Eqs (3)
and (4) subject to the constraints (5).

Control Structures Involving both L and V
as Composition Manipulators
Consider the "energy balance" or "conventional"
control scheme where the top composition y and the
bottoms composition χ are controlled by the reflux
flow L and the boilup V, or by any two independent
variables directly affecting the reflux flow and
the boilup (V could, e.g., stand for heat input
to the reboiler in all equations that follow).
A steady-state model for this structure is
Ay = K ^ A L + Κ ^ Δ ν
Δχ = AL + AV
If some secondary variables, such as two tempera-
tures, are controlled instead of the product com-
positions, the gains in Eq. (4) cannot be calcu-
lated by Eq. (5), because the gains of the model
replacing Eq. (3) are no longer related to compo-
sition. In this case the gains in Eq. (4) must
be determined in the same way as the four gains
between the controlled and manipulated variables.
After that, however, the variable transformations
presented below can be used even if the variables
controlled are not product compositions. If only
the four gains between the variables controlled
and manipulated are given, the model is simply not
(3a) complete enough for arbitrary variable transform-
ations even if the steady-state data are known.
(3b) This, of course, applies to all control struc-
tures.

The gains Κ in the model may have been obtained, As a second example of dual-composition control
e.g., experimentally, through simulation, or, if schemes involving both L and V as manipulators,
an analytical model is available, through calcu- consider a scheme suggested by Ryskamp (1980), in
lation of partial derivatives. which the composition manipulators are D/(L+D)
and V. The steady-state model for this scheme is
Combination of Eqs (2) and (3) gives
AR
K AL + AVK Ay <R (7a)
AD (4a)
DL DV Ax =KAR + AVK
K AL + AVK xR xV
(7b)
ΔΒ (4b)
BL BV
where we here denote
where
D K + B K R = D/(L+D) (8)
V
K
DL BL
;
L X L
(5a) Combination of Eqs (2) and (7) gives
L L
K DK
yVW
+ BK „
xV AD = K j A R
R + KJSV
AV (9a)
(5b)
DV " BV y -χ
ΔΒ = K g A R
R + KJj„AV (9b)
Equations (4) and (5) are fundamental steady-state
relations between the composition manipulators L,
V and the inventory manipulators D, Β in the con- where
 Steady-state Properties of Distillation Control System Structures 245

D K + B K nipulators D and V should be included in the model

yR xR
DR BR
+Β Κ
Κ ν χν
Κ -K
DV BV
Equations (9) and (10) define fundamental steady-
state relations between the composition manipu-
lators R,V and the dependent variables D,B in the
(R,V) scheme. The relations are completely anal-
ogous with the relations (4) and (5) for the (L,V)
scheme.
The examples above can be generalized. For every
structure where two independent functions of L and
V (and possibly D or B) are used as composition
manipulators, relations analogous with Eqs (4) and
(9), with the consistency relations (5) and (10),
respectively, apply. These relations have impli-
cations analogous with those discussed above
(Haggblom and Waller, 1986).
for the material balance scheme.
(10a)
Another control structure discussed in the litera-
ture (McAvoy, 1983; Shinskey, 1981; Tsogas and
McAvoy, 1981) is a (D/V,V) scheme, usually cred-
(10b)
ited to Ryskamp (1980). The scheme is usually
intended to be implemented as a (D/(L+D),V)
scheme, but it is sometimes analyzed as a (D/V,V)
scheme (e.g. Tsogas and McAvoy, 1981). The reason
for this is that the relation V = L + D is
assumed to hold. (This assumption is very general
in the literature.)
Ryskamp's (D/(L+D),V) scheme earlier considered
in this paper and the (D/V,V) scheme considered
below are, however, generally not equivalent.
The model related to the (D/V,V) scheme is
(14a)
(14b)

Control Structures Not Involving both L and V

as Composition Manipulators Now

rAD
Consider a structure where the product composi- rAV (15)
tions are controlled by manipulation of D and V,
a so called material balance scheme (Shinskey,
1984). The steady-state relations between the Combining Eqs (2), (14) and (15) and considering
product compositions and the manipulators are that V and D in effect are independent, give the
given by following relations between the gains in Eq. (14).

K AD + K ( x)y
(16a)
Δy yD ïv
K AV
(11a) xD/V " Β yD/V " Β '
Δχ =Κ
ΔΟ +ΚΔν ,D/V DK„D/V D y-x
(11b) (16b)
ϋθ χν xV Β yV Β V

Combination of Eqs (2) and (11) gives Thus, only two of the gains in the (D/V,V) model
are independent and the model does not give any
D K+ B K DK+ B K information about the effect of L on the system.
yD xD- 5v xV This is an analogy between the (D/V,V) and (D,V)
1 + AD + — ^ — AV 0 (12) schemes on the one hand, and the (D/(L+D),V) and
(L,V) schemes on the other.
Because .D and V are independent variables in the
material balance scheme, Eq. (12) must be valid
regardless of the values of AD and AV. This means RELATIONS BETWEEN GAINS IN
that DIFFERENT CONTROL SCHEMES
V
In the previous section consistency relations
1 +
DK
n + ΚxD = 0 between the open-loop gains of some control
D y -Dχ schemes were derived, such as Eq. (5) for the
conventional control scheme and Eq. (10) for
D K ., + B K Ryskamp's control scheme.
xVV
yv = 0
If in a control scheme only the four independent
gains related to the composition dynamics are
given, which is the usual situation, consistency
J2 ^ ΣζΛ (13a) relations like Eqs (5) and (10) give the four
"xD Β yD Β gains between the inventory manipulators and the
K K
D „D composition manipulators. When these eight gains
(13b) and the proper steady-state data are known, the
xV B yv
process gains for arbitrary control schemes can be
Equation (13) expresses consistency relationships calculated. An example illustrates.
between the gains in the model (11) for the ma-
terial balance scheme (D,V). Consider the model (3) for the conventional con-
trol scheme (L,V). Elimination of AL by Eq. (4a)
Thus, there are only two independent process gains gives
in the model (11). The four gains in the model
should therefore be determined subject to the
constraints (13). AV (17a)
It should be noted that the model (11) does not DL DL
give any information about the effect of L on the V L
product compositions, not even when the material K L KKV
balances are taken into account. This means that Ax = ^ A D ( K xV ° ' (17b)
the model as such cannot be transformed to a model +yV L
K; )AV

containing L as a composition manipulator. To DL DL

overcome this the relation between L and the ma-
246 Kurt-Erik Haggblom and Κ. V. Waller

which now actually is a model for the material Column C (z = 0.50, y = 0.99, x = 0.01, L/D = 4.51 )
balance scheme (D,V). Comparing Eqs (17) and (.11)
and taking Eq. (5) into account, the following Γ Ay] _ Γ 0.471
7 94 -0.495] TAL]
relations between the gains in the material bal-
ance scheme and the conventional scheme are ob- [ΔΧ J L °- -0.832 J |_AV J
tained. [Ay] Γ -0.685 0.0085] Γ AD]
[Δχ J
=L-0-505 -0.0094 J [AVJ
(y-χ) κ
yL (18a)
yD McAvoy (1983) has later improved the accuracy of
DL D L K xL+
; i
the gains in the material balance scheme. The
X B ( KK K K ) differences are not large enough to affect the
K =K DV.V iL xV" xL ^V result of the study here. We use the gains above
yV yv
D VK + BV K (18b) since they are expressed in units more suitable
^DL yL xL for our analysis.

ν Numerical values for D and Β are not given by

K (y-χ) κ xL
McAvoy and Weischedel (1981) but can be estimated
xL
xD = V
D VK + BV K (18c)
as follows. Since the separations are symmetrical
DL yL xL D = B. Equation (13a) can be used to estimate D
(and B). The result is that D and Β are in the
D K K K K ) range of 0.75 to 0.82 (expressed in units consist-
^D
=Κ ( yLVxV xLx W (18d) ent with the given data). In the sequel the value
xV xV
D K . +| D = Β = 0.75 is used. The conclusions would be
^DL xL
yL the same if another numerical value in the range
Naturally, the relations (18) satisfy the consist- were used.
ency relations (13).
The process gains and the steady-state data given
In a similar manner the relations between the above can be used to check the often used assump-
gains of Ryskamp's scheme (D/(L+D),V) and the tion of Eq. (6), i.e.,
conventional scheme (L,V) can be derived. In
fact, the gains of any control scheme can be de- Κ -K -1 (or AD = -AL for V = const.)
DL BL
rived from the gains of any control scheme invol- and
ing both L and V as composition manipulators
(Haggblom and Waller, 1986). KL
DV = - K g V= -1 (or AD = AV for L = const.).

Control schemes not involving both L and V as Equation (5) gives the following results for the
composition manipulators can be used to derive the three columns.
gains of control schemes involving the same compo-
sition manipulators as the initial scheme (note Column A Column Β Column C
that D and Β are always interchangeable). For
example, the gains of the (D/V,V) scheme previous- -0.81 -0.71 -0.93
DL BL
ly discussed can be obtained from the gains of the K K
material balance scheme (D,V). 0.81 0.71 1.02
DV " BV

Summing up, exact relations between process gains

in different control schemes can be derived, pro-
vided that certain compatibility requirements are
fulfilled. As an example, such relations between
the conventional control scheme (L,V) and the
material balance scheme (D,V) are given in Eq.
(18). However, it is not uncommon that reported
gains do not satisfy relations like (18). One
example from the literature illustrates the point.
As seen, for more realistic distillation column
models, like the ones studied by McAvoy and
Weischedel (1981), the assumption AD = -AL (for
V = constant) and AD = AV (for L = constant)
can be considerably in error. This also provides
the answer to a question by Hashimoto (1981) in
the discussion after the presentation of the paper
in question.

Another interesting question is the validity of

APPLICATION TO PUBLISHED DATA
McAvoy and Weischedel (1981) use the following
gains in a comparison between energy balance and
material balance control for three columns. The
gains were obtained by numerical simulation of
nonlinear process models.
Column A (z = 0.50, y = 0.93, x = 0.07, L/D = 1.30)
general recommendations concerning suitable con-
trol system structures for distillation which are
based on the assumptions of Eq. (6) (e.g.
Shinskey, 1984).
The consistency of the gains given above for the
conventional and the material balance schemes can
also be checked by means of Eq. (18). If the
gains in the conventional scheme are (arbitrarily)

m•[
assumed to be correct, the following gains in the
0.580 -0.450 material balance scheme can be calculated.
0.350 -0.480 ][»] Column A
[Ay] Γ-0.7104 02 0.140] [ A D ]
= 'Ay] Γ-0.715423 0.130] T A D ]
[Δχ J |_-°· -0.130j|_AVJ =
_Δχ_ [_-°- -0.130j|_AVJ

Column Β (z = 0.50, y = 0.98, x = 0.02, L/D = 1.71)

Column Β
0.562 -0.516] [AL"
lay] . Γ-0.794
][S]
0.344 -0.394J |_AV_ 0.048
[Δχ J _ - 0 .486 -0.048

[3:
805
465
0.055
-0.055 Its]
 Steady-state Properties of Distillation Control System Structures
Column C
Ay] _ Γ-0.504
820 0.017 1 Γ AD Ί
AxJ |_-°- -O.OUJLAVJ
As seen, the differences between the gains ob-
tained by simulation and those calculated by means
of the relations (18) are considerable. As could
be expected, the differences are largest for Col-
umn C, which is the most nonlinear of the ones
studied. The values calculated for and
for Column C differ from the ones simulated by
about 100 %. Also the differences between the
simulated and calculated gains K^p and for
Column C are strikingly large. In fact, if vari-
able pairing for material balance control of Col-
umn C were made on the basis of relative gain
analysis (RGA), the calculated process gains would
suggest another pairing than the simulated ones
(McAvoy, 1983; Shinskey, 1984). The relative
gains of the model calculated and the one obtained
by simulation are namely 0.39 and 0.60, respect-
ively.
The problem of minimizing the number of simula-
tions required numerically to evaluate the el-
ements of the relative gain matrix for various
control schemes was recently treated by Mijares
and co-workers (1985). Fundamental consistency
relations like Eqs (5) and (13) and relations con-
necting the gains in different control schemes,
like Eq. (18), have not, however, been recognized
by Mijares and co-workers. This means that they
have estimated e.g. the gains and directly
by simulation without considering the constraints
due to the relations mentioned. This has led to
small inconsistencies in the process gains and the
relative gains obtained. (Mijares and co-workers
do not report the gains obtained, but the consist-
ency of the gains can be checked from the relative
gains they report for different control schemes -
see Haggblom and Waller (1986).)
These examples show the importance of obtaining
correct process gains if variable transformations
are to be made. Process gains should at least
satisfy the consistency relations, e.g. Eq. (5)
for the conventional scheme and Eq. (13) for the
material balance scheme. If the process gains of
two different control schemes are estimated from
experimental or simulated data, the gains should
also satisfy the relations existing between the
gains of two schemes, i.e. Eq. (18) in the case
considered here. It is therefore, in general,
necessary to determine the process gains by some
optimization procedure, where relations like Eqs
(5), (13) and (18) are used as constraints.
CONCLUSIONS
Consistency relations between process gains in
individual distillation control schemes and rela-
tions connecting gains in different control
schemes have been derived for some control schemes
used in binary two-product distillation. In the
same manner such relations can be derived for
arbitrary control schemes. Compact matrix nota-
tion makes the relations simple and perspicuous.
Due to space limitations of this paper the treat-
ment is presented elsewhere (Haggblom and Waller,
1986).
The relations between the gains in different con-
trol schemes can be used to calculate the gains in
other control schemes when the process gains and
proper steady-state data are available for one
DCCR-Q
247
scheme. Thus, it is possible to make transform-
ations from one set of manipulators to other sets.
A restriction is, however, that if models relating
only the product compositions to the manipulators
are used, only control schemes involving both the
reflux flow and boilup (or functions of them) as
composition manipulators can be used for arbitrary
variable transformations. This means that the
gains in, e.g., the conventional control scheme
(L,V) cannot be calculated from the gains in the
material balance scheme (D,V). The reason is that
a model for the (D,V) scheme does not contain
information about the effect of L on the system.
This information cannot be derived from the ma-
terial balances either. Mathematically this turns
up as a singular matrix which cannot be inverted.
The general treatment also results in expressions
for relative gains (RGA) for various control
schemes as well as in exact relations between the
relative gains for different control schemes.
Also variables that make the control loops com-
pletely decoupled can be readily derived. These
results are given in Haggblom and Waller (1986).
The equations presented in this paper are exact
analytical relations based on fundamental first
principles. These relations can be used, e.g.,
to increase the accuracy in parameter estimation.
The relations are then used as constraints in the
parameter estimation procedure.
ACKNOWLEDGEMENT
Financial support from the Waldemar von Frenckell
Foundation (to K-E H) and to the Neste Foundation
is gratefully acknowledged.
NOTATION
Β bottoms flow
D distillate flow
F feed flow
Κ process gain, K ^ = f f |
c = .c o n s t
L reflux flow
R D/U+D)
V boilup
χ bottoms composition
y top composition
ζ feed composition
Δ denotes deviation from nominal steady-state
REFERENCES
Haggblom, K-E. (1986). In preparation.
Haggblom, K-E., and K.V. Waller (1986). In
preparation.
Hashimoto, Y. (1981). Discussion remark. Proc.
8th IFAC World Congress, 5_, 2752.
McAvoy, T.J. (1983). Interaction Analysis, ISA.
McAvoy, T.J., and K. Weischedel (1981). Â dynamic
comparison of material balance versus conven-
tional control of distillation columns.
Proc. 8th IFAC World Congress, 5, 2773-2778.
Mijares, G., C D . Holland, R. McDaniel, C.R.
Dollar, and S.E. Gallun (1985). Analysis and
evaluation of the relative gains for non-
linear systems. Comp. & Chem. Eng., 9, 61-70.
Ryskamp, C.J. (1980). New strategy improves dual
composition column control. Hydrocarbon
Processing, 59 (6), 51-59.
Shinskey, F.G. (1981). Predict distillation
column response using relative gains.
Hydrocarbon Processing, 60 (5), 196-200.
Shinskey, F.G. (1984). Distillation Control,
2nd ed. McGraw-Hill, New York.
Tsogas, Α., and T.J. McAvoy (1981). Dynamic simu-
lation of a non-linear dual composition con-
trol scheme. Proc. 2nd World Congress of
Chemical Engineering, 5_, 365-369.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

INDEPENDENT SINGLE-INPUT-SINGLE
OUTPUT CONTROL SCHEMES FOR DUAL
COMPOSITION CONTROL OF BINARY
DISTILLATION COLUMNS
f
A. C. Kridiotis* and Christos Georgakis**
^Department of Chemical Engineering, Massachusetts Institute of Technology,
Cambridge, MA 02139, USA
**Chemical Process Modeling and Control Research Center, Lehigh University,
Bethlehem, PA 18106, USA

ABSTRACT The steady-state interaction measure of dual composition control

schemes for binary distillation columns is examined under the assumptions of constant
relative volatility and constant molar overflow. An exhaustive enumeration is made of
the control schemes that use column flows or their ratios as manipulated variables and
product compositions as the controlled variables. Using a redundancy criterion, it is
shown that 34 classes of dual composition control schemes have independent values for
the interaction measure. This number reduces to 17 when the feed is a saturated liquiJ
or gas. For high purity columns, the interaction measure of the above 34 control
schemes approaches the values for the (L,V), (D,V), and CL.B) control schemes. In this
case none of the control schemes has a desirable interaction measure.

KEYWORDS. Distillation Control, Interaction Measure, Multivarable Control,

Process Control, Decoupling.

1977; Jafarey et al., 1979). These approxi-

INTRODUCTION
The problem of efficient distillation control has
attracted much interest in the last decade, par-
ticularly as its potential for reducing the ener-
gy requirements of distillation columns became
apparent. Since better control would allow for
narrower safety margins on product composition,
wasteful over purification would no longer be
necessary in order to ensure that products do not
fall below specifications. The energy savings
from such an improvement would be especially
substantial in high purity distillation, where
the energy requirements increase rapidly with
increasing product purity. Unfortunately, it is
here that the problem of interaction, between the
two control loops, that regulate the product
purities, is most severe.
This interactions is quantified, to a large e x -
tent, through the steady state Relative Cain
Array CRGA) (Bristol (1966)). In the case of
binary distillation control the controlled varia-
bles often chosen relate directly to the mole
fractions x , x of the more volatile compo-
D B
nent in the top and bottom product, respectively.
However, several flow rates or combinations of
them (L, V, D, L/D, etc.) exist as choices of
manipulated variables and each choice gives rise
to a different control scheme.
Previous work on this problem has focused on the
introduction and evaluation of new control
schemes that aim to minimize loop interaction.
Various choices for manipulated variables have
appeared, such as reflux t o boilup schemes (Rij-
nsdorp, 1965a; 1965b), material balance schemes
(Shinskey. 1979), hybrid methods (McAvoy, 1977),
and reflux ratio schemes (Ryskamp, 1980). In
addition, approximate analytical techniques have
been studied for the estimation of the interac-
tion measure of some control schemes (Shinskey,
f
Correspondence author.
mations have shown that, for symmetric ( l - x - x )
and high purity separations, none of the tradi-
D w
tional choices of control variable OL, D, V* or
B) provide a satisfactory strategy (Georgakis and
Papadourakis, 1981).
The present work aims to examine all possible
dual composition control schemes in which the
manipulated variables are either flow rates or
simple ratios of them. By means of a simple
criterion, it is possible to cast these schemes
into groups with identical values of the steady
state interaction measure (RGA). It is shown
that 34 independent control schemes exist. Sur-
prisingly, however, these schemes collapse into
the traditional patterns of steady state intera-
ction behavior as the product purity approaches
the limit of perfect separation. It is therefore
deduced that, for high purity distillation c o -
lumns, the 34 control structures examined here do
not provide any better choice of control strategy
than the CL,V), CD,V) and CL,W) control
schemes, none of which performs satisfactorily.
INTERACTION MEASURE
Under the assumptions of constant relative vola-
tility and constant molar overflow, the steady
state of a binary distillation column can be
calculated by the group method described by King
(1980). A schematic representation of the disti-
llation column and its flows is shown in Figure
1. We denote the L and V* column flows by vector
d,
T
d - (D,V) (1)
and observe that the remaining flow rates can all
be derived in terms of dj and d through
2
steady-state material balances. A set of 7 alge-
braic equations describing the column are ob-
tained (Kridiotis, 1983). They can also be writ-
ten in vector form as follows.

249
250 A. C. Kridiotis and C. Georgakis
FEED
il
Fig. 1: Schematic Diagram of different flows in a
distillation column and their symmetries cha-
racteristics. Solid lines denote variables that
can be manipulated.
f (x. d) - Ο
By defining suitable deviation variables and
linearizing around a steady-state, one then ar-
rives at the following expression for the steady
state sensitivity of the column state vector χ
to changes in flow rates, d:
1
δχ - A" C ôd.
The matrices A, C are the Jacobians of the
function f in eq. (2) with respect to χ and
d, and can be derived analytically by
differentiating the governing equations. If the
variables F, x , q, x , x , α and δ,
F D B
characterizing a column design, are given then
the matrix:
-A"
can be calculated.
Next consider the relation of the D and V flows,
originally selected as the process Tinputs, to the
manipulated variables m - (m m ) of
lt 2
another control scheme we wish to consider.
d - g (m)
This relationship might be either linear or non-
linear. If it is nonlinear, we linearize around
the steady-state and in both cases we obtain.-
ôd - H ôm,
where H j - dg/dmj. Substitution of eq. (6)
t and V , results in 21 variables consisting of six
into eq. (3) yields:
δχ - Q H ôm S Ôm
Then the steady-state interaction measure between
the (ni! -> x ) and the (m -» x ) loops of the
D 2 B LVV, LVB, VVB, L/L\ V/V\ L, V. V/B, L/V\
examined control scheme is given by:
3 3
3 33 3 As a result of the constant molar overflow assum-
S1 62
51 62 52 S1
It is worth remarking here that matrix Q does
dependent on the column design characteristics,
while matrix H changes each time we consider a
different control strategy.
EQUIVALENCE CRITERION
We will now establish a criterion which indicates
whether two control schemes are equivalent to
each other with respect to their steady state
interaction characteristics quantified by value
of RGA. Equivalence will be used here to imply
both identical and complimentary values of the
interaction measure \ . The RGA values of two
control schemes are called complimentary when
they add up to one. This implies that a change
in the input-output pairing of one of the control
schemes will make the RGA's equal to each other.
It should however be remembered that RGA equi-
valence does not imply that the control schemes
will have identical dynamical characteristics.
Consider two control schemes that use as pairs of
input variables (m , m ) and (mj\ m ) ,
t 2 2
respectively. The necessary and sufficient con-
dition, CKridiotis, 1983), for equal interaction
measures is either:
(2) (9)
mj m
2
(10)
(3)
An important corollary of eqs. (9) (10) is that
the following control schemes (X. Y). (1/X. Y).
(X, 1/Y) and (1/X, 1/Y) are RGA equivalent. Such
RGA equivalent control structures will be repre-
sented by one of their members. This greatly
reduces the number of schemes that need to be
considered and simplifies the analysis. For two
different control schemes the use of equation (6)
gives:
(4) Ôd - H Ôm - H* ôm*. (11)
Then the relationship between the two s e t s of
manipulated variables, m and m* is given by:
ôm LI-l H* ôm*- U ôm* (12)
Therefore we need only examine matrix U to
find whether conditions (9) or (10) are met and
(5) determine whether the two control schemes using
the m and m* variables as inputs are equi-
valent or not.
An exhaustive examination is performed of the
different control schemes that utilize either the
(6) distillation column flows or their ratios as
manipulated variables. Enumeration of all possi-
ble combinations of the six flows, D, B, L, L\ V
flows plus ( ) - 15 ratios. Three of
2
these variables, however, L\ V and LVV, cannot
(7)
be directly controlled, since they consist of
internal flows. The remaining 18 possible con-
trol variables are: D, B, D/B, V/L, V/D, L/D,
L/B, D/L- and D/V\
(8)
ption, the flows in the column, and consequently
the control variables, exhibit considerable sym-
metry. This symmetry is apparent in the governing
equations, where one can interchange D with Β, V
with L\ V with L, x with x and finally q
D B
with (1-q) to obtain the corresponding relation
for the other half of the column. This observa-
tion can be made more graphic by viewing the two
halves of the distillation column in Figure 1 as
mirror images of each other. With this defini-
tion of symmetry, it can stated that: D, L, V
 Independent Single-input-single-output Control Schemes 251

are symmetric to Β, V and L' respectively. The above selection procedure, along with the
criterion of equivalence based on equations (9)
The concept of symmetry can be extended to any and (10), was applied to the 4 5 possible control
group of flows. Two expressions are symmetric if schemes. This resulted in 34 RGA independent
one can be derived from the other by substituting groups of control schemes, listed in Table 3.
each flow by its symmetric counterpart. The 18
control variables obtained above can thus be cast (Table 3 is given at the end of the paper)
into the symmetric pairs listed in Table 1.
The interaction measure given in the same table
TABLE ltSymmetry Between Manipulated Variables was calculated for x - 0 . 9 8 , x «0.02, a - 2 . 1 ,
D B
δ-1.25, q-0.8 and x - 0 . 4 5 , and serves as a
F
confirmation that all 34 groups of schemes are
D <—> Β
D/B <--> B/D (*) characterized by different interaction measures.
V/L 1 <—> L'/V
L/D <--> V'/B
The number of independent control schemes changes
L/L <__> v/v when the feed is a liquid at its boiling point
L <--> V»
V/B <—> L'/D (q*l) and when the feed is a vapor at its dew
L/V <--> point (q*0). For q=l, the flows V and V* become
L/B <--> V'/D
V/D <--> L'/B identical, and the control variable V / V ls eli-
minated. Furthermore, schemes that differ only
(*) Self Symmetric in that one uses V where the other involves V
become identical with each other. Similarly,
The criteria described in eqs. (9) and (10) can the flows L and L' are equal to each other when
now be applied to determine which of the identi- q-0. For these special values of q, the groups
fied control variables satisfy them. Bearing in of independent control schemes drops to 17.
mind that the results should again be symmetric, These new groups are listed in Table 4 in terms
we obtain the 10 independent groups of manipu- of their constituent original groups.
lated variables listed in Table 2.
TABLE 4: RGA Independent SISO Control Schemes
TABLE 2;Independent Manipulated Variables and for Feed Quality Equal to q«l or q»0.
Ratios with their Symmetry Relations
Member Groups (Refer to Table 3)

No. q = 1 q = 0
[ D or Β or D/B ] (self-symmetric)
1 1, 5, 7 2, 5, 6
[ V/L or V/D or L/D ] <--> [ L'/V or L'/B or V'/B ] 2 2 1
3 3 4
f L or L/L' ] <—> [ V or V / V ] 4 4 3
5 6 7
V/B <—> D/L» 6 8, 10, 15, 18, 28 8, 17, 11, 12, 31
7 9, 24, 27 14, 21, 22
L/B <--> D/V 8 12, 30, 33 18, 32, 33
9 14, 20 9, 26
L/V (self-symmetric) 10 16, 29 13, 34
11 17 10
12 19 19
13 21, 23 24, 25
14 22 27
15 25 23
RGA INDEPENDENT CONTROL SCHEMES 16 26 20
17 31, 34 28, 29
We now consider the combinations of the 10 inde-
pendent control variables, two at a time, into
control schemes. The total number of possible
pairings is (*°) - 45. Notice that schemes HIGH PURITY COLUMNS
of the form (X, Y) and (Y, X) do not have to be
considered separately, since their interaction The calculation method for the interaction mea-
measures add to unity. Of the 4 5 possible sure already described can be used to examine the
schemes, some must be eliminated on the grounds behavior of λ for various design conditions. Of
of one or more of the following restrictions: particular interest is to calculate the interac-
tion measure (RGA) for the case of high purity
i) The scheme cannot be implemented in practice separations. The interaction measure was calcu-
because both of its loops use only one mani- lated for all 34 schemes and plotted as a a
pulabie variable. For example, in the function of the distillate purity e - l - x . As
the top purity wes increased so was the bottom
D
scheme CD.V/D), the second control variable
can only be adjusted through the manipula- one by reducing the value of x . In particu-
B
lar, we let x » ( i - x ) / o ) , where ω is the
tion of D, because V cannot be set directly. B D
However D is already used in the first loop separation symmetry factor defined by Georgakis
and this control scheme is not viable. and Papadourakis (1983)and was kept constant. As
e -> 0, one observes that 7 of the control
ii) The scheme attempts to vary D and Β indepen- schemes, (No. 1-7),
dently. These two flows are related through converge to the same value of \ which is between
the steady-state overall material balance, D 0 and 1, while all other schemes diverge to ±oo.
• Β « F and cannot be varied independently. This behavior is shown in Figures 2 and 3, and
persists for all other values of the design para-
iii) No flow is left to control the level in the meters. Furthermore, it was found that the high
accumulator or the reboiler. For example, purity limit of λ for Schemes 1-7 coincided
in the scheme QO.V/L), the second control exactly with an asymptotic expression of
variable must be adjusted through changes in X . and X
L. This uses up both available flows in the D|V LBi derived analytically
by Georgakis and Papadourakis, (1982). This
rectifying section, D and L, and leaves none approximation, derived by assuming q-1 and ε -+
by which to control the accumulator level. 0, is as follows*.
Similarly, Β and V (symmetric to D and L)
cannot be used simultaneously for compo-
sition control.
 252 A. C. Kridiotis and C. Georgakis

manipulate column flows or their ratios and have

^ωχ/ - J Cl 3)
F different interaction characteristics. Special
feed quality conditions (q-1 or q-0) reduce this
6 xF number in half, leaving only 17 independent
schemes for composition control.
The agreement between numerically calculated
values of λ and the results of eq. (17) was Furthermore, this work suggests that none of the
explored for other values of α, δ, x and ω initial 4 5 control schemes examined none has a
F
and was found to be very accurate. The different desirable interaction characteristics at high
behavior of these two groups of control schemes purity separations. This might also imply that
for high purity designs is primarily due to the changes in the way the measured variables are
use of an "external" flow CD or B) as manipulated utilized might be necessary for the design of
variable in the first group (Nos. 1-7), while all noninteracting control structures for high purity
other schemes use only "internal" flows (L, V, L' columns.
or V).

0.6 0 I • • M""l • • Ί " "| • ' • • I|MH| ' ' • I'M.l

NOTATION

B: Bottom product flow rate (mol/s)

D: Distillate flow rate (mol/s)
0.5 0
Fj Feed flow rate (mol/s)
L: Reflux flow rate (mol/s)
q: Feed quality (molar fraction of feed in
liquid phase)
0.4 0 -
R: Reflux ratio (L/D)
RJQ : Minimum reflux ratio
V: Internal vapor flow rate (See Figure 1)
(mol/s)
0.30 Î x: Mole fraction of the more volatile component
Creek Symbols

α Relative volatility
0.20h
δ: Ratio of actual reflux ratio to minimum
reflux ratio
\t Interaction measure
5 A 3 2 1 φ: State variable as defined by eq. ( 8 - 8 3 ) in
η inl • •• ' • ' • • •!•'»» • ' » • '"nil
Ref.4
' 10" 10~ 10' 10' 10" 10 °
ω: Separation symmetry factor ( l - x ) / x
D B
Fig. 2: Interaction measure vs. e for schemes 1 to 7
and the following values of the parameters: Superscripts
(q-0.8, α-1.7, δ-1.3, x - 0 . 4 5 , ω-0.7)
F Β·· In bottom product stream
D: In distillate stream
F: In feed stream
f: On feed plate

REFERENCES

Bristol, E.H. (1966). IEEE Trans. Autom. Control,

A C - l l . p.133.
Georgakis, C. and A. Papadourakis, (1982). Chem.
Eng. Set., 37, p. 1584.
Jafarey, Α., T.J. McAvoy, and J.M. Douglas, (1979)
Ind.Eng.Chem.Fund., JJ L p.181.
King, C.J., (1980). Separation Processes,
2nd ed., McGraw-Hill. New York, Ch. 8.
Kridiotis, A.C., (1983). M.S. Thesis, MIT.
Rijnsdorp, J.E., (1965a). Automat tea, L, p. 15.
Rijnsdorp, J.E., (1965b). Automatica. L, p.29.
Ryskamp, C.J., (1980). Hydroc. Proc, 60, p.5l.
Shinskey, F.G., (1977). Distittation Control for
Productivity and Energy Conservation,
McGraw-Hill, New York. Chs. 8 and 10.
Shinskey, F.G., (1979), Process Control Systems,
2nd ed., McGraw-Hill, New York, Ch. 8.
Fig. 3: Interaction measure vs. e for schemes 8 -
34 and the following values of the parameters:
(q=0.8, α=1.7, δ=1.3, x =0.45, ω=0.7)
F

CONCLUSIONS

In this communication we examined the question of

independent control schemes in dual composition
control of binary distillation. Using a simple
criterion, it was possible to establish that
there are 34 independent control schemes that
 Independent Single-input-single-output Control Schemes 253

TABLE 3:RGA Independent SISO Control Schemes and

their RGA Values for a Refernce Design.

No.? Schemes RGA No Schemes RGA

"Ί
1 : (B,V/L); (B.L/D); | 0.294 16 ( L 7 V , L / B ) ; L7B,L/B) -4.594
(D/B,V/L); (D/B,L/D) (V/B,L/B)

( D , L 7 V ) ; (D,V'/B); | 0.503 17 1
(L7V,D/L'); -0.429
(D/B, L ' / V ) ; (D/B,V'/B) (V7B.D/L )
T
(B,L); (B,L/L ); f | 0.382 18 (L7V,D/V); (V7B,D/V)| -0.790
(D/B,L); (D/B,L/L )
19 ( L , V ) ; (L,V/V); 7.587
( D , V ) ; (D,V/V); ! 0.416 j (L/L\V); (L/L',V/V)
( D / B , V ) ; (D/B,V/V )
20 j (L,V/B); (L/L\V/B)
( D , L / V ) ;f ( B , L / V ) ; | 0.222 f ; 2.390
(D/B,L/V ) 21 j ( L , L / V ) ; (L/L ,L/V) -0.859
I
(B,L/B); (D/B,L/B) \ 0.454 22 I (L,L/B); (L/L',L/B)
f 3.905 j
( D , D / V ) ; (D/B,D/V ) ; 0.309 23; (ί,ϋ/ν'); (L/L',D/V)
-2.599
(V/L,L'/V); (V/L,L'/B); 1.697 24; ( V , L / V ) ; ( V / V , L / V )
(V/L.V7B); ( V / D . L V Vf
);. 25 ( V , L / B ) ; ( V / V ,L/B) -0.670
(V/D,V7B); (L/D,L7V ); ( f f 6.986!
ι
(L/D.L7B); (L/D,V/B) jj
! 26 j (V,D/L ); (V/V,D/L )
-0.744 \
!
( V / L , V ) ; (V/L,V/V); 2.401 |
( V / D , V ) ; (V/D,V/V);
} 27 (V,D/V); (V/V,D/V) J -1.681j
( L / D , V ) ; (L/D,V/V) ' j
28: (V/B,L/V) -0.367 j
10 (V/L,V/B); (L/D,V/B) 1.642 J ί
. i 29ί (V/B,L/B)
11 (V/L,L/V); (V/D,L/V); -2.199 ! -3.585 j
(L/D, L / V ) j 30 ( L / V , L / B )
( 1.524 *
12 (V/L,L/B); (L/D,L/B) 2.000 ,| 31 (L/V,D/L')
: 16.633
il
13 (V/L,D/V); (V/D,D/V); 14.420 \ 32 ! ( L / V , D / V )
(L/D,D/V) j 2.776
33| (L/B,D/V)
( L 7 V , L ) ; (L7V,L/L'); -1.570 ί -1.161
14
(L7B,L); (L7B,L/L»); 1 ! 34; (D/L\D/V)
(V7B,L); (V7B.L/L ) j 3.131

(L7V,L/V); -0.393
15
( L 7 B , L / V ) ; (V/B,L/V)j
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

THE JACOBI EIGENVALUE CRITERION:

A DYNAMIC EXTENSION AND
STABILITY THEOREM
G. Mijares, J. D. Cole, H . A. Preisig and C. D. Holland
Department of Chemical Engineering, Texas A&M University, College Station,
TX 77843, USA

Abstract. The Jacobi Eigenvalue criterion is a recently developed, steady state,

scaling independent tool for the selection of variable pairing for multiloop control
systems (Mijares et al., 1 9 8 6 ) . This criterion is based on the Jacobi Iteration method
for solving sets of linear equations or obtaining the inverse of a matrix. A dynamic
extension of the Jacobi Eigenvalue criterion which consists in evaluating the spectral
radius of the Jacobi Iteration matrix in the frequency range of interest is presented.
The Jacobi Eigenvalue criterion is inherently a diagonal dominance 1 measure, and as
such, is readily applied to develop an extension of Rosenbrock s Nyquist stability
theorem. The extended Jacobi Eigenvalue criterion is then compared to different inter-
action and weak coupling measures used for design of decentralized control structures
which are based on conditions for generalized diagonal dominance and Η-matrices. Two
examples are used to illustrate the use of the Jacobi Eigenvalue criterion and its
relationship with the other weak coupling measures.

Keywords. Multivariable control systems; Nyquist criterion; iterative methods; inter-

action; stability criteria; computer control.

INTRODUCTION stability theorem with equally or less restrictive

The pairing of controlled and manipulated variab-
les in a multiloop control system or the design of
maximum decentralized control structures continues
to represent a major challenge for control engi-
neers .
Methods which use the concept of diagonal domi-
nance like those proposed by Rosenbrock (1974)
(INA, DNA) have been widely applied to assess
interaction and design decentralized control
structures. In these methods the amount of domi-
nance is used as an indicator of the accuracy of
the nominal decoupled process. As originally
proposed by Rosenbrock these methods were sensi-
tive to scaling of the transfer function matrix.
Further refinements in the dominance conditions
were presented by Araki and Nwokah ( 1 9 7 5 ) , Nwokah
(1979, 1 9 8 0 ) , and Limebeer (1982) which made the
methods scaling invariant and provided sharper
bounds for stability.
In this paper, a dynamic frequency domain exten-
sion of the Jacobi Eigenvalue criterion (Mijares
et al., 1 9 8 6 ) is presented. By evaluating the
Jacobi iteration matrix at different frequencies,
the effectiveness of the control structure at
the frequency range of interest is assessed.
First, a brief description of the steady state
Jacobi Eigenvalue criterion is given and the ex-
tension to include the dynamics of the system is
presented. Application of the extended criterion
as a diagonal dominance measure leads to1 the deve-
lopment of an extension of Rosenbrock s Nyquist
dominance conditions than those required in the
extensions by Limebeer ( 1 9 8 2 ) and Nwokah (1980).
Subsequently, two theorems are stated to aid in
the understanding of the relationship between the
Jacobi Eigenvalue criterion and weak coupling
measures based on conditions for generalized dia-
gonal dominance or Η-matrices. Finally, two exam-
ples selected from the literature are presented to
illustrate the use of the dynamic extension of the
Jacobi Eigenvalue criterion and its relationship
with the weak coupling measures discussed above.
THE JACOBI EIGENVALUE CRITERION
The Jacobi Eigenvalue criterion as originally
presented by Mijares et al. ( 1 9 8 6 ) is based on the
difficulty of finding the inverse of the steady
state gain matrix. Since the best pairing of
variables or decentralized structure is the one
which makes the system most closely resemble a set
of independent single-loop systems, it is also the
one whose off-diagonal elements make the smallest
contribution to the inverse of the steady state
gain matrix. Consider the following multivariable
system:
y = G m (1 )
where G is the η χ η gain matrix and y and m are
the vectors containing the controlled and manipu-
lated variables, respectively. The general servo
control problem can be reduced to obtaining the
necessary values of the manipulated variables to
reach a desired new set of controlled variables.

255
256 G. Mijares et al

0 aa a
For fixed values of y and G, a solution to Eq. (1)
!2 i 3 "·· m
exists provided that the inverse of G exists and a 0 a
if the inverse exists, then Eq. (1) may be solved
21 23 2n
for m to give: (10)
A = -

(2)
η ,n-1
For a system free of interaction, G is a diagonal
matrix and its inverse is equal to1 the Ί
inverse of where
its diagonal elements; that is G = D , where

ij g
1/g 11

1/g.22 Thus Eq. (9) reduces to:

1 1
(Ν)
(3) Β [I + A + A + + A V . G- (11)

The necessary and sufficient condition for the

matrix series given by Eq. (11) to converge is
For a system having interaction, let Β denote an that the largest eigenvalue of A be in modulus
approximation of G . Therefore, the inverse of G less than unity, How "big" or how "small" Δ Β is
Τ
may be represented as follows: depends on the specific norm used to obtain a
quantitative measure of ΔΒ An examination of
Β + ΔΒ W the characteristics of the different matrix norms
and shows that there is no unique or "best way" for
measuring Δ Β .
(Β + AB)G = I (5) Τ However, Young (1971) concluded
that as Ν approaches infinity, the closer G is to
•1 D, the smaller Δ Β is, and the faster the method
Thus, how well Β represents G depends on the Τ
contribution of ΔΒ to B. An expression for ΔΒ in converges. Young (1971) defined the rate of con-
terms of G and Β may be found by use of an itera- vergence as follows:
tive procedure in which better and better approxi-
mations of Β are obtained. Postmultiplication of R(A) = - log S(A) (12)
both sides of Eq. (5) by Β followed by rearrange-
ment yields where S ( A ) is the spectral radius of A (the eigen-
value of A which is largest in modulus) and R ( A )
ΔΒ GB = (I- BG)B (6) is the rate of convergence. Thus, a quantifica-
tion for the pairing that makes the system most
With continued improvements in the choice of B, GB closely resemble a set of single-loop systems (or
approaches I. Thus, the pairing whose off-diagonal elements are least
() 1 ( i )
U) significant) has been obtained; namely, it is that
ΔΒ = (I - B G)B (7) pairing whose corresponding Jacobi iteration mat-
rix A has the smallest-largest eigenvalue in
and the first improved value of Β is given by modulus of all the possible pairing.
() 1 () 0
B = B°
+ 0 )- ( ) 0
ΔΒ
(
DYNAMIC EXTENSION OF THE JACOBI
EIGENVALUE CRITERION
[I
+(I - B G)] B (8)
The dynamic extension of the Jacobi Eigenvalue
To initiate the correction process, the 1 first criterion which makes it possible to analyze the
approximation Β is taken equal to D . After Ν control structure at specific frequencies or fre-
iterations or improvements on Β , one obtains quency ranges follows. By starting with the
linear dynamic model of the multivariable system,
()N ()0 N-1 () 1 ( ) 0
B . B l ΔΒ - B • ΔΒ
Τ y(s) = G(s)m(s) (13)
(9)
i=0
where setting s = ίω, and following the same development
presented above, it is found that the best pairing
at any frequency ω is that pairing whose Jacobi
N-1
iteration matrix has the smallest spectral radius
() ί ( 0 )
of all possible pairing. The Jacobi iteration
l ΔΒ = ΔΒ = [(I - B G) + ... +
Τ matrix for any frequency ω is given by:
( 0 0)
) ( )ί ω 3
(I - B G)V 0 a (1ω) 1η
() 0 1 ( )
i a ) a
Since B = D~ and G = D(I - A ) , it follows that a (iu>) 0 2n
21
(I - B ^ G ) = A, where A is equal to the Jacobi Α(ΐω) = (14)
iteration matrix,

a (iu)) 0
nl
 The Jacobi Eigenvalue Criterion 257

Thus by evaluating the spectral radius of A at Since p(J(Q(s))) < 1 for all s on C it follows
different frequencies, it is possible to analyze that the real part of |l - J ( Q ( s ) ) | is positive
the effectiveness of a desired pairing or control and less than unity for all s on C. Hence the
structure over the frequency range of interest. mapping of C by |l - J(Q(s))| can not encircle the
origin, or N = 0. Thus,
P
THE JACOBI EIGENVALUE CRITERION AND N N
SYSTEM STABILITY Ν N_ (21)
Q " D " %1 Q2 qn

Based on the characteristics of the Jacobi where Ν . is the number of encirclements of the
qi
Eigenvalue criterion, an extension of Rosenbrock's origin by the mapping of C by q ^ t s ) .
Nyquist stability theorem with a less restrictive
dominance condition can be developed. The condi- Theorem 1 , Consider the general feedback system
tion that for a general system matrix G the spec- with a feedforward path Q(s) = L(s)G(s)K(s) and a
tral radius of the Jacobi iteration matrix, feedback compensator F , where F and K(s) are dia-
p(J(G(s))), be less than unity (diagonal dominance gonal matrices. Let C be a simple closed curve
according to the Jacobi Eigenvalue criterion) may which encircles all the poles and zeros of |l +
be used in lieu of the more restrictive general- Q(s)F| lying in the right-half plane. Let
ized diagonal dominance or Η-matrix conditions to p(J(G(s))) < 1 for all s on C. Futhermore, let
determine asymptotic stability for a multivariable the mapping of C by q L for k
kk' N ts i 2, m e .n,
system from the characteristics of the nominally
encircle the point -f.k 1 ' q k clockwise.
decoupled system. Before presenting the theorem,
Then the closed-loop
closed-lo system is asymptotically
two lemmas are presented which are used in the
stable if and only if
proof of it.
η
Lemma If G(s) has a Jacobi iteration matrix I Ν qk (22)
with a spectral radius less than unity for all s k=1
on a simple closed curve C, then for any nonzero
diagonal matrices K(s) and L(s), Q(s) where P is equal to the number of poles in the
L(s)G(s)K(s) has a Jacobi iteration matrix with a
Q
right-half plane of the open-loop system.
spectral radius less than unity for all s on the
simple closed curve C. Proof. Following the same approach as Rosen-
1 number of encirclements of the
brock's (1974), the s i
Proof. The Jacobi iteration matrix of G may be point (-f , 0 ) by the mapping of C by Q ( s )
i 1 i i
stated as (for simplicity the Laplace operator the same as the number of encirclements of the
has been removed): origin by the1 mapping of C by 1
f. + q..(s). Then

because |F~ + Q ( s ) | = jF"~ | 11 + Q(s)F|, the

J(G) I - D 'G (15)
Q result follows by application of Lemmas 1 and 2
and Corollary 3-4.1 of Rosenbrock (1974).
where Ό is a diagonal matrix with the diagonal
η
elements of G and J(G) is the Jacobi iteration
matrix of G. Similarly, THE JACOBI EIGENVALUE CRITERION AND WEAK
COUPLING MEASURES
J(Q) = I - D 'Q (16)
Q
First, the following definitions are necessary:
where
nn x(Berman and Plemmons,
Definition 1979). The mat-
LGK (17) rix G e C is generalized row diagonally dominant
if there exist an xeR , χ > 0 such that:
Substituting in Eq. (16) we have

J ( Q ) = I - K~ V ''GK (18) x
(J
|δ Ι i > I ISijI X j for i = 1, 2,...η (23)
Thus,
η

λ [J(QJ] = X[J(G)] (19) nn x

A matrix G e C is generalized n
column diagonally
Lemma 2. Let Q(s) = {q. }eC . Let D (s) be a dominant if there exists an xeR , x>0 such that
Q
diagonal matrix containing the diagonal elements
of Q(s). Let all the right-half plane poles and χ Ν
zeros of |Q(s)| and |D (s)| be enclosed by a
Q ISiil ι > Σ ISjil X j for i = 1,2,...η (24)
simple closed curve C. Further suppose that for
all s on C, p(J(Q(s))) < 1. Then the mapping of C j-l
by IQ(s)I and | D Q ( S ) | encircle the origin the same n x n
number of times. Definition (Varga, 1976). nnxG e C , its com-
For
parison matrix M(G) = ( m _ ) e R is defined by:
Proof. Let N,
Q' ' V
and Np denote the number of m g= ; =
m g r
f Jo ) (
encirclements of the origin by the mapping of C by
ii l iil ij " l ijl ^
|Q(s)|, |D(s)|, and II - J(Q(s))|. Since
and the set of equimodular matrices associated
|Q(s)| |D ||I - J ( Q ( s ) ) | with G is defined as:
Q
then, W 1N
Ν = Ν + Ν (20)
Q % P
258 G. Mijares et al.

1 1
ft(G) = { B (b..) ε C
±J
XCJ(G)]
1* i,j^n}

The relationship of the Jacobi Eigenvalue criter-

ion to the conditions for generalized diagonal p(J(G)) = X
dominance (Limebeer, 1982) and to the dominance
and,
pF CJ(M(G))]
conditions in terms of Η-matrices (Nwokah, 1980) Therefore, the conditions required for a matrix to
are summarized in the following theorems: be generalized diagonal dominant are equivalent to
nn x those required for it to be an Η-matrix. For 2x2
Theorem 2. For any G = [ g ^1] ε C with all cases the interaction index defined by Nwokah
diagonal entries nonzero ihe following are (1979), X .[J(M(G))], is the same as the spectral
DC
rr
equivalent:
radius of the Jacobi iteration matrix and the
best pairing or structure according to the Jacobi
a) G is generalized row diagonally dominant
b) Ί
G is generalized column diagonally dominant
Eigenvalue criterion is that one for which the
matrix is most generalized diagonal dominant. For
c) There exist a diagonal S>0 such that S ϋΒ
higher order systems, requiring that the spectral
row dominant
-1 radius of the Jacobi iteration matrix be less than
d) There exist a diagonal ¥>0 such that VGV
unity is a looser dominance condition than that of
column dominant
requiring the matrix to be g e n e r a l i z e d diagonal
e) G is an H-matrix
dominant or an H-matrix.
f) For any Β ε Ω ί Ο , p(J(B) ) £ p ( | J(B) ( ) =
p(J(M(G))) < 1, where J(.) is the Jacobi EXAMPLES
iteration matrix of (·).
In this section the use of the proposed extension
Proof. The equivalence of conditions (a), (b), of the Jacobi Eigenvalue criterion is illustrated
(c), and (d) was proved by Limebeer (Theorem 2, by solving two examples from the literature con-
1982). The equivalency of (a), (b), (e) and (f) sisting of two distillation columns. The results
was proved by Varga (Theorem 1, 1976) or Berman obtained by applying the Jacobi Eigenvalue
and Plemmons (Theorem 5.14, 1979). 1 compared with those obtained by
criterion are
using Nwokah s interaction index or the equivalent
Theorem _3. For a 2x2 system the spectral radius concept of generalized diagonal dominance.
of the Jacobi Iteration matrix of G, p(J(G)), 1 is
equal to the Perron-Frobenius eigenvalue, of Example This example consists of a model of a
PF
J(M(G)). distillation column containing 10 stages which was
presented by Tung and Edgar (1977). The system is
Proof described by

χ = Ax + Bu Ky
'12
3
J (G) [y
11 r y 2] = [χ.,, x
1L Q u = [u , u ]
1 2
5
where,
21
5
x^ = MeOH concentration in the bottom product
22
*1Q= MeOH concentration in the top product
u^ = reboiler steam pressure
XCJ(G)] = ±, u = reflux rate.
2
(27)
For numerical values and information, see Tung and
Edgar (1977). Figure 1 presents the spectral
*12'
radius of the Jacobi iteration matrix for the
control structures {y^-u^/y ~u } and [ y u / y
2 2 r2 2
J(M(G)) 5 u }. As expected the Perron-Frobenius eigenvalue
of the corresponding matrix J(M(G)) for each
21 I
> structure is equal to the spectral radius of the
Jacobi iteration matrix. These results recommend
= 22'
the structure {y^-u /y^-u^} for all frequencies.
1
Since the spectral radius of the Jacobi iteration
A[J(M(G))] = ± (28) matrix for this structure is less than unity for
all frequencies, the structure is dominant accord-
ing to the Jacobi Eigenvalue criterion and Theorem
8 g 1 may be used to infer system stability from the
Thus if, 12 21 > ο properties of the individual single-loop systems.

Example 2. A complex distillation column with two

product sidestreams presented by Elaahl and Luyben
then, p(J(G)) = λ
ρρ [J(M(G))]
(1985) was selected to illustrate the use of the
dynamic extension of the Jacobi Eigenvalue pairing
On the other hand.If < 0, then criterion for larger systems. In this example,
the product purities of four output streams are to
 Control of a Distillation Column with a Heat Pump 259

composition of propane in the distillate (XD3), interaction problem. Although the procedure is
the composition of isobutane in the second side- simple to apply, it gives a precise quantitative
stream (from bottom) (XS2), the composition of n- comparison of the accuracy of the nominal decoup-
butane in the first sidestream (from bottom) led models for a given set of control schemes.
(XS1 ), and the composition of isopentane in the Based on the characteristics of the Jacobi Eigen-1
bottoms product (XB3). The manipulated variables value criterion, an extension of Rosenbrock s
are the reflux flow rate (R3), the two sidestream Direct Nyquisy Array stability theorem with equal-
flow rates (LS1 and LS2), and the heat input to ly or less restrictive dominance conditions than
the column (QB3). It is intended to analyze five those required by similar extensions presented in
control schemes (see Table 1) proposed by Elaahl the literature was developed. A similar develop-
and Luyben (1985). ment exists for Rosenbrock's Inverse Nyquist Array
stability theorem.
Although McAvoy (1986) has shown that the system
as described by the linear model with controllers REFERENCES
including integral action is unstable for all five
control schemes, Elaahl and Luyben (1985) found Araki, M., and 0. I. Nwokah (1975). Bounds for
control scheme Β to be the scheme for which the closed-loop transfer functions of
best control could be achieved. They used a non- multivariable systems. IEEE Trans. Autom.
linear simulation program to test the five dif- Control, AC-20, 666-670.
ferent schemes and found scheme Β to be the best
control structure. Elaahl and Luyben explained Berman, Α., and R. J. Plemmons (1979). Iterative
this by suggesting that the high degree of non- methods for linear systems. Nonnegative
linearity of the system resulted in a shift to a Matrices in the Mathematical Sciences,
stable region for scheme B. Although the con- Academic Press, New York, pp. 165-207.
clusions presented by McAvoy may be obtained by
used of the Jacobi Eigenvalue criterion in terms Elaahl, Α., and L. Luyben (1985). Control of an
of integral controllability (Mijares et al., energy-conservative complex configuration of
1986), it is not the purpose of this example to do distillation columns for four-component
so. Instead, it is desired to show that despite separations. I&EC Process Pes. Dev., 24,
the fact that all schemes are nominally unstable, 368-376.
scheme Β is the scheme closest to being dominant
according to the Jacobi Eigenvalue criterion. Limebeer, D. J. N. (1982). The application of
Also, it is desired to show how the spectral generalized diagonal dominance to linear
radius of the Jacobi iteration matrix compares to system stability theory. Int. J. Control,
the Perron-Frobenius eigenvalue of the correspond- 36, 185-212.
ing matrix J(M(G)).
McAvoy, T. J. (1986). Comments on "control of an
Table 2 shows the spectral radius of the Jacobi energy-conservative complex configuration of
iteration matrix and the Perron-Frobenius distillation." Ind. Eng. Chem. Process Pes.
eigenvalue of the corresponding matrix, J(M(G)), Dev., 25, 325-326.
for each of the studied schemes at steady state.
These results show that schemes Β and D have the Mijares, G., J. D. Cole, Ν. W. Naugle, H. A.
smallest spectral radius, and consequently should Preisig, and C. D. Holland (1986). New
be selected as the most desireable structures criteria for the pairing of control and
according to the Jacobi Eigenvalue criterion. manipulated variables. Accepted for
Figure 2 presents the dynamic results for this publication in AIChE J.
example. For simplicity only schemes Β and D are
shown. The spectral radius of the other three Nwokah, 0. D. I. (1979). The convergence and
schemes is larger than those of schemes Β and D. local minimality of bounds for transfer
Note that despite scheme D having a smaller spec- functions. Int. J. Control, 30, 195-202.
tral radius than scheme Β at steady state, when
the dynamics are considered scheme Β is always Nwokah, 0. D. I. (1980). A recurrent issue in the
closest to being diagonal dominant according to extended Nyquist array. Int. J. Control,
the Jacobi Eigenvalue criterion. 31, 609-614.

Figure 3 serves to compare the spectral radius of
the Jacobi iteration matrix and the Perron-
Frobenius eigenvalue of the corresponding matrix
J(M(G)) for scheme B. As expected the spectral
radius of the Jacobi iteration matrix is smaller
than the Perron-Frobenius eigenvalue for all fre-
quencies .
CONCLUSIONS
Rosenbrock, Η. H. (1974). Multivariable systems.
Computer Aided Control System Design,
Academic Press, London, pp. 117-179.
Tung, L. S., and T. F. Edgar (1977). Dynamic
interation analysis and its application to
distillation column control. Proceedings of
IEEE Decision and Control Conference, 1107"
1112.

The results obtained in the above examples demon- Yarga, R. S. (1976). On recurring theorems on
strate the applicability and usefulness of the diagonal dominance. Linear Algebra and its
proposed extension of the Jacobi Eigenvalue crit- Applications, 13, 1-9
erion. The transparency of the development
leading to the Jacobi Eigenvalue criterion gives
the practitioner insight into the pairing and
260 G. Mijares et al.

Young, D. M. (1971). Linear stationary iterative

methods. Iterative Solution of Large Linear
o = p ( J ( G ) ) for s c h e m e [1-1/2-2]
Systems, Academic Press, New York, pp. 84-89.
• = p(J(G)) for s c h e m e [1-2/2-1]
for s c h e m e [1-1/2-2]
= λ ρ for s c h e m e [1-2/2-1]
Ρ
TABLE 1 Control Schemes Analyzed for Example 2

Control Scheme Variable Pairing

X D3 - R3
XS2 - LS2
XS1 LS1
XB3 QB3

X D3 R3
XS2 LS2
XS1 QB3 -2.0 -1.0 0.0 1.0 2.0
XB3 - LS1
log cj 10
Fig. 1. p(J(G)) and λ
ρρ for the two control
X D3 LS2 schemes of Example 1.
XS2 QB3
XS1 R3
XB3 LS1

X D3 QB3
XS2 LS2
XS1 R3
XB3 LS1

X D3 R3
XS2 QB3
XS1 LS2
XB3 LS1

p(J(G)) for s c h e m e Β
p(J{G)) for s c h e m e D
TABLE 2 Steady State Results for Example 2 0.5
-2.0 -1.0 0.0 1.0 2.0
A
Control Scheme p[J(G)]
PF log cj 10
Fig. 2. p(J(G)) for control schemes Β and D of
Example 2.
A 3.94 5.85

Β 1.19 1.86

C 2.62 3.43

D 1.13 1 .80
1.5
Ε 1 .28 2.00

— p(J(G)) for s c h e m e Β
- _ λρρ for s c h e m e Β
0.5 . .. ι ι

-2.0 -1.0 0.0 1.0 2.0

log cj 10
Fig. 3. p ( J ( G ) ) and X
pF for control scheme Β of
Example 2.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

CONTROL OF A DISTILLATION COLUMN

WITH A HEAT PUMP
Y. Naka and E. O'Shima
Research Laboratory of Resources Utilization, Tokyo Institute of Technology,
Yokohama 227, Japan

Abstract. S t a t i c and d y n a m i c o p e r a b i l i t i e s of two types of d i s t i l l a t i o n

s y s t e m s w i t h a heat p u m p are i n v e s t i g a t e d b a s e d on d i g i t a l s i m u l a t i o n . A
heat p u m p is not i n s t a l l e d b e t w e e n a c o n d e n s e r and a r e b o i l e r but between
a p l a t e in the c o o l i n g z o n e of a d i s t i l l a t i o n c o l u m n and a r e b o i l e r (SC/R
System). One of the SC/R s y s t e m s is a v a p o r - s p l i t t y p e s y s t e m and the
another e m p l o y s a p a r t i a l c o n d e n s e r - t y p e s y s t e m as a side-cooler. Both
systems have the b r o a d r e g i o n s of f e a s i b l e o p e r a t i o n of the the manipu-
lative variables, r e f l u x r a t i o and heat to a r e b o i l e r . The l a t t e r s y s t e m
is s u p e r i o r to the a n o t h e r in t e r m s of c o n t r o l p e r f o r m a n c e .

Keywords. Process control; petro-chemical control; PID controllers;

d y n a m i c r e s p o n s e ; m o d e l s ; d i s t i l l a t i o n ; heat p u m p

INTRODUCTION v o l v e d in t h e zones can s a v e energy. The

heat e x c h a n g e b e t w e e n a c o n d e n s e r and re-
A c o n c e p t of a p p l i c a t i o n of a heat p u m p to boiler is an e x a m p l e of the above men-
a d i s t i l l a t i o n to r e d u c e e n e r g y load to a tioned. In g e n e r a l , typical possible pat-
reboiler has b e e n e s t a b l i s h e d . Some such terns of heat e x c h a n g e b y use of a heat
combined s y s t e m s h a v e been a c c e p t e d in an p u m p a r e c l a s s i f i e d in a r e a l s e n s e t o :
e x i s t i n g p l a n t and t h e i r heat p u m p is used (i) s u p p l y of heat e n e r g y r e m o v e d from a
merely b e t w e e n a c o n d e n s e r and a r e b o i l e r . condenser to a side-heater added to an
Due to this limited application, those intermediate plate in the heating zone
s y s t e m s cannot be w i d e l y u s e d . This reason (C/SH t y p e )
is b e c a u s e a heat p u m p has s e v e r a l serious (ii) s u p p l y of heat e n e r g y r e m o v e d through
restrictions; the temperature difference a side cooler added to an intermediate
b e t w e e n the top and t h e b o t t o m s , which in- plate in the c o o l i n g zone to a reboiler.
cludes the m i n i m u m a p p r o a c h t e m p e r a t u r e of (SC/R t y p e )
a heat e x c h a n g e r , is less than about 3 0 ° C ,
due to thermal decomposition of working The SC/R s y s t e m is d i s c u s s e d b e l o w . Each
f l u i d s the b o t t o m t e m p e r a t u r e is lower t h a n p a t t e r n has two m a n n e r s of heat exchanging
130°C and so o n . Recently, thermodynamic in a s i d e - c o o l e r . The s i d e - c o o l e r is c a l l e d
a n a l y s i s of d i s t i l l a t i o n s y s t e m s has prog- an e v a p o r a t o r , in a heat p u m p :
ressed. Engineers can now more easily (1) A p a r t of the v a p o r s t r e a m is cut at
determine the s u i t a b l e p o s i t i o n on an in- s o m e p l a t e in the c o o l i n g zone and is con-
termediate heat e x c h a n g e r or a heat pump densed in the e v a p o r a t o r in a heat pump
into a distillation system (Naka, 1980, s y s t e m , w h i c h is d e n o t e d as S C - l / R t y p e .
1 9 8 3 ) , and then r o u g h l y d e s i g n a heat i n t e - (2) A w h o l e v a p o r s t r e a m is t a k e n a w a y from
grated system. For a d i s t i l l a t i o n s y s t e m a suitable p l a t e in the c o o l i n g zone and
with a heat pump, remaining significant fed into an e v a p o r a t o r . A part of the
p r o b l e m s are of o p e r a t i o n , that is o p e r a b l e vapor stream is c o n d e n s e d by a working
region, dynamic behavior and so on. To f l u i d in t h e heat p u m p s y s t e m . B o t h liquid
develop distillation systems with a heat and v a p o r are r e t u r n e d to the c o l u m n , w h i c h
p u m p , o p e r a t i o n a l p r o b l e m s are d i s c u s s e d in is d e n o t e d as S C - 2 / R t y p e .
t e r m s of s t a t i c c h a r a c t e r i s t i c s and d y n a m i c
The schematic s t r u c t u r e s of b o t h systems
b e h a v i o r b a s e d on s i m u l a t i o n in this p a p e r .
are s h o w n in F i g s . 1 and 2, r e s p e c t i v e l y .

SC/R D i s t i l l a t i o n Systems
S T R U C T U R E S OF D I S T I L L A T I O N S Y S T E M W I T H HEAT
PUMP SC-l/R. The f l o w r a t e of v a p o r s t r e a m to
be cut at a s u i t a b l e p l a t e is f i x e d . This
Though the p o s i t i o n w h e r e a heat p u m p has p l a t e is d e t e r m i n e d so that t h e t e m p e r a t u r e
b e e n i n s t a l l e d in a d i s t i l l a t i o n c o l u m n is of the s a t u r a t e d liquid c o n d e n s e d from the
determined, t h e r e m a y be t w o p o s s i b l e p a t - vapor s t r e a m is near the d e s i r e d heat ex-
terns for use of a heat p u m p . Every dis- changing temperature. The S C - 1 e v a p o r a t o r
tillation s y s t e m has c o o l i n g and heating in a heat p u m p s y s t e m is the total con-
zones. A limited amount of heat e n e r g y can denser in a d i s t i l l a t i o n s y s t e m . As the
be r e m o v e d from the c o o l i n g zone and can be difference of the v a p o r t e m p e r a t u r e s bet-
s u p p l i e d to the h e a t i n g z o n e . A condenser ween t h e inlet and o u t l e t is great, the
and a r e b o i l e r b e l o n g s to the c o o l i n g zone large heat e x c h a n g i n g a r e a s h o u l d not be
and the heating zone, respectively. To p r o v i d e d in the e v a p o r a t o r . B u t , w h e n the
combine two heat loads respectively in- temperature profile is changed by some

261
262 Y. Naka and E. O'Shima

Evaporator

F i g . 2 S c h e m a t i c d i a g r a m of SC-l/R
Fig. 1 S c h e m a t i c d i a g r a m of SC-l/R
system
System

disturbance, heat e x c h a n g i n g load m a y v a r y i m p o r t a n t c o n t r o l l e r s in S C - l / R and SC-2/R,

very much. The liquid in the r e b o i l e r is respectively. H e r e , a r o t a r y c o m p r e s s o r is
c i r c u l a t e d a r o u n d the c o n d e n s e r in the heat installed and has two m a n i p u l a t i v e varia-
pump system. If the c i r c u l a t i o n f l o w r a t e bles, the n u m b e r of r o t a t i o n s and a t h r o t -
is s u f f i c i e n t l y large and the v a p o r i z a t i o n tle valve w h i c h are used to control the
r a t i o of the c i r c u l a t e d s t r e a m is m u c h less vapor f l o w r a t e of the w o r k i n g fluid and
than 1, the v a p o r c o m p o s i t i o n g e n e r a t e d is pressure. So, for s i m p l i f i c a t i o n of c o n -
in e q u i l i b r i u m w i t h the liquid c o m p o s i t i o n . t r o l p r o b l e m s , it is a s s u m e d that the p r e s -
If the c i r c u l a t i o n r a t e is too small and sures of the e v a p o r a t o r and the condenser
the r a t i o is o v e r 1, the v a p o r c o m p o s i t i o n are w e l l - c o n t r o l l e d by u s i n g t h e s e m a n i p u -
equals to the liquid c o m p o s i t i o n in the lative variables. But the f l o w r a t e must
reboiler. be c h a n g e d .

SC-2/R. The SC-2 is a kind of partial DESIGN AND OPERATING CONDITIONS

condenser in a d i s t i l l a t i o n column. The
temperature difference between the dew The design c o n d i t i o n s are summarized in
point of the v a p o r s t r e a m s u p p l i e d to the TABLE 1. B o t h types of d i s t i l l a t i o n s y s -
evaporator and that of the vapor stream tems c o n s i s t of a c o n d e n s e r , a r e b o i l e r and
r e m o v e d from the c o n d e n s e r is less than the 25 p l a t e s w h i c h are n u m b e r e d from top to
t e m p e r a t u r e d i f f e r e n c e of S C - 1 . The o u t p u t bottom. The n o m i n a l v a l u e s of the t e m p e r a -
streams of v a p o r and liquid g e n e r a t e d in t u r e of a c o n d e n s e d s t r e a m at the outlet of
SC-2 are returned to the "side-input" the e v a p o r a t o r w a s given as 105 °C. The
plate. The amount of the c o n d e n s e d s t r e a m R 1 1 3 is used as a w o r k i n g f l u i d . Its tem-
is d e t e r m i n e d b y the r e l a t i o n a m o n g UA and peratures in the e v a p o r a t o r and in the
the temperatures and flow rates of the
v a p o r and the w o r k i n g fluid at the inlet of
the evaporator. Here, U and A mean the TABLE 1 Design Conditions
overall heat t r a n s f e r c o e f f i c i e n t and heat
exchanging area, respectively. This s y s t e m Components : E t h a n o l and A c e t i c acid
a l s o has the c i r c u l a t i n g s t r e a m a r o u n d the
c o n d e n s e r in the heat p u m p s y s t e m .
SC-l/R SC-2/R
A w o r k i n g fluid r e c e i v e s heat e n e r g y from
the evaporator and its temperature is
raised by the use of a compressor. And (Distillation system)
then it t r a n s p o r t s heat e n e r g y to a re- •Plate : Total 25 25
b o i l e r , that i s , a c o n d e n s e r in a heat p u m p : Feed 8 5
system. In this p a p e r the R a n k i n e c y c l e is : Side-input 6 3
u s e d as the heat p u m p c y c l e . But, as the : Side-cut V(9) V(4)
whole working fluid p a s s i n g through the •Feed :Comp. 0.182 0.182
e v a p o r a t o r m a y not be c o m p l e t e l y v a p o r i z e d , : F. R. 50 50
a v a p o r - l i q u i d s e p a r a t o r s h o u l d be p r o v i d e d *Top :Comp. 0.900 0.900
so as not to d a m a g e the b l a d e s in a com- •Bottom:Comp. 0.011 0.011
p r e s s o r and the n o n - v a p o r i z e d w o r k i n g fluid •Atomospheric Pressure
is r e t u r n e d to the inlet of the e v a p o r a t o r .
V ( j ) : V a p o r F. R. from the j -th plate
Figures 1 and 2 i n v o l v e the l o c a t i o n of the F. R. : F l o w r a t e (kg-mol/hr)
 Control of a Distillation Column with a Heat Pump 263

condenser are f i x e d as 99.8 °C and 115.2 plates bounded by the heat pump system
°C, r e s p e c t i v e l y . The r e f l u x and t h e b o i l - a p p r o a c h e s the t o t a l r e f l u x c o n d i t i o n .
up flow r a t e d e p e n d on the heat load re- Systems w i t h no heat load to the side of
m o v e d from t h e e v a p o r a t o r , even t h o u g h t h e columns are c o n v e n t i o n a l systems. After
s i d e - c u t and s i d e - i n p u t p l a t e s are f i x e d . heat exchangers in a heat p u m p are de-
signed, the heat load c h a n g e in normal
operation is not large e x c e p t during the
M O V A B L E R E G I O N OF M A N I P U L A T I V E VARIABLES p e r i o d of s t a r t - u p .

A distillation system with a heat pump C H A N G E IN F E E D C O M P O S I T I O N . Due to the

seems to be v e r y c o m p l i c a t e d at a glance. c h a n g e s of the feed c o m p o s i t i o n , the v a p o r
In g e n e r a l , given the s p e c i f i c a t i o n s of the load from the r e b o i l e r and the c o n d e n s e r in
top a n d b o t t o m c o m p o s i t i o n s and t h e oper- the heat p u m p and the r e f l u x r a t i o c o r r e s -
ating pressure, no d e g r e e of f r e e d o m re- ponding to e x c h a n g i n g heat load at the
mains. B u t , in t h e c a s e of i n t r o d u c t i o n of side-input plate are shown in the same
heat exchanging o p e r a t i o n at s o m e plate, figures. For b o t h s y s t e m s , w h e n the e t h a -
one d e g r e e of f r e e d o m is a d d e d . The rela- n o l feed c o m p o s i t i o n i n c r e a s e s , the t e m p e r -
tionships b e t w e e n the r e m o v e d heat e n e r g y , a t u r e at the s i d e - c u t p l a t e d e c r e a s e s . As
vapor f l o w r a t e g e n e r a t e d in the reboiler the h e a t load at the e v a p o r a t o r d e c r e a s e s ,
and the c o n d e n s e r of t h e heat pump and the v a p o r f l o w r a t e in the r e b o i l e r m a y be
reflux r a t i o are i n v e s t i g a t e d in t e r m s of generated more. The c h a n g e s of the v a p o r
static characteristics. The r e s u l t s are load for S C - 2 are v e r y s m a l l and linearly
shown in F i g s . 3 for SC-l/R and 4 for S C - 2 changed, even if the feed composition
for S C - 2 / R . varies. M e a n w h i l e , t h e c h a n g e of t h e r e f -
lux r a t i o s in b o t h s y s t e m s are a l m o s t i d e n -
H E A T L O A D C H A N G E IN E V A P O R A T O R . For b o t h tical .
s y s t e m s , in g i v i n g t h e feed c o m p o s i t i o n , as
heat load r e m o v e d from SC increases, the CHANGE IN S P E C I F I C A T I O N OF B O T T O M C O M P O S I -
v a p o r load r e q u i r e d in the c o l u m n i n c r e a s e s TION . In t h i s c a s e , the v a p o r loads and
but the reflux ratio decreases. It is r e f l u x r a t i o s h a v e the same c h a r a c t e r i s t i c s
because in the c a s e of large exchanging as t h o s e w i t h the feed c o m p o s i t i o n c h a n g e s .
heat load the o p e r a t i n g c o n d i t i o n in the But the vapor flow rate varies greatly

150f

140

1 1 1 !
' ' 0
0.0 .1 .2 .3 .4
Feed composition Feed composition

Fig. 3 R e l a t i o n s h i p s a m o n g s i d e - c u t heat Fig. 4 R e l a t i o n s h i p s a m o n g s i d e - c u t heat

l o a d , v a p o r f l o w r a t e and r e f l u x l o a d , v a p o r f l o w r a t e and r e f l u x
r a t i o for S C - l / R r a t i o for S C - 2 / R

DCCR-R
264 Y. Naka and E. O'Shima

c o r r e s p o n d i n g to a v e r y s m a l l c h a n g e in the tion and the b o i l - u p flow rate to the b o t -

set point. The r e a s o n is t h a t , as both tom composition, the transfer functions
distillation systems have a p i n c h point were determined by using step responses.
near the b o t t o m u n d e r the given separating The i n i t i a l o p e r a t i n g c o n d i t i o n s are shown
conditions, very s m a l l a l t e r n a t i o n of the in T A B L E 2. T h e y are a little deference
p r o d u c t s p e c i f i c a t i o n c a u s e s the c h a n g e of f r o m T A B L E 1 due to the d e g r e e of accuracy
the gradient of t h e o p e r a t i n g l i n e . design parameters involved in t h e evapo-
r a t o r and the c o n d e n s e r . The c o m b i n a t i o n s
of the m a n i p u l a t i v e and c o n t r o l l e d varia-
DYNAMIC MODEL AND DYNAMICS bles offer implicit decoupling controllers
(Ryskamp, 1981) in the case of a conven-
Dynamics Models tional column. T h e step r e s p o n s e s to the
reflux ratio in 3% increments from the
The d y n a m i c m o d e l of the d i s t i l l a t i o n sys- nominal state were very smooth for both
tems is d e s c r i b e d b e l o w in b r i e f . cases. B u t , the step input to the b o i l - u p
8flow rate w i t h the same increment 3% for
Distillation column. The d y n a m i c simula- both systems caused slightly inverse
tion of a d i s t i l l a t i o n s y s t e m w i t h a heat responses. As the m a g n i t u d e of the inverse
p u m p b e t w e e n a c o n d e n s e r and a r e b o i l e r has r e s p o n s e w a s v e r y s l i g h t , it m a y h a r d l y a f -
been e x e c u t e d b y F e r n a n d e z ( 1 9 8 1 ) . As t h e r e f e c t e d p e r f o r m a n c e of c o n t r o l s y s t e m s .
is no c o n d e n s e r and no r e b o i l e r w h i c h use
outer u t i l i t i e s and the p r e s s u r e c h a n g e m a y A l l of the s i m u l a t i o n case s t u d i e s on d y n a -
affect a heat p u m p s y s t e m and a distilla- mics and p e r f o r m a n c e of control systems
tion column, he u s e d a p r e s s u r e v a r i a t i o n were executed m a k i n g use of the program
model. But, in t h i s p a p e r , as the heat p a c k a g e , "Dynamic Process S i m u l a t i o n ( D P S ) "
p u m p has an e x t e r n a l w o r k i n g f l u i d , such a d e v e l o p e d b y Japan Union of S c i e n t i s t s and
pressure m o d e l is not n e c e s s a r y u n d e r the Engineers.
pressure control with the manipulative
variable of the c o o l a n t f l o w rate. Mass
b a l a n c e e q u a t i o n s and the w e i r e q u a t i o n as C O N T R O L S Y S T E M D E S I G N A N D THEIR PERFORMANCE
a plate m o d e l are considered. The heat
e x c h a n g e r s of the c o n d e n s e r and r e b o i l e r in Control system design
the d i s t i l l a t i o n s y s t e m are of a complete
mixing unit. Control systems are d e s i g n e d b y use of a
method of c o n t r o l s y s t e m d e s i g n based on
Evaporator and condenser in a heat pump partial k n o w l e d g e of c o n t r o l l e d processes
system. In g e n e r a l , as the temperature (Kitamori, 1978). This method tunes con-
difference between hot and cold streams trol p a r a m e t e r s so that the r e s p o n s e to be
should be s m a l l in the e v a p o r a t o r and the d e s i g n e d is a d j u s t e d to the r e f e r e n c e m o d e l
condenser, the heat e x c h a n g e r m o d e l with w h i c h has a d e s i r e d r e s p o n s e . The r e f e r e n -
s e q u e n t i a l s t i r r e d - t a n k s are e m p l o y e d . ce m o d e l is d e n o t e d as ( α Q = Q^=1 )
As the heat p u m p s y s t e m seems to h a v e f a s t -
er response, its m o d e l is described in 2 2
algebraic equations. G
r f e= l/( a 0+ a 1a s + a 2a s +...) (1)
W h e r e σ is the r i s e - t i m e and the v a l u e s of
DYNAMICS {α·} are to s a t i s f y d e s i r e d p e r f o r m a n c e of
control in general as follows: a =0.5,
To design two conventional control loops, .15,o^-O.03, ...
2
A t r a n s f e r function
the r e f l u x r a t i o to the d i s t i l l a t e c o m p o s i - can be w r i t t e n in g e n e r a l form as f o l l o w s :
2
TABLE 2 Operating Conditions G(s) = l / ( a ' + a ' s + a ' s + . . . ) (2)
0 1 2
W o r k i n g fluid ( W f ) : R113 As the g e n e r a l s t r u c t u r e of PID controller
is given b y

S C - 1/R S C - 2/R 2
G (s) = (c +c s+c s )/s (3)
c ( )1 2
The t r a n s f e r f u n c t i o n of the s e t - p o i n t to
(Distillation s y s t e m ) the c o n t r o l v a r i a b l e is
•Feed :F. R. 50 50 1
:Comp. 0 .182 0 .182 a' +a'-s+a ~s~+...
*Top :Comp. 0 .901 0 .899 G*(s) = l/(l+s
n ) (4)
: Reflux r a t i o 4 .33 4 .65
*Bottom:Comp. 0 .014 0 .011
C +C S +S +C
:Total V - F . R. 132 .1 130 .1 0 1 2 " * * *
: Steam Q 8 .50 8 .71 The c o n t r o l p a r a m e t e r s can be obtained
Gs ; so
as for G * ( s ) to be e q u a l to
(Heat p u m p s y s t e m ) for PI a c t i o n
r f (e )
•Evaporator
:Wf-F. R. 130 .0 140 .0 a O0t
a / ά+σ2 ()2 aί α 0= ) (
: Exchanged Q 22 .06 21 .56 * 2^*' 0~ 2 ' 1 ' θ 2 ~ 3^
:P 44 .03 44 .03 The control p a r a m e t e r s are d e t e r m i n e d by
•Condenser u s i n g the m i n i m u m and p o s i t i v e v a l u e of the
:Wf-F. R. 103 .9 114 .8 r o o t s of σ o b t a i n e d from e q . (1) :
.-Circu. F. R. 121 .5 121 .1 c = /a a =c
: Exchanged Q 23 .99 23 .27
:P 62 .37 62 .37 0 'û ' i a'oiaya^-a^l/a

F.R.: F l o w rate
5
(kg-mol/h)
c
2 = a' {a' /a' -aa a' /a'
0 2 0 2 2 12 0
Q : Heat load (x 1 0
6 kJ/h)
+ σ ( ( α ) - α ) }/σ (6)
Ρ : Pressure (χ 1 0 Pa) 2 3
 Control of a Distillation Column with a Heat Pump 265

CLOSED-LOOP RESPONSE the side-cut plate became lower. The tem-

perature difference between the hot and
As mentioned above, there are some d i f f e - cold streams in the evaporator became
r e n c e s of t h e s t a t i c r e l a t i o n s h i p s between smaller w i t h o s c i l l a t i o n and the heat ex-
the r e f l u x r a t i o and the v a p o r l o a d . Let's c h a n g i n g load w a s r e d u c i n g . A l s o , the v a p o r
investigate their differences between SC- flow rate from the c o n d e n s e r in the heat
l/R and S C - 2 / R . M a i n d i s t u r b a n c e s to t h e pump was reducing with small oscillation
s y s t e m s w e r e c o n s i d e r e d as f o l l o w s : c h a n g e s from 9 4 . 7 k g - m o l / h t o 8 6 . 5 k g - m o l / h for 2
in the feed c o m p o s i t i o n and f l o w r a t e . In min and f i n a l l y to 85.4 k g - m o l / h . Conse-
addition, t h e c h a n g e in the s i d e - c u t flow q u e n t l y , the v a p o r to be g e n e r a t e d f r o m the
r a t e w a s i n t r o d u c e d to S C - l / R . reboiler increased corresponding to the
c h a n g e s of t h o s e v a r i a b l e s . Although the
C h a n g e in feed c o m p o s i t i o n . Let's investi- vapor f l o w r a t e r e v e r s e l y d e c r e a s e d at the
gate the c a s e that the feed composition beginning of the response, the control
changes by 2 . 0 % i n c r e a s e of the ethanol performance was very well.
composition from the nominal state. In t h e
case of no c o n t r o l , the top c o m p o s i t i o n w a s C h a n g e in feed f l o w r a t e . The p r o d u c t c o m -
shifting by 3% in 30 min. For SC-l/R, p o s i t i o n s for b o t h s y s t e m s w e r e controlled
though the bottom composition was well- v e r y w e l l as s h o w n in F i g s . 7 and 8. For
controlled, the top composition was gra- SC-l/R, the vapor flow rate generated in
d u a l l y d r i f t e d b y a b o u t 0.4% in 30 m i n . as the r e b o i l e r first d e c r e a s e d f r o m 34.5 k g -
s h o w n in F i g . 5. The side-cut vapor stream mol/hr to 32.8 kg-mol/hr and then in
w a s c o m p l e t e l y c o n d e n s e d in t h e evaporator reverse, r a p i d l y i n c r e a s e d in a b o u t 8 m i n .
of the heat p u m p . The munipulative v a r i - Meanwhile, the v a p o r f l o w r a t e for S C - 2 / R
ables, the r e f l u x r a t i o and the heat load slightly d e c r e a s e d and t h e n g r a d u a l l y in-
to the r e b o i l e r w e r e c h a n g i n g v e r y s l o w l y . creased. For S C - l / R , the vapor flow rate
The v a p o r f l o w r a t e from the r e b o i l e r did f r o m the c o n d e n s e r in the heat p u m p varied
not c h a n g e for a b o u t 20 m i n . The reflux up and d o w n v e r y s l o w l y and t o o k t h e l o w e r
ratio w a s d e c r e a s i n g as w a s e x p e c t e d from value than the i n i t i a l v a l u e in 30 min.
the s t a t i c c h a r a c t e r i s t i c s . The c o r r e s p o n d i n g v a l u e of S C - 2 / R monoto-
For SC-2/R, when the feed composition neously decreased. The r e f l u x r a t i o of S C -
varied from e t h a n o l 1 8 . 2 % to 21.8%, both l/R v a r i e d d o w n and u p v e r y s l o w l y . Mean-
product compositions were well-controlled while t h e r e f l u x r a t i o of S C - 2 / R t o o k the
as s h o w n in F i g . 6. The r e s p o n s e s of the m i n i m u m v a l u e in about 10 m i n . and recov-
munipulative variables were much faster e r e d s l i g h t l y b y 0 . 1 0 % a f t e r 20 m i n .
than those for SC-l/R. Especially the
r e f l u x r a t i o r a p i d l y r e a c t e d to t h e d i s t u r - Change in s i d e - c u t f l o w r a t e . The side-cut
bance. As the ethanol composition in- vapor f l o w r a t e in S C - l / R m a y n o t b e pre-
creases, the t e m p e r a t u r e of the v a p o r from cisely controlled. In the c a s e of d i s t u r -

0.98 4.5
.___Reflux ratio

0.94
eu
ο ε
o
4.0
C4-1 I I
h en
Ο w I Dis t i l i â t e J •H M
TJ 0.90 O «—·
4-» Ω
•Η 3.5
r/3
Ο 40
Û4 0.08
ε
ο
υ Boil-up 35
w
ε
ο 0.04 Fig. 5 C l o s e d r e s p o n s e of
S C - l / R to c h a n g e in
ο Bottoms 30
m feed c o m p o s i t i o n
0.00
10 15 20 25 30 m i n

0.98]
4.5
to
a o
Η 0.94 ρ ε
Reflux ratio I I
.h en 4.0
•H M
û 0.90 Ο <—»
Distillate
40 3.5
ο o
0.08

Fig. 6 C l o s e d r e s p o n s e of
S C - 2 / R to c h a n g e
2 w
Boil-up
35
•H
-M
U
rd
Ο o 0.04
in feed c o m p o s i t i o n
υ o
0.00
Bottoms
30 H
4-1
«
10 15 20 25 30 m i n
 266 Y. Naka and E. O'Shima

Fig. 7 Closed r e s p o n s e of
SC-l/R to c h a n g e in
feed flow r a t e

0 5 10 15 20 25 30 m i n

Φ 0.98
+-> -α
to
H "s
\ 3 H
I Ο
•H 0.94 Reflux ratio H ε
4-1 •H
Ο Cn
I
•H
W
CQ X
û ι—ι
Distillate
0.90
w
o 45
ε Boil-up
o 0.08
u 40
w
ε
o
Fig. 8 C l o s e d r e s p o n s e of 0.04 35
S C - 2 / R to c h a n g e in
feed f l o w r a t e Bottoms
0.00 • • . 1 , 30 .
!0 15 20 25 30 m i n

b a n c e on it w i t h 1 0 % i n c r e a s e , b o t h p r o d u c t r i o r to S C - l / R b e c a u s e the f o r m e r can cor-

compositions could be sufficiently con- responds to d i s t u r b a n c e m o r e r a p i d l y than
trolled but the v a p o r f l o w r a t e generated the latter. In the case of the reflux
in the c o n d e n s e r of the heat p u m p system ratio and heat load to the b o t t o m as main
decreased in l a r g e from 34.5 k g - m o l / h r to manipulative v a r i a b l e s to c o n t r o l two p r o -
20.0 k g - m o l / h r and t h e n g r a d u a l l y i n c r e a s e d duct compositions, two SISO controllers
to 29 k g - m o l / h r . The f l u c t u a t i o n of the could o f f e r good p e r f o r m a n c e against the
side-cut flow rate m a y induce the extremely disturbances on the feed c o n d i t i o n s except
big c h a n g e in the v a p o r load from the re- for the set p o i n t c h a n g e of the products.
boiler. In r e v e r s e , the c i r c u l a t i o n f l o w To m a k e t h e c l o s e d - l o o p r e s p o n s e around the
rate from the p r o c e s s s t r e a m or the flow b o t t o m c o m p o s i t i o n f a s t e r , the f l o w rate of
rate of the v a p o r i z e d w o r k i n g fluid s h o u l d the circulation process stream through the
be carefully controlled. c o n d e n s e r in the heat p u m p or the flow rate
of the w o r k i n g fluid s h o u l d be carefully
In all c a s e s t h e b o t t o m c o m p o s i t i o n s were controlled.
almost i n v a r i a n t . The r e s u l t s w e r e c a u s e d
b y the large d e g r e e of the c i r c u l a t i o n r a t e
a r o u n d the c o n d e n s e r in the heat p u m p sys- REFERENCES
tem. For e x a m p l e , w i t h l a r g e amount of i t ,
the bottom c o m p o s i t i o n held the desired Fernandez, E. (1981). Feasibility
value but a l i t t l e amount of the circula- analysis, dynamics and control of
tion flow made operation very difficult. distillation columns with vapor
If the set p o i n t is c h a n g e d , the circula- recompression. Ph. D. Thesis, The
tion flow r a t e s h o u l d be carefully con- Luisiana State Univ.
trolled. N a k a , Y., M. T e r a s h i t a , S. H a y a s h i g u c h i and
T. t a k a m a t s u (1980). An intermediate
heating and c o o l i n g m e t h o d for a d i s -
CONCLUSION tillation column. J. Chem. Eng.,
Japan, 13, 123-129.
The s t a t i c and d y n a m i c c h a r a c t e r i s t i c s for N a k a , Y., K. B a b a , T. S a t o h and T. T a k a m a -
two types of a d i s t i l l a t i o n w i t h a heat t s u ( 1 9 8 3 ) . A t h e r m o d y n a m i c approach to
p u m p w e r e i n v e s t i g a t e d b a s e d on s i m u l a t i o n . distillation system d e s i g n for e n e r g y
B o t h of S C - l / R and S C - 2 / R p r o p o s e d in this c o n s e r v a t i o n . P r o c . of P A C H E C , K o r e a .
paper have the b r o a d r a n g e s of feasible Ryskamp, C. J. (1981). Explicit vs.
operation, respectively. But, the patterns Implicit Decoupling in Distillation
of o p e r a t i o n in t e r m s of s t a t i c characte- Control. P r o c . of C h e m . Process C o n -
ristics are different with each other as t r o l , Sea I s l a n d , U . S . A .
shown in F i g s . 1 and 2 . As the results
f r o m the d y n a m i c b e h a v i o r , S C - 2 / R w a s s u p e -
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

DYNAMICS AND CONTROL OF AN

ADIABATIC FIXED BED REACTOR WITH
VARYING CATALYST ACTIVITY
+
L. Goldschmidt , L. Hallager* and S. B. Jorgensen
Instituttet for Kemiteknik, Technical University of Denmark,
DK 2800 Lyngby, Denmark

Abstract. The dynamic behavior of an experimental adiabatic catalytic fixed

bed reactor is characterised by the relative importance of its two main dyna-
mical features: the enthalpy transport due to the heat capacity of the gas,
and the enthalpy transport due to gas phase reactant transport. The latter
enthalpy is released by the reaction. Within the operating condition space
five regimes of dynamic behaviour are identified and illustrated with experi-
mental open loop results. The experimental results reveal slow reversible
variations of the reaction rate within two of the regimes. This variation is
probably due to product inhibition.
The characterization of the dynamic behavior is used for development of a
sparse discrete time model structure for multivariable adaptive control within
the faceted operating condition space. The performance of the adaptive control
is illustrated by setpoint tracking experiments covering several of the regi-
mes with significantly varying catalyst activity. The performance of the
adaptive control is very satisfactory, even though the setpoint changes are
performed in relatively rapid succession.

Keywords. Reactor dynamics, Discrete time model structure, multivariable

adaptive control.
INTRODUCTION
During startup and other transient types of opera-
tion a fixed bed reactor may pass through a
number of operating areas where the dynamic cha-
racteristics change significantly. In order to
control the reactor operation during such severe
transients it is advantageous to use feedback
control. In general the control problem in these
cases will be multivariable and require a careful
design in order to be able to handle the varia-
tions in the dynamics. A number of different con-
trol design procedures can be applied to this type
of control problem, but they will probably all
require knowledge of the structure of a process
model and some will require parameter values as
well. One approach may be to develop an a priori
control design, which is robust towards the expec-
ted parameter variations, using the structured
singular value approach (Doyle and Morari (1986)).
Another approach would be to precalculate control-
lers for different operating regions and then
switch between these controller settings as the
process evolves. For both these approaches a de-
tailed model of the process is required, including
knowledge of the expected parameter variations. In
many cases such knowledge is not available. In
this paper an emtirely different approach is taken
in order to aim at obtaining a robust control
design; namely adaptive control.
The adaptive identification and control utilises
recursive estimation of the parameters in a pre-
specified modelstrueture followed by recursive
calculation of the control parameters according to
a selected control design criterion. The advanta-
ges of this procedure are thus twofold: The model-
structure may be specified using general chemical
engineering knowledge in a relatively simple pro-
cedure, and the control design criterion may be
Present addresses:
+ Danish Energy Agency. * CELL Automation.
Landemaerket 11. Djursvang 7b.
DK 1119 Copenhagen Κ DK-2620 Albertslund.
chosen to suit actual needs, e.g. a robust design
migth be chosen and the uncertainties in the para-
meter estimates used for locating caution in the
control actions where it is needed. In this paper
an optimal control design is applied, where the
the weigths in the control criterion are user
specified, thus the user must choose the weigths
to obtain a cautious control design.
The purpose of this paper is for a specific fixed
bed reactor to investigate types of dynamic beha-
viour experienced over a large range of operating
conditions, and to investigate the behaviour of an
adaptive controller tested under relatively deman-
ding transient conditions. The structure of the
paper is as follows: the following section con-
tains a brief description of the reactor setup,
and a description of the essential dynamic fea-
tures of the experimential process illustrated by
selected transient open loop responses. The adap-
tive control design is presented in the following
section, where a relatively simple discrete time
model structure is given. This structure is deve-
loped based upon knowledge of the essential dyna-
mic features of the reactor. Then identification
and control results during a sequence of setpoint
changes are given and discussed. Finally the con-
clusions are drawn.
REACTOR SETUP AND REACTOR DYNAMICS
First the experimental process is briefly presen-
ted; then the principal process dynamics are dis-
cussed using a dynamic quantity interaction diag-
ram (DYQUID). The operating region is subdivided
into five operating regimes with qualitatively
different dynamic behaviour. The basis for the
subdivision is the reaction rates in the reactor.
The dynamic behaviour within three of the regimes
are illustrated by experimental results.

267
268 L. Goldschmidt, L. Hallager and S. B. J0rgensen
The experimental investigation is carried out in a
nearly adiabatic pilot plant fixed bed reactor
shown in figure 1. Hyrogen in large excess is
oxidised irreversibly over Platinum on an alumina
carrier. The operating region covered in this
paper is inlet temperatures between 35 and 90.,
inlet Oxygen concentrations between 0.0 and 3.0
mole%2 , and a total mass flow rateof 1.2-3.8
mg/cm /sec. The pilot plant is described in more
detail in Hansen and Jorgensen (1976a).
Thermocouple
Gas sample
outlet
ΣΙεΛ : Sketch of experimental fixed bed reactor
set up. Length 0.50m and diameter of 0 098m 15
thermocouple measurements and gas concentration
measurements from two out of six gas sample out-
lets.
The dynamic behaviour of this reactor with the
highly exothermal oxidation reaction may be de-
scribed by considering the essential quantities in
the process: Oxygen mass, Enthalpy and total mass
and their interactions. Of the quantities listed
above the residence time is very short for the
masses in the gas phase, while the effective ther-
mal residence time is significantly larger and
thus dominates the dynamic response. The effective
thermal residence time may be calculated based
upon a pseudohomogeneous enthalpy balance for the
gas and solid phases. For the above mentioned
operating region the thermal residence time varies
from 10 to 36 min, mainly dependent upon the total
mass flow rate. The interactions between the sig-
nificant quantities may be qualitatively il-
lustrated in a DYQUID (DYnamic Quantity Interac-
tion Diagram) as shown in Fig.2a. The main inter-
actions between the Oxygen mass flow and Enthalpy
is naturally due to the exothermic reaction. The
interaction between the total mass flowrate, oxy-
gen, and enthalpy are also shown. This DYQUID
clearly illustrates that inlet disturbances in
total mass and Oxygen mass flow rates will affect
the reactor exit almost instantaneously, whereas
disturbances in inlet temperature will have to
propagate to somewhere into the reactor before it
through the coupling (i.e. the reaction) to the
Oxygen mass may affect the reactor exit. The main
effect of an inlet temperature disturbance may
however not be seen until the thermal wave gene-
rated by the inlet temperature disturbance reaches
the exit. Clearly the the strength of the interac-
tions and thus the actual effect of the dis-
turbances depend upon the reaction rate profile in
the reactor.
For this experimental reactor no simple accurate
reaction rate expression is known. An n'th order
irreversible rate expresssion is shown in the text
of Fig.3. In this expression it is known that the
dimensionless activation energy may decrease 30%
over the reactor with a consequent increase in the
Damkohler number (Hansen and Jergensen 1976b). At
steady state conditions a set of constant para-
meter values may however be used to obtain appro-
ximate reaction rate profiles for the reactor.
Oxygen
k Enthalpy^ • ^
+
FiR.2 : DYQUID (DYnamic Quantity Interaction
Diagram) for nearly adiabatic fixed bed reactor
where Hydrogen in large excess is oxygenated.
a: Insignificant water adsorption.
b: Significant water adsorption and reaction
rate inhibition
The quantities are labelled: F Total mass flow
rate. W gas phase water mass flow. Q adsorped
water. 0_ gas phase Oxygen mass flow. H pseudo-
homogeneous gas and solid phase enthalpy. Quantity
flows are shown in thick line. Quantity holdups as
oblong boxes to indicate the axial coordinate.
Interactions in thin line.
Four profiles calculated from experimental data
are shown in Fig.3. The dimensionless reaction
rate applied in this figure, may be exemplified by
a fictitiously constant reaction rate of 1 over
0.1 of the reactor length which would give a
conversion of 0.1 of the reactant. A number of
experimental steady state temperature profiles are
shown in Fig.4. Now a number of regimes of dynamic
behaviour for sections of - or the entire - reac-
tor may be sketched based upon the magnitude of
the reaction rates and the dynamic characteristics
of the reactor. The relationship between the dy-
namic behaviours and the reaction rates depends
upon the exothermicity (the adiabatic temperature
rise) and thermal sensitivity of the reaction rate
(the activation energy). Especially when high
reaction rates are encountered more regimes may be
present simultaneously in the reactor.
A: Extinguished reaction: R < 0.1
At very low reaction rates only a small part of
the reaction enthalpy is released; thus the reac-
tion rate stays nearly constant , and the reactor
behaves nearly like a thermal delay.
B: Moderately sensitive reaction: 0.1 < R < 0.5
Axial temperature profiles in this regime appear
nearly linear as seen in Fig.5a. A steady state
gain of nearly one in Τ /T. for all ζ is seen by
comparing curves 1 and 2. Pulseresponses to inlet
disturbances in this regime are shown in Fig.6a.
It is clear that the thermal delay dominates the
reactor dynamics with only a small inverse respon-
se visible. The sensitivity towards mass flow and
Oxygen concentration changes are small. The dis-
tributed effects of the latter two changes show
in the immediate temperature changes along the
reactor axis.
 Adiabatic Fixed Bed Reactor with Varying Catalyst Activity 269

0 .0 5 Π ΣΟ âTO
l CXrrol
in e
Fig.4 : Qualitative sketch of reaction rate re-
Fig.3 : Reaction rate as a function of reactant gimes within the operating condition space, at a
concentration and temperature. Isotherms are shown mass flow rate of 2.6 mg/cm /s. Labels for the
through origo. The curves show approximate reac- reaction rate regimes are:
tion rate paths from the rector inlet to the A: Extinguished or
outlet, where the path is terminated. The numbered E: Burned out reaction : 0.1 > R(C,T)
paths are calculated from steady state temperature Β : Moderately sensitive r.: 0.5 > R(C,T) > 0.1
profiles and inlet concentration, using the rate C: Sensitive reaction : 2.5 > R(C,T) > 0.5
expression: D: Labile reaction : R(C,T) > 2.5
R(C,Τ)=D *exp(g*(1-1/T))*C**n C: The sensitive reaction: 0.5 < R < 2.5
m In this range of reaction rates the enthalpy
released by the reaction plays an important role
for the sensitivity of the reactor, as seen from
the more nonlinear axial temperature profiles in
Fig.5b and c. Curves 1 and 2 in Fig.5b show that
1
the steady state thermal gain Τ /T. varies betwen
one and four, with a maximum close ?© ζ — 0.6. The
pulse responses in Fig.6b show now clearly the
inverse response characteristics, both for tempe-
rature and mass flow rate disturbances. The sensi-
tivity towards the distributed forcings has in-
creased significantly. The positive mass flow
disturbance cools the first part of the reactor,
whereas more enthalpy is released in the late part
of the reactor due to the resulting increased
Oxygen concentration there. The inverse response
characteristics may be illustrated by using the
DYQUID in fig.2a. When say a positive thermal
wave, which travels slowly due to the significant
thermal capacitance, has reached a region in the
reactor with high reaction rate, then more reac-
tant will be consumed than before. Hence the down-
stream part of the reactor receives less reactant,
since the reactant travels fast, and thus less
enthalpy is released, temporarily giving a lower
temperature.
The steady state profiles from regime C in Fig.-5b
are obtained at low inlet temperatures where the
reaction rate appears to be of low order, leading
to a rapidly increasing reaction rates towards the
outlet. At these inlet conditions the product
water is known to adsorp upon the carrier and
believed to inhibit the reaction. This adsorption
and inhibition diminishes clearly with increasing
temperature. Step responses to inlet temperature
disturbances under fairly similar conditions are
shown in Fig.7 a and b. The positive step change
in Fig.7a shows response characteristics similar
0 Nornwljjw p 7.
l axUl diitlaro 0.
5 to those of the pulse response in Fig.6b, except
F for a overshoot with a peak about 60 min after the
ig-5 : Experimental steady state axial tempe- step has been applied. In the response to the
negative step disturbance in Fig.7b the inverse
Β :1 1 2 3 4 characteristics are much more pronounced than in
T. °C 82 87 82 82 the positive step response, furthermore a slow
c" .25 .25 .30 .25 transient dropoff of the outlet temperature is
1 2 3 seen. Both these effects are probably due to va-
T.
1 °C 70 79 70 rying catalyst activity. With the high inlet tem-
c" .98 1.02 1.51 perature the catalyst is more reactive, giving the
1 2 3 much more pronounced inverse response. When the
T.
1 °C 35 35
4 5 6 7
temperature drops water is adsorped upon the cata-
35 35 35 35 35
c" .06 .93 1.55 1.77 lyst, thus decreasing the activity and then the
in
270 L. Goldschmidt, L. Hallager and S. B. J0rgensen

tempereature drops further, but more slowly. The
peak in Τ- Q in Fig.7a may be due to released heat
of adsorption, in this case however there does not
seem to be an increased catalytic activity, within
the time horizon shown.
D: The labile reaction 2.5 < R
High reaction rates are only possible in a small
part of the reactor, giving relatively steep tem-
perature profiles in this part. Steady state pro-
files including this range are shown in fig.5b and
c. Fig.7c shows responses to a positive step in
the inlet temperature at various axial positions
in the reactor. A relatively high sensitivity is
seen at ζ = 0.8 where the steady state thermal
gain is almost 10. Higher sensitivities may easily
be encountered, e.g. under the conditions of curve
7 in fig.3c the steady state thermal gain in
nearly twice the above gain. Under these condi-
tions the inevitable small variations in inlet
conditions make open loop operation of the reactor
difficult, and the measurements become much more
noisy. At ζ = 0.8 a slow transient is also seen,
again probably due to adsorption and inhibition
effects giving rise to a shift in the temperature
profile within the reactor. This shift will only
be effective in the zone where the catalyst adsorp
water.
The eksperimental results reveal that product
adsorption and inhibition play a role for the
dynamic description of the reactor. A DYQUID in-
corporating this effect is shown in fig.2b. Here
the water is assumed to be produced adsorped upon
the catalyst surface. The water adsorption, and
hence the amount of adsorped and gas phase water,
depends upon temperature, and hence upon the en-
thalpy. The adsorption also releases enthalpy. The
total flow rate depends upon the amount of water
being ad- or desorped. In order to investigate the
reversibility of the reaction rate inhibition an
80 hour experiment was performed covering the
regimes with water adsorption effects. The inlet
temperature was first stepped up allowing the
reactor to reach steady state after each step
change, once full conversion was reached the pro-
cedure was repeated in reverse order. The step
responses in fig. 7 are from this experiment. Fig.
7a followed 36 hours after fig. 7b. The main
result of the experiment is that the water ad-
sorption effects are indeed reversible. Their
dynamics however are relatively slow, in some
cases transients lasted up to 14 hours, before a
new steady state was reached.

Ln

Fig.6 : Pulse responses to upstream disturbances

in a:temperature, b:Oxygen concentration and
c:mass flow rate. Steady state conditions: T. -
82°C and F = 2.66 mg/cm /s.
n
I) C. = 0.25 II) C. - 0.75
Units?T IK, C: 0 . o l % m o l e , F: 0.1 mg/cm /s

E: The burned out reaction.

When all reactant has been consumed by the reac-
tion, the reactor again behaves like a themal
delay in this part. The step response at ζ — 1.0
in fig.7c is clearly from a nearly burned out part
of the reactor, where the inverse response charac-
teristic caused by the upstream section of the
reactor is clearly visible, but the steady state
thermal gain is near one. The thermal sensitivity
to a concentration disturbance is determined by
the adiabatic temperature rise, with a delay cau-
sed by the upstream length of the burned out
section. The thermal sensitivity to flow changes Fig.7 : Step responces in inlet temperature at
will be zero untill the flow is increased suffici- n
C = 1.40 and F = 2.66 mg/cm / s . a) T. 56 to 51
ently to extend the reaction front to the measure- ° C b) T. 51
n
to 56, and c) T. 61 to^o. l
ment position. Units:Τ 2&? C: 0.01 %mole
 Adiabatic Fixed Bed Reactor with Varying Catalyst Activity 271

ADAPTIVE CONTROL DESIGN shown in Table 2. An open loop identification is

The purpose of control in this paper is to be able first carried out with PRBS pertubation of the two
to carry out a setpoint tracking experiment on the inputs, after 90 min the loop is closed and iden-
fixed bed reactor starting at relatively mild tification continued for 50 min. Then the pertuba-
operating conditions and working towards a labile tions are removed and the reactor allowed to reach
reaction. Since the process dynamics change signi- a stedy state. Thereafter the setpoint changes are
ficantly during such an experiment, this is one applied as changes in the desired values for the
example where adaptive control seems to be a well integral states. The changes are filtered through
suited approach. The adaptive controller used first order filters with a time constant of 0.95.
performs RELS estimation in structured ARMAX mo- Thus giving a rise time for the setpoints which
dels and assymptotic deterministic optimal control is about twice the thermal residence time at these
using the estimated model parameters as desribed conditions. When the reactor reaches the desired
by Hallager et al, 1985. The procedure for desig- setpoints, new changes are introduced.
ning measurement, control and ARMAX model
Representative experimental measurements are shown
structure has been given in Jorgensen et al. 1985.
in Fig. 8a. Prediction errors, outputs and forget-
This procedure is based upon the information con-
ting factors for three of the outputs are shown in
tained in a DYQUID, and knowledge about approxi-
Fig. 8b, and parameter values for the six'th
mate ranges for delays present in the process. A
output model equation in Fig. 8c. Initially during
model structure for the fixed bed reactor has been
the identification the prediction errors for the
developed in the above reference, therefore this
three outputs shown decrease, but once the loop is
step will not be repeated. Below the applied model
closed fairly large control signals are generated
structure is given. This structure is based upon
which added to the pertubations, creating very
measurement of six axial temperatures concentrated
noisy outputs and measurements. When the per-
towards the outlet of the reactor in order to be
T T
able to describe the reactor behavior with steep
tubations are removed the desired setpoints are
reached quite fast compared to the thermal resi-
profiles there: Τ τ T 0.7'Jo.9i l . û dence time. The output variances during the later
and upon usage of two manipulated variables Τ and
part of the identification are well explained by
C . The measurements and inputs are filtereS by
the model as judged from the prediction error
using differences between two samples adjacent in
variances.
time. The ARMAX model is :
Then a combined change of the two setpoints is
y(t) = A ( t - 1 ) + A y(t-2) + A y(t-3) +
iy 2 3 introduced, the inlet temprature setpoint is re-
duced by -5 Κ and the exit increased by 40 K. The
+ B ^ t - l ) + e(t) + C ( t - 1 )
ie controller is able to bring the reactor to the
desired values. The response is however relatively
where e(t) is discrete time white noise. The sluggish, due to the opposite changes in the two
structure of matrices are as follows: setpoints and to the effects of the adsorping
water. The negative change in the inlet tempera-
A A A Β C ture creates a negative thermal wave through the
y
1 . . .7 7 T.. XT X. ... . reactor, which the relatively simple setpoint
y Χ Χ Χ Χ .Χ controller does not predict, thus causing the
y^ XX Χ Χ Χ . .Χ. . . delayed response of the outlet temperature. The
y^ .XX Χ Χ. . . .Χ . . .Χ. . final slow approach to the setpoint is probably
y^ . .XX Χ Χ. . .Χ Χ. due to a simultaneous deactivation of the catalyst
yj? . .XXX Χ. . . .XX. .Χ Χ by the adsorpbing water. This inhibition is indi-
cated by the continued increase of the inlet Oxy-
where Χ represents parameters to be estimated, and
gen concentration. This response must be consi-
dot (.) represents parameters preset to zero. The
dered to be quite satisfactory since the outlet
model is developed for a sample interval of 2
temperature increases with almost 200% based upon
minutes. The model structure aims at describing
information from a linear adapting model. The main
convective enthalpy transport with superimposed
parameter changes are around 20% and occur in the
dispersion in the first subdiagonal elements in
second column of B^ for the three last outputs.
the Α-matrices, and at describing the inverse
When the setpoint is reached a new stepoint change
response effects with the second and third subdia-
of 20K is introduced this time only in the outlet
gonal elements in the A matrices. The effect of
temperature. The response of the outlet tempera-
the short gas residence time is also seen in the
ture is relatively fast but oscillatory. The last
secomd column in Bl, which describes the in-
part of the reactor is now in regime C, and thus
fluence of the concentration control. The effect
quite sensitive. However at 565 min a data trans-
of slow water dynamics is with this model
mission in the sampling system failed. The reini-
structure to be handled by slow changes in the
tiallization an restart of the sampling took
relevant parameter estimates.
around 20 min and introduced disturbances in the
inlets, the effect of which can be seen in fig.8,
EXPERIMENTAL RESULTS AND DISCUSSION after the constant period which shows the fallout
For the sequence of setpoint changes to be carried period. The controller brings the process to a
out two outputs are integrated: y_ and y^, repre-
senting the in- and outlet temperatures respecti-
vely. Thus the inlet temperature may only be used
Table 2:
for dynamic compensation during the setpoint chan- OPERATING AND PERFORMANCE DATA FOR SERVO EXPERIMENTS.
ges. The control weigths and other tuning parame- Dist.rejc. Disturbance rejection steady state control.
ters are given in Table 1. The experiment is T n . n + n n to C h a n g e of s e t p o i n t for T n . n t a k i n g the
initiated with the reactor in open loop steady p r o c e s s to t h e l i s t e d s t e a d y s t a t e .
state operation in regime B. This condition and
those towards the end of each step change are Expra. A c t i v i t y ., O p r . c o n d .
Tinl. Cinl.
Rise
Time.
Resp.
time.
Over- Decay
shoot. Ratio
uOA cm cp .l
Table 1:TUNING PARAMETERS FOR SERVO EXPERIMENTS. °C %mole min min °C c
Time Vo Control Weights Identification 34.1 0.60 0
1 Tl.0+40 to 29.5 1.68 204 204 0 crit 0
min Outputs Intgr.Outputs Τ
TO.0-5
0.0 0.0
2 Tl.0+20 to 29.3 1.85 70 - 7 -
001 6*5. 6*0.001 0.25 0.25 4.0 2.0 Dist.rejc. 29.3 2.09 210 34 3/4 0
690 6*0.001 0.25 0.25 8.0 4.0 3 Tl.0+38 to 29.1 2.37 49 84 5 1/3 1
720 6*0.001 0.25 0.25 2.0 1.0 4 Tl.0+40 to 29.1 2.50 42 70 k 1/1 8
780 6*0.001 0.25 0.25 1.0 0.5
272 L. Goldschmidt, L. Hallager and S. B. J0rgensen
stable operating point with a very oscillating re-
sponse with a decay ratio of 2/3. These oscil-
lations may in part be due to the size of the
disturbances created by the fallout, and the ina-
bility af a linear controller to handle such dis-
turbances. The controller weigths were changed
three times during this transien with- out much
effect. After each change the Riccati equation was
solved for ca. 16 iteration before the new control
weigths were applied. The parameter values for
outputs 5 and 6 change significantly during this
setpoint change and fallout disturbance, as seen
in fig.8c. During the fallout disturbance some
parameters, e.g. ^ undergo significant varia-
tions. This is also seen in the color of the model
r •— ν H I tHi
^ ^ A/\k Jlllllh-i^yiijKJlJuJstat.
! ,-AV- - -ν ipy • »
_/Ι/ΙΛΑ -
KA-r,
: -3
iil|tl«|J -ι - τ
f operating conditions with large variations in the
300 6CO
Fig.8 : Closed loop responces to a succession of
set point changes. The initial identification is
carried out in open loop. The control is initiated
at Lc. The set point changes and inlet conditions
are shown in Table 2.
a) Selected measured variables. All temperature
curves have common zero. Units:T: 10K, C:0.2 %mole.
b) Output, prediction error and forgetting factor
for measurements at ζ = .5, .9, and 1.0.
Units: Outputs (T(t)-T(t-l)) IK, pred. err. IK,
Forg. fact. 0.1.
c) Parameter estimates for six'th output equation.
Units: 0.1
noise i.e. the c parameter for equation 6. Note
that the prediction error for both output 5 and 6
seem to decrease just prior to the fallout distur-
bance and after the fallout disturbance has been
handled seem to obtain almost equal or lower val-
ues now with some of the parameters, e.g. a^ ^,
having quite different values. This behavior may
be an indication of not having a pertubation which
is sufficiently rich over the frequency range of
interest, or of an overparameterization of the
model.
The responses to the following two setpoint chan-
ges take the reactor outlet section into the dyna-
mic regime D. Taking the fairly large changes of
40 Κ into consideration, the first response is
quite satisfactory, being relatively fast and
exhibiting dampened oscillations. The parameter
changes during this setpoint change are most sig-
nificant for output 6. These are seen in fig. 8c
where the parameters vary more than before but
seem fairly stable.
The final setpoint change is accomplished by a
relatively small change of the inlet concentra-
tion, indicating the very large sensitivity of the
outlet temperature in this regime. Here the output
variance is increasing as is the noise level, see
the prediction error in fig.8b for the six'th
output. The forgetting factor also indicates that
a significant amount of unexplained information is
arriving since the filter is opening to the limi-
ting value of 0.8 fairly often. The parameter va-
lues for this output exhibit variations not seen
earlier to the same extent. In this response the
1 1 output temperature oscillates with relatively
ipwriwm ^ short period, this, oscillation may be due to the
6 output feedback. At this point the erxperiment was
terminated.
CONCLUSIONS
The adiabatic fixed bed reactor setup exhibits
relatively large changes in the process dynamics
over the investigated operating condition space.
This is evidenced by the large gain variations,
A
?
-
• ·
-
yS ψψ
«6E
5 T
ι ^
the variation of the inverse response and and the
A T
slow reversible influence of the adsorbed product.
The experimental investigation of the relatively
. r n f i - j i - n ," · rapid succession of setpoint changes demonstrates
that the employed model structure is indeed feasi-
- State ble for adaptive control over a wide range of
process dynamics. The experiment also demonstrates
that the applied control indeed is quite robust,
being able to handle the fairly large fallout
disturbance. In addition the experiment shows that
the fixed bed reactor setup indeed seems to be a
reasonable test bench for multivariable adaptive
control.
REFERENCES
Doyle J.C. and M.Morari (1986): A Unifying
Framework for Control system Design under
Uncertainty and its Implications for Chemical
Process Control. Chemical Process control III,
Asilomar California.
Hallager L. and S.B.Jergensen (1981): Experi-
mental investigation of self-tuning control of
a gas phase fixed bed catalytic reactor with
multiple inputs. Preprint 8. IFAC World Con-
900 min gress Kyoto, paper 106.2.
Hallager L., L.Goldschmidt, and S.Β.Jergensen
(1984): Multivariable adaptive identification
and control of a distributed chemical reactor.
Large Scale Systems, vol. 6, pp323-336.
Hansen K.W. and S.B.Jergensen (1976a-c): Dyna-
mic Modelling of a Gas Phase Catalytic Fixed-
Bed Reactor - I-III. Chem.Engng.Sci., 31, 579-
598, and 473-479.
Jergensen S.B., L.Β.Goldschmidt, and L.Halla-
ger (1985): Sparse process modelling for ro-
bust multivariable adaptive chemical process
control. 7. IFAC/IF0RS Symp. on Ident. and
System. Parm. Est.,York.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

ESTIMATION OF THE DYNAMIC BEHAVIOUR

OF A FIXED-BED REACTOR THROUGH
FILTERING
C. Canavas
Institut fiir Systemdynamik und Regelungstechnik, Universitat Stuttgart,
Stuttgart, FRG

Abstract. During operation of fixed-bed reactors in industrial practice real-time in-

formation about the spatial temperature and concentration profiles is necessary for
optimization, safe reactor operation and control. Extended Kalman filtering is a power-
ful method for on-line estimation on the basis of a priori nonlinear models and experi-
mental measurements. This method is applied to estimate the dynamic behaviour of a
fixed-bed reactor. The reactor is described by a quasi homogeneous nonlinear distributed
parameter model. The spatially discretized reactor model is used for the design of a
continuous extended Kalman filter with temperature measurements. Efficiency of the fil-
ter is tested in the case of error in the initial conditions or in the model parameters.
The influence of spatial discretization and sensor location is investigated. Implica-
tions for possible simplifications, effective realization of the estimator and reduc-
tion of computational effort are then discussed.

Keywords. Chemical industry; distributed parameter systems; filtering; fixed-bed reac-

tor; Kalman filters; modelling; nonlinear filtering; nonlinear systems; partial diffe-
rential equations.

INTRODUCTION measurement technique is the fact that a priori

The operation of distributed parameter systems in
industrial practice is often characterized by the
necessity to determine the spatial profiles of
state variables in a real-time scheme. Applied to
the fixed-bed reactor this prerequisite consists
in defining the axial and radial profiles of tem-
perature and product concentrations. Real-time in-
formation about the above variables is necessary
for yield optimization, safe reactor operation and
control. If the source of this information is the
experiment itself, two major difficulties are en-
countered. The first results from the fact that
measurement data generally contains experimental
error due to sensor location, sensor failure,
noise during signal processing etc. The second dif-
ficulty is related to the availability of experi-
mental data along the catalyst bed. Most industri-
al reactors are equipped with a limited number of
sensors (in most cases only one or two thermocoup-
les). It is virtually impossible to measure the
concentration of reactants and reaction products
along the catalyst bed due to difficulties in
sample acquisition. Even if fluid samples are avail-
able (e.g. at the reactor outlet), usually they
can only be processed with long time delays - for
instance, a gas chromatograph analysis of a single
sample often takes several minutes. This time in-
terval is longer than the time transients of the
reactor; therefore chromatographic measurements
can not be used on-line, to control reactor dyna-
mics,
A powerful tool to reconstruct the transients of
the spatial profiles of all state variables is the
Kalman filter (Kalman, 1960; Gilles, 1985). This
method compensates both the restraints in collec-
ting experimental data and the problem of experi-
mental error by utilizing a priori knowledge of
the process in the form of a mathematical model.
A characteristic feature of this model-aided
stochastic information, concerning 1) modelling
inaccuracies, 2) measurement errors, 3) deviations
in the initial conditions, is accounted for in the
estimation algorithm. The theory of state estima-
tion by means of filtering is relatively well de-
veloped (Gelb, 1974), including also nonlinear fil-
tering (Krebs, 1980) and distributed parameter
state estimation (Ray and Lainiotis, 1978). How-
ever, it is often difficult to apply these methods
in chemical reaction systems with distributed pa-
rameters, such as the fixed-bed reactor. This is
due to the fact, that the complex, strongly nonli-
near nature of such processes considerably compli-
cates sufficiently accurate modelling. If a rigo-
rous model of the spatially distributed interac-
tion between chemical reactions and physical
transport phenomena is constructed, the resulting
order of the estimator is so high, that computa-
tion effort exceeds the limits set by a real-time
realization scheme. Therefore, it is necessary to
introduce appropriate simplifications. Computation
effort also depends on the numerical method used
to solve the resulting nonlinear partial differen-
tial equations. Hence, the procedure chosen for
approximating the axial profiles of model and fil-
ter variables is an important factor for the im-
plementation of the estimator. Modelling assump-
tions and spatial approximation should also be
considered in selecting the appropriate filter al-
gorithm and in formulating the goals of the fil-
ter implementation. An additional aspect in the
design of a filter for the fixed-bed reactor is
the spatial positioning of the sensors. The loca-
tion of the sensors should be selected in a way to
ensure maximal amount of information about the
current state of the reactor. However, this "opti-
mal" location refers always to the spatial appro-
ximation used. Therefore the filter efficiency
should be tested with respect to the interaction
among the factors mentioned above.

273
274
In the present work a stochastic estimator based
on the extended Kalman filter technique is de-
signed for a nonadiabatic fixed-bed reactor using
a nonlinear, quasi homogeneous model of the process.
First the spatial profiles of temperature and con-
centration are piecewise approximated by second
order parabolas (Eigenberger and Butt, 1976). Then
the extended Kalman filter algorithm is used to
estimate the dynamics of the reactor during start-
up. Our aim is to test the efficiency of the fil-
ter in the case of error in the initial conditions
of the state variables, as well as in the presence
of a deterministic error in the model parameters.
In both cases the number and the location of both
the sensors and the grid points are taken into
consideration as factors possibly influencing fil-
ter performance. In order to assure the practical
relevance of the study, experimental measurements
obtained from a pilot-plant fixed-bed reactor are
used for all filter tests.
REACTOR MODEL
The reactions to be considered are the partial oxi-
dation of methanol to formaldehyde by means of air
oxygen and the consecutive oxidation of formalde-
hyde to carbon monoxide:
CH 0H + \ 0 -> HCHO + H 0 + -AH )
3 2 2 R1
HCHO + £ 0 CO + H 0 + (-AH )
2 2 R2
Both reactions are strongly exothermic and are
conducted in an oil-cooled, stainless steel reac-
tor tube. The reactor is packed with pellets of an
iron oxide-molybdenum oxide industrial catalyst.
The formulation of a mathematical model for the
system considered depends on various chemical or
physical phenomena, which are accounted for in the
model. If the two-phase character of the fixed-bed
reactor is to be preserved, mass and energy con-
servation equations should be formulated separate-
ly for the catalyst solid and the reactants gas
phase. However, our preliminary experimental data
indicated that in the investigated range of opera-
tion conditions there is no considerable resistan-
ce to mass and heat transfer between the phases.
We assume therefore that the reactor can be de-
scribed satisfactorily as a quasi homogeneous
medium. This assumption agrees with experimental
and computational results published by other
authors (Schwedock, 1983) and reduces the complex-
ity of the reactor model.
In comparison to the relatively slow temperature
transients, mass transfer and reaction phenomena
can be considered as quasi stationary. Therefore,
the dynamic character of the model is expressed
in the energy balance equation, whereas the mate-
rial balance equations for methanol and formalde-
hyde are assumed to be in a pseudo-steady state.
According to a common practice for industrial re-
actors with high gas velocities, a plug-flow of
the gas through the reactor tube is assumed and
diffusion in the material balance equations is
ignored. Further, orthogonal collocation is app-
lied to approximate radial temperature and concen-
tration profiles. Thus, an one-dimensional reactor
model is obtained, which comprises one partial
differential equation of the parabolic type for
the energy balance and two quasi stationary mate-
rial balance equations. In terms of normalized,
dimensionless variables and parameters these equa-
tions are: 2 (5) in the axial direction (Eigenberger and Butt,
8X, 3X 3 X
1 ( xx ) =r D a +
Da r
at 3Ζ Ί7 l~ lK l l 22
C. Canavas
3x
x ( 0 , t ) = x ( t ) ; - ^n = 0; x (z,0)=x (z)
1 l o 1 1
9X
2= -Da r
3 1 ; x 2(0,t) = x 2( ot ) (4)
3X
0 r D ra x( 0 = î X (t t
Da (5)
4l " 5 2 > 3 ) 3o )
The reaction rates are given by the following
equations:
a) main reaction (methanol oxidation)
r ( x , x ) = e χι x
1 1 2 2
b) consecutive reaction (formaldehyde oxidation)
x
1
1
r (x^,x ) (7)
2 3 e x
0
The identification of the parameters of the above
model has been the subject of previous studies
(Canavas, 1984a, 1984b). The reactor operation
conditions considered as reference for the present
study and the corresponding values of the normali-
zed, dimensionless model parameters are given in
(1) Table 1. In Fig. 1 temperature profiles at various
time instants are given for the starting-up process
(2) of the reactor. The simulation results match very
well the experimental data and show that the model
is able to describe the dynamic reactor behaviour
with a sufficient accuracy. A large amount of con-
centration measurements collected for comparison
confirms the validity of the model over a wide
range of reactor operation conditions.
FILTER DESIGN
Reported investigations concerning filter estima-
tion of the dynamic behaviour of fixed-bed reac-
tors are mostly based on linearization of the mo-
del at the steady state (S^rensen, J^rgensen and
Clement, 1980; Clement, Jtfrgensen and Stfrensen,
1980; Jutan et al., 1977) or on an early linear
approximation of the nonlinear reaction rate terms
by using measured data (Alvarez and Stephanopoulos,
1982). Such procedures, however, may considerably
reduce the validity range of the estimators,
especially if reactor transients during start-up
or during strong disturbances of the operation
conditions are to be evaluated. In a recent publi-
cation Windes, Ray and Cinar (1984) formulated the
filter estimation problem using the full nonlinear
distributed parameter model (i.e. before spatial
approximation), derived an extended Kalman filter
by approximation about the current estimate and
then solved the equations by using orthogonal col-
location. This procedure, however, resulted in a
large number of differential equations for the
filter algorithm. In order to assure a realistic
computing effort for possible real-time applica-
tions the authors solved the filter equations off-
line at a nominal steady state. Therefore only
steady-state gains were used.
In this paper an extended Kalman filter with time
varying gains is designed for the spatially dis-
cretized reactor model. First a finite difference
method with non-equidistant space steps (a modi-
fied Crank-Nicolson technique) is applied to dis-
cretize the partial differential equations (3) -
1976). An advantage of this method is the saving
of computation time, since a set of tridiagonal
(3)
matrices has to be inverted instead of a set of
 Dynamic Behaviour of a Fixed-bed Reactor Through Filtering
full matrices as in the case of orthogonal collo-
cation. Further, the axial grid points can be po-
sitioned according to the needs of the problem
considered, whereas in the case of orthogonal col-
location they are determined by the zeros of the
orthogonal polynomial, which are not necessarily
associated with the expected solution profiles. In
applying the modified Crank-Nicolson technique fi-
nite difference approximations for the axial dif-
ferential quotients are derived by assuming that
the profiles of the dependent variables can be
piecewise approximated by second order parabolas.
Thus one ordinary nonlinear differential equation
for the energy balance is generated at each grid
point, whereas the material balances are expressed
by a set of coupled nonlinear algebraic equations:
dx = ; 0=
(8)
HT 1ΐ(*1·*2·*3*Κ> *1< ) *1
dx
2
dt" 2^-1 '—2 ) (9)
dx
3 (10)
The reactor model is used to determine a continu-
ous-discrete extended Kalman filter. The discrete-
time measurements are given by the equation
1,2,... (H)
{v.} is a white randon sequence of zero mean
gaussian random variables with associated co-
variance matrices { R ^ K
Formulation of the filter equations is based on
the energy balance of the process, which is ex-
pressed by the nonlinear stochastic differential
equation
dx
n ]
+ w (t)
3 t " " ίΐ^1'^2'-3'-Κ x
x^O) = (12)
w,(t) is a zero mean gaussian noise having spec-
tral density matrix Q if t). Model error wj(t) ac-
counts for simplifica tions during modelTing, para-
meter uncertainties a nd noisy process inputs. The
material balances are considered in the state
estimate propagation, but are not accounted for
explicitly in calcula ting the gain matrices for
the update equations of the temperature x.i. If no
disturbances of the inlet concentration or the
reactants are probabl e to occur, this aspect of
the filter design is not expected to have any sig-
nificant influence on the performance of the esti-
mator.
The error covariance propagation is obtained1by
linearizing ]f,
: about the current estimate x =
[*{ *2 * 3
dP^t)
= F (x, x ^ P ^ t ) P ^ t J F J i x ^ + Q ^ t ) (13)
dt L steady state gains would be too slow in dealing *
where 9
_ ίι(χ.»*κ) ι
(14)
The initial values of the reactor temperatures are
assumed to form a gaussian normal random vector
XV with mean x? and covariance matrix P, . Further
it is assumed that measurement errors ana model
uncertainties are not correlated. The resulting
gain matrix is
275
K H + 1
(15)
k = v - K < w - > k v "
The update equations are
* (+) = * H (16)
l k l k + K k[ y k- H kx l ( k- ) ]
P (+) = [I - K H ] P (-) (17)
k kk k ad n
The estimates of the concentration vectors x?
>u are then updated independently by means of
ïqs. (9) - (10).
IMPLEMENTATION AND RESULTS
Efficiency of the estimator developed in the pre-
vious section is tested in the case of error in
the model parameters and in the initial conditions
of the state variables. The values of the filter
parametersa used,
n c ai.e.
er the
v neelements
n a ue of the matrices
k* ^ l ^ ) * ^lo 9^ ^ T ^ 2. The perfor-
mance of the estimator is judged by comparison to
the actual reactor transients, which have been de-
termined by means of simulation based on the true
model and have been confirmed by the large amount
of experimental data available. Experimental mea-
surements from the pilot-plant reactor are used
for all filter implementations.
Error in the initial conditions
The estimator is first implemented by assuming an
incorrect initial temperature profile, as well as
an incorrect value of the inlet temperature
(30 C above real temperature). By considering
measurements of a single thermocouple in the sen-
sitive reactor region (z = 0.11) the filter esti-
mates approach rapidly the real temperature pro-
files (Fig. 2). For the given location of the
thermocouple, the number of grid points (LM) has
no significant influence on the estimation of the
axial profiles of the state variables, provided
that the profile form can be adequately approxi-
mated. In this case LM can be reduced to 11 equi-
distant points (i.e. 2 at the boundary and 9 in
the interior), or up to 7, if an appropriate non-
equidistant grid is chosen. If less steep profiles
are studied (e.g. at high gas velocities), even
6 points can be sufficient.
The contribution of update to the estimate can be
understood by examining the behaviour of the co-
variance matrix (Fig. 3). During the 1st min error
covariance and, consequently, filter gains still
have relatively high values; therefore the contri-
bution of update is significant. This effect can
be observed in Fig. 4, for instance at the fast
approach of the estimates of inlet temperature to
the true value. However, error covariance and fil-
ter gains diminish rapidly to low steady state
values. Therefore, estimate propagation practi-
cally dominates after 1 min, whereas the influence
of update is not significant any more. This fur-
ther implies that a filter with time invariant,
with the estimation problem examined here.
If a greater influence of theestimator downstream
of the hot spot is desirable, a second thermo-
couple in this reactor region is necessary (e.g.
ζ = 0.83). Due to the spatial approximation app-
lied, this influence is less pronounced on remote
grid points. Thus, by using a non-equidistant
grid with 5 interior points ( LM = 7) we obtained
better estimates for the crue ial 1st min than by
using LM = 11 points in equal distances (Figs.5
and 6).
276 C. Canavas
The filter estimates of the long term reactor dy-
namics using LM = 7 (non-equidistant) are almost
identical to the actual profiles. An obvious addi-
tional advantage in using small LM values is the
lower order of the resulting filter and, therefore,
the reduction of computing effort.
It should be stated that in the case of 7 equidi-
stant grid points the update "forces" the reactor
inlet estimates for small t to move away from the
actual values for both sensor configurations men-
tioned above (1 or 2 thermocouples). Since no dif-
ficulties arise during simulation of the reactor
transients with 7 equidistant points, the source
of the above problem is obviously associated with
the relative position of the thermocouple(s) in
respect to the grid points.
Model parameter error
The values of model parameters obtained through
identification methods may deviate from the real
values of the process parameters. Thus an additio-
nal error is introduced into the model, which
causes the estimator to yield a larger rms error
than that predicted by the filter covariance com-
putations. In order to reduce the divergence bet-
ween true and calculated estimation error, the
process noise wi(t) considered in the process mo-
del is increased accordingly.
To test performance of the estimator the coeffi-
cient of heat transfer to the reactor wall is di-
minished by 10 %. If only one thermocouple (z=0.11)
is used, the estimator has a tendency to yield a
lower inlet temperature and to "shift" the tempe-
rature profiles downstream. Further, update has
little influence on the error covariance: the pro-
pagation equation imposes an increase of the error
covariance, even after the reactor has reached the
steady state. This behaviour does not depend on LM.
This problem can be solved satisfactorily by con-
sidering a second thermocouple downstream of the
hot spot but still in the region of high tempera-
tures (z = 0.29). As in the case of error in ini-
tial conditions, the estimator using a grid with
LM = 7 nonequidistant points yields axial profi-
les of the state variables, which are remarkably
close to the real ones (Fig. 7).
Fig. 8 shows that steady-state condition for esti-
mation error covariance is achieved in 2 min, i.e.
before critical (in terms of system accuracy) pro-
files are reached. The increase through the error
covariance propagation is compensated by the up-
date.
CONCLUSIONS
The dynamic behaviour of a catalytic fixed-bed re-
actor can be estimated by means of a continuous
extended Kalman developed on the basis of a spa-
tially discretized nonlinear quasihomogeneous re-
actor model. In the cases studied in this paper
errors in the initial conditions, the inlet tempe-
rature and the parameter value of the heat trans-
fer to the reactor wall were considered. Taking
into account measurements of one or two thermo-
couples and using a non-equidistant grid with 2
boundary and 5 interior points to discretize the
distributed parameter model, reliable estimates
of the state variables were obtained. By means of
heuristic criteria the location of thermocouples
as well as of grid points can be defined in a way
that ensures good performance of the estimator and
reduces computation effort. Whereas the design of
a constant gain filter in the case of model error
diminishes computation effort without affecting
filter performance, time variable gains are neces-
sary in the case of error in the initial condi-
tions of the state variables.
REFERENCES
Alvarez, J. and G. Stephanopoulos (1982). An esti-
mator for a class of non-linear distributed
systems. Int. J. of Control, 36, 787-802.
Canavas, C. (1984a). State and parameter estima-
tion through nonlinear filtering. An applica-
tion to a fixed-bed reactor. Proceedings of
the TELECON '84 (IASTED International Confe-
rence), Aug. 27-30, 1984, Halkidiki (Greece),
pp. 456-459.
Canavas, C. (1984b). On the identification problem
of fixed-bed reactor parameters. Proceedings
of the 3rd Mediterranean Congress of rh'emical
Engineering, Nov. 19-21, 1984, Barcelona,p.85.
Clement, K., S.B. Jjorgensen and J.P. S0rensen
(1980). Fixed bed reactor Kalman filtering -
II. Experimental investigation of discrete
time case without stochastic disturbances.
Chem. Eng. Sci , 35, 1231-1236.
Eigenberger, G. and J7¥. Butt (1976). A modified
Crank-Nicolson technique with non-equidistant
space steps. Chem. Eng. Sci., 31, 681-691.
Gelb, A. (Ed.) (1974). Applied Optimal Estimation.
M.I.T. Press, Cambridge (Mass.).
Gilles, E.D. (1985). Application of model-based
measuring techniques in chemical engineering,
Proceedings of the International Conference
"On Industrial Process Modelling and Control",
June 6-9, 1985, Hangzhou, China.
Jutan, A. et al. (1977). Multivariable computer
control of a butane hydrogenolysis reactor.
AIChE J., 23, 732-758.
Kalman, R.E. (TÏÏ60). A new approach to linear fil-
tering and prediction problems. Trans. ASME,
series D, J. Basic Engng., 82, 35-45.
Krebs, V. (1980). Nichtlineare TTlterung.
Oldenbourg, Mlinchen.
Ray, W.H. and D.G. Lainiotis (Eds.) (1978).
Distributed Parameter Systems: Identification,
Estimation and Control.M. Dekker, N.Y.
Schwedock, M . J . (1983). Mo'delling and identifica-
tion of a catalytic packed bed reactor.
Ph.D. Thesis, Univ. of Wisconsin, Madison, USA.
Stfrensen, J.P., S.B. Jtfrgensen and K. Clement
(1980). Fixed bed reactor Kalman filtering and
optimal control - I. Computational comparison
of discrete vs. continuous time formulation.
Chem. Eng. Sci., 35, 1223-1230.
Windes, L.C., W.H. Ray"and A. Cinar (1984).
Dynamic estimation of temperature and concen-
tration profiles for control of a packed bed
reactor. Preprints of the 1984 Annual AIChE
Meeting, San Francisco, Nov. 25-30, 1984.
ACKNOWLEDGEMENT
This work was supported by the Bundesministerium
fur Forschung und Technologie. The author would
like to thank Prof. E.D. Gilles for his constant
interest, as well as Mr. R. Wahl for his help in
the computation work.
 Dynamic Behaviour of a Fixed-bed Reactor Through Filtering 277

TABLE 1 Reactor operation conditions 300

and model parameter values

a) Operation conditions
Inlet temperature (T ) 244°C STEADY STATE
0
Reactor wall temperature (T„) 240°C
Air flow rate 1600 Nl/h Τ
/ 120S\ \
Feed weight fraction of methanol 5.0 % (°C1

b) Dimensionless, normalized model parameter

values
L—O 24S
Activation energy
methanol oxidation A} 119.8
240 1 0S — .o""-- *j
formaldehyde oxidation A 138.0
2
Coefficient of heat transfer
12.86 Fig. 1. Temperature transients
to the reactor wall Β during start-up.
Coefficient of axial heat
conductivity ε 0.033
Damkohler numbers Da 148000
1
Da 25000
2
Da 148000
3
Da 138800
4
Da 16460
Reaction order
c
methanol oxidation 0.75
formaldehyde oxidation 1.00

TABLE 2 Filter parameter values

All filter parameters are assumed time invariant.

Fig. 2. Estimates in the case of error in
a) Error in the reactor initial conditions initial guess and inlet temperature.

Model error
2 2 2 J }
Q = d i a g ^ ,σ ,.. ,σ.
1 2 LM-1
σ / = 0.1 exp(-4.6z..) ; i = 1,2,...,LM-1
2
+ exp(-4.6z.) (0.47°C) /s

2
Initial error covariance
P = I - (46.2°C) I
l0
Measurement error 2
R = 0.01 (4.6°C) 2 (1 thermocouple)
R = 0.01 I -» (4.6°C) I (2 thermocouples)

b) Model parameter error

Model error Fig". 3. Diagonal elements of P(+) in the case
2 2 2 ,cr ) of error in initial guess and inlet
2 = diag{a
1 2 2
,σ ,..,σ^
' LM-l temperature (LM = 11).
σ Ζ = 0.01 + (0.15°C)
2 /s
χ 3 00
σ. = 1.00 + (1.5°C) /s ; i = 2,...,LM-1
a c t u a l profiles
Initial error covariance 2 : filter estimations
TIC]
0.01 I (4.6°C) I ο measurements
LO

Measurement error 2 initial guess

270 f

w-vT
R = 0.01 (4.6°C) 2 (1 thermocouple)
R = 0.01 I + (4.6°C) I (2 thermocouples)
— — .-60^.

240
: ~oT~"
000 0.11 1 00

Fig. 4 Estimates in the case of error in

initial guess and inlet temperature
(LM = 11).
278 C. Canavas

3 00 300

actual profiles actual profiles ( B = 12.86)

— filter estimations filter estimations (EU 11.64)
TIC]
a measurements TIC] simulation (B = 11 64)
ο measurements

27Cf

240 240
0 00 0.11 0.00 0.11

Fig. 5. Estimates in the case of error in

initial guess and inlet temperature Fig. 7. Estimates in the case of model para-
(LM = 11). meter error (LM = 7 ) .

3 00

actual profiles
TIC) filter estimations
« measurements

initial guess

240
0.00 0.11 0.00 0.11

Fig. 6. Estimates in the case of error in ini- Fig. 8. Diagonal elements of P(+) in the
case of model parameter error (LM = 7).
tial guess and inlet temperature (LM = 7 ) .
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

MODEL REDUCTION FOR NONLINEAR

SYSTEMS WITH DISTRIBUTED PARAMETERS
U. Epple*
Institut fiir Systemdynamik und Regelungstechnik, Universitàt Stuttgart,
Stuttgart, FRG

Abstract : Reduced nodeb- of nonlinear distributed systems which are accurate

•:or a wide range of .operating conditions, are difficult tc obtain. This paper
;.'oa)s '.;ith a method which uses the detailed inferential balance equations
•:-.s α starting point and result3 finally in a low ordc?r lumped parameter
r.odel. The main idea is to profit by a self organization process, which
physically confines the variety of the possible systembehaviour considerably,
he result are structured solutions with a characteristic form and dynamic.
One of the widespread and most simple types is the moving wa^e front
solution. It is surprising, that such a moving wavefront is a characteristic
structure not only for one, but for many totally different systems. Starting
with a short phenomenological view of some examples of moving wavefront
solutions the mathematical procedure to carry out the reduction is described
in detail. In a very special case it is possible to show the physical meaning
of the used assumptions by an analytical investigation. This property is very
helpful for the clarity of the method and the resulting reduced model. It is
well known, that the reactionzone in a fixed-bed reactor can be interpreted
as a moving wavefront solution. As an application example for the reduction
technique a complete reduced model of a catalytic fixed-bed reactor is
implemented.

Keywords; Modelling; model reduction; fixed-bed reactor; distributed

parameter systems; nonlinear systems; waveform analysis.

solutions can be seen as to be quasistationary

INTRODUCTION
The mathematical modelling of a physical system
often results in nonlinear partial differential
equations. Due to the infinite dimension of a PDE, a
wide variety of solutiontypes would be expected.
But, from an empiric point of view, it is well
known, that in many cases the behaviour of such a
complex system can be characterized by only a few
parameters. The state of a fixed-bed reactor for
example is mainly determined by the height and the
location of the temperature maximum. The reason of
this restriction of the system behaviour is a
selforganization in the system. The selforganization
leads tc the formation of a dynamic asymptotic
solution. This means, that the systembehaviour
splits into two part3, - α fast approach from
initial or disturbed conditions to the asymptotic
solution and a 3low motion according to the slow
dynamic changes of the asymptotic solution. The
prior condition for the existence of such a
splitting is a noticeable nonlinear term and ' an
internal feedback mechanism in the system. The
decree of the splitting is directly coupled to the
rate of r.onlinearity. In the case of a very strong
nonlinearity, the approach to the asymptotic
solution is very fast. Assuming the deviations of
the asymptotic solution vanish immediately, the
coupled to the asymptotic solution. This situation
is of course an ideal base for a reduced description
of the system behaviour. A great number of models
now find a later physical justification in this
property.
But moreover, the structuring of the system can be
taken as a starting point of a reduction method. In
this way it is possible to develop a reduced model
systematically. The first step is a description of
the characteristic form of the asymptotic solution Φ
by a nonlinear expression of the spacecoordinates ζ
and the timedependent formparameters jj(t).
Φ = Φ (p.(t), ζ) (1)
Each of these forms is generally applicable to a
whole class of solutions. The wavefront is in the
form of an S-profile moving along the z-coordinate.
To get the dynamic equations of the formparameters
p_(t) a method analogous to the method of weighted
residuals can be used.
The procedure results in low order lumped parameter
models. It becomes evident, that the new state
variables p(t) show only a siightly nonlinear
dependency on the input variables. The strong
nonlinearity in Φ is mainly introduced by the
formexpressior (1).

The example used here is the catalytic fixed-bed

reactor. The fixed-bed reactor is of great technical
importance and shows all the characteristic
(*) Present address: properties - but also the difficulties - arising
IN PLT, Bayer AG, 4047 Dormagen, Germany. with the concrete application of the reduction
The research was performed when the author was a method.
member of the Institut fur Systemdynamik und
Regelungstechnik, Universitàt Stuttgart,Ger-Tiny.

DCCR-S
279
280 U. Epple

PHENOMENOLOGY

Wave propagation processes occure in different

forms. A very frequent one is the moving front.
Figure 1 shows the principle shape of this
structure.

Fig. 2: The catalytic fixed-bed reactor

Λ gas flow streams through the tube which is filled

with catalytic pellets. The reactant, carried by the
gas-flow, diffuses into the pores of the pellets and
reacts on the catalytic layer. In the case of an
exothermic reaction, the turnover causes a
heatproduction in the corn. This heat is lead off by
heat exchange with the gasstream and by direct
heat-conduction to neighbouring pellets. Due to the
direct heat-conduction an internal feedback
mechanism arises. The interaction of the strongly
nonlinear reactionrateterm with the internal heat
Fig. 1 Moving wavefront feedback in the fixed-bed can result in marked
reaction zones. Reaction zones are small regions of
In a small regies of the space-coordinate ζ the space in which the reactant is nearly completely
variable χ changes monotonically from a low to a consumed. Fig. 3 - 5 show the typical wavefront
high level. The transition takes place in the form behaviour of the reactionzone in different
of an S-profile. Due to this S-profile, the first situations. The decrease of the temperature behind
derivative of x, 3x/3z shows a high maximum at the the reaction front is caused by the cooling of the
2
middle 2 of the front and the second derivative wall of the tube.
3 x/3z a maximum at the left and a minimum at the
right side of the front. The main dynamic behaviour
of a wavefront is the movement along the space
coordinate with a constant velocity w. During this
shift, the shape of the front remains nearly
unchanged. The velocity is completely determined by
the parameters of the system and independent of the
initial conditions. Moving wavefronts can be seen as T[°C]
the source of the characteristic behaviour of
important systems in different disciplines of
science.

Well known are the wavefronts in population genetics

(Fisher 1937), nerve propagation (Hodgkin and Huxley
1952), or the Shockwaves in fluiddynamics (Burgers
1948).

Some of the most impressive and technically most Fig.3: A moving reactionzone after a decrease in the
important examples arise in the region of chemical feed temperature
engineering. The limitation of energy and
mass-transport causes gradients in temperature and In Fig. 3, after a step decrease of the inlet
concentrations. Consequently the systems have to be temperature, the reactor goes from an ignited state
regarded to have distributed parameters. to an extinct state. The dynamic transition occurs
in form of a reaction-front moving downstream
Additionally the nonlinearities arising for example through the fixed-bed.
with the reactionrateterms or the expressions of the
thermodynamic equilibrium effect the behaviour of
the system essentially. Well known are the fronts in
adsorption columns especially in chromatographic
packages, ignition fronts in flame propagation,
thermic fronts on catalytic wires, or concentration
fronts in reaction-diffusion systems. Lately it has
become obvious, that the regions of mass exchange in
distillation and absorptioncolumns even can be
interpreted as wavefronts (Marquardt 1986). A
typical example for a wavefront is the moving
reaction zone in a fixed-bed reactor.

The catalytic fixed-bed reactor is the most

important continuously operated chemical reactor. It
is technically seen as the obvious means to carry
out heterogeneous catalysis. In industry the
fixed-bed reactor is applied to acid-catalyed
reactions, hydrogénation, oxidation and other
reactions. For example the synthesis of ammonia, Fig. 4: Extinction wave caused by a deactivation of
methanol or vinylacetate and the oxidation of the catalyst (Weng, Eigenberger, Butt 1975)
methanol, SCh , ethylene or naphthalene are well
known. Fig. 2 shows one tube of a multitube reactor. Fig. 4 shows another form of an extinction-wave. In
this case, the extinction is caused by a
deactivation of the catalyst.
 Model Reduction for Nonlinear Systems with Distributed Parameters
T[°C]
Fig. 5: Transient behaviour after a stepdecrease
the feed temperature
Fig. 5 demonstrates, that even in the case of the
transition between two stationary ignited states,
the wavefront dynamic remains the dominant part of
the system behaviour. Although the temperature and
the velocity }f the front are changing due to the
strong influe ice of the upstream boundary condition,
the main form remains r.early unchanged.
MATHEMATICAL MODELLING
The taste of a mathematical model is to describe the
behaviour of a system in a formal way. Such a model
has to fulfil several demands:
A: Qualitiy: The main qualitative behaviour of
the system should be reproduced by
the model.
Quantitatively a sufficient
accuracy is required.
These two properties have to be
valid in a wide range of operating
conditions.
B: Systematic: It should be possible to deduce the
model by a general method.
The model should allow to
C: Continuity:
reconstruct the whole space-profile
in the case of a distributed
system.
D: Relatedness: To understand the systems behaviour
it is important, that the state
variables of the model correspond
to quantities with a physical
meaning.
The basic method to get a mathematical model is to
use the physical laws of energy-, impuis-, and
masspreservation. Formulating the balances of a
concrete system results in a set of differential
equations.
In the case of the wavefront system considered here,
the set consists of nonlinear partial equations of
the relaxation type.
3x
jl — - Ν (x(z,t)) = 0 (2)
at
Mathematical models of this type are known for the
most important systems of technical interest. By
high order simulation it can be shown, that the
results given by these models are in good accordance
281
with the experimental data. Problems arise if the
calculation methods of the control theory are
applied to such a model. It becomes obvious, that
the methods of control theory don't conform to a set
of nonlinear partial differential equations. Thus,
if we want to design a controller, to construct an
observer or to carry out an optimization, it is
desirable to have a model with the property Ε
additional to A - D.
E: structure: - The mathematical model should
consist of a low order lumped
parameter system.
- The dependency of the state
variables on the input variables
should be just barely nonlinear.
To fill the gap betv?een the T.odel (2) - presumed to
be known - and the desired low order lumped
parameter system, a model reduction has to be
carried out. The basic idea of the reduction method
applied here is to take advantage of the self-
organization process occuring in systems with wave
propagation. Due to this selforganization process,
solutions of different initial conditions converge
quickly to a wavefront solution. This asymptotic
aspired wavefront solution is characterised by a
constant shape and only slowly changing
formparameters. As we will see in the r-ixt chapter,
this splitting of the dynamic into a fast part - the
asymptotic approach to the wavefront - and a slow
part - the changing of the formparameters of the
wavefront - can be proved analytically.
It is evident that the physical seIforganization
process for a modelreduction is to be used. The
reduction is carried out in three steps:
I Quasistationary asumption
In most applications, the fast dynamic approach to
the wavefront solution is of minor interest.
Neglecting these quickly vanishing parts reduces
extremely the variety of the possible system-
behaviour.
Under the quasistationary asumption the behaviour of
the system is approximately characterized by the
behaviour of the wavefront solution. This means the
problem of finding a model describing the variety of
solutions reduces to the problem of finding a model
of the wavefront solution.
II. Description of the form
A model of the wavefront solution consists of two
parts, an expression describing the constant shape
of the solution and a set of ordinary differential
equations describing the timedependency of the
formparameters.
The main problem is the mathematical description
of the shape of the front. The shapo depends on
the special nonlinearity and is in almost all
cases not known.
As a good approximation, the expression (3) can be
used.
Φ (p_,z) =
-4q, (z-z, )
e +1
(3)
(ζ, , q )
x
This expression contains only two parameters ζ ,
t
the location of the front and q , the gradient in
the front.
x
282 U. Epple

III. Dynamic equations for the formparameters A possibility to keep the calculations easy is to
choose the weightfunction p(z) as a sum of M
In the scalar case, the approximation of the δ-functions.
wavefront gives an expression of the form
M
χ - Φ (p_(t),z) (4) p(z) = l δ(ζ-ζΓ) (12)
3=1
The timedependent formparameters jp_(t) have to be
interpreted as the state variables of the reduced In this case the components of the vector RS and
model. χ is a nonlinear function of p(t). At the matrix A can be calculated by sums of
this point an essential difference to the usual algebraic expressions. M must be at least equal
formalistic modelreduction technique becomes to . But to achieve results which are independent
visible. of the choice of the collocationpoints ζ,* , M
should not be under 10-20. This number of points
The formalistic modelreduction technique uses an does not complicate the procedure at all - only
expansion cf the solution χ in the form. some more components of the summation need to be
calculated but this number of points has an
η essential stabilising effect.
χ = I a j (t).g , (z) (5)
i=l Solving the linear problem (9) we get the desired
system of ordinary differential equations for the
g (z) belongs to a system of orthogonal functions. formparameters p.
t
The solution x in (5) is a superposition of items 1
which are linearly dependent on the statevariables ρ = A" (p,z*). RS (p,z*) (13)
a ,(t).

The dynamic equations for the state variables of STRUCTURE ANALYSIS

(5) a , (t) can be achieved systematically by the
method of weighted residuals (Finlayson 1972). One of the basic equations producing a wave front
This method has been primarily developed for an solution is given by (14)
expansion of the form (5), nevertheless it can be 2
also applied to nonlinear expressions of the form 3x 3x
(4).
ε 2 = f(x) (14)
3t 3z
The procedure is as follows:
Substituting equation (4) into the partial dif- where f(x) is a nonlinear function.
ferential equation we get.
A series of papers deal with the existence,
3Φ 3Φ
stability and dynamic behaviour of the solutions
Ν(Φ) = j j - Ν(Φ) = Res(p,p,z) (6) of (14). Famous contributions are for example
3t 3fî Kolrnogorov, Petrovski and Piskunov (1937), Aronson
and Weinberger (1975) or Fife (1979).
To determine the η free variables ρ we minimize From their work, we will use the important fact,
the quadratic error function Res(p,p,z) with that there exists a unique stable wavefront
respect to p. solution if f(x) is a nonlinearity of type (15).
1 2 !
J(Res (ζ,ρ,ρ)) . p(z)dz = Min (7)
0 f(x)

ρ is a weightfunction to be chosen. If Φ would be

an exact solution, the Integral would vanish
identically. (Res = 0!).
(15)
By a differentiation of (7), it is possible to
obtain ii conditions to determine the components of
the vektor p.

1 3Φ . !
j — . Res (ζ,ρ,ρ) .p(z)dz = 0 i=l,.·,n (8)
0 3D.
Fig. 6 Form of f(x) for an equation of the Huxley
Due to the relaxationform of the partial type.
differential equation, the residuum (6) depends
linearly on the components of p. This equation is known as Huxley type. It is used
Thus equation (8) can be written as for example to model the nerve propagation process
(Hodgkin a. Huxley 1952), or ignitionfronts on
A(p.p(z)).p = RS (p p) (9) catalytic wires (Barelko et al (1978)).
f
It is a very helpful attribute of this equation,
with:
that in the special case
1 3Φ 3Φ Τ
A(p,p(z)) = / — .(—).p(z)dz (10) f(x) = - k x(l-x)(a-x) (16)
0 3p. 3p.
the wavefront solution is explicitly known.
i 3Φ
1
RS(p,p(z))= J — .Ν(Φ).ρ(ζΜζ (11)
Φ(ζ) = (17)
0 3p
— (Z-Zo-W.t)
and ρ can be determined by solving the
2ε
problem (9).
e + 1
 Model Reduction for Nonlinear Systems with Distributed Parameters 283

APPLICATION
The velocity of the wavefront is given by
A good example to test the efficiency of the
w = \/2£k' (cc-0.5) (18) reduction technique is the catalytic fixed-bed
reactor.
With help of this solution it is possible to
investigate, analytically, the spectrum of the In the easiest case of a reaction A — > products
operator liniarised around the solution. The the behaviour of the reactor can be described by
results are summarized: an energy - and a massbalance.
2
3T 3T 3T
- B(T-T ) + Da.r (T,c)
1. Using a quadratic norm with weight func- K
tion 3t 3z 3z
(20)
(2d-l) (l-2ct> 3c
ρ = Φ .(1-Φ) (19) - Da.r(T,c)
3z
(This weight function corresponds to the weight
function ff = 1 in the self-adjoint system), The normalized state variables Τ and c are defined
the spectrum consists of a continuous and a in the region 0<z<l. The boundary conditions arc
discrete part.
3T
= T = 0 C x„
2. The discrete part of the spectrum exists of the lu 3z
eigenvalue X =o and - in the region 0,25 < a <
t
0,75 - of a second eigenvalue X . Except for z=0 z=l z=0 (21)
2
X and X the discrete part of the spectrum is
t
empty.
2 Comparing simulation results and measured data a
good agreement is achieved (Ruppel 1980 , Canavas
1984 ). This quantitative agreement exists also
in the region of moving reaction front solutions.
^ Mi It is necessary to emphasize that the nonlinear
Re effects of selforganization are dominant only in
the front region. Outside the front, the
behaviour of the reactor is mainly influenced by
Fig. 8 Spectrum for a - \/2 additional effects like boundary conditions,
cooling and the residual reaction of the extinct
state. In these regions it is to be expected that
the nonlinearities are of little significance.
3. The eigenfunction corresponding to λι causes a With this in mind it seems suitable to split up
shift along the z-coordinate. The value of the model system into two subsystems, a basesystem
Xi=0 is characteristic for all wave front SI and a wavefrontsvstem S2.
solutions. It reflects the fact, that a
wavefront solution can be displaced at any
location along the infinit spacecoordinate
without changing form or velocity. The
eigenfunction corresponding to X leads to a
2
variation of the gradient in the front.

A comparison of this analytic result with the

mathematical model of the last chapter shows:

1. The chosen expression (3) for the shape of the

wavefront model becomes an exact solution in
the special case (16).

2. The splitting of the dynamic behaviour of the

model into the fast approach and the slow
changes, corresponds to the decomposition of
the spectrum of the linearized system into + c B T Tc
+ Dr a
continuous and discrete parts. Moreover there âT ^- âz^- iV i) - t r i)
is a one to one correspondence between the
dynamic of ζ , and the first eigenfunction and
the dynamic of q and the second eigenfunction. SI
t (22)
Without entering into details it should also be T 0 C 1
mentioned that in the case of the Burgers equation 2 0=
1 I = zu ΤΓ I « •· 1 zu
z=o
an exact solution is known having the same form as zil
(17). The discrete spectrum of the wave front
solution here too contains the dominant eigenvalue
X =0 corresponding to a shift of the profile. The
t
Burgers equation is important as a basic equation -rtT^))

of fluid dynamics and very similar to the basic (23)

3C2 -DaCrtTj+T^Cj+CgJ-rtTj.Cj)}
equations of countercurrent exchange processes. S2
3Z

The results of the structure analysis can be seen

as a hint that there is a physical meaning in the T2 I = 0 ;
Z I = 0 ; c
3= 0
chosen form and formparameters of the reduced
2
z=o z=l
2I
z=o
model.
Fig. 8 Splitting of the system in a basesystem SI
and a wavëfrontsystem S2.
284 U. Epple
It is easy to see, that with x=x 1 +x , y=y 1 +y 2 ,
2
by an addition of. the two parts SI and S2 the
split system passes exactly to the original
system. The consequence of the nonlinearity is a
unidirectional coupling of the basesystem to the
wavefrontsystem. This coupling occurs via the SI
dependent parameters of the reaction rate term and
transfers the input-influences to the
wavefrontsystem S2.
To model the slighty nonlinear basesystem is no
problem. In most cases simple linear term3 are
sufficient to describe the transferfunctions.
Here we take a model of order two to describe the
response of Τ (z-z ) to a change in the cooling
t 7
temperature and the inlet temperature.
The modelreduction of the wavefront system is
carried out as described. The adequate expression
for the reaction front is in the case of the
cooled reactor.
1 -0(z-z, )
T (24)
2 = T2
-7(z-z , )
+1
The additional formparameters T
20 and β describe
the magnitude of the temperature rise in the front
and the decrease of the temperature behind the
front which is caused by the cooling of the wall.
With ζ , r,T , β and the two state variables of the
7 20
basesystem we get a model of order 6.
Figure 9 shows a comparison between the reduced
model and the detailed model. Form and velocity of
the front is reproduced very well by the low order
model
Fig. 9 Temperature profiles of a moving front
reduced model
— detailed model
CONCLUSION
We started with a phenomenological view to systems
with wave propagation. Systems with wavepropaga-
tion represent an important class of the nonlinear
systems with distributed parameters. They are of
great technical interest. To apply the methods of
control theory to such a system it is desirable to
have a low order model, which is able to describe
the behaviour of the technical system in the whole
region of possible operating conditions. It is
surprising that such a good fitting model can
really be achieved. The physical reason of this
possibility is a selforganization process reducing
the variety of possible system behaviours extreme-
ly. This selforganization process leads to a
dynamic asymptotic solution. All other solutions
approach very quickly to the asymptotic solution
and follow the slow dynamic changes of this solu-
tion quasistationary. For a special system the
existence of the asymptotic solution as well as
its dynamic property can be proved analytically.
In the case of the wavefront systems the characte-
ristic feature of the asymptotic solution is an
S-shaped profile which moves along the space coor-
dinate. While the physics and the balance equa-
tions of the different wavefront systems are to-
tally different, the existence of the asymptotic
solution with the S-form and the moving along the
space coordinate is a common property of all these
systems. This gives the possibility to develop a
reduction method generally applicable to the whole
class of wavefront systems. In the first step the
part of the system determined by the wavefront is
separated from the rest-system. The rest-system or
base system contains effects influencing the beha-
viour mainly outside the front where the nonlinea-
rities are of small magnitude. The wavefront
system can be described by its form and the dyna-
mic change of the formparameters. To produce the
S-shaped front the expression for the form con-
tains an exponential term in the denominator. The
dynamic dependency of the formparameters can be
achieved ' by a generalization of the method of
weighted residuals.
As shown in the example of a fixed-bed reactor,
the resulting reduced model is in good accordance
with reality. A very helpful property of the model
is that there remains only a slightly nonlinear
dependency of the state variables from the opera-
ting conditions. The strong nonlinearity, visible
in the solution x, is introduced by the transfor-
mation χ = Φ(ρ_,ζ) in the formexpression.
Thus such a model seems to be an ideal base for
controller design, modelbased measurement tech-
niques or optimization calculations.
REFERENCES
Aronson D.G., Weinberger H F . (197S). Nonlinear
f genetics, combustion
diffusion in population
and nerve propagation. Lecturenotes in
mathematic 44b. Springer, Berlin, 5-49.
Bare1ko V., Kurochka I.I.. Merzhanov A.G.
Shkadinskii K.G. (1978). Investigation of
travelling waves on catalytic wires. Chem.
Enq. Sci. 33. 805-811.
Burgers J.M. (1948). A mathematical model
illustrating the theory of turbulence. Adv.
Appl. Mech., 1, 171-199.
Canavas C. (1984). Modelling and simulation of a
fixed-bed reactor. International 84 AMSE
Conference "Modellinq and simulation",
Athen.
Fife P.C. (1978). Asymptotic states for equations
of reaction and diffusion. Bulletin of the
American Mathematical Society, Vol. 84,
Nr. 5 693-726.
α
Finlayson, B.A. (1972). The method of weighted
residuals and variational principles.
Academic Press, New York.
Fisher R.A. (1937). The advance of advantageous
genes. Ann. Eugen. 7, 355-369.
Gilles E.D. (1974). Quasi-stationares Verhalten
von wandernden Brennzonen. Chem, Eng. Sci.,
29, 1211-1216.
HodgkinA.L., Huxley A.F. (1952). A quantitative
description of membrane current and its
application to conduction and excitation in
nerves. Journal of physiology, 117, 500-544.
Kolmogorov A.N., Petrovski I.G., Piskunov (1937).
A study of the equation of diffusion with
increase in the quantity of matter and its
application to a biological problem. Bjul.
Moskovskovo Gos. Univ. 1 no. 7, 1-72.
Mavquardt W. (1986). Nonlinear model reduction for
binary distillation. DYCORD 86, Bournemouth.
Ruppel W. (1980). Eine mathematische Beschreibung
wandernder Reaktionszonen in Schiittschich-
ten. Dissertation. Universitàt Stuttgart.
WengH.S., Eigenberger G., Butt J.B. (1975).
Catalyst poisoning and fixed-bed reactor
dynamics. Chem. Eng. Sci., 30, 1341-1351.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986
MULTIVARIABLE CONTROL OF A
CATALYTIC TUBULAR REACTOR USING
BOTH WIENER-HOPF CONTROLLER DESIGN
AND INTERNAL MODEL CONTROLLER
DESIGN APPROACHES
D. J. Kozub*, J. F. MacGregor* and Joseph D. Wright**
*Dept. of Chemical Engineering, McMaster University, Hamilton, Ontario,
Canada L8S 4L7
**Xerox Research Centre, Mississauga, Ontario, Canada
ABSTRACT
A practical application of advanced model-based multivariate
control to a pilot plant packed bed reactor carrying out highly
exothermic butane hydrogenolysis reactions is presented. The
system was first stabilized by using butane flow rate to control
the reactor hotspot temperature. Propane production and
butane conversion were then controlled using multivariate
controllers which manipulated the hotspot temperature
setpoint and hydrogen flow rate. An approximate linear
dynamic-stochastic model was developed using multivariate
time series and process identification methods.
Two multivariate controllers were applied. The first was a
simple version of Internal Model Control (IMC) utilizing a
stabilized approximation of the model inverse and a tunable
exponential filter. The second was a constrained minimum
variance controller design using pulse transfer function
m o d e l s to c h a r a c t e r i z e t h e p r o c e s s d y n a m i c s a n d
autoregressive-moving average models to characterize the
stochastic disturbances. The optimal control solution was
arrived at via a spectral factorization solution to the
Wiener-Hopf equations. The results indicated that both
designs, if well-tuned, provide adequate performance and
robustness over a wide region of operation.
K e y w o r d s : Multivariate Control; Internal Model Control;
Linear Quadratic Optimal Control; Wiener-Hopf; Catalytic
Reactor.
1. INTRODUCTION
An important problem in many chemical processes is that of
controlling the product stream compositions from catalytic
reactors in which one is carrying out multiple reactions
involving multiple species. A high degree of coupling among
the product compositions due to the reaction stoichiometry
makes the control problem a truly multivariable one. Since
the reactions are often highly exothermic and have high
activation energies, stability of the reactor is also a major
concern. Furthermore, reaction kinetics are usually highly
nonlinear in both reactant concentrations and temperature. A
controller design strategy for such reactors must therefore be
capable of h a n d l i n g c o m p l e x i n t e r a c t i o n s b e t w e e n
input/output pairs and must be robust to nonlinearities across
the operating region.
Advanced optimal controller design using a state space model
formulation of the multidimensional problem has received
much attention in control literature. Successful application of
this theory has been demonstrated by Jutan, Tremblay,
Wright, and MacGregor (1977), Sorensen, Jorgensen, and
Clement (1980, a,b), and Wallman, Silva, and Foss (1979) to
packed bed control problems using rigorous mechanistic
modelling that led to convenient state space model forms.
This approach to a packed bed reactor control problem was
also found to be successful by MacGregor and Wong (1980)
using empirical transfer function model identification
followed by transformation to state space form. Optimal
linear-quadratic (LQ) controller design is usually associated
with these state space techniques. However, when one has
measurements, or inferred values, of the outputs to be
controlled, there are many a d v a n t a g e s to d e s i g n i n g
285
controllers using input-output transfer function type models.
The resulting controllers are then expressed directly in terms
of past inputs and outputs, and the Internal Model structure of
the controllers makes any subsequent analysis simpler, and
provides a better understanding of the resulting controller.
In this paper we present the design and results of a practical
application of two multivariate controllers to a pilot plant
tubular reactor c a r r y i n g out the h i g h l y e x o t h e r m i c
hydrogenolysis of butane over a nickel on silica gel catalyst. A
favourable choice of input control variables is demonstrated
which leads to an open-loop stable configuration through use
of an internal hotspot controller. In this work we employ
simple linear empirical input/output models developed from
time series analysis and process identification techniques on
data collected directly from the reactor. The two multivariate
designs presented are constrained minimum variance control
(LQ output) and internal model control. Both designs are well
suited for output control using a pulse transfer function
process model. The designs are straightforward for practical
chemical process industrial application, and take full
a d v a n t a g e of t h e t r u e i n t e r a c t i v e n a t u r e of a
m u l t i d i m e n s i o n a l s y s t e m to a c h i e v e good c o n t r o l l e r
performance. Furthermore, they allow one to incorporate into
the designs a considerable degree of robustness to mismatch
between the model and the process.
2. MULTIVARIATE CONTROL
2.1 Multivariate Model Representation
Frequency response representation of a multivariate process
has become a popular structure for controller design which
n a t u r a l l y r e s u l t s from i n p u t / o u t p u t i d e n t i f i c a t i o n s .
Multivariate controller design with pulse transfer functions,
unlike single input/single output design, is hampered by
messy matrix computations and the noncommuting property
of matrices. A discrete, linear multivariate process can
always be represented in a right matrix fraction form
l
Y(t) = ( ù ( z - i ) 8 ( z - i ) - z - k U ( t ) + θ ( ζ - ΐ ) φ ( ζ - ΐ ) - lA(t) (1)
which leads to some simplification in the controller design
equations. Y(t) represents the vector of η observable outputs
at sampling interval t. U(t) is the vector of m controllable
inputs at sampling interval t. Then nX m matrix ω(ζ - !) and
1
the m X m matrix δ(ζ~!) are polynomial operators in the
backward shift operator z - with nonrational elements.
δ ( ζ - ! ) is diagonal with the coefficient of z° set to I. The
minimum whole periods of delay is given by k. The zeros of
the process are defined to be the roots of
|ω(ζ-ΐ)| = 0 (2)
A multivariate process is invertible if all roots of equation 2 lie
inside the unit circle. The poles of the process are defined to be
the roots of
|δ(ζ-ΐ)| = 0 (3)
A process is open-loop stable if all roots lie within the unit
circle.
286 D. J. Kozub, J. F. MacGregor and J. D. Wright
The second term on the right side of equation 1 represents the
observable but uncontrollable contributions to the process
outputs which are assumed to be stochastic. A(t) represents a
vector of η independently distributed random shocks with
covariance matrix Σ. An autoregressive integrated moving
average model in a right matrix fraction form is used to
characterize the nature of the process disturbances. The n X n
1 1
matrices θ ( ζ - ! ) and φ ( ζ - ! ) are polynomial operators in z -
l
having nonrational elements. The zO coefficient of both
_ 1
matrices is set to I and φ ( ζ ~ ) is diagonal. Nonstationarity of
1
a process output is accounted for by a (1 — z ) factor in the
corresponding element of the φ ί ζ - ) matrix. We can also
represent deterministic disturbances with equation 1 if A(t) is
a vector of shocks having a probability distribution with
non-zero values occurring only infrequently (MacGregor et al.,
1984).
Right matrix fraction representations are not unique. A
simple and convenient method for transforming rational
transfer function matrices into a right matrix fraction form is
by choosing the i'th diagonal element of δ(ζ~!) or φ(ζ-*) to be
the greatest common factor of the denominators of the i'th
column of the rational transfer function matrix or ARIMA
matrix respectively.
2.2 Constrained Minimum Variance Control
This section presents the results of the linear quadratic
optimal control problem using a spectral factorization solution
of the Wiener-Hopf equations. Details of the solution are given
by Wilson (1970) and Harris and MacGregor (1986).
The performance index chosen in the constrained minimum
variance (CMV) or linear-quadratic (LQ) output controller
design is the minimization of
T T
+ V U ( t ) Q V U(t) + U ( t ) Q U ( t ) }
2 3
where Qi, Q2, and Q 3 are positive semi-definite weighting
matrices, and e(t) = Y ( t ) - Y ( t ) . In equation 4 we may
s p
penalize the variance of the inputs about a mean level using
weighting matrix Q 3 or the change in the control action, given
by VU(t) = U ( t ) - U ( t - 1 ) , w i t h w e i g h t i n g m a t r i x Q .
Disturbances encountered in chemical processes are normally
nonstationary with noticeable drifting output mean levels.
R e g u l a t i o n of p r o c e s s e s w i t h c h a n g i n g m e a n l e v e l s
theoretically requires infinite input variance. For this reason
we choose to p e n a l i z e the c h a n g e in control a c t i o n
VU(t) = U ( t ) - U ( t - 1 ) . It can be s h o w n t h a t if the
disturbances are nonstationary and VU(t) is penalized that
integral action will be present in the optimal controller
design. For the case of the number of inputs exceeding the
number of outputs we must penalize only the change in control
action of a choice of inputs equal to the number of outputs in
order to acquire reset action. The remaining inputs may be
penalized about a mean input level using weighting matrix
Q3.
Harris and MacGregor (1986) show that the general LQ
optimal feedback control solution can be expressed in the
internal model controller form:
U(t) = H ! ( z - i ) [ G ( z - i ) U ( t ) - Y(t)l + H ( z - i ) Y (t)
m 2 sp
where
Hi(z-i) = G (z-i)Fi(z-i)
c and setpoints enter the control loop as noninteractive
H (z-i) = G (z-i)F (z-i)
2 c 2 present. In this section the results of a simple IMC design
The control block diagram is shown in figure 1. G ( z - !) is an
m
n X m process model transfer function and Y ( t ) is the
sp
setpoint. G ( z - ! ) is a square m X m matrix. When a
c
multivariate system is square G ( z - l ) can be physically
c
related to an invertible approximation of the process model
1
inverse. F i i z - ) and F ( z - ! ) are n X m disturbance and
2
setpoint change filters. The optimal approximate inverse for
the cost function given by equation 4 leads to
G (z-1) = δ(ζ-ΐ)Γ(ζ-ΐ)-ΐ (8)
1 c
1
where IXz- ) is an m X m nonrational polynomial in matrix
z - having no zeros outside the unit circle. Π ζ - ) is
determined from the solution to the spectral factorization
equation
G)T(z) QiQ(z-l) + δΤ(ζ) Q a S i z - l )
+ δΤ(ζ) (1 - z) Q ( l - ζ - 1 ) δ(ζ - 1 ) = ΓΓ(ζ) Γ (ζ - 1 ) (9)
2
The choice of Q equal to 0 with the special case of a square
2
ω ( ζ - ! ) having all zeros inside the unit circle leads to
minimum variance control with Γ(ζ~!) = ω(ζ-!). The optimal
filters corresponding to minimization of equation 4 are
d e p e n d e n t on the a p p r o x i m a t e i n v e r s e and on t h e
disturbances and setpoint models respectively. Both filters
can be expressed as
F(z-i) = Τ(ζ-ΐ)θ(ζ-ΐ)-ι (10)
_ 1
where T ( z ) , a nonrational polynomial matrix, is determined
by solving the set of linear equations obtained by equating
coefficients of ζ in
coT(z)Q kO(z-l) = 0 ( ζ ) φ ( ζ - ΐ ) + ΓΤ( )Τ(ζ-ΐ) (11)
lZ ζ
with the z - ° matrix coefficient of C(z) set to 0.
Equations 5, 6, 7, and 8 can be combined to calculate control
actions at each sampling interval as a function of past inputs
and outputs as
[Γ(ζ-ΐ) - Fi(z-i)a>(z-i)]6(z-i)-iU(t)
(4)
= F ( z - l ) Y (t) - F i ( z - l ) Y(t) (12)
2 sp
As an alternative, we may retain the separate block structure
of figure 1. The advantage of a separate block structure is
inherent reset windup protection since actual implementated
U(t)'s are entered into the model predictor block. Another
advantage of the separate block structure is that differences
between process outputs and model predicted outputs can be
recorded to keep track of the severity of mismatch.
2
Constrained minimum variance controllers will always be
stable for any choice of input weighting matrices provided no
model mismatch is present. When model mismatch is present
stability cannot be guaranteed for any choice of input
weighting matrices. The robustness of constrained minimum
variance controllers can be improved by increasing the
magnitude of input penalty weights (detuning). Harris and
MacGregor (1986) show that there exists a Q which is
sufficiently large to guarantee robustness to arbitrary large
2
m o d e l l i n g e r r o r s p r o v i d e d t h a t Re{A(j)[co(z = 1)
Γ(ζ = 1 ) - 1 ] } > 0 ; j = 1,2, . . . ,n, where λ(])[Α] is the j'th
eigenvalue of A. It has been shown by Bergh and MacGregor
(1986), and Harris and MacGregor (1986) that constrained
m i n i m u m v a r i a n c e c o n t r o l l e r s e x h i b i t an e x c e l l e n t
compromise between performance and robustness.
2.3 Internal Model Control
(5)
The concept of internal model control (IMC) is very simple and
is easily applied to multivariate systems. The detailed
development of this design is given by Garcia and Morari
(6) (1985,a,b). In the IMC design we assume that disturbances
(7) deterministic step changes and no stochastic disturbances are
approach for square multivariate systems will be presented
which leads to a decoupled multivariate closed-loop response.
 Multivariable Control of a Catalytic Tubular Reactor 287

The IMC block diagram is represented by Figure 1. The

_ 1 1 - f.
II
closed-loop transfer functions relating the output and input F ( z ) = diag
response to disturbances and setpoint changes are 1 -f.z" (17)

Y(t) = N(t) + G ( z - i ) G ( z - l ) ( F ( z - l ) Y (t)

p c 2 sp where each filter parameter lies in the range of 0 ^ % < 1 .
Each filter parameter can be physically related to the
-F (z-i)N(t)) (13)
t closed-loop time constant of a setpoint change or disturbance
rejection by
U(t) = G ( z - i ) ( F ( z - i ) Y (t) - F i ( z - i ) N ( t ) ) (14)
c 2 sp
when no model mismatch is
1
present.
These equations show f.. = exp
(18)
that selection of a stable G c U - ) matrix will lead to cl.

closed-loop stability provided that the process is open-loop

stable. In the IMC design independent selections of G ( z - ! ) ,
1c
F j ( z - l ) , and F ( z - ! ) are made, unlike in the constrained
1 2
minimum variance controller design where F j i z - ) and
F2(z~ ) are dependent on the specification of G ( z - ! ) through
c
equation 11. The a d v a n t a g e of the IMC d e s i g n over
conventional feedback controller designs operating on the
error signal is that controller stability is addressed directly
through a stable choice of G ( z - i ) . Stability in conventional
c Re{À(j)[G (z = 1) G ( z = l ) - i ] } > 0 for j = 1,2
feedback blocks is examined indirectly through a complicated
return difference equation.
The only requirement for the specification of G ( z - i ) is that it
c stable G ( z - M is specified. This is contrary to the constrained
leads to a stable controller. Hence the designer is free to make
any stable selection. The optimal choice of G ( z - 1 ) depends on
c procedure would have to be resolved for any new choice of Q
the objective function to be optimized and the properties of the
multivariate process. The simplest design is that given by
Garcia and Morari (1985, a,b) which leads to a decoupled
closed-loop response. In this design G ( z - *) is an approximate
model inverse given by
c multivariate process is balanced. A multivariate process is
1 b a l a n c e d if the rows or c o l u m n s of G ( z - ! ) can be
G (z-1) = ( 5 ( z - l ) ( z - l ) - z k ) V ( z - l ) v (z-l)
c W + 1 +2
The first expression in brackets on the right of
1
side
equation 15 is the inverse of the process model. V + ι ί ζ - ) is a
diagonal matrix with each diagonal element of the form of
z-s(j). This matrix eliminates prediction terms in the inverted
process model. S(j) is equal to k plus the highest positive
1
power of ζ in column j of the inverted ω ( ζ - ! ) matrix. This
choice of V i ( z - ) leads to perfect decoupling with output y(j)
+ _ 1
having the minimum s(j) periods of delay in the closed-loop
1
response. V + ( z ) is a rational or nonrational polynomial in
2 1
z - which removes or replaces unwanted zeros in | ω ( ζ - !)| that
will appear in the controller. Zeros of | ω ( ζ - ) \ that are outside
the unit circle yield an unstable controller. The designer can
choose to simply cancel unstable zeroes or cancel and replace
unstable zeroes by their image or any other desired values.
Replacement of unstable zeros with their image can be shown
to be optimal in a least squares sense, and arises naturally in
1
the unconstrained CMV design (Q = Q 3 = 0). Stable zeros of
2 1
Ι ω ί ζ - ) ! located near (—1,0) lead to unfavourable input
ringing which should be removed using ν (z- ).
+2
The selection of G ( z - ! ) , as given by equation 15, will lead to
c
the closed loop output response
Y(t) = N(t) + V + i ( z - i ) v ( z - i ) [ F ( z - 1 ) Y (t)
+ 2 2 sp
•Fi(z-i)N(t)]
when no model mismatch is present. Equation 16 shows that
the IMC design requires that time delays and nonminimum
phase c h a r a c t e r i s t i c s become part of the c l o s e d - l o o p
specification.
If F i ( z - l ) and F ( z - l ) are removed from the control block
2 C H + H -»2CH
diagram in figure 1, the IMC design using equation 15 will
lead to deadbeat control actions for step changes. Deadbeat
control actions can lead to harsh actuator manipulations and
severe rippling in the true process output. It can also be
shown that the stability of the closed-loop input and output
1
transfer functions are very sensitive to mismatch between
Gmiz- ) and G ( z - ! ) . The filter blocks serve to relax the
p
specification of perfect control and deal with these two
1
important problems. The simplest and most common choice
for the filter blocks F i i z - ) and F ( z - ! ) are diagonal
2
(decoupled) first order exponential filters with a gain equal to
I as givetvbeVov*
where T is the sampling interval and T i.. is the closed-loop
s c
time constant for output i. A filter factor corresponding to
each output can be used to shape the closed-loop response and
reduce the severity of input manipulations. It has been shown
by Garcia and Morari (1985a) that there exists ΐα for each
output which is sufficiently large to guarantee robustness to
arbitrary large model mismatch provided
n. The
p m
appealing feature of the IMC design is that a single, physically
meaningful filter tuning factor for disturbance and setpoint
changes for each output can be used for on-line tuning once a
c
minimum variance design where the Wiener-Hopf design
and Q3.
2
The decoupled IMC design is of the correct structure if the
m
manipulated such that the minimum dead time in each row
(15)
appears on the diagonal. High performance control of an
unbalanced multivariate process is not possible with the
decoupled G ( z - ! ) design, especially if there is a large
c
difference between the minimum dead time and the minimum
settling time. The search for a least squares optimal inverse is
far from trivial when the process is not balanced. An easy
approach to solve this problem is to use the results of the
constrained minimum variance controller design for a
nonstationary noise model. The right matrix spectral factor,
determined by equation 9, provides a stable least squares
approximation to the inverse of ω ( ζ - ! ) and is a function of
only the process model parameters. The control block in the
IMC design is then given by equation 8. Severe harshness of
the input movements from the model inversion and input
ringing can be prevented by a suitable specification of the Q
input weighting matrix. IMC filters can then be tuned on-line
2
in the usual way.
3. PILOT PLANT REACTOR SYSTEM
The process under investigation is a pilot plant packed bed
reactor carrying out highly exothermic butane hydrogenolysis
r e a c t i o n s over a n i c k e l on s i l i c a gel c a t a l y s t . The
hydrogenolysis of butane can be represented by the following
overall reactions:
(16)
C Hio + H - > C H
4 2 3 8+ CH4
C Hi
4 0+ H ->2C H
2 26 (19)
C H
3 8+ H -*C H
2 2 6 4- CH
4
26 2 4
Only three of the reactions shown above are stoichiometrically
independent, but in the usual operating region of the reactor
only two are practically independent (MacGregor and Wong,
1980). Therefore, only two output or product species need to be
controlled.
The reactor consists of a 1 1/2 inch Schedule 80 stainless steel
pipe mounted concentrically within a carbon steel pipe The
inner pipe contains the catalyst bed and the outer pipe forms a
cooling jacket l'or a heal transfer oil circulation loop The
288 D. J. Kozub, J. F. MacGregor and J. D. Wright
length of the catalyst bed is 89 cm. The feed gases, hydrogen
and butane are mixed up stream and enter the reactor with
back pressure set to 14 psig. Thermocouples positioned on the
axis of the reactor provide temperature profile readings along
the length of the bed. The region of operation is set by
maintaining a constant wall temperature setpoint. On/off
control of an air source to a shell and tube heat exchanger in
the oil circulation loop is used to regulate wall temperature.
An on-line gas chromatograph system sampling the effluent
provides concentration measurements at 3 minute intervals.
The process control configuration of the pilot plant is shown in
figure 2. Percent butane conversion and propane production
(cc/sec) are the outputs controlled by a multivariate outer loop
controller which computes control actions at three minute
intervals from the gas chromatograph readings. The
multivariate control variables are hydrogen flow rate setpoint
(cc/sec) and hotspot setpoint (C). Hotspot setpoint is sent to an
internal hotspot control loop which uses the butane flow rate
setpoint to control the difference between the reactor wall
temperature setpoint and the hottest temperature detected by
the thermocouples in the bed. Control actions are taken by the
hotspot loop at 30 second intervals. Temperature gradients
exist in the reactor bed due to the highly exothermic reactions
and the limited ability to remove heat from the center of the
reactor through the reactor wall. A typical temperature
profile along the axis of the bed is given in Figure 3. The
purpose of the internal hotspot loop is to provide a safe
open-loop stable configuration for multivariate control.
Hydrogen and butane flow rate setpoints are determined by
the multivariate control loop and the hotspot inner loop
respectively. The flow rate of the two feed streams are
controlled once every 2 seconds by discrete PI controllers.
The hotspot temperature controller is a "one-step optimal" LQ
controller tuned by using the implicit self-tuned controller of
Clarke and Gawthrop (1975). The self-tuning approach was
used simply as a convenient way to tune the temperature
controller. It was not used in a continuously adaptive mode.
Since the response of hotspot temperature change to change in
butane flow rate is v e r y n o n l i n e a r , a l o g a r i t h m i c
transformation of the hotspot temperature was used to
linearize the temperature control loop. The controlled output
after the linear transformation becomes
Y(t) = in(hotspot (t)) - €n(hotspot setpoint (t)) (20)
where
hotspot(t) = hotspot temperature (t) — wall temperature (t) (21)
This nonlinear transformation was found to greatly improve
the performance of the self-tuning regulator. The choice of the
input magnitude constraint factor in the one-step optimal
algorithm was used to provide a trade off between acceptable
hotspot control to stabilize the system and the requirement of
smooth variations of the butane feed rate. The second
requirement was necessary since the end objective was to
control the reactor outputs rather than hotspot temperature.
Details of this controller design are given by underwater,
MacGregor, and Wright (1986).
The pilot plant is operated through a distributed control
network. The computer hardware consists of a DIGITAL
VAX11/750 host computer connected to a DPM-23 Front End
minicomputer. A VISTA 6000 gas chromatograph and a
VISTA 401 GC control station is interfaced to the computer
hardware. Communication between the user and the field is
handled by real time software running in the VAX11/750 and
DPM-23. I/O and low level control operations are performed in
the DPM-23 Front End while high level control, operator
communication, and data logging are performed by the VAX
host.
4. MODEL IDENTIFICATION
A theoretical investigation of the process using fundamental
material and energy balances can be performed to develop a
mechanistic dynamic model Jutan (1976) has shown that this
will lead to four simultaneous, non-linear partial differential
equations. Detailed knowledge of the reaction kinetics and
process parameters together with lengthy mathematical
model reduction operations are required to arrive at a useable
model for control.
In this work an empirical transfer function approach was
selected to model the dynamics of the system based on
input/output data. This technique eliminates the tremendous
effort required in mechanistic modelling at the expense of
losing some important insight into the process. Multivariate
time series analysis combined with process identification
techniques were used to develop an empirical linear model.
The details of this identification technique are covered in
great depth by Wilson (1970) and Box and Jenkins (1970).
This approach was used to identify a linear transfer function
and noise model which approximates the dynamics of the
process in the region of operation.
Input/output data were generated by pseudo random binary
perturbations of the input variables. Hydrogen flow rate was
switched between 68 and 83 cc/sec. Hotspot setpoint was
switched between 5 and 9 C above the wall temperature
setpoint. Wall temperature setpoint was kept constant at
232 C during the data gathering. Two switching frequencies
were used in the identification. A switching frequency of
6 minutes was taken in the first half of the data set to allow
adequate identification of the fast output dynamics associated
with changes to inlet concentration and flow rate. A
20 minute switching frequency was chosen in the second half
to allow proper identification of the slow output dynamics with
change in temperature.
The input/output and time series data was used to identify an
interactive model in the Wilson canonical form
Y(t) = V ( z - l ) U ( t ) + N(t) (22)
4>(z-i)N(t) = 9(z-l)A(t) (23)
where
1
V(z - !): η X m rational polynomial matrix in z -
1
φ(ζ-ΐ): is an n X n diagonal polynomial matrix in z -
having nonrational elements and z0 coefficient
equal to I
1
θ(ζ-ΐ): is a full n X n polynomial matrix in z - having
nonrational elements and z0 coefficient equal
toi.
Maximum likelihood parameter estimation and diagnostic
checking of resulting fitted model led to the approximate
linear model
Y(t) •
1 2
1.8 - 6.643z + 5.158z 0.5384z"
1 2
-0.7498 - 3.398z - 1.012z
l-0.6143z"
U(t-3)
(1-z )
A(t)
(1-z )
(24)
where
y( 1 ,t) : propane production - 24 (mmol/sec) 1
y(2,t): butane conversion - 29.6%
u(l,t): (hydrogen flow rate - 75 cc/sec} X 1 0 -
u(2,t): hotspot setpoint - 7 (C)
and the covariance matrix for the innovations A(t) was
 Multivariable Control of a Catalytic Tubular Reactor
? — I >
\mJ
I 0.043 0.200 ]
This process model can be shown to display nonminimum
phase response with respect to propane production. The model
is not invertible and is not balanced with respect to time
delays. Furthermore, there was evidence of a fairly high
degree of process nonlinearity over the wide range of outputs
and inputs to be considered in the control studies. Therefore,
this model could only be considered to be a rough linear
approximation to the true process dynamics, and in the design
of the controllers robustness to process model mismatch was
an important consideration.
5. C O N T R O L STUDIES
Constrained m i n i m u m variance and internal model
controllers were designed using the model transformed into a
right matrix fraction form. The control objective for both
designs was to achieve good regulatory and good setpoint
performance, while at the same time providing adequate
robustness to the large degree of model mismatch that will be
present due to nonlinearities and changing catalyst activity.
5.1 Constrained Minimum Variance Control Runs
The cost function used for designing constrained minimum
variance controllers for the reactor was
J = E{eT(t) Qi e(t) + VUT(t) Q V U(t)} (26)
2
The model, given by equation 24, contains a commonly
occurring nonstationary integrated moving average
disturbance model. A short open-loop run on the reactor
showed that this disturbance led to substantial drifts in the
conversion and propane production rate away from
steady-state values. The controller was to be designed to
compensate for this nonstationary stochastic disturbance, and
to accommodate step changes in the setpoints of either output.
The model for randomly occurring step changes in setpoints is
given by
VY (t) = a(t) (27)
1 sp
where V = (1 — ζ- ) I. Since the structure of this model for
setpoint changes is very close to that for the stochastic
disturbances in the process it was considered sufficient to
1 _1
design the C M V controller for the IMA disturbance model in
24 and take Fjiz- ) = F ( z ) in the controller given by
2
equations 5, 6, and 7. This, of course, would be true if the
disturbance models were identical (Harris and MacGregor,
1986). The C M V controller used was therefore the single
degree of freedom controller
U(t) = Hi(z-i) [G (z-i) U(t) - [Y(t) - Y (t)]] (28)
m 8p
The form of the cost function and the disturbance model will
lead to a controller with integral action. The output weighting
matrix Qi was set to I for all controller designs and Q was
2
adjusted to provide the required amount of input constraint
and controller robustness.
Figure 4 shows a control run on the same catalyst charge used
for the model identification. A n input constraint matrix of
Q = diag(100,10) was selected for this run. The dashed lines
2
in figure 4 indicate operating setpoints. The abscissa on these
figures and others to follow refers to the sampling interval,
where each sampling interval represents 3 minutes. The solid
line of the input settingfigurerepresents hydrogen flow rate
setpoint (cc/sec χ 10-*) and the dashed line represents hotspot
setpoint (C). Excellent regulation can be observed in the first
5 hour section of data in Figure 4. In the remaining part of the
run a variety of conversion and propane production rate
setpoint changes were made, some in the same direction, some
in opposite directions and some where only one setpoint was
changed. In all cases the setpoint change performance of the
controller was good, and it appeared to be robust over the wide
range of operating conditions studied No significant
289
closed-loop interaction can be observed when the final two
setpoint changes were made where one of the two setpoints
was held fixed. Figure 4 shows that input manipulations were
smooth and well behaved during the entire control run.
Figure 5 shows a control run on a new catalyst charge using
the same Q = diag(100,10). The outputs, especially percent
2
conversion, of the new charge are noticably noisier than those
in figure 4. However, regulation and setpoint change
performance was still good.
A more tightly tuned constrained minimum variance control
run is shown in figure 6 with less constraint on the input
variances using Q = diag(50,5). This control run was
2
operated with a wall temperature of 233 C and all setpoint
changes were raised 5 units relative to the previous run. The
adjustment was made to compensate for significant observed
changes in the catalyst activity. A high degree of oscillations
were observed with the inputs and propane production. The
problem became more severe when a setpoint change was
made to a region of low propane production and high
conversion. The unexpected poorer performance relative to
the previous run can be attributed to loss of robustness by
tighter tuning which made the controller more sensitive to
model mismatch and nonlinearities.
5.2 Internal Model Control Runs
A simple decoupled IMC controller was designed using the
identified transfer function model in equation 24. A stable
model inverse G (z-!) was designed according to equation 15.
c
A n unstable pole resulting from a non-invertible zero was
reflected inside the unit circle, and a ringing pole arising from
a negative real axis zero was shifted to the origin. Both these
adjustments were accomplished using v + (z-*) in
2
equation 15. Identicalfirstorder diagonalfilters(17) were
used for Fi(z-1) and F (z- !). Details of the design procedure
2
are outlined by Kozub (1986).
A control run with thefilterconstants f^ tuned to diag(0.7,0.7)
is shown infigure7. This run was performed immediately
following the C M V run infigure5 with the new catalyst
charge. The identical pattern of setpoint changes were made
in all of these runs in order to facilitate comparisons. Setpoint
change performance and regulation were good for thefirsttwo
setpoints. The second setpoint change demanding high
conversion with low propane production led to very poor
controller performance. Large oscillations observed with the
inputs and propane production suggest controller instability
in this region of operation. Complete recovery by the
controller can be seen when a setpoint change was made in
propane production rate back to a more stable region of
operation.
In order to increase the robustness of this IMC controller to
accommodate such setpoint changes the controller was
detuned by using largerfilterconstants in F(z-*) namely
f
F = diag(0.9,0.9). The control run is shown in figure 8.
Detuning led to substantial reduction in the variation of the
input manipulations. Regulation still appeared to be good over
the full region of operation. Improved robustness to the
process nonlinearities can be observed by the stable controller
performance when the setpoint changes were made to the
region of high conversion and low propane production. No
significant interaction was observed when thefinalsetpoint
change was made to propane production with conversion
setpoint held fixed.
Thefilterblock was retuned to improve the setpoint change
response of butane conversion. A control run with
F = diag(0.9,0.8) is shown infigure9. This control run was
performed at a higher wall temperature of 233 C and all
setpoints were shifted 5 units higher relative to the previous
runs. The process disturbance noise was noticed to be slightly
higher with this change in operating conditions. Again,
regulation was adequate over the entire region of operation
although slightly increased deviation of propane production
was observed in the region of low propane production and high
butane conversion Substantial improvement in the setpoint
290 D. J. Kozub, J. F. MacGregor and J. D. Wright
change response of butane conversion can be observed in
figure 9 relative to the previous controller, as a result of the
smaller filter constant (0.8 vs. 0.9).
5.3 Discussion of Experimental Results
The results indicate that the tightness of control for both
controller designs was restricted by severe model mismatch,
mainly due to severe process nonlinearities and changing
catalyst properties. Trial and error adjustment of the tuning
matrices led to robust designs with acceptable regulation and
setpoint change performance. I m p r o v e m e n t in the
performance of these controllers would require use of a
- 1
complicated adaptive mechanistic nonlinear model in the
G m ( z ) block to eliminate some of the significant and time
varying sources of model mismatch.
The quality of performance of both controller designs was
approximately the same in this investigation, although
slightly improved propane production control was evident in
the c o n s t r a i n e d m i n i m u m v a r i a n c e c o n t r o l r u n s .
Improvement of propane production control in the IMC design
might have been achieved by further on-line adjustment of the
filter block.
The results also indicate that the inner hotspot control loop
design provided a safe stable configuration for application of
good multivariate control. Even when marginally stable
multivariate control actions were present, as in figures 6 and
7, the reactor did not runaway.
6. CONCLUSIONS
An application of multivariate constrained minimum variance
and internal model controller designs to composition control of
a packed bed butane hydrogenolysis reactor was considered to
be very successful. A simplified linear empirical input/output
pulse transfer function model was used in these designs. Both
linear designs, when well tuned, were robust to process
nonlinearities and provided good regulatory and servo control
over a wide region of operation. Performance of these
controllers is limited by model mismatch between the
approximate linear empirical model and the true nonlinear
process, indicating that incorporation of a rigourous
mechanistic model in these controller designs might be
desirable if tighter control were necessary. The multivariate
control loop proved to be successful in stabilizing the reactor
system for applicaton of safe multivariate control.
REFERENCES
Bergh, L.G., and J.F. MacGregor (1986). Constrained
Minimum Variance Controllers: Internal Model Structure
and Robustness Properties, submitted to Ind. Eng. Chem..
Proc. Pes. Dev.
Box, G.E.P., and G.M. Jenkins (1970). Time Series Analysis:
Forecasting and Control. Holden Day, San Francisco.
Clarke, D.W. and B.A.Gawthrop (1975). Self-Tuning
Controllers. Proceedings IEE, 122, pp. 929-934.
Garcia, C.E. and M. Morari (1985). Internal Model Control. 2.
Design Procedure for Multivariate Systems. Ind. Eng. Chem.
Process Pes. Dev.. 24, 472-484.
Garcia, C.E. and M. Morari (1985). Internal Model Control. 3.
Multivariable Control Law Computation and T u n i n g
Guidelines. Ind. Eng. Chem. Process Pes. Dev.. 24,484-494.
Harris, T.J. and J.F. MacGregor (1986). Design of Discrete
Multivariable Linear Quadratic Controllers Using Transfer
Functions. Submitted to AIChE Journal.
Jutan, A. (1976). State Space Modelling with Multivariate
Stochastic Control of a Pilot Packed-Bed Reactor. Ph.D.
Thesis, McMaster University, Hamilton, Canada.
Jutan, Α., J.P. Tremblay, J.F. MacGregor and J.D. Wright
(1977). Multivariate Computer Control of a Butane
Hydrogenolysis Reactor: Part III. On-line Linear Quadratic
Stochastic Control Studies. AIChE Journal, 23, 5.
Kozub, D.J. (1986). Advanced Model-Based Multivariate
Control Design and Application to Packed Bed Reactor
Control. M.Eng. Thesis, Dept. of Chem. Eng., McMaster
University, Canada.
MacGregor, J.F. and A.K.L. Wong (1980) Multivariate Model
Identification and Stochastic Control of a Chemical Reactor.
Technometrics. 22,4,435-464.
MacGregor, J.F., P.A. Tayor and J.D. Wright (1984).
Advanced Process Control, An Intensive Short Course on
Digital Computer Techniques for Process Identification and
Control. McMaster University, Hamilton, Ontario, Canada.
Onderwater D., J.F. MacGregor and J.D. Wright (1986)
Nonlinear Temperature Control of a Catalytic Tubular
Reactor Using Self-Tuning Regulators. Submitted to Can. J.
Chem. Eng.
Sorensen, J.P., S.B. Jorgensen and K. Clement (1980) Fixed
Bed Reactor Kalman Filtering and Optimal Control - I
Computational Comparisons of Discrete vs. Continuous Time
Formulation. Chemical Engineering Society, 35, 1223-1230,
Pergamon Press Ltd.
Sorensen, J.P., S.B. Jorgensen and K. Clement (1980) Fixed
Bed Reactor Kalman Filtering and Optimal Control - II
Experimental Investigation of Discrete Time Case with
Stochastic Disturbances. Chemical Engineering Society. 35,
1231-1236, Pergamon Press Ltd.
W a l l m a n , P . H . , J . M . S i l v a , and A . S . F o s s ( 1 9 7 9 )
Multivariable Integral Controls for Fixed Bed Reactors. Ind.
Eng. Chem Fundam., 18, 4, 392-399.
W i l s o n , G.T. (1970) Modelling Linear S y s t e m s for
Multivariate Control. Ph.D. Thesis, University of Lancaster,
England.
 Multivariable Control of a Catalytic Tubular Reactor

N(t)
(t) U(t)
H
j-
G (z-')
F (z-')
2 c G (z )
p

FIGURE 1. I n t e r n a l Model Controller Block

Diagram

U
H
2Setpoint

Hydrogen

Butane

= Controller

FIGURE 2. Control Configuration

24S.9

! Ιββ
.

Reactor L e n g t h (cm)

FIGURE 3. A x i a l T e m p e r a t u r e P r o f i l e With
Wall Temperature = 233 C
 292

iiàimnîàiï
,.,..·: V: V. V . — '

<s/-ioww) 'aoyd ε ο
ïtiiiïmmï

NoisaaANoo κ
Ό. J . Kozub, J . F. MacGregor and J . D. Wright

S0NIJLJ3S XndNI

SONIJLJLSS xndNi
isa»*»»»!

FIGURE 4. CMV Run With Old Catalyst Charge FIGURE 5. CMV Run With New Catalyst Charge FIGURE 6. CMV Run With New Catalyst Charge
Using Q = Diag(100,10) Using Q = Diag(100,10) Using Q = Diag(50,5)
 Multivariable Control of a Catalytic Tubular Reactor 293

IMC Run With New Catalyst Charge

τ

Using F = Diag(0.9,0.8)
T
I 5
5
.,Λ.

FIGURE 9.
imïïiiimï 5ÎÎÎÎÎÎIÎHÎ5 5

<S/~IOWN) 'aoad so N0ISd3AN00 κ S0NI113S XHdNI

IMC Run With New Catalyst Charge FIGURE 8. IMC Run With New Catalyst Charge
Using F = Diag(0.9,0.9)

S0NI113S XRdNI

;;; ί
Τ Γ.*τ r.f : : : : : : : .·
Using F = Diag(0.7,0.7)

.^••·:·:·:·:·:·:* ·········· ,

'•'••••eï/*"
.......

Ϊ
ι•ÎI.Ί

Vf.- 8 ω S en
;
j
FIGURE 7.

2222222222222
cs/ioww) "aoyd εο N0ISd3AN00 κ S0NIJ.13S JJIdNI
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

DESIGN AND CONTROL OF TUBULAR

AUTOTHERMAL REACTORS:
AN EVOLUTIONARY APPROACH
R. W . Chylla, Jr., R. A. Adomaitis and Ali Çinar
Department of Chemical Engineering, Illinois Institute of Technology, Chicago,
IL 60616, USA

A b s t r a c t . A systematic procedure for the design of a reactor system and controller for an au-
tothermal process is proposed. Use of Structural Dominance Analysis eliminates the need for
many detailed simulation runs to determine best reactor and controller configuration. Multistage
wall-cooled reactors with cocurrent coolant are found to be a superior design. Reactor control
is achieved using cold feed bypass gas to regulate in inlet temperature to each reactor bed. A
controller is designed using the Internal Model Control structure. Performance and robustness
are investigated using a first-order diagonal filter.

K e y w o r d s . Autothermal reactors; control system analysis; filtering; Internal Model Control;

model reduction.

industry. It is only with the accumulated experience of

1. I N T R O D U C T I O N
Tubular autothermal reactors are of significant inter-
est industrially as well as academically. Ammonia and
methanol synthesis and hydrocarbon oxidation reactions
are examples of industrial processes which operate in
an autothermal manner. The inherent thermal feedback
mechanism is a source of consternation to both the design
engineer who must insure the reactor operates economi-
cally and the control engineer who is ultimately responsi-
ble for designing a control system to operate near design
specifications in spite of process disturbances and load
changes. Recently, several studies have examined these
reactors with various reactor designs, controller config-
urations and differing control objectives (Metchis, 1982;
Bonvin, Rinker and Mellichamp, 1983; Mandler, Sein-
feld and Morari, 1985). This study attempts to present a
methodology whereby the design problem and the control
problem are unified and examined in a logical step-wise
procedure where all the design and control options can
be identified and evaluated based on the objectives for re-
actor operation and control. Several powerful techniques
recently developed for analysis of large-scale systems in-
cluding model reduction and identification of the domi-
nant system properties are applied with excellent results.
The controller design is handled using the Internal Model
Control (IMC) framework which is ideally suited for re-
actor systems of this type.
2. C O N F I G U R A T I O N I D E N T I F I C A T I O N
There are many potential configurations for the reactor
and controller that must be evaluated by the design engi-
neer. While it can be said that there is no "best* configu-
ration for all applications, much of the diversity in design
is due in part to a lack of fundamental understanding of
the inherent system properties and their interaction with
one another. This is by no means through lack of effort;
indeed, the evaluation of such energy integrated systems
is a task that is not at all obvious as demonstrated by
the number of different process designs currently used in
these designs, new analysis methods recently developed
and current computer technology that the computational
effort can be reduced. With the reactor bed itself, pos-
sible designs include single reactor beds cooled from the
wall using incoming feed gas, staged adiabatic beds with
intercooling or some combination of the two. This is rep-
resented in Fig. 1. If wall-cooling is being used, the feed
gas can be passed either countercurrently or cocurrently
leading to potentially two very different operations. The
intercooling between beds in multi-bed configurations can
be accomplished through cold-shot injection of feed gas
or feed-effluent heat exchange. Then there is the choice
of manipulated and measured variables for the particular
configuration used. Obviously, constructing a detailed re-
actor model for each possibility and simulating over the
entire range of operating parameter space would be an
impossible task. Yet the problem can still be tackled sys-
tematically.
3. P H A S E P L A N E A N A L Y S I S
Determining the potential region of operation for a given
set of reactor parameters is relatively easy for an adia-
batic reactor without wall cooling. Integration of a simple
plug flow performance equation serves quite nicely.
When the reactor is being cooled from the wall using in-
coming feed gas the problem becomes considerably more
complicated. A technique recently developed by Chylla
and coworkers (1986) handles the case of countercurrent
flowing coolant rather well. A reactor phase plane shown
in Fig. 2 is computed for a set of reactor design param-
eters using the inlet coolant temperature as the absissa
and the inlet reactor temperature as the ordinate. Four
lines are determined using a simple algorithm explained
in the above reference. These lines explicitly determine
the region of safe operation and potential autothermal
operation given the reactor length, diameter, flowrate
and overall heat transfer coefficient. Lines (A),(B) and
(C) envelope the region of potential safe operation. The

DCCR-T
295
296 R. W. Chylla, Jr. R. A. Adomaitis and A. Çinar
line designated as the autotherm approximately indicates
the resulting exit coolant temperature for a given inlet
coolant temperature and reactor conditions. If this line
falls within the safe operating zone than autothermal op-
eration is possible. Because the algorithm can be quickly
implemented on a computer, a large range of reactor pa-
rameters can quickly be evaluated for maximization of
the range of safe operation. Because the minimum inlet
coolant temperature is explicitly given, it becomes im-
mediatly known whether preheating or additional feed-
effluent heat exchange will be required. Two phase planes
for a CO oxidation reactor are shown in Fig. 3. For an
inlet concentration of 1% CO, the majority of the au-
totherm fall outside envelope and autothermal operation
is not feasible. This is verified by the actual points ob-
tained using a detailed reactor model. When the inlet CO
concentration is reduced to 0.6%, however, autothermal
operation becomes feasible. Because the minimum inlet
coolant temperature is 40-70 °C above ambient temper-
ature, some additional feed preheating is indicated. This
method is extremely useful at the early design stage in
selecting D/L ratios, flowrates and examining the feasibil-
ity of autothermal operation. Extension of the technique
to cocurrent operation is currently underway.
4. S T R U C T U R A L D O M I N A N C E A N A L Y S IS
Dynamic reactor models necessary to describe reactors
consist, in general of nonlinear partial differential equa-
tions. The number of equations can differ from two for a
case of an adiabatic bed, to seven for a completely het-
erogeneous reactor model with wall cooling. In order for
these models to be solved numerically, axial and/or ra-
dial discretization is required. The technique of orthogo-
nal collocation has been used extensively in many studies
successfully. This technique leads to large systems of or-
dinary differential equations that when linearized, can be
expressed in state-space form as
χ = Ax + B u (1)
These systems can range from 16th order for plug-flow
reactors with 8 interior collocation points to 120th or-
der for complex multi- bed reactors. Two facts become
immediately apparent:
1) A method capable of evaluating various reactor
configurations and manipulated variables without solving
these models over various ranges of operating conditions
would be highly desirable.
2) In order for any model-based control technique to
be used, model reduction is necessary.
The Structural Dominance Analysis technique (SDA) in-
troduced by Litz and developed by Bonvin and Mel-
lichamp is extremely useful in the evaluation of large-
scale systems and reduction of high-order models. The
technique is outlined clearly elsewhere (Bonvin and Mel-
lichamp, 1982) and only the results applied to autother-
mal reactor systems will be given here. SDA provides two
useful measures which characterize a state-space system.
The normalized coupling coefficient q i is a quantitative
am
measure of the effectiveness of the tth input on the 5 t h
state variable acting through the mth mode. This can be
used to evaluate the effectiveness of various inputs on the
autothermal reactor without detailed simulation studies.
The generalized dominance measure d i indicates the
am
dominance of the tth input on the sth. state variable
acting through the mth mode. The difference between
dominance measures is that d i is time weighted to ac-
am
count for the fact that slower modes tend to dominate
the system dynamics. For each input, therefore, the im-
portant modes (eigenvalues) of the reactor system can be
immediately identified and the selection of a reduced or-
der model of appropriate order is facilitated. It has been
shown (Bonvin and Mellichamp, 1983) that SDA can also
identify models of 'poor internal structure' in the sense
that these models cannot be accurately reduced. The
inclusion of second order dispersion terms in the reac-
tor model equations has been found to yield models with
good internal structure which are reduc ible even though
the presence of these terms is negligible in the numerical
results of the simulation.
The results of SDA for several configurations for a single
reactor bed is shown in Tables I, II and III. Table I con-
tains the complete SDA results for cocurrent operation
so that the reader can see the total results for one repre-
sentitive case. Table II contains only the more important
results for the remaining cases. The effect of four inputs
and/or disturbances on the reactor operation was exam-
ined:
1) Inlet concentration
2) Preheating the feed
3) Total flowrate
4) Inlet coolant temperature.
For all test cases examined (except the adiabatic reac-
tor case), the dynamics of the outer wall dominates the
system as can be seen from the d-values in Table II. Ex-
amination of the corresponding q elements for the outer
wall temperature in Table I indicate, however, that this
state variable is not sensitive to any of the inputs evalu-
ated. Herein lies the difficulty in controlling these reac-
tors. The conduction of heat through the outer wall de-
termines the speed of system response but cannot readily
be controlled. The inlet variable which has the most di-
rect impact on the outer wall is the temperature of the
coolant flowing adjacent to it. It should be noted that al-
though the outer wall temperature determines the system
dynamics, it has very little influence on the steady-state
profiles. This was observed by comparing model results
neglecting the outer wall equation and is noted in Chylla
and coworkers (1986). From the q elements associated
with the inlet coolant temperature, it can be seen that it
has a large overall effect on the system. Clearly, the reac-
tor inlet temperature is the most effective input variable.
Because of the dominating effect that the outer wall and
especially the inner wall have on distributing the heat ax-
ially along the reactor, there is not a marked difference
between countercurrent and cocurrent cooling flow. This
heat distribution effect was noted in regard to conduction
along thermocouple wells in ammonia reactors by Bad-
dour and coworkers (1965). Recent studies by Degnan
and Wei (1979) determined that a reactor was much less
parametrically sensitive to cocurrent coolant flow than
countercurrent coolant flow. Their theoretical and exper-
imental results are based on a reactor design that does not
incorporate metal reactor walls and a packed-bed which
both serve to distribute the energy through conduction
and heat loss to the environment through the outer wall
and insulation. It will be shown later that using cocur-
rent coolant flow is advantageous from a control point-
of-view. The reactor system is not particularly sensitive
to the total flowrate of the system except for the case of
countercurrent coolant at 407UT inlet temperature. This
operating point was chosen because independent studies
by Adomaitis and Cinar (1986) indicated this point to be
a bifurcation point with respect to the total flowrate, that
is, a point in which a slight decrease in the flowrate would
 Design and Control of Tubular Autothermal Reactors
cause the reactor to ignite to the upper steady-state. The
large jump in the q values for this case associated with
the total flowrate support these findings. Another inter-
esting finding is that the most sensitive state variables
are the reactor temperatures and surface concentrations
just before the hot spot in the system. This result sup-
ports the contention of Cinar (1984) that temperature
measurements for control purposes should be measured
just upstream of the hot spot rather than the more com-
mon practice of taking measurements at the hot spot.
These state variables upstream of the hot spot are more
sensitive to small load changes or disturbances and detect
dynamic changes sooner than state variables located at
the hot spot. This can be crucial to the quality of control
in industry when the number of measurements is limited
and costly. Since the most difficult process element to
control in any process is dead-time, early detection can
make the difference between success and failure.
The speed at which a particular input affects a process
can be gauged by which modes are excited by that input.
The eigenvalues closest to the origin will be the slowest
and hence, inputs which effect state variables principally
through these modes will take the most time to influence
the system behavior. The inlet coolant temperarature
and the inlet concentration excite the faster modes in the
system and influence the behavior rapidly. The net effect
of a particular input on the system can be determined
by examining the static gain of the input on that state
variable; this is done by simply summing q i over all
am
modes m:
Ν mi q 2 profiles corresponding to these two cases are shown in
= Σ * ()
m =l
The static gains for the adiabatic beds are shown in Ta-
ble III.
Naturally, the results presented for this reactor are not
universal for all reactor systems. Individual reactor sys-
tems are strong functions of the reaction kinetics, heat
transfer properties, tube and particle diameter and op-
erating conditions. More important is the methodology
through which these reactor systems can be evaluated.
Without numerically integrating large systems of stiff
equations over many sets of operating conditions, the in-
herrent system properties were evaluated using a simple,
transparent yet powerful technique. Nevertheless, the au-
thors contend that these results are valid, in general, for
the majority of autothermal reactor systems and a re-
silient design can be proposed from these results.
5. P R O P O S E D R E A C T O R D E S I G N
The ideal reactor system would achieve high conversion
with mild temperature profiles, be resilient to process
disturbances, and be easily controllable with a minimum
amount of interaction. While a system exhibiting all of
these properties cannot be realized in practice, using the
SDA results and some engineering common sense, a re-
actor system can be designed approaching the ideal.
The proposed reactor design is depicted in Fig 4. Because
of the high degree of exothermicity in these processes,
relatively large temperature gradients have to be toler-
ated in the reactor bed if staged adiabatic beds are to
be used. Many stages are necessary if both mild profiles
and high conversion is desired. Since SDA shows that in-
terstage cooling by quench injection is the most effective
method, a large control effort is required to maintain all
297
the beds at the desired operating point and the potential
for undesirable interaction effects is high since increasing
the bypass quench flow to one bed can have detrimental
effects on the others. For these reasons, wall-cooled reac-
tors are chosen in spite of the fact that they tend to be
parametrically sensitive.
Structural Dominance Analysis shows that this paramet-
ric sensitivity can be reduced greatly by independently
controlling the inlet feed temperature to the reactors.
Rather than allowing disturbances in the coolant stream
propagate through the reactor as the coolant is turned
around into feed gas, independent temperature regula-
tion of the inlet temperature will isolate the disturbances
locally. This temperature can be regulated by bypass-
ing some cold feed and mixing after the feed leaves the
coolant annulus. Since changing the total flowrate has
little effect on the system, this control scheme is quite sat-
isfactory. When the reactor is operated in an open-loop
manner, the reactor multiplicity exhibits itself in a "feast
or famine" operation where either very low conversions
with mild temperature profiles or complete conversion
with unacceptable hot spots occur. This can be readily
seen by recalling the phase plane for 1% inlet CO in Fig.
3. Allowed to operate open-loop the possible operating
states correspond to a lower steady-state with only 16%
conversion or an upper steady-state where the autotherm
lies above the upper operating zone and the temperature
profiles are too severe. If the inlet temperature was in-
dependently manipulated by only 10 °C, however, stable
operation is possible. The temperature and conversion
Chylla and coworkers (1986),
High overall conversion can be acheived by staging only
two or three of these wall-cooled beds in series and con-
trolling the inlet temperature to both reactors by inter-
stage cold-shot injection. Earlier it was noted that SDA
indicated there was little difference between countercur-
rent or cocurrent cooling on the operation of the reactor
because of the distribution of heat due to the conduction
up and down the metal reactor walls. There is a signifi-
cant difference in terms of the ease of controllability be-
tween the two operations, however. When countercurrent
cooling is used, disturbances in the second reactor bed
will reach the first reactor bed through the coolant tem-
perature. In cocurrent operation, however, downstream
reactor upsets cannot travel upstream and the reactor in-
teraction is only one-way.
Finally, the inlet coolant temperature was shown to be an
important state variable. Indépendant control of the in-
let bed temperature lessens the effect but not completely.
Therefore, the inlet coolant temperature to the first re-
actor bed should also be closely regulated. While inlet
coolant temperature regulation in the second bed would
be desirable, it is impractical from an operating stand-
point. If reasonable control is attained in the first reac-
tor bed, manipulation of the inlet coolant temperature
to the second be should be unnecessary. Because Phase
Plane Analysis indicated that the inlet coolant temper-
ature to the system is above ambient, feed-effluent heat
exchange between the effluent from the second reactor
and the coolant can be used. The temperature of the
coolant entering the reactor must be regulated using a
bypass stream, however.
The resultant configuration achieves high conversion with
relatively mild temperature profiles, is easy to control and
exhibits only a small amount of interaction.
298 R. W. Chylla, Jr. R. A. Adomaitis and A. Çinar
6. C O N T R O L L E R D E S I G N
A ΜΙΜΟ controller must now be designed to control the
inlet temperature for each bed as well as the inlet coolant
temperature. The controller is designed using the IMC
technique. Internal Model Control has received consid-
erable attention recently because the IMC structure al-
lows a rational controller design procedure where control
performance and robustness are addressed directly (Gar-
cia and Morari, 1982). Before any model-based control
scheme can be used, model reduction is necessary to re-
duce the plant matrix (120th order) to a practical size.
Model Reduction
An excellent review of modal-based model reduction tech-
niques by Bonvin and Mellichamp (1982b) demonstrated
the power and accuracy of the technique proposed by Litz
(1980). Modal techniques continue to be of interest since
they permit they permit closed-form formulation. The
reduced model is derived directly from the linear large-
scale system through algebraic relationships. One of the
principal advantages of using the Litz technique is that
the reduction technique is independent of the state vari-
ables to be retained. This is a distinct advantage in that
state variables may be chosen which are best from a con-
trol and/or measurement viewpoint. The Litz technique
is described in detail elsewhere (Bonvin and Mellichamp,
1982b; Litz, 1980) and is beyond the scope of this paper.
Essentially, the contributions of the fast modes ("nondom-
inant') to the reduced order model are approximated by
an optimal linear combination of the slow modes:
z
2= E*i (3)
where and z \ are the canonical 'nondominant* and
'dominant* state vectors, respectively. The matrix Ε is
determined optimally through Lagrange multipliers so as
to minimize a weighted intergral of the square of the error
between z i and the reduced order approximation z ^ . The
reduced order system model is expressed as
xr = Α χ
Λ Λ+ B Ru (4)
IMC Design
Internal Model Control is reviewed by Garcia and Morari
(1982); design procedure for ΜΙΜΟ systems are dis-
cussed by Garcia and Morari (1985a) and Economou
and Morari (1986) and tuning guidelines by Garcia and
Morari (1985b). The design procedure used is summa-
rized here. The controller design is conveniently done
in the θ-domain, therefore the process transfer function
G(s) is defined
1 tentially achieve high conversion with relatively mild tem-
G(s)=C {sl-A )~ B .
R R R (5)
One of the properties of IMC is that the only require-
ments for a closed-loop configurartion to be stable are
that both the process G(s) and the controller G (s) be
c
open-loop stable. It has been shown (Garcia and Morari,
1982) that the "perfect" controller is implied by the con-
trol law
G = G~\ (6)
C A controller was designed using the IMC structure with
This controller is "perfect" in the sense that such a con-
troller which produces a stable IMC loop will result in
zero offset. Thus, integral action is achieved automat-
ically by making the steady-state gain of the controller
the inverse of the steady-state gain of the process. Be-
cause this controller is generally not realizable in practice,
the IMC design procedure is accomplished in two steps.
First, the system transfer matrix is factorized into the
product of two matrices
G = G+G_. (7)
G+ contains the time delays and and other nominimum
phase elements and is equal to an identity matrix at
steady-state. The other matrix C?_ has a stable realizable
inverse and its inverse gives the controller
l
G = GZ . (8)
c
The second step involves detuning the controller with a
diagonal first-order filter
η * ) = * Μ ψ ϊ \ \ (»)
This filter accounts for the fact that controller (6) is
highly sensitive to modeling errors or uncertainties. Un-
certainties caused by changes in plant parameters or by
simplification of model structures can be represented by
additive uncertainties, making this design procedure very
attractive for robust controller design for the autothermal
reactor system. Thus the performance vs. robustness is-
sue is addressed in a very direct and transparent man-
ner by adjusting the diagonal filter parameter for each
loop i,r^. The onset of oscillations and/or instability
in the loops implies that the existing plant/model mis-
match does not allow high performance and the rji's are
increased until the stability of the system is guaranteed.
The resulting controller and performance vs. robustness
results will be available for presentation.
7. S U M M A R Y A N D C O N C L U S I O N S
A rational systematic procedure was proposed to design a
reactor system and controller for autothermal operation.
The potential design alternatives were identified and eval-
uated using several techniques including numerical simu-
lation. Phase Plane Analysis was shown to be extremely
useful in early stages of design for identifying potential
regions for autothermal operation in countercurrent wall-
cooled reactors. Structural Dominance Analysis was used
to gauge the relative parametric sensitivity of several de-
signs, determine the most effective manipulated variables,
identify the most dominant modes for model reduction,
and ascertain the benefits and problems of several reactor
designs.
Staged cocurrent wall-cooled reactors were shown to po-
perature profiles. Independent regulation of the inlet
temperature to each reactor bed was necessary to suc-
cessfully control the reactor; this temperature control
was achieved by manipulating a cold bypass stream in-
jected before each bed. When the number of process
measurements is limited, reactor temperature measure-
ments should be made at locations just before the reactor
hotspot.
a reduced-order model obtained using the Litz model re-
duction technique. This controller was augmented with
a filter containing one adjustable tuning parameter for
each loop. This tuning parameter directly influences the
robustness of the controller thus simplifying on-line per-
formance adjustments as well as giving an indication of
the model accuracy with respect to current operating con-
ditions.
 Design and Control of Tubular Autothermal Reactors 299

REFERENCES Economou, C.G. and M. Morari, (1986). Internal model

Adomaitis,R.A. and A. Cinar, (1986). The bifurcation be-
havior of an autothermal tubular reactor. Submitted
for presentation at the 10th World Congress of IFAC.
Baddour, R.F., P.L.T. Brian, B.A. Logeais and J.P.
Eymery, (1965). Steady- state simulation of an am-
monia synthesis converter. Chem. Eng. Sci, 20, 281-
292.
Bonvin, D. and D.A. Mellichamp, (1982a). A general-
ized structural dominance method for the analysis of
large-scale systems. Int. J. Control, 3 5 , 5, 807-827.
Bonvin, D. and D.A. Mellichamp, (1982b). A unified
derivation and critical review of modal approaches
to model reduction. Int. J. Control, 35, 5, 829-848.
Bonvin, D. R.G. Rinker and D.A. Mellichamp, (1983).
On controlling an autothermal fixed-bed reactor at
an unstable state-Ill, Chem. Eng. Sci, 38, 4, 607-612.
Chylla, R.W., R.A. Adomaitis, and A. Cinar, (1986). Sta-
bility of tubular and autothermal packed bed reac-
tors using phase plane analysis. Submitted to Ind.
Eng. Chem. Process Des. and Dev.
Cinar, A. (1984). Controller design for a tubular catalytic
reactor. Can. J. Chem. Eng., 6 2 , 746-754.
Degnan, T.F. and J. Wei, (1979). The Co-current reactor
heat exchanger Part I-Theory. AIChE J.,25, 2, 338-
344.
T A B L E I Largest q i
am E l e m e n t s
Cocurrent Coolant at
INPUTS
control. 6. Multiloop design. Ind. Eng. Chem. Pro-
cess Des. Dev., 25, 411-419.
Garcia, C.E. and M. Morari, (1982). Internal model con-
trol. 1. A unifying review and some new results. Ind.
Eng. Chem. Process Des. Dev., 2 1 , 308-323.
Garcia, C.E. and M. Morari, (1985a). Internal model con-
trol. 2. Design procedure for multivariable systems.
Ind. Eng. Chem. Process Des. Dev. 24, 472-484.
Garcia, C.E. and M. Morari, (1985b). Internal model
control. 3. Multivariable control law computation
and tuning guidelines. Ind. Eng. Chem. Process Des.
Dev., 24, 484-494.
Litz, L. (1980). Order reduction of linear state-space
models via optimal approximation of the nondom-
inant modes. IFAC Symposium on Large-Scale Sys-
tems Proceedings, Toulouse, Fr., 1980.
Mandler, J.A., M. Morari and J.H. Seinfeld, (1985). Con-
trol system design for a fixed-bed methanation re-
actor. Paper presented at AIChE Annual Meeting,
Chicago, IL. 1985.
Metchis, S.G. (1982). Averting extinction in operating
autothermal catalytic reactors. PhD Thesis, Univer-
sity of California at Berkeley. 1982.
of a n A u t o t h e r m a l R e a c t o r
inlet temp. 0.6% CO

State Inlet Inlet Bed Total

Inlet Coolant
Variables Concentration Temp. Flowrate3
Temp.
Solid Temp. -0.5 2.7 - Ι Ο "239.0
Outer Wall Temp. 2.1 10.3ίο-3 20.0
Inner Wall Temp. 0.1 3 0.1ίο- 19.0
Bulk Gas Temp. - Ι Ο "5 - 0 . 13
- 0 . 22 95.0
Coolant Temp. -ΙΟ" 10~ ίο- 21.0
Surface Cone. 46,000 - 8 8 92
1312 -707
Bulk Gas Cone. 21.0 -10" ίο- -4.3
N o t e : Values are listed for largest element in each state variable vector

T A B L E I I S t r u c t u r a l D o m i n a n c e A n a l y s i s of a n A u t o t h e r m a l R e a c t o r
Countercurrent Coolant at 407UT inlet temp. 0.6% CO
Largest Corresponding Largest C orresponding
INPUTS q — elements State Variables d — elements State Variables
Inlet Cone. 50 X 150 5 ow
Inlet Bed Temp. 7263 rpQ
5.32 χ 10 y9
1682 5
8
ow
Total Flowrate 17 OW
Inlet Coolant Temp. 2.45 χ 10 T? 9291 OW

Adiabatic Bed at 400K inlet temp. 0.6% CO

Largest Corresponding Largest Corresponding
INPUTS q — elements State Variables d — elements State Variables
Inlet Cone. 29 1.3 X
439 3
x !
Inlet Bed Temp T?
g 1476
x t
ψ9
8
Total Flowrate 4.9 χ 1 0 ~ Y 0.4
9
9 T?
N o t e : T and X denote the solid temperature and the solid concentration at the
8 8
ninth collocation pt., respectively.
T A B L E I I I Static Gains Qaifor a n A d i a b a t i c R e a c t o r
A07K inlet temp. 0.6% CO
INPUTS Tsolid Tga8
3 Xsolid Xbulk
Inlet Cone. - 4 . 4 x 10" -1.8 χ 10~ 2.54 1.83
Inlet Bed Temp. 18.3 1.94 4 -0.91 -0.20
Total Flowrate - 4 . 2 x 10" -4.8 χ 10~ 1.2 χ 10" - 3 . 8 X 10"
300 R. W. Chylla, Jr. R. A. Adomaitis and A. Çinar

UPPER IGNITION REGION

C \ upper ignition Line

Feed-Effluent
Heat Exchanger UPPER OPERATING^*

{3b— Autotherm

Reactor Π Reactor I
-{5b- ZD
NEGLIGIBLE
REACTION
REGION

Quench Line Lateral Ignition

Line

LOWER OPERATING
ZONE
Bypass

Fig. 1 Multi-bed co-current configuration with

interstage feed-effluent heat exchange.

INLET COOLANT TEMPERATURE

Fig. 2 Reactor Phase Plane with Definitions

500, 500
5000 SCCM 5 0 0 0 SCCM
= 0,010
y c o YC =0,006
0
2
.4501 Autotherm 450
60% Adiabatic Efficiency
1-° O
LU 1 0 0 % \ ^
CC \
y Autotherm
4001 £ 400. / h /
û.
y /
Lu
Ι- / 0 0
Ο /
3501 £ 350.

I- /
LU /
Δ Model Ignition Pts _l
/ Δ Model Upper Stable Pts
/ Ο Model Lower Stable Pts.
Ο Model Stable Pts. /
300 300 / ν Δ
300 350 400 450 500 300 350 400 450 500

INLET COOLANT TEMPERATURE t, <K>

INLET COOLANT TEMPERATURE "t, W

Fig. 3 Phase Planes for Autothermal CO Oxidation Reactor

Feed-Effluent
Heat Exchanger

Reactor I ~**- Reactor II —*-

t
I
ι

~i
Bypass
<TC> Quench Feed Gas

M M
Preheater

Fig. 4 Proposed Reactor Design

Acknowled gement
Partial financial support from the U.S. Dept. of Energy (DE-FG02-84ER13205) and the American
Chemical Society Petroleum Research Fund (PRF 15627-G5) is gratefully acknowledged.

300
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

ON THE DYNAMICS OF A BENCH-SCALE

CALORIMETER
P. de Valliere and D. Bonvin
Technisch-Chemisches Laboratonum, ETH-Zunch, Switzerland

Abstract. A dynamic model for a bench-scale heat-flow calorimeter is

derived in order to support the accurate reconstruction of the heat
evolution in the reactor. The importance of describing the heat
transfer between the reactor contents and the heating/cooling medium
is described. The degree of discretization of the inner wall is
deduced from the comparison of experimental and model-based frequen-
cy responses. The nonlinear behavior of the calorimeter is modelled
and verified over a large range of operating conditions. Finally a
qualitative assessment of the validity of such a dynamic model for
both parameter and state estimation is provided.

Keywords. Modelling; Calorimetry; Dynamic Analysis; Frequency

Response; State Estimation; Parameter Estimation;

INTRODUCTION

This work deals with various modelling

aspects for the calorimeter from both a
theoretical and an experimental viewpoint.
It sets the basis for subsequent identifi-
cation and estimation studies:
- On-line estimation of selected time
varying parameters,
- State estimation with simulated or known
chemical reactions,
- On-line state and parameter estimation
with poorly known chemical kinetics.
Safe and optimal operation of a batch or
semi-batch reactor requires good knowledge
of the system's chemical state, especially
if the reaction system is complex or exhi-
bits large thermal effects. Ideally, such
chemical information should be avai-lable
on-line, i.e. during the reactor opera-
tion. When concentration measurements are
not practically feasible, on-line recon-
struction of the necessary information
concerning the state of the reaction sys-
tem is of great interest.
Bench-scale calorimetry is an ideal tool
for deriving thermal information about a
reaction system under conditions similar
to the ones existing in industrial reac-
tors. The bench-scale calorimeter deve- Fig. 1 Schematic view of the BSC-81
lopped by CIBA-Geigy (BSC 81) is well
suited to the task of running batch reac-
tions under tight thermal supervision. It
is schematised in Fig. 1. Since it has a noninsulated lid, a glass stirrer, an
already been discussed intensively in the electric calibration heater and a Pt-100
literature (Karlsen, 1982 and 1984), only temperature sensor.
its main functions and components are The accurate investigation of a chemical
summarized here. state clearly requires the ability to
The 2.5 liter stirred glass reactor is separate the dynamic effects associated
surrounded by a jacket in which a heat with the reactor and its operating mode
transfer fluid (silicone oil) circulates from the direct effects caused by the
at a very high rate. In typical reaction chemical reaction. Hence, a detailed dyna-
conditions, a controller adjusts the tem- mic model of the reactor without reaction
perature of the circulation fluid so that is a prerequisite for the successful
the heat transfer through the reactor wall application of the calorimeter in
balances out the heat evolution in the identification and estimation work.
reactor. The reactor is equipped with

301
302 P. de Valliere and D. Bonvin

CALORIMETER MODELLING Cal

(UI (6)
The amount of heat which evolves in the *Cal'
dt
calorimeter is determined by measuring the
heat flux through the reactor wall (heat- and the product of current and voltage
flow calorimetry). Hence, the temperature over the calibration heater is treated as
measurements have to be accurate and truly one input variable. Through linearization
representative of the conditions existing centered on an operating point of interest
in the various balance volumes. For (without chemical reaction the operating
example, one has to guarantee that the point is not necessarily time dependent) a
reactor and jacket zones are thermally state-space model of order (N+2) with 5
homogeneous. In the jacket this is inputs or disturbances and one output is
achieved by letting the heat transfer obtained.
fluid circulate at a very high flow rate. s =( qT T T T T) ) ( 7
Consequently, the temperature difference » ' Cal' ™ ' w ' ··· "
between oil jacket inlet and outlet re-
mains very small (typically 0.1 K ) . Inside u ( , q , N , UD- (8)
the reactor, a well-designed stirrer gua- Tj VR s
rantees that the reactor zone can be con- y = Τ (9)
sidered homogeneous as well. A second Two remarks concerning the state-space
objective of the present study is to be model are necessary:
able to follow fast thermal processes. 1) The BSC 81 calorimeter has two main
Therefore, the accumulation of heat within temperature control modes: T and T con-
the wall has to be accounted for and in R
n J
T
addition, the thermally inhomogeneous wall trol. In either case the valve which regu-
needs to be modelled as a true distributed lates the mixing of cold and hot heat
system. transfer fluid is manipulated, resulting
The basic energy balance for the calori- in a corresponding value for T j which can
meter reads: be measured. Whereas Tj is a manipulated
d T input for the calorimeter, it is a mea-
R q +=q + q + q sured input (or disturbance) for the pre-
(1) sent model, where no information on tem-
"VPR ~ R F Cal Sec
at perature control loops is available.
2) It may seem unusual to include chemical
where q is the rate of heat production
R reaction in the model by considering its
due to the chemical reaction, qp the heat heat production as a known disturbance,
however this allows testing of the calori-
flux through the wall, q
C lathe heat in- meter dynamics with a known chemical reac-
putted by the calibration heater and q tion without having to include the highly
S ce nonlinear terms typical of chemical kine-
the sum of all secondary heat effects (see tics in the model. For example, the effect
the nomenclature section for an explana- of chemical reaction on composition-depen-
tion of other symbols). dent parameters such as the heat transfer
The heat flux which is defined as: coefficients or the heat capacity inside
A ( TT the reactor can be modelled without a
(2) highly nonlinear term in the state-space
R" WR
model.
requires for its computation the knowledge The values of the model parameters were
of Tp^, the wall temperature in contact
either taken from the literature (material
with the reactor contents. Since the wall property parameters or parameters specific
is thermally inhomogeneous, its tempera- to the calorimeter) or estimated off-line
ture is described by the following heat through appropriate calibration procedures
conduction equation: (Karlsen, 1984). The dynamic investigation
of the calorimeter filled with 1 to 2 kg
2 of water was performed in the temperature
3T .
T P C 3 T.
1 T 3 (3) range of 10 to 80 °C. Table 1 indicates
at the numerical values or the correlations
W PW used for the main parameters.
with the boundary conditions:

3T . = H ( T
T TABLE I Calorimeter parameters including
31 1=0 R WR- V (4) their
stirrer
temperature
speed
(in K) and
(in rpm)
dependencies.
3T .
r h_ <T_ - T _ ) (5)
J J r
WL c == 4180
C p R 41 1 4 9 3 CJoule/kg Kl
5
31 1=L
CJoule/kg K3
PW · V ·
and T^k being the wall temperature in
h j = 2.02 T j - 138.9
contact with the heating/cooling medium. 2 3/n w a t t / i A a
The partial differential equation (3) is hp = (1.04 T - 209.5) N CWatt/irAa
discretized in space by applying the me- R g
thod of orthogonal collocation, using Ν \ = 0.0005 (T + Tj) + 0.827 CWatt/m K3
interior collocation points (N being a
w R
q
design parameter at this stage of the S =e -(0.0119
c T - 2.35)(T -
R R V +
development) . The electric calibration + 0.0095 N + 4.113 CWattl
heater is modelled as a first-order pro- s
cess :
 On the Dynamics of a Bench-scale Calorimeter 303

The heat-flow calorimeter can be characte- Table 1 ) . The convergence of the dominant
rized dynamically by its ability to res- eigenvalues when Ν is increased is shown
pond to perturbations (or exitations) in Fig. 2. As expected, the dominant
either generated within the reactor or eigenvalues do not change much with an
stemming from the jacket section. Table 2 increasing discretization in the reactor
summarizes the important time constants of wall. Each additional collocation point
the system. brings one additional fast eigenvalue to
the system. A complex eigenvalue is pre-
sent with 7 or more collocation points.
Since the calorimeter data are sampled
TABLE 2 Principal time constants of the every 2 seconds, the sampling and Nyquist
calorimeter frequencies are ω = ir rad/s and ω. ττ/2
σ Ν"
rad/s, respectively.
reactor contents: 600 - 800
inner wall : 25 s 0.5
calibration heater: 20 s
T T- control loop: 30-s
Temperature sensors: 2.5 s

MODEL VALIDATION

The eigenvalues of the linearized approxi-

mation provide useful information concer-
ning the dynamics of the system at a given
operating point. Table 3 lists the eigen-
values evaluated for two temperature le-
vels (20 and 60 °C) and for several values
of N, the number of interior collocation
points in the wall.

Fig. 2 Eigenvalues of the model vs. the

number of collocation points.
TABLE 3 Eigenvalues at 20 and 60 C for 1
to 5 collocation points
This clearly indicates that the dynamic
T R= 293 Κ T R= 333 Κ processes characterized by eigenvalues
with an absolute value larger than about 2
are too fast to be of any practical signi-
Ν = 1 -0., 0015 -0.,0017 ficance during transient operation. Fig. 2
-0.,0500 -0.,0500 shows that 2 to 3 collocations points in
-0..1513 -0., 1589 the wall are sufficient to have all of the
eigenvalues with |λ| < 2 very close to the
Ν = 2 -0., 0015 -0.,0017 converged values.
-0.,0500 -0.,0500 There is little point in modelling a
-0..1597 -0.,1681 system in more detail than can be verified
- 0 . ,6760 -0.,6952 or tested by experiments. It was therefore
decided to first measure the frequency-
Ν = 3 -0..0015 -0.,0017 response of the transfer function
-0.,0500 -0.,0500 T R(s)/Tj(s) using Fast Fourier Transform
-0..1666 -0.,1757
-0..7690 -0.,7860 techniques. The objective here was prima-
-1..6085 -1.,6419 rily to investigate the dynamics asso-
ciated with the heat flow through the
Ν = 4 -0..0015 -0.,0017 wall. Fig. 3a and 3b compare the measured
-0..0500 -0., 0500 frequency responses to the one predicted
-0,.1667 -0.,1758 by the model. The agreement in the fre-
-0.,8397 - 0 . ,8588 quency domain of interest (below ω = 0.1
-1.,9470 -1.,9567 rad/s) is excellent. Above that frequency
-2..9514 -3.,0061 the sampled nature of the real process is
apparent. The good agreement at low fre-
Ν = 5 -0.,0015 -0.,0017 quencies is indicated by the location of
-0..0500 -0..0500 the low frequency asymptote. Two main
-0.,1666 -0.,1757 difficulties were encountered with the
-0..8374 -0.,8577 frequency-domain test: The generation of
-2..2222 -2..2346 appropriate test signals and the sampled
-3..7470 -3..7039 characteristics of noisy measurements.
-4..7199 -4.,8259 1) The first problem was partially solved
by using approximate pseudo-random binary
test sequences for the excitation of the
The slowest eigenvalue is related to the set point temperature of the jacket fluid.
dynamics of the calorimeter contents while However limitations arose from the thermo-
the second slowest deals directly with the static unit itself which of course cannot
lag in the calibration heater. The eigen- handle absolutely square signals. In fact
values associated with the inner wall are the thermostatic unit behaves like a first
2 orders of magnitude faster than the order lag with a time constant of about 30
first one. It is observed that the reactor seconds. Thus the amplitude of the test
is slightly faster at higher temperatures, signals at high frequencies is barely
a result, of larger heat transfer coeffi- higher than the one induced by noise.
cients and secondary heat effects (see
304 P. de Valliere and D. Bonvin

steady-state (room temperature for T_.) T T

is lower then
Ρ in order to compensate
for the energy generated by the stirrer;
at the final steady-state T j is higher
than Tp in order to compensate for the
heat lost tô the surroundings. The agree-
ment between the measured and simulated
temperatures is excellent since their
maximum difference is less than 0.2 °C.
Other tests were performed but are not
reported here. On the basis of the results
of the time as well as frequency domain
testing, the model appears appropriate to
support the on-line estimation of certain
parameters and states.
350.

3a Amplitude
function T (s)/Tj(s)
R
function of frequency measured

simulated

1000. 2000. 3000. 4000. 5000. 6000.

Fig. 4a Measured and simulated
temperatures (in K) vs. time (in
seconds)

Fig. 3b Phase of the transfer function

vs. frequency

This results in very imprecise Bode plots

fir the experimentally derived system
transfer function. A great enhancement of
quality is achieved by making several
measurements, transforming them separately
and taking the mean of the frequency plots
as it is done in NMR-spectroscopy, laser-
anemometry and similar FFT applications.
2) For higher frequencies the aliasing
phenomenon gets more and more important
and distorts the experimental frequency-
response curve. As the sampling period of
the calorimeter is fixed at 2 seconds,
it does not seem possible to move the
aliasing area to higher frequencies. How-
ever, the calorimeter logs the Tj and T R
data with a sampling time of 0.2 seconds.
It is therefore possible to implement a
digital filter to remove some of the high
frequency noise. An even better way to
limit the band of the temperature signals
would consist of implementing an analog
filter before the A/D-conversion.
Further testing was performed within the
time domain using the nonlinear model.
Fig. 4a and 4b show the measured and pre-
dicted reactor temperature when the calo-
rimeter is operated in Tj-control mode.The
jacket temperature is made to follow a
ramp from approximatively room temperature
to 70 °C. Notice that even at steady-state
the measured reactor and jacket tempera-
tures differ slightly: At the initial
0. 1000. 2000. 3000. 4000. 5000. 6000.
Fig. 4b Difference between measured and
simulated reactor temperature
vs. time
PARAMETER AND
STATE ESTIMATION
For the investigation of an imperfectly
known chemical reaction, it is necessary
to include the chemical state (e.g. con-
centration) in the state vector. Since a
batch reactor does not have a constant
operating point, linearization techniques
centered on a fixed reference point are
not applicable. The reference point has to
move along with the process and can be
chosen, for example, as the last sampled
value of the variables. If necessary, the
linearized model is recalculated at each
sampling time. The influence of the chemi-
cal reaction on some thermal parameters
 On the Dynamics of a Bench-scale Calorimeter
such as the heat transfer coefficient at
the reactor-wall interface or the heat
capacity of the solution is quite diffi-
cult to model. Furthermore, certain kine-
tic parameters may be known only poorly.
Hence, the idea of an extended observer
for the reconstruction of states and para-
meters has been investigated. The prin-
ciple is based on the recalculation of the
model at each sampling time. It is
possible, with a new pole placement algo-
rithm (Roppenecker, 1981), to create an
observer with constant error dynamics,
even when working with a nonlinear model.
Consequently, with every sampling the
model is not only recalculated but the
pole placement is updated as well. The
difficulty with this method is to overcome
the problem of local unobservability. As
it is very hard to prove global observabi-
lity for nonlinear systems, the problem is
circumvented by making the pole placement
algorithm drifting and not absolutely
deterministic (Kreisselmeier, 1980) (Lya-
punov-synthesis). With this approach one
can avoid the unbounded growth of certain
elements in the observer design matrix if
by chance a calculated state is very near
an unobservable state.
Another problem of the observer described
above is the fact that the reactor on
which all "the modelling is based has quite
slow dynamics. The thermal capacity of the
solution inside the reactor is large com-
pared to that of the reactor wall. When
one tries to observe the time constant of
the calibration heater, for example, the
reactor is so slow at responding that it
cannot "see" dynamically the heat produc-
tion at the heating coil. In other words,
the time constant of the reactor being
approximately 600 seconds, the oberver
cannot easily detect a time constant of
about 20 seconds. If the time constant of
the electric calibration heater is defined
as an extended state, the observer tends
to converge towards an estimated time
constant of zero seconds as if it had
"seen" a rectangular heat production.
Clearly, the sensitivity of the observer
is limited in this case by the way the
equipment has been designed or is ope-
rated. A first possibility to improve
these results would consist of reducing
the thermal capacity of the reactor con-
tents .
CONCLUSIONS
A dynamic model for a bench-scale calori-
meter has been derived and tested in the
frequency as well as time domain. The
agreement between measured and predicted
temperatures is excellent, considering
that no parameter fitting has been per-
formed. The results of this investigation
will be used to try to further improve the
value of certain parameters or corre-
lations. The noise level and the fixed
sampling rate of the calorimeter place an
upper limit on the frequencies that can be
observed and followed. The model derived
in this study appears well suited to
support the on-line estimation of both
parameters (e.g. heat transfer coeffi-
cient, thermal capacity, kinetic para-
meters) and states (e.g. chemical composi-
tion) of the calorimeter.
305
REFERENCES
Karlsen, L.G. (1982). Heat flow
calorimeter, Progress report no. 1,
Instituttet For Kemitektik, Danmarks
Tekniske Hojskole, Lyngby.
Karlsen, L.G. (1984). Optimal data
aquisition for heat flow calorimeter,
Thermochim. Acta, 72, 83-54.
Kreisselmeier, G. (1980). Adaptive control
via adaptive observation and asymptotic
feedback matrix synthesis, IEEE Trans,
on Aut. Contr., AC-25, 717-722.
Roppenecker, G. (1981). Polvorgabe durch
Zustandsrueckfuehrung,
Reqeluncrstechnik, 29, 228-233.
NOMENCLATURE
Heat transfer area Crn^H
Heat capacity of the reactor
PR
contents CJoule/kg KI
Heat capacity of the reactor wall
-pw
CJoule/kg Kl
Heat transfer coefficient on the
2
jacket side of the wall CWatt/m K3
Heat transfer coefficient on the
2
reactor side of the wall CWatt/m K3
Lenght coordinate in the wall CmU
Wall thickness Lml
Thermal conductivity of the wall
CWatt/m Kl
Mass of the reactor contents Ckgll
Ν
Number of interior collocation
points in the wall
Ne
Stirrer speed CrpmH
ω :
q Frequency Crad/s3
C a l * Heat evolution through the
calibration heater CWattl
Heat flux through the wall CWattI
Heat generated by chemical reaction
^Sec* Secondary heat effects 3
:
PW Density of the wall Ckg/m l
t : Time Csl
Time constant of the calibration
heater Csl
Jacket temperature CK3
Reactor temperature CK3
Temperature of the surroundings
Wall temperature
Wall temperature on the jacket side
•"•WJ
Τ Wall temperature on the reactor side
WR
Input vector
u
Electrical power inputted into the
UI calibration heater CWattl
χ State vector
y Measured output
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

A NEW APPROACH FOR PARAMETER

ESTIMATION APPLICATION TO
HETEROGENEOUS CATALYSIS
H. Piet-Lahanier*, E. Walter* and J. Happel**
^Laboratoire des Signaux et Systèmes, CNRS/Ecole Supérieure d'Electricité,
Plateau du Moulon, F91190 Gif-sur-Yvette, France
**Department of Chemical Engineering and Applied Chemistry, Columbia University,
New York, NY 10027, USA

Abstract . In this paper , a ne w methodolog y base d on membershi p se t theor y is

presente d for estimatin g the parameter s of a mathematica l mode l whe n the
dat a ar e corrupte d by a n unknow n bu t bounde d noise . The methodolog y is
describe d usin g a simulate d exampl e which illustrate s its robustnes s to
outliers . It is the n applie d to the stud y of the isobutene-isobutane-hydroge n
reaction s by mean s of transien t isotopi c tracing . Severa l model s ar e
discusse d an d thei r performance s compared . The valu e of this approac h for
the developmen t of furthe r rat e expression s is discussed .

Keywords . Membershi p se t estimation ; paramete r uncertainty ; robus t

estimation ; transien t tracing ; isobutene-isobutane ; heterogeneou s catalysis ;
compartimenta i modelling .

output s ar e availabl e for eac h measurement .

INTRODUCTION
Estimatin g the parameter s the n mean s determinin g
the se t in the paramete r spac e tha t contain s all
Estimatin g the paramete r vecto r θ of a give n
model s of a give n structur e tha t ar e compatibl e with
mathematica l mode l usuall y mean s obtainin g a suitabl e
this hypothesis . Characterizin g this se t provide s a t
numerica l value of It is of cours e importan t to go
the sam e time a n estimat e of the valu e of θ an d of the
furthe r an d to evaluat e the uncertaint y induce d on the
rang e within which it ca n vary . Howeve r mos t of the
paramete r estimate s by the nois e corruptin g the data .
technique s presente d s o far ar e limited to the stud y
With traditiona l methods , suc h a s maximu m likelihood
of model s dependin g linearl y on thei r parameters . Fo r
estimation , paramete r estimate s ar e mos t ofte n
model s tha t ar e nonlinea r in thei r parameters ,
obtaine d by minimizing a weighte d su m of the square s
Belforte an d Milanes e (1981 ) propose d a n algorith m
of the differenc e betwee n the dat a an d the
which relie s on linearization s a t variou s point s of the
correspondin g mode l outputs . The paramete r
membershi p se t to determin e the extrem e value s
uncertaint y Is the n compute d by calculatin g the squar e
reache d by eac h paramete r ove r the set .
root s of the diagona l entrie s of the invers e of the
Fishe r informatio n matrix .
The metho d presente d her e provide s two mai n
improvement s ove r classica l membershi p se t
Thes e technique s ar e computationall y simpl e bu t
estimators . Firs t it ca n be use d to estimat e
suffe r thre e majo r drawbacks . First , the y requir e
parameter s of model s whos e output s ar e nonlinea r in
detaile d (althoug h ofte n implicit) assumption s on the
the parameter s withou t requirin g an y linearization .
statistica l natur e of the nois e corruptin g the data .
Secondly , it ca n handl e dat a containin g a larg e
Secondly , the us e of the Fishe r Informatio n matri x
percentag e of outliers , i.e . dat a for which the valu e of
relie s on asymptoti c results , s o tha t its validity for
the erro r canno t be mad e acceptabl e becaus e of
smal l set s of dat a ca n be questioned . Finally, leas t
unexpectedl y larg e measuremen t errors . Thes e
square s criteri a ar e well know n for no t bein g robus t
to outliers . outlier s ar e detecte d an d ca n be discarde d if
appropriat e upo n furthe r consideration .
In orde r to overcom e thes e limitations , we shal l
In orde r to illustrat e the propertie s of the metho d
introduc e a n alternativ e approach , which belong s to
unde r controlle d conditions , It is presente d usin g dat a
the categor y of membershi p se t estimation . This
obtaine d by simulatin g a mathematica l mode l with
approach , which ca n be trace d back , in the field of
known parameters . It is the n applie d to rea l dat a
stat e estimation , to the work s of Schwepp e (1968 ) an d
obtaine d In the stud y of the isobutene-isobutane -
Bertseka s an d Rhode s (1971) , ha s bee n expande d to
hydroge n syste m usin g transien t tracing .
paramete r estimatio n by Foge l (1979) , Foge l an d Huan g
(1982 ) an d Milanes e an d Belfort e (1982) . Recen t
development s ca n be foun d in Belfort e an d Bon a (1985 )
METHODOLOGY
an d Norto n (1985) . Membershi p se t estimatio n only
require s the hypothesi s tha t uppe r an d lowe r bound s
of the accepXabt e error s betwee n the dat a an d mode l To introduc e the ne w methodology , let us conside r the
following (ver y simple ) parametri c mode l :

307
308 H. Piet-Lahanier, E. Walter and J. Happel

dx/d t = -ej x + θ 2u, x(0 ) = 0 , Pvfinition 2

A paramete r vecto r ά. is consistent with the objec t
ba r B 0(k) if
Ynf « *'

wher e u is the Input , y nf is the noise-fre e outpu t an d f(â,k) e B 0(k),

6j an d θ£ ar e the parameters .
wher e f(e,k ) is the outpu t of the parametri c mode l
Suppos e tha t the outpu t measure d a t time t k is give n with paramete r vecto r e , associate d with the kth
measurement .
by
) =ν ( ) 1χ (+ β
Definition 3
Υ(^ ηί ^ ^» The membership set S is the se t of all the
paramete r vector s tha t ar e consisten t with all the
wher e e k satisfie s objec t bars .

e k€ [-e m
, em
], In orde r to characteriz e the membershi p set , let us
introduc e a s a criterio n to be maximize d
an d e m is known . j(e) = percentag e of objec t bar s with which e
is consistent .
The proble m is to estimat e the valu e of Θ from a séri e
of value s of y an d u. In this simulate d example , thes e Whe n the dat a contai n no outliers , the optima l valu e of
dat a hav e bee n compute d a s follows : this criterio n is 100 % an d maximizin g j(e ) allows us to
reac h a poin t within the membershi p set . Determinin g
u is a unit ste p function , the membershi p se t is the n equivalen t to
θ|= 0.5 , θ 2=3 , characterizin g the isocriterio n j(â>100 % (i.e . the se t
e m= 0.2 , of all paramete r vector s tha t correspon d to this valu e
t k= k/2 for k=l to 40 . of the criterio n j). To do so , we searc h in rando m
direction s for point s on the boundar y of this
isocriterion . Fo r mor e detail s on the algorith m used ,
The following definition s will be use d whe n processin g se e (Walter , Piet-Lahanie r an d Happel , 1986) .
thes e dat a (Walter , Piet-Lahanie r an d Happel , 1986) .
Whe n ther e ar e no mor e tha n two o r thre e
Definition 1 parameters , it is possibl e to dra w a two - o r
The object bar B 0(k) is the se t of all value s tha t ar e three-dimensiona l representatio n of the boundar y of
considere d acceptabl e for the mode l outpu t associate d S . Figur e 2 thu s describe s the membershi p se t
with the kth scala r measurement . associate d with ou r simulate d example . As expected ,
the tru e valu e of the parameters , indicate d by a cros s
Fo r the example , we hav e on the figure , belong s to S .

B 0(k) = { y | ym
i n ( t k ) i y i ym
ax<tic>}. h
(

3.47
wher e
/
/
y ( t k)
Vmin«k> = .

an d

y ( t k)
Ymax^) - e ·
î- m / /

The correspondin g objec t bar s ar e give n in Figur e 1.

2.9\

Fig. 2 . Membershi p se t S

Whe n ther e ar e mor e tha n thre e parameters , it

become s impossibl e to obtai n a visua l representatio n
of the result . The proble m of characterizin g S als o
occur s for linea r membershi p se t estimator s an d
severa l solution s hav e bee n proposed . It Is possibl e
to determin e the equation s of the boundar y of a se t
tha t would contai n S an d would be of minimu m size .
Foge l an d Huan g (1982 ) determin e boundin g ellipsoids .
Milanes e an d Belfort e (1982 ) calculat e boundin g
""I"' |M, ,, , ,....,.-,. . ι orthotopes . An alternativ e metho d combinin g thes e
>.· 8 4 8 6 9 21 1
4 1
6 1
8 1· 2 two approache s ha s bee n develope d b y Belforte an d
Bona (1985) . Her e we shal l determin e the rang e within
Fig. 1. Objec t bar s which eac h paramete r varie s ove r S .
 New Approach for Parameter Estimation 309

Fo r the simulate d example , the uncertaint y interval s
ar e
θ! e [ 0.47 ; 0.58] ,
θ 2 e [2.91 ; 3.47] .
Figur e 2 evidence s the fac t tha t the actua l surfac e of S
is muc h smalle r tha n the surfac e of the rectangl e
associate d with the abov e ranges .
To illustrat e this propert y in a rathe r extrem e
situation , let u s replac e 25 objec t bar s of the
simulate d exampl e picke d a t rando m b y a s man y
outlying bars . Processin g the resultin g dat a se t with
OMNE, we find tha t the optima l valu e of j(e_) is equa l to
37. 5 %. The uncertaint y interval s becom e
θ 1 € [ 0.41 ; 0.67] ,
θ 2 e [2.54 ; 4.] .

In orde r to allow a compariso n of the measure d The interval s ar e wide r tha n thos e obtaine d previousl y
proces s outpu t with the se t of all the output s an d still contai n the tru e value s of the parameters.Th e
generate d by model s belongin g to S , we introduc e objec t an d mode l bar s ar e give n in Figur e 4 , which
evidence s the fac t tha t the outlyin g bar s hav e bee n
Definition 4 correctl y discarded .
The model bar B m
( k ) is the se t

( k ) = {f(e J k) I â. € S }.
Bm

Whe n ther e ar e no outliers , all the mode l bar s mus t be

within the correspondin g objec t bars . Wheneve r a n
extremit y of a give n mode l ba r correspond s to the
extremit y of the associate d objec t bar , it mean s tha t
this objec t ba r contribute s to the definitio n of the
boundar y of S . Fo r the example , the mode l bar s ar e
give n in Figur e 3 .

Fig. 4 . Objec t an d mode l bar s (outlier s case )

CHEMICAL REACTION EXAMPLE

The isobutene-isobutane-hydroge n syste m is

representativ e of a wide variet y of catalyti c
12 16 processes . It is of chemica l interes t becaus e it ca n be
convenientl y studie d experimentall y in bot h forwar d
an d revers e directions . The dat a ar e isotopi c
transient s obtaine d in a gradientles s recirculatin g
Fig. 3 . Mode l bar s reacto r syste m in a wa y simila r to tha t describe d in
(Happe l an d colleagues , 1982) . After stead y stat e of
Robustnes s to outlier s
hydrogénatio n o r dehydrogenatlo n is reache d with a
give n mixtur e of isobutane , isobuten e an d hydrogen ,
It ma y happe n tha t som e dat a do no t satisf y the the fee d strea m is rapidl y switche d to on e containin g
hypothese s mad e on the bound s of the error . Whe n exactl y the sam e proportion s of isobutane , Isobuten e
usin g classica l weighte d least-squares , the presenc e 1 3
an d hydrogen , bu t with eithe r isobutan e o r isobuten e
of suc h outlier s 1s boun d to drasticall y affec t the marke d with C . The specifi c activitie s of the
value s of the estimate d paramete r vector , an d suc h
termina l specie s ar e the n compute d a s a functio n of
method s ar e know n no t to be robus t to outliers . Som e
time from mas s spectra l data .
technique s hav e bee n develope d In the literatur e to
robustif y maximu m likelihood estimator s - an d in
particula r leas t square s estimator s - to outlier s (se e Fo r the dehydrogenatio n reactio n
e.g . Martin , 1978) .
CH <> =H
C + H
4 10 4 8 2»
Sinc e we do no t assum e her e the availabilit y of
detaile d informatio n on the nois e statistica l the following mechanis m ca n be proposed :
properties , thes e technique s for dealin g with outlier s y
do no t apply . We shal l se e tha t the methodolog y
describe d in this pape r nevertheles s present s ver y +1
attractiv e robustnes s properties . If outlier s ar e (i) C 4H 1( 0
9) + 2 ( H ^vz ± c 4H 9^ ) +H< «
presen t in the dat a set , the mode l need s no longe r be
-1
consisten t with all the objec t bars . Maximizing the
criterio n j ca n the n be see n a s minimizing the numbe r
of objec t bar s considere d a s outliers . This is why we
v
hav e name d this estimato r Outlier s Minimal Numbe r +2
Estimato r (OMNE). The robustnes s propertie s of OMNE (11) C 4H 9( « + ( £ ) ^ C 4H 8^ ) + H ^ ) f
rely mostl y on the fac t tha t a n objec t ba r is eithe r
take n into accoun t o r totall y discarded .
310 H. Piet-Lahanier, E. Walter and J. Happel
v
Isobutan e an d Isobutene . Two type s of inpu t condition s
+3
ar e considered , correspondin g to a n isobutan e
(iii) C 4H 8W ^v C 4H 8( 9 ) + (£) . step-u p (u A=0.984 , υβ= 0.011) , o r to a n isobuten e
step-u p ( u A
= 0.011 , u B
=0.984) .
-3
To describ e the dynami c behavio r of ^ C , the If an y of the thre e mechanisti c step s is assume d to be
four-compartmen t mode l give n in Figur e 5 ca n be at equilibrium , the compartment s involve d in the
associate d with this mechanism . equilibriu m mus t be lumped . Thre e two-compartmen t
an d thre e three-compartmen t model s ar e thu s
possible . Deuteriu m tracin g experiment s conducte d
previousl y (Happe l an d colleagues , 1973 ) indicat e tha t
adsorbe d butène s ar e clos e to equilibriu m with the ga s
phase , s o tha t the las t two compartment s of the four -
compartmen t mode l ca n be combined .

The resultin g three-compartmen t mode l satisfie s

Fig. 5. Four-compartmen t mode l
1
βC out in
Compartmen t A correspond s to isobutan e in the ga s Ad x A FA FA
phase . Compartment s Ai an d Bfc respectivel y
correspon d to adsorbe d C 4H g ( ^ an d C 4H B' ^ . Finally (— ) — . ( = 1V) x +A +_ vl + X+
£ A u
W dt W W
compartmen t Β correspond s to isobuten e in the ga s
phase . The stat e spac e equation s of this mode l ar e
obtaine d by materia l balance s on eac h compartmen t :
C dx fi vX v =+ v χ + vX
A£ +1 A - < -1 +2 > Α β - 2 B»
?C Ad x A FA
ou t FA
1n dt
( ) - ( + v=1 )+x A+ v H x M + uA
,
1N
W dt W W
(β C B C) d x B+ vx =F Bou t v X+ + F
B
L dx Α β B£ + 2 Ai ~ ( -2 ^ B %
W dt W W
A2' - v+ i x A- (v_j+ v + ) 2x M + v _ 2 χ Β, β
dt
wher e x B is the proportio n of marke d tertiar y carbo n
atom s in isobuten e adsorbe d an d in the ga s phase . Its
paramete r vecto r is
Cu = v +2 x Ai -f ( v _ 2+ v + ) 3xM + v _ 3 x B, τ
dt
fi = [ v + , 1v + , 2C A , £C B, iβ ] .

pC out
B dx B FB"
Befor e attemptin g to estimat e the parameter s of suc h
( — ) —
3vxM =-+( + v _ 3 ) x B+ u B,
a mode l from input-outpu t data , it is advisabl e to
W dt W W
chec k tha t ther e is only on e paramete r vecto r θ tha t
correspond s to a give n input-outpu t behavior . This
wher e W is the weigh t of catalys t (g) , inC Aan d Cin
Bar e correspond s to the notio n of structura l identifiabillty,
the ga s phas e fraction s of A an d B, F A an d F B ar e an d the method s describe d for exampl e in (Walter ,
the inlet flow o t urate s ooft u Isobutan e an d isobuten e 1982 ) ca n be use d to prov e tha t this three -
(ml/min), F A an d F B ar e the outle t flow rate s of compartmen t mode l is structurall y globall y
identifiable .
Isobutan e an d Isobuten e (ml/min) . The stat e variable s
Xj, i=A, Aft, B£, Β, ar e the proportion s of marke d
tertiar y carbo n atom s in compartmen t i, u A an d u B It wa s desire d to furthe r investigat e whethe r one of
the simple r two-compartmen t models , in which the
ar e the percentage s of marke d tertiar y carbo n atom s
two surfac e specie s wer e als o a t equilibriu m with eac h
in the inlet flows of isobutan e an d isobutene . The
other , would als o be abl e to describ e the data . This
paramete r vecto r Is
τ two-compartmen t mode l satisfie s
θ = [ v +f 1v + 2, v + , 3C A , £C B , Êβ ] ,
p C Ad x A F .ou t
A° - rA
»
( ) _ ( + V | ) X + V_| X + U ,
wher e v+j, 1=1,2,3 , ar e the direc t velocitie s (ml/g-ca t + =A B A
W dt W W
min), C A £ an d C B
£ ar e the surfac e concentration s of
Ail an d B£ (ml/g-cat ) an d β is the dea d spac e (ml). The 0 U
invers e velocitie s v_j ar e compute d from the direc t β CB dx B FB *
one s v +i by the relatio n
( — + C Ê) — = v +| x A- ( — + v_ , ) x B+ u B,
W dt W W
v-i= v +i - V,

wher e V is the overal l rat e of reactio n (ml/g-ca t min) . wher e Cg is the surfac e concentratio n of the adsorbe d
The initial condition s ar e Xj(0) = 0.011 , i=A, Afi, Bfi, B, specie s Afi an d BE . The paramete r vecto r θ for this
mode l is
which correspond s to the natura l abundanc e of ^ C . τ
The variable s accessibl e to measuremen t ar e the
proportion s x Aan d x B of marke d carbo n atom s in the θ = [ v +, 1C f , cβ ] .
 New Approach for Parameter Estimation 311

This two-compartmen t mode l is structurall y globall y Sinc e dat a set s 1 an d 2 correspon d to differen t
identifiabl e too , an d we will stud y its behavio r an d tha t operatin g condition s of the reactor , the y ar e
of the three-compartmen t mode l on the basi s of associate d with differen t value s of the parameter s of
preliminar y data . the model s to be considered . We therefor e will stud y
eac h of thes e dat a set s separately .
Two set s of dat a wer e considere d correspondin g bot h
to a dehydrogenation . Dat a se t 1 correspond s to a n
3 Two-compartmen t model . Fo r dat a se t 1 , it is possibl e
isobutan e 3
step-u p o uwher
t e CA= 9o. 9u3 3tx l 0 ~ , to find two-compartmen t model s tha t ar e coheren t
i n 1χί(Γ ,1
=4.625
CB n F A =0.7164 , 3F B = 0 . 3 3 3 6 , with 95. 6 % of the objec t bars . The mode l parameter s
hav e the following rang e ove r the membershi p se t
F A« 0 . 7 8 2 6 , F B - 0 . 2 6 7 4 , V=|.80x10~ , u A
«0.98 4 an d 3 3
ug»0.0 1 Κ Dat a se t 2 correspond 2
s to a n isobuten 3e v +| € [4.34x10~ ; 4 . 5 4 x l 0 " ],
o u pt
step-u wher e CAout
=î.057x1<T , CB in
=3.991x10~ ,
3 3
i n= 0 . 7 6 2 3 ,
FA F B =0.2877 , F A= 0 . 7 9 1 6 , C £ € [4.13x10~ ;4.97x10~ ],

F B= 0 . 2 5 8 4 , V=7.96x10""* , u A
=0.01 1 an d u B
=0.984 .
β € [117. ; 120.] .
Figure s 6 to 8 give the objec t bar s correspondin g to
the origina l data . Fo r dat a se t 2 , the two-compartmen t mode l manage s
1 to be coheren t with 100 % of the objec t bars . The mode l
parameter s hav e the following range s
I ι ι I II ι I II I
I

3 3
v +1 € [1.03x10~ ; 1.28xl0" ],
3 3
C £ € [2.20x10~ ; 4.78x10" ],

β € [103. ; 120.] .

Sinc e no two-compartmen t mode l with the structur e

considere d her e is coheren t with 100 % of the data , we
shal l no w conside r the three-compartmen t model .
-Ί I • "I ι • ι
99 195 129 135 159

TIRE (9.1mln
Three-compartmen t model . Fo r dat a se t 1 , ther e is no
Fig. 6 . Objec t bar s for dat a se t 1: Isobutan e Improvemen t of the valu e of the criterio n j. The
range s for the estimate d parameter s ar e
3 3
v +1 € [6.25x10~ ; 6.72X10" ],
3 2
v+2 € [9.59x10~ ; 1.11x10~ ],
3 3
C A£ e [2.07x10~ 3; 2.28x10" 3],
£ 49
1 1 11 C B£ € [2.07x10~ ; 2.39x10~ ],
I I I I I I I
« 39

Ι Ι β € [119. ; 120.] ,
11
an d the tota l concentratio n of adsorbe 3 d specie 3s is
foun d to belon g to the rang e [4.13x10~ ; 4.54x10~ ].
99 195 129 135 159

TIME <9.1xnin>
As coul d be expected , it is als o possibl e to find
Fig. 7. Objec t bar s for dat a se t 1: Isobuten e three-compartmen t model s coheren t with all objec t
bar s associate d with dat a se t 2 . The range s for the
estimate d parameter s ar e
1 1 4 3
Ι ι I I I I ι v +1 e [8.97xl0~ ; I.77x10"" ],
11 ' 3 3
11
v +2 € [1.20x10~ ; 3.26x10" ],
3 3
C A €f e
[1 .25X10" ; 5.49x10~ ],
3 3
£ 39 C Be f [i1 . 2 5 X 1 0 " " ; 3.86x10~ ],

β € [100. ; 120.] ,
iI
an d the tota l concentratio n of adsorbe 3 d specie 3s is
195 129 135 159
foun d to belon g to the rang e [2.69x10~ ; 8.20x10" ].
Ftç . θ. Otyec t bar s for dat a se t 2 : Isobuten e

DCCR-U
312 H. Piet-Lahanier, E. Walter and J. Happel
Two differen t problem s mus t be addresse d whe n
severa l mode l structure s compet e for the descriptio n
of the sam e data . The first on e is the estimatio n of the
parameter s of an y give n structur e assumin g tha t it is
correct . The secon d on e is to selec t the bes t
structure .
As fa r a s estimatio n is concerned , we ca n chec k the
value s obtaine d for bot h dat a set s an d mode l
structure s b y comparin g the estimate d tota l
concentratio n of adsorbe d specie s with tha t obtaine d
3
from independen t experimen t (approximatel y equa l to
4.08x10"" ml/g-cat) . In bot h cases , the estimate d
range s for the tota l concentratio n of adsorbe d specie s
ar e compatibl e with this value .
On the basi s of preliminar y data , the
two-compartmen t mode l shoul d be selecte d becaus e It
is the simplest . Howeve r the paramete r value s
obtaine d for the three-compartmen t mode l structur e
ar e als o reasonable , s o tha t it shoul d be wort h
retainin g this structur e for furthe r consideratio n .
The modelin g procedur e describe d in this pape r
furnishe s a n interestin g ste p in the constructio n of
rat e equation s for heterogeneou s catalyti c reactors .
The parameter s obtaine d will enabl e a decisio n to be
mad e a s to whethe r a singl e rate-controllin g ste p
exist s in a reactio n mechanism . If this is the case ,
conventiona l rat e expression s suc h a s thos e following
the familia r Langmuir-Hinshelwood-Hougen-Watso n
(LHHW) development s ca n be employed . If not , mor e
complicate d derivation s ma y be necessar y for optima l
modelin g of reactio n rates .
DISCUSSION AND CONCLUSIONS
The methodolog y describe d in this pape r appear s to be
an attractiv e complementar y approac h to the
conventiona l weighte d least-square s estimator . It
require s minima l assumption s abou t the statistic s of
the erro r an d prove s particularl y robus t to dat a tha t
do no t satisf y them . It ca n be use d wheneve r bound s
ar e availabl e on the acceptabl e erro r betwee n the dat a
an d the mode l outputs .
Dependin g on the applicatio n bein g considered , the
bound s definin g the objec t bar s ma y be determine d in
variou s ways . The y ma y correspon d in som e case s to
som e a priori knowledg e concernin g the natur e of the
measuremen t error . The experimenter' s intuition
coul d als o be use d to defin e wha t is consioere d a s a n
acceptabl e error . In an y cas e the definitio n of the
objec t bar s is a crucia l par t of the procedur e tha t
define s the siz e an d shap e of the membershi p se t S
entirely . This se t mus t the n be characterized , an d
the metho d presente d her e for this purpos e ha s the
advantag e of bein g abl e to handl e model s tha t ar e
nonlinea r with respec t to thei r parameter s withou t
linearizin g them . This is of specia l importanc e for
kinetic studie s wher e the mode l outpu t Is almos t
alway s nonlinea r with respec t to the parameter s to be
estimated .
In the example s discusse d here , we wer e concerne d
with model s describe d b y ordinar y differentia l
equation s tha t ar e linea r with respec t to the inpu t an d
time-invariant . The metho d would appl y a s well to
nonlinea r and/o r time-varyin g model s an d to model s
describe d by partia l differentia l equation s suc h a s ar e
encountere d with plug-flo w reactors .
Acknowledgement
This work wa s supporte d by a NATO Collaborativ e
Researc h Gran t 85/0083 . Joh n Happe l is als o gratefu l
for suppor t from Nationa l Scienc e Foundation , Gran t
CPE-83-13757 .
REFERENCES
Belforte , G. an d B. Bon a (1985) . An improve d
paramete r identificatio n algorith m for signal s
with unknown-but-bounde d errors . Preo . 7th
IFAC/IFOR S Svmp . on Identificatio n an d Syste m
Paramete r Estimation . York, 1507-1512 .
Belforte , G. an d M. Milanes e (1981) . Uncertaint y
interva l evaluatio n in presenc e of unknown -
but-bounde d errors . Non linea r families of
models . Proc , 1s t 1ASTE P Int, Symp . Modeling ,
Identification . Control , Davos,75-79 .
Bertsekas , D. P . an d I. B. Rhode s (1971) . Recursiv e
stat e estimatio n for a set-membershi p
descriptio n of uncertainty . IEEE Trans . Automat .
Contr. . AC-16 . 117-128 .
Fogel , E. (1979) . Syste m identificatio n via membershi p
se t constraint s with energ y constraine d noise .
IEEE Trans . Automat . Contr. . AC-24 . 752-758 .
Fogel , E. an d Y. F. Huan g (1982) . On the valu e of
informatio n in syste m identificatio n - Bounde d
nois e case . Automatica . J8 , 229-238 .
Happel , J . , H. Y. Cheh , M. Otarod , S . Ozawa , A. J .
Severdia , T. Yoshida , an d V. Fthenaki s (1982) .
Multiple isotop e tracin g of methanatio n ove r
nicke l catalyst : II. Deuteromethane s tracing .
Journa l of Catalysis , 3H-328 .
Happel , J . , K. Kamholz , D. Walsh , an d V. Strangi o
(1973) . Kinetic s of the isobutane -
Isobutene-hydroge n syste m usin g tracers . Ind.
Eng . Chem . Fundam. . J2 , 263-267 .
Martin, R. D. (1978) . Robus t estimatio n for time serie s
autoregressions . In R. L. Laune r an d G. N.
Wilkinson (Eds.) , Robustnes s in Statistics .
Academi c Press , New York. pp . 147-176 .
Milanese , M. an d G. Belfort e (1982) . Estimatio n theor y
an d uncertaint y interval s evaluatio n in presenc e
of unknow n bu t bounde d errors : linea r families of
model s an d estimators . IEEE Transaction s on
Automati c Control . AC-27 . 408-414 .
Norton , J.P . (1985) . Identificatio n an d applicatio n of
bounde d paramete r models . Prep . 7th IFAC/IFOR S
Svmp . on Identificatio n an d Svste m Paramete r
Estimation . York, 1197-1202 .
Schweppe , F. C. (1968) . Recursiv e stat e estimation :
unknow n bu t bounde d error s an d syste m inputs .
IEEE Trans . Automat . Contr. . AC-13 , 22-28 .
Walter , Ε., H. Piet-Lahanier , an d J . Happe l (1986) .
Estimatio n of no n uniquel y identifiabl e
parameter s via exhaustiv e modelin g an d
membershi p se t theory . Proc . 1MACS - IFAC
Svmp . Modelling an d Simulatio n for Contro l of
Lumpe d an d Distribute d Paramete r Systems .
Villeneuv e d'Ascq , 313-317 . To appea r In Math ,
an d Comp . in Simulation .
Walter , E. (1982) . Identifiabilitv of Stat e Spac e Models .
Springer-Verlag , Berlin Heidelber g New York.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

FAST ADAPTIVE CONTROL ALGORITHMS IN

pH MEASUREMENT
J. J. Minambres and M . de la Sen
Departamento de Fisica, Facultad de Ciencias, Universidad del Pais Vasco, Leioa,
Aptdo. 644 de Bilbao, Spain

ooiatro l p r o e e s s . Im th e secon d metho d (s o

A b s t r a c t . A f a s t adaptiv e contro l
o f pH measurement s i n a c i d i c was -
tewate r i s implemented . I t i s b a -
se d upo n th e us e o f numerica l me -
thod s to accelerat e th e converge n
ce i n parameter-adaptiv e contro l
a l g o r i t h m s . The one-paramete r c a -
se i s exhaustivel y investigated .
In suc h a cas e th e l i m it valu e o f
the update d paramete r i s uniqu e (
roughl y b e i n g equivalen t to a n i-
d e n t i f i c a t i o n scheme ) s o tha t th e
h y p o t h e s i s from Numerica l Analy -
s i s to chang e th e b a s i c e r r o r s e -
quenc e b y a n accelerate d e r r o r s e -
quenc e (b y changin g th e adaptiv e
law ) remai n s t a n d a r i z e d .
Keywords , Adaptiv e c o n t r o l ; f a s t
adaptiv e a l g o r i t h i n s ; pH measure -
ments ; numerica l a n a l y s i s .
INTRODUCTIO N
Adaptiv e Contro l o f i n d u s t r i a l processe s
i s o f a n obviou s i n t e r e s t , i n chemica l indu s
t r y whe n th e P A R A M E T E R S O F t h e P R O C E S S ar e
p a r t l y unknow n o r t i m e - v a r y i n g (Goodwi n
an d S i n , 1984) · One o f th e mos t importan t
drawback s on e cope s whe n d e s i g n i n g adapt i
ve contro l algorithm s i s t h e i r l a c k o f s -
pee d i n updatin g th e tru e p r o c e s s . Thi s
i s du e to th e o s c i l l a t o r y c h a r a c t e r i s t i c s
o f th e parameter-estimatio n scheme . Seve -
r a l approache s hav e bee n g i v e n to dea l
w i th t h i s problem . I n (De l a S e n , 1984a ;
De l a S e n , 1984b ) th e s o - c a l l e d optimiz a
t i o n an d adaptiv e samplin g approache s ar e
formulated . I n th e f i r s t on e a n equiv a -
l e n t nea r l i n e a r mode l ( " a d - h o c " f o r qua -
d r a t i c optimization ) i s derive d s o tha t a
pproximatel y describe s th e o v e r a l l adapt i
ve p r o c e s s . The fre e parameter s o f th e b a
s i c algorith m ar e updated , w i t h in th e s t a
b i l i ty domain , i n real-tim e i n v o l v i n g
bot h p r e d i c t i o n s an d c o r r e c t i o n s o f th e
t r a c k i n g / r e g u l a t i o n e r r o r measurements .
This updatin g work s a s a secon d adapta -
t i o n l e v e l i n a h i e r a r c h i c a l contro l con -
tex t an d i t i s performe d throug h a " p r o -
p e r " t r a n s l a t i o n o f th e o p t i m i z a t i o n r e -
sult s on "nea r l i & e a r q u a s i e q u i v a l e n t "
syste m approximatin g th e o v e r a l l adaptiv e
c a l l e d : th e adaptiv e samplin g approach) ,
th e h i e r a r c h i c a l secon d adaptatio n l e v e l
i s implemente d b y a real-tim e updatin g o f
the samplin g i n t e r v a l s o f th e d i s c r e t e
p r o c e s s . To achiev e th e a s s o c i a t e imple -
mentatio n a p r e v i o u s mode l i s derive d f o r
a d i s c r e t e syste m with state-dependen t
samplin g i n t e r v a l s . T h i s mode l ( f o r m a l l y )
work s a s a z-transfor m (polynomia l AP-
proach ) o f th e input/outpu t d e s c r i p t i o n
s o tha t a p a r a m e t r i c a l , althoug h time -
v a r y i n g , r e a l i z a t i o n o f th e o v e r a l l d i s -
cret e proces s i s allowed . I n bot h methods ,
a considerabl e improvemen t o n th e t r a n -
s i e n t performance s o f th e parameter-ada £
t i ve scheme s wer e r e g i s t e r e d . I n Minam -
b r e s an d De l a Se n ( 1 9 8 6 ) , anothe r conce £
t u a l l y d i f f e r e n t approac h i s developped .
While th e two abov e method s dea l w i th i n -
d i r e c t approache s o f th e proble m b y mean s
o f a secon d adaptatio n l e v e l , th e l a s t o -
ne m o d i f i e s d i r e c t l y th e parameter-adapt i
ve algorith m b y changin g th e adaptiv e s -
chem e b y u s i n g numerica l a n a l y s i s method s
to improv e th e rat e o f convergenc e o f th e
t r a c k i n g e r r o r to a l i m it while m a i n t a i -
n i n g th e scheme' s s t a b i l i t y ,
NUMERICAL ANALYSIS PRELIMINARIES
Give n i s a r e a l sequenc e generate d
as χ = g ( x . . ) , t = 0 , 1 , . . . with a r b i t r a r y
x q, whic h i s convergen t to a f i x e d p o i n t
x *. The rat e o f convergenc e i s , i n gene -
ral , linear .
The i t e r a t i v e schem a
x a s x £ g ( x t) - χ. 3 ) G ( x
t4i t - t
g ( g ( x t) ) - 2 g ( x t) +1 x t (1 )
i s know n a s th e S t e f f e n s e n s metho d (Blum
1972 ; Orteg a an d R h e i n b o l d t , 1 9 7 0 ) . The
two f o l l o w i ng s e t s o f c o n d i t i o n s (th e
f i r s t on e b e i n g prove d i n Blum , 1 9 7 2 ; an d
the secon d on e i n Minambre s an d De l a Se n
1986 ) bot h ensur e tha t
x x x 2 0 )
l t « . - * l ~ K - * \
Se t o f C o n d i t i o n s I

313
314 J. J. Minambres and M. de la Sen
Eqn. (2) stand s f o r sequenc e (1) i f
(1.1) g ( x ) ha s a f i x e d p o i n t x *.
(1.2) g ( x ) ha s continuou s t h i r d d e r i v a t i -
ve i n a neighborhoo d centere d a t x *.
(1.3) The f i r s t d e r i v a t i v e o f g ( x ) a t x *
i s d i f f e r e n t from u n i t y .
Se t o f Condition s I I
Eqn . (2) stand s f o r sequenc e ( l ) if
(11.1) G(x ) ha s a f i x e d poin t x *.
(11.2) The f i r s t d e r i v a t i v e o f G(x ) a t x *
is zero .
(11.3) The f i r s t d e r i v a t i v e o f G(x ) i s
continuou s an d th e secon d on e e -
x i s t s an d i t i s bounde d i n th e i n -
t e r i o r o f a neighborhoo d centere d
at x * .
It wa s prove d i n Minambre s an d Be l a Se n
(1986 ) tha t ( I . i ) ( 1 5 3 ) 4 . ( I I . l ) ^ ( l I . i)
( 1 ^ i £ 2) s o tha t th e se t o f Condition s
I I (assumin g tha t I I . l stands ) i s implie d
by th e se t I .
Objectiv e i n adaptiv e contro l
Ou r objectiv e i n adaptiv e contro l i s to
modif y th e adaptatio n sequenc e o f th e a -
daptiv e c o n t r o l l e r s o a s to mak e th e tra c
k i n g / r e g u l a t i o n e r r o r e ( t ) to v e r i f y (2)
(e*s=0) b y mdifyin g th e adaptiv e contro -
l l e r parametrica l adaptatio n by u s i n g a n
expressio n l i ke eqn . ( l ) while m a i n t a i -
n i n g th e c l o s e d - l o o p s t a b i l i t y o f th e o -
v e r a l l adaptiv e scheme .
Remark s
(1 ) The S t e f f e n s e n * s method , l i ke show n
i n ( 1 ) , i s applicabl e f o r s c a l a r sequen -
ce s s o tha t th e curren t experimentatio n
ha s bee n performe d o n on e a d a p t i v e - p a r a -
mete r (o f th e adaptiv e c o n t r o l l e r c a s e ) .
Althoug h wor k i s i n proges s abou t th e me -
t h o d ' s g e n e r a l i z a t i o n to th e m u l t i - p a r a -
mete r case , extension s ar e n o n t r i v i a l .
This i s b a s i c a l l y motivate d b y two r e a -
s o n s , namely :
( l a ) Steffense n s metho d impose s a u -
niqu e l i m it to th e i n v o l v e d (accelerated )
sequenc e while adaptiv e contro l s u p p l i e s
l i m i ts o f convergenc e o f th e adaptiv e
c o n t r o l l e r parameter s whic h ar e dependen t
(excep t i n th e one-paramete r case ) o n
the i n i t i a l c o n d i t i o n s an d o n th e choice s
o f th e fre e parameter s (namely , thos e
whic h ca n be f r e e l y chose n b y th e d e s i g -
ne r w i t h in th e s t a b i l i t y bounds) ,
( l b ) To ou r knowledge , ther e i s no a -
v a i l a b l e d i r e c t g e n e r a l i z a t i o n o f th e
S t e f f e n s e n ' s metho d to th e vecto r cas e
whic h prevent s a g a i n s t change s o f th e
foresee n l i m it o f th e accelerate d s e -
quence .
(2) The a p p l i c a b i l i t y o f th e propose d me
tho d ha s a wide characte r f o r parameter -
adaptiv e a l g o r i t h m s . The onl y "a p r i o r i "
c o n d i t i o n i s to us e a n algorith m havin g
th e t h e o r e t i c a l p o s s i b i l i t y o f real-tim e
m o d i f i c a t i o n o f a fre e paramete r while
maintainin g s t a b i l i t y .
(3) The p r a c t i c a l r u l e s whic h mus t be t a -
ke n i n t o accoun t ar e then :
(3a) To Commutat e ( i n p r a c t i c e to r e s -
tart ) to th e nomina l value s o f th e fre e
paramete r whe n n o t r a n s i e n t behaviou r im -
provemen t i s experimentall y observed .
(3b) To implemen t a " t e s t " s u p e r v i s o r y
l e v e l i n th e multi-paramete r cas e s o a s
to preven t a g a i n s t th e scheme d s t a b i l i t y .
I n th e nex t s e c t i o n , th e metho d i s applie d
to th e standar d adaptiv e contro l algorith m
o f th e P r o j e c t i o n metho d o f Goodwi n e t a l .
(1980) (Se e a l s o , Goodwi n an d S i n , 1984),
whic h ha s bee n modifie d f o r th e b i l i n e a r
mode l a r i s i n g from a schem e o f ρH measu -
remen t i n a c i d i c wastewater .
ADAPTIVE CONTRO L FOR pH MEASUREMEN T
I t mus t be pointe d ou t tha t th e bette r o -
ne i s abl e to p r e d i c t th e outpu t o f a sy s
tern , th e bette r on e i s abl e to contro l i t .
Thu s i f th e syste m respons e i s dominate d
by n o n l i n e a r c h a r a c t e r i s t i c s , i t i s prob a
b l y b e t t e r to us e a n o n l i n e a r mode l rathe r
tha n a n approximat e l i n e a r model . The pr o
ble m o f c o n t r o l l i n g a c i d i t y i n a c o n t i -
nuou s flow o f i n d u s t r i a l wastewater" u s i n g
a n o n l i n e a r adaptiv e c o n t r o l l e r i s consid e
red . The mode l a i s tha t o f McAvoy, Hs u
an d Loiltentha l (1972). A s t r o n g a c i d flows
i n t o a tan k an d i s ther e thoroughl y mise d
with a s t r o n g bas e inwar d rat e o f flow i s
c o n t r o l l e d i n suc h a wa y a s to produc e a
n e u t r a l outwar d flow from th e tank . Beca u
s e th e a c i d an d th e bas e ar e s t r o n g , eac h
i s completel y d i s s o c i a t e d , an d a l s o th e
d i s s o c i a t i o n o f th e wate r ca n be d i s r e g a r
ded . The equatio n d e s c r i b i n g t h i s mode l i s
v -±JL
at
P t( a - y t) - u t( b 4 y t) (3)
whe re
y t= £*r] - [θΗ~3 i s th e distanc e from
neutrality .
V== volum e o f th e tan k
Ρ =rat e o f flow o f th e a c i d
t
a = concentratio n o f th e a c i d
u^= rat e o f flow o f th e bas e
b= concentratio n o f th e bas e
y ca n be determine d from th e ρH valu e
"D
p ( t ) b y th e no n l i n e a r transformatio n (
Quagliano , 1958) ;
y t- 1 0 - * < *> -1θΡ<*> (4)
K
wher e k «wate r e q u i l i b r i u m constan t C±
w
10"* ( u s i n g gram-ion , l i t e r u n i t s ) .
I t i s assume d tha t b (known ) an d a (unk -
 Fast Adaptive Control Algorithms in pH Measurement 315

nown ) ar e bot h f i x e d , i s o n - l i n e mea - V=l l i t e r ; T=9 s e c ; a=b= 2 m o l / l i t e r l ( l l a )

surabl e an d tha t u . ca n be g i v e n a s s i g n e d
value s w i t h in c e r t a i n l i m i t s . I n c o r p o r a -
t i n g me sûremen t an d inpu t actuato r e r r o r s
we hav e th e f o l l o w i ng approximat e d i s c r e -
te-tim e mode l (th e n o t a t i o n i s h e l d a s a -
bov e )
't+l - t T ~f~ t y a -y )-u (b4 -y )3 (5a )
y +
t t
& 10. Scheme' s no n l i n e a r i t i e s mak e th e S t e -
A
= 4t 0
w i th
wher e denote s th e distanc e from neutr a
l i t y, an d Τ i s th e samplin g i n t e r v a l .
I t i s assume d tha t two parameter s o f (6b )
(namel y & 2, > ) ar e known . Foν r an y bounde d
desire d outpu t sequenc e ^ £~° · t= t +nT f
η i n t e g e r } th e inpu t i s generate d b y
mente d b y assumin g a n o i s e - f r e e p r o c e s s .
0 < u i 4 liters/minut e l(llb)
F t«=0.1125+0. 1 s i n ( t σ / 2 5 ) l i t e r s / m i - j
nut e -J
t
Figur e 1 correspond s to © ^ ( O j ^ l O , e(0) =
(5b ) f f e n s e n ' s algorith m ( S ) to tak e abou t 1 0
samplin g step s to trac k th e reference .
Figur e 2 i s r e l a t e d to s i m i l a r experimen -
(6a ) t a t i o n w i th @ 1( 0 ) = 5 , c ( 0 ) = 0 . 0 8 5
Note tha t i n b o t h s i t u a t i o n s th e goodnes s
o f th e propose d metho d toward s a f a s t com
(6b ) vergenc e o f th e a l g o r i t h m . Not e a l s o tha t
th e f a s t algorith m doe s no t s u f f e r s o f o £
c i l l a t i o n s aroun d th e n e u t r a l i t y durin g
the t r a n s i e n t . Forbidde n d e v i a t i o n s ca n
be c o n t r o l l e d b y approriat e f i x a t i o n o f
the upper/lowe r bound s f o r th e rat e o f
flow o f th e bas e whic h i s i n j e c t e d to th e
tan k i n orde r t o achiev e th e pursue d neu -
t r a l i t y . The s i m u l a t i o n s hav e bee n imple -

(7 ) CONCLUSIO N
T h i s pape r ha s deal t with th e a p p l i c a t i o n
w i th th e choic e to a chemica l proces s o f pH measuremen t
r.. o o f a f a s t adaptiv e c o n t r o l algorith m
u. if 0 « y, u (7a ) whic h i s base d o n th e S t e f f e n s e n ' s metho d
t ^ t '/ ms ma x o f numerica l a n a l y s i s . R e g i s t e r e d p e r f o r -
u(t)« imax
i f u t ' / ma x (7b ) mance s hav e bee n show n to be c o n s i d e r a b l y
improve d w i th respec t to th e standar d
0 i f u; ο (7c ) ornes .
é

Fo r th e standar d p r o j e c t i o n algorith m o f ACKNOWLEDGMENTS

Goodwin , Ramadg e an d Caine 1 s ( 1 9 8 0 ) , eqn .
(1 ) become s f o r th e k** update d paramete r The author s ar e ver y g r a t e f u l to Comisi6 m
Asesor a de I n v e s t i g a c i o n C i e n t i f i c a y Se e
sk £k J*<ê )-8 3 t t n i c a b y i t s p a r t i a l suppor t o f t h i s wor k
(ο) (projec t 9 6 8 / 8 4 ) .
β ( β ( $ ; » - 2 β( $ ; ) + 8
w i th REFERENCE S

\l\ Blum . E . K . ( 1 9 7 3 ) , Numerica l A n a l y s i s

9 + an d Computatio n Theor y an d P r a c t i c e .
t-l
° t ft* + φ
t-d
(9 )
Addiso n Wesley , Reading , Massachu -
sett s

with d (delay ) bein g u n i t y . £ 2 ^ De l a S e n , M. (1984a) . O n - l i n e optim i

s a t i o n o f th e fre e pa r cone t e r s i n d i s
The p H i s obtaine d from th e n o n l i n e a r cret e adaptiv e c o n t r o l systems . I E E
transformatio n (Quagliano)a s f o l l o w s ; Proceeding s (Par t D ) , V o l . 1 3 1 . ρρ·
146-157 .
1 -z k ; ζ k + z , y , - 1 = 0 (10a )
t w' ww t t [p3 De l a S e n , M. (1984b) . A metho d f o r
+ 4 k improvin g th e adaptatio n t r a n s i e n t
z * t W t w
u s i n g adaptiv e sampling . I n t . J .
(10b ) Control , Vol. 4 0 , p p . 639-665 .
t= 2k
4 4 k (VJ Goodwin , G . C . , P . J . Ramadge. ; an d P .
*t w E . Caine s (1980) · D i s c r e t e time m u l t i
=log(z.)=lo g (10c ) v a r i a b l e adaptiv e c o n t r o l " , IEE E
2k
Trans . Autom . C o n t r o l , V o l . AC-25 ,
Experimenta l exampl e N o . 3 , ΡΡ· 449-456 .
The numerica l v a l u e s ar e no w {VJ Goodwin , G . C , an d K . S . S i n (1984) .
Adaptiv e f i l t e r i n g p r e d i c t i o n an d
316 J. J. Minambres and M. de la Sen

control. Prentice Hall, New York.

\β\ McAvoy, T.J., E. Hsu, and S. Lowen-
thal (1972). Dynamics of pH in a con
trolled stirred reactor. Ind. Eng.
Chem. Process Pes. Dev., Vol. 13t
No.l
{Y^ Mitiambres, J.J., and M. de la Sen
(1936). Application of numerical me-
thods to the acceleration of the con
vergence adaptive control algorithms.
The one dimensional case. Computers
and Mathematics with Applications.
1986 (in press).
\s\ Nuyan, S., and R.L. Carrol (1979).
Minimal order arbitrarily fast adap-
tive observers and identifiers. IEEE
Trans. Autom. Control, Vol. AC-24,
No. 2.
| V ] Ortega, J.M. and V/.C Rheinboldt
(1970). Iterative solution of nonli-
near equations in several variables.
Academic Press, New York.
\l6\ Quagliano, J.V. (1958). Chemistry,
Prentice-Hall, Englewood Cliffs N.J.

— tuff,.,*
— ca<fc
\i !I 1
\ \ 1
ι ϋ ! ί ! 1 ί
! ! ! I !
\\\\\ ! ! Ί!
f

m
l i- i 1
! ! i! ·
! ! ! i-' : !; !ϋ!
! !i!
!ι·; !; !ί!'
!; U!'

Ii
ϋ !'
Η!
ί !

jj
I •

! I

-ΤΑΜ Ρ Le.-

EPKOR.
cejç

• \t\\
\ I.
ίΐ 5 Ν ! ,
T î !5 • 2 ί ί J • ,
ί» ;!ι » 2 ϋ ϋ Ζ • S { ι
ci 'Ί ι i' i! i' i' ii 'i ?· 5 ι I Ί
~\ . iiiiiiiiiiii-iiilii!if H ,
1
ϋ •'· ϋϋϋϋ S Ϊ i Μι1φ '
!· I Ί Ϋ ϋ Ϋ 5 Ϋ Η
Ϋ ÎL'H 5 Π 51ί (
«Ί 5 Ί ' S 5 Π Ι !
Ί i' Π 111
11
Ί Ί!

to ffi/IPLEf
a» jl ο 5 (Ο 7ο 1 1

fi *
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

DESIGN OF A MULTIVARIABLE
SELF-TUNING CONTROL ALGORITHM FOR
DISTRIBUTED SYSTEMS
C. McGreavy and J. A. A. de M . e Castro
Department of Chemical Engineering, University of Leeds, Leeds LS2 9JT, UK

Abstract. Optimal control strategies are very dependent on the accuracy of the model
and this assumes some importance in operating plant when there is a change in
characteristics. Consequently, in order to ensure good and stable performance a
detailed description of the process is required. Since it is not common to incorporate
tunable properties in the model to compensate for any process changes it is difficult
to realize the benefit of the sophisticated strategies typical of the successes in the
aerospace industry. Nevertheless, some significant improvements are possible. An
example of the kind of problems which arise in respect of distributed parameter
processes, is the fixed-bed catalytic reactor. Detailed models are not useful for
control because it results in loss of simplicity and excessive computing requirements.
Moreover, it is generally difficult and costly to identify such models because of the
problems associated with the measurement of some of the relevant variables e.g. rate
constants and/or catalyst activity. To be attractive, overall controller design must
be based on a simple but robust control algorithm which allows for an approximate form
of model. Consideration is given to the use of an adaptive closed-loop control law
based on a multi-input-multi-output system having random process inputs and the
observations subjected to noise. A comparison is made with the performance of a
multivariable optimal regulator using a quadratic performance index. The effects of
model inaccuracies, sampling times and other design parameters are also considered.
The need for this basic structure is shown to be important in the case of the tubular
reactor to allow for the approximation of multidimensional lumped parameter
representation of the process.

Keywords. Adaptive control; Distributed parameter systems; Multivariable control

systems; Optimal control; Parameter estimation; Temperature control

INTRODUCTION facilitate the implementation of a simpler but

robust control algorithm.
Conventional optimal control strategies have shown
to be very satisfactory when working well. Typical process plants are non-stationary and so
Nevertheless, they are model dependent, ie, they the dynamic behaviour changes with time. Clearly,
are very dependent on the accuracy of the model and any controller where parameters are not
this is obviously an important factor in operating periodically adjusted so that any changes are
plant when there is a change in characteristics. A estimated and compensated for will suffer in terms
detailed description of the process is often of overall performance. In classical two or
required to ensure good and stable performances. three-term controllers, retuning is often achieved
Commonly, there are no tunable properties in the by experienced personnel when major changes in
model to compensate for the continuous process operating conditions are detected or planned. When
changes suffered in most industrial situations. this retuning is automatic, the controller is
This has been an important deterrent in applying considered to be adaptive. This might involve an
such strategies in the chemical and petrochemical adjustment in either the controller parameters or
industries, although the potential can be seen from of the controller configuration while still
the successes in the aerospace industries. carrying out normal operation.
Although the problems are rather different it
nevertheless suggests that useful improvements may An important class of such controllers is based on
be gained within the rather more resource and self-tuning algorithms. Different schemes have
financial constraints imposed by the process been developed where a recursive estimation
industries. procedure is used to update the parameters of a
simple pre-selected model of the plant, and these
This becomes even more limiting when consideration are then used to determine new values for the
is given to the problem of controlling distributed controller constants in accordance to some specific
parameter processes, such as the fixed-bed design rule. The new values of the controller
catalytic reactor. Use of detailed models parameters are then employed calculate the
reflecting the distributed parameter character is appropriate action based on a suitable feedback
not feasible as it results in lost of simplicity control law.
and excessive computing requirements. In addition,
it is generally difficult and costly to identify There is a vast range of efficient self-tuning
such models because of the problems associated with schemes, the main differences being related to the
the measurement of some of the relevant variables choice of the identification algorithm or of the
(e.g. rate constants and/or catalyst activity). control structure together with some performance
The overall controller design must therefore be criterion. As far as model parameter estimation is
based on very much simplified approximations to the concerned, the techniques used range from recursive
behaviour. This will require a more frequent least squares (Boehm, 1982) to maximum likelihood
updating ot t\\e model but nevertheless will (Isermann, 1982) and Kalman filtering (Hubbard,

317
318 C. McGreavy and J. A. A. de M. e Castro

1982). In relation to the control criteria,
theorectical and/or experimental applications of
linear quadratic control laws (Schumann, 1982;
Bucholt, 1979b), the fundamental minimum variance
(Ko, 1982) or some of its suboptimal derivatives
(El-Sherief, 1982) have been reported in the
literature. Many of these self-tuning techniques
(Astrom, 1973; Clarke, 1975) have been referred to
single-input single-output systems and therefore
may only be applied to multivariable systems when
decoupling procedures are feasible. Recently,
there has been proposed a number of algorithms
extending the previous schemes to include the
control of some classes of multi-input multi-output
systems (Bucholt, 1979a; Morris, 1979).
Optimal control schemes are based on finding a
procedure which will cause a performance index to
approach an optimum value. The resulting control
law is of the type
Κ' χ (2)
where Κ' is a (ncxnv) controller gain matrix
resulting from the optimization of the desired
performance index. A common approach is to
minimize a scalar quadratic objective function
including both the states and the control variables
in the form

An important characteristic of most adaptive N-l +

control procedures is that only the process model
parameters are considered for estimation, whether
this is performed in open- or closed-loop with the
process and controller. This is equivalent to
regarding the plant output as only being required
to update the model, with the controller
coefficients being calculated by a separate
algorithm. However, there is no reason why the
controller parameters should not be included in the
identification procedure so that a scheme based on
a completly closed-loop mode can be employed.
The work reported in this paper can be regarded as
a preliminary step in attacking the complex problem
of designing multivariable self-tuning regulators
for systems typified by the fixed-bed catalytic
reactor. In such systems the effect of dead times
characterizing the behaviour of these processes can
seriously affect the performance of the control
system and therefore the overall performance of the
plant operation. This is because of the delay in
receiving the information about the state of the
system and/or the delay before any corrective
action can take place. On the other hand, the
requirement for simplicity in such designs have led
to the use of first or second order plus time delay
transfer functions as approximate models for
process control purposes (Awda, 1980). However,
these are limited to single-input single-output
systems and unless complex (and sometimes not very
effective) decoupling schemes are used they cannot
be applied to multi-input multi-output systems.
Moreover, in some cases (Morris, 1981) the
parameters of such models are strongly dependent on
the nature of the test signals used. Nevertheless,
such explicit inclusion of the delay is not the
only alternative to using the complete distributed
model of the process.
In this paper an adaptive closed-loop control
procedure is developed for a multi-input
multi-output system subject to random process
inputs and the observations are noisy. The
fundamental procedure is based on Kalman filtering
techniques so that the performance can be compared
with that of a multivariable optimal regulator
using a quadratic performance index. The effects
of model innacuracies, sampling times and other
design parameters are also considered.
I <5k Q, *k> (ΐ&Α) (3)
^NI^N k=0
ea r
where J?, O^j and £ 2 symmetric positive definite
matrices. For systems described by a discrete
version of Eq. (1) it has been shown that the
optimal gain matrix K' minimizing the scalar J
can be obtained by recursive solution of the matrix
equations, details of which are presented elsewhere
(Awda, 1980; Vago, 1974). The solution requires
-
knowledge of system matrices A and Β and of the
weighting matrices JF, 0^ j a n d O ^ . although the
solution is optimal™ it" strongly depends on the
accuracy of the process model and then on matrices
A and _B.
Taking note of this, a multivariable self-tuning
regulator can be implemented. It consists in
combining an identification scheme based on Kalman
filtering to estimate the main parameters of A and
_B and an optimal control algorithm which calculates
ΚΓ' from the new values of A and J3. The control
uction can then be determined by^ using Eq. (2)
where the states are substituted by the best
estimates. With such a procedure, the model
parameters are updated and compensate for the
changes in the process or the innacuracy of the
initial model. Figure 1 illustrates the sequence
of calculation for the self-tuning strategy.
MODIFIED SELF-TUNER
The strategy proposed here differs significantly
from that presented above. Although the control
law is of the same type (Eq. (2)), the essential
feature is that it does not make use of any
separate optimization scheme for the determination
of the controller gain matrix. Instead, it uses
the standard estimator to update the controller
coefficients which are chosen a priori in
accordance with some design rule. The updating of
the parameters is then performed alternately with
the estimation of the parameters of the approximate
model. This strategy is basically the same as that
used for the Kalman filter to estimate parameters
in ordinary differential equations since the
mathematical formulation used for the
identification scheme is that of the closed-loop
configurât ion.

OPTIMAL SELF-TUNER To illustrate this consider a system represented by

a discrete version of Eq. (1)
Consider a system where dynamic behaviour can be
represented by
s Φ (xf)
*k l +D ( x f) (4)
+
A χ + B u (Ο
where the superscripts (s) and (pm) refer to state
where A is a (nvxnv) system matrix and Β is a and model parameters. The (nvxnv) transition
(nvxncf" matrix relating the nv states to ^ h e nc matrix Φ and the (nvxnc) input matrix D are
control inputs. functions^of the model parameter vector. If a
proportional control law of the type
 Multivariable Self-tuning Control Algorithm for Distributed Systems 319

pC S becomes
1 2
= -κ' ( x ) x (5) S
Γ Φ ô Ί r- -i
2fc+l
k "LE

is used (the superscript (pc) refers to controller

parameters ie to the elements of matrix Κ) , then
—
the following closed-loop state equation (12)

vp m 0 1
xs
XFE+I Ψ
pm
φ pc=.
( s
(6) _-k+l _ _-k
u
k ,XK £
K where Φ- Φί and jD result from appropriate
can be obtained, where the new (nvxnv) transition factorTzation^of the "model parameters in the
matrix ψ is given by original equation model given by Eq. (1). On the
other hand, the control vector can be given by an
extended form of Eq. ( 9 ) , which when substituted in
(7)
Eq. (12) leads to the following vector differential
equation

As stated above, this closed-loop configuration can

then be used in the identification procedure. The Φ, X . (13)
alternating parameter estimation strategy is -k+1 =k —k
employed here and has to be related to the type of 1
where the ((nv+npm)x(nv+npm)) transition matrix
parameters to be updated. However, in each case Φί is of the form
the parameters are estimated simultaneously. This
means that, for instance, all model parameters are
estimated at one time and that at next time step
2 m
the estimation procedure is switched to the m C
controller parameters.
S> (*l )
l m
- D(x£ ) K* + K°(x£ )
i (x£ )
(14)
It is convenient to work with reduced values for
the parameters. Thus, a reduced gain k^j relating
the state variable x£ to the manipulated action
u^ , can be defined as For the case where the procedure estimates the
controller parameters together with the state
variables the following matrix equation results
k. . - k7 .
k. . (8)
k*. Γ" s

where k^j is an appropriate reference value for

this gain, From this, Eq. (5) can be written in
the form 15)

c
Κ + Κ o£ > (9) pc
-k -k+1 3c

Factorization of the second term of the right hand In this situation the control action is given by
side of Eq. (9) allows the control vector to be Eq. (10), which when substituted into Eq. (15)
also written as gives

pc
ς s x (10) ^k+1 - |k^k 06)
I 1 I 2 ^k

where the constant matrix is only a function of where the (nv+npc)x(nv+npc) transition m a t r i x ^
the reference values for "the gains and S 2 is a
matrix with elements which are state variables or
zeros. It has a general block diagonal structure m F m
resulting from the above factorization and from the
need to maintain valid the basic matrix operations £ ( x P ) - D ( xpm)
N
in the second term of right-hand side of Eq. (10). (17)

The need for a similar parallel structure of this

algorithm requires the formulation of the
estimation problem to be made in accordance with
the type of parameter being estimated. This From Eq. (16) the state vector and the controller
results in two alternative procedures with parameters can therefore be estimated and, by
different mathematical forms but nevertheless direct substitution in Eq. (10), an appropriate
having the common feature of being implemented in control action can be determined.
closed-loop.
The sequence of calculations of this method is
Consider, first, the case of estimating the model schematically represented in Fig. 2 where the
parameters. The extended state vector is: alternating feature of the estimator is also
illustrated. Here, the environment is the whole
Xs pm
X 1 closed-loop, although the estimation is applied to
the controller and to the process model once each
1 1 (11)
time step. It can be seen that in the case of the
—k 1 —k controller coefficients, the updated information
and therefore the corresponding state-space model i.e. the gain matrix K_ and the state vector is
DCCR-V
320 C. McGreavy and J. A. A. de M. e Castro

fed directly into the controller to determine the response characterized by large off-set.
action. On the other hand, when only the model is Therefore, careful specification of the weighting
updated, the control action is obtained by using matrices in the control performance index is
the new estimates of the states and the previous necessary.
best values of K.
It should be noted that the above results
correspond to a situation where both algorithms are
DISCUSSION OF RESULTS subjected to severe changes in conditions. This is
especially true for the large value used for the
The method reported here for closed-loop adaptive sampling time. Values of one tenth of the dominant
control has the characteristic that it can easily time constant are often used in practice. The
be applied to multivariable systems typical of many results presented above correspond to a sampling
chemical processes. In this section, the main period which equals the minimum time constant of
objective will be to assess the performance of this the real process. Clearly, the slow sampling rate
strategy which does not involve the time consuming is a major factor in the unsatisfactory performance
calculation of the gains every time the model of the controller. Thus, when the initial model
parameters are updated. Inclusion of the control parameters are given as pj - p2= 3.0 and p^= 4.0
loop in the identifier allows the controller both algorithms diverge. Very poor state estimates
parameters to be also updated by the estimator so are produced as a consequence both of the
as to minimize the variance in relation to an excessively large sampling time and of the large
optimum mean value. errors in the initial estimates of the model
parameters. In addition to that, the poor state
To illustrate the essential features of this estimates give rise to inappropriate corrective
approach, consider the above algorithm when applied action even when these are highly penalized
to the two tank heating system shown in Fig. 3, according to Eq. 3.
which can be regarded as a discrete approximation
to a distributed parameter system. An energy Reducing the sampling time to a more reasonable
value leads
balance on each tank leads to the following
differential equations representing the dynamics of
= to an improved performance as
illustrated in Fig. 6 where pj = p£= 3.0 and p^- 4.0
the whole system and ^ 2 0.01 1 for a sampling period of six
seconds. Although the MST gives a slightly more
oscillatory response, mainly due to the closed-loop
dT ( p +)=
P T +T +T nature of the estimation procedure, it can cope
more satisfactorly with the problem of the off-set.
1Γ - i 3 1 Pl o P
3s This is because, compared with the OST the modified
scheme does not depend on the values of _F.

T t 19 The advantage of the proposed method can be seen

IF " * i - P 2 < > when inaccurate process models are used. Figure 7
2 2 illustrates this for the case where the initial
where pj= p^= p^- 1.0. It is desired to maintain estimates of the parameters are Pj = p = 6*0 and
2
p^= 8.0. These estimates represent errors of 500%
the temperatures Tj and at specified set-values
by manipulating the inlet and steam temperatures. in the values of the residence times in both tanks
It is assumed that only temperature can be and an error of 700% in the ratio between the heat
measured and that this is corrupted by a random transfer coefficient and the thermal capacitance in
disturbance having a standard deviation of 0.5 K. the first tank. The response obtained with the OST
The process noise in the inlet and steam is clearly unsatisfactory while an effective
temperatures has a standard deviation of 1 Κ and control of the system is possible using the MST.
the sampling period is taken as 30 s. For Another important feature of this technique which
identification and optimal control calculations, greatly contributes to its reliability is the
the following data have been used unless stated relative insensitivity with respect to most of the
control design parameters, when compared with that
otherwise
p m 0 1 of the OST.
=
This can be seen by changing the
values of £2 (MST: £ = 0.01 1; OST: £ 0.04 1 )
^(states) = P Q ( x ) = P Q C / ) = Ι Ο " I
= 2 2
to those corresponding to Fig. 8 where ~Q? = 0.001
3 2 3 is used for both algorithms. In thTs case the
£ =âj = 10 I; ^ assumed values between 10-10 I
modified self-tuner also presents a convergent,
though oscillatory response. This arises from the
When the system is subjected to a load change of 10
fact that these small values of (£2 give rise to
1
Κ in the inlet temperature, the performances of the
two control algorithms (Optimal Self-Tuner, inappropriate initial coefficients in matrix K
referred to as 0ST and Modified Self-Tuner referred which are used to initialize the identification ~ô£
to as MST) can be seen in Fig. 4. The initial the controller parameters. It corresponds to using
estimates of the model parameters equal those of poor initial estimates of the parameters in the
the process ie pj= p2= P3 - 1.0. Although the control identification procedure and therefore
control responses are comparable, there is a leads to less satisfactory behaviour. However, in
slightly faster convergence for the optimal scheme. spite of this, the response converges to an
This is not surprising since the optimal gains are acceptable value which is not the case for the
calculated every time step, while for the MST the optimal self-tuner procedure. Indeed, the greater
gains are updated only after each two iterations. sensitivity to this design parameter results in it
diverging.
The same results are not obtained when the initial
process model is innacurate. To demonstrate this a The system used here for the evaluation of the two
similar situation may be considered, the only adaptive control techniques has a linear structure
difference being in the initial estimates for the with respect to the state variables, although it
parameters in the model. Thus, if the set of can also exhibit intrinsic nonlinear
values pj = p = 1.5 and p^= 2.0 is used, the characteristics. This happens when the model
2
corresponding performances can be observed in parameters are considered as state variables for
Fig. 5. Despite errors of 50 and 100 per cent in identification. These nonlinearities have been
the initial estimates, the MST is able to maintain reasonably well overcome in the proposed adaptive
the system under effective control. This does not algorithm by incorporating a simple factorization
happen with the 0ST which gives a deteriorating procedure in the estimation strategy (Castro,
 Multivariable Self-tuning Control Algorithm for Distributed Systems 321

1983). This may not be sufficient if systems with discrete approximations to distributed parameter
very high nonlinearities among the main variables systems. Applying it to the fixed bed system, it
are to be considered. However, special is necessary to consider a model of slightly higher
linearization techniques, such as that using an dimensionality to take account of the equivalent
iterated form of the mean-value linearization lumped stages and this can be done by using a
theorem (McGreavy, 1972; McGreavy, 1981) have state-space model of a suitable higher order. Such
proved effective in such situations. It it then models implicitly incorporate the effects of the
possible to use this technique where locally time delay in the system response and so can be
linearized models are satisfactory. included in adaptive control algorithms of kind
considered here.
REFERENCES
CONCLUSIONS
Astrom, K.J., and B. Wittenmark (1973) On self
The conventional self-tuning regulator can be tuning regulators. Automatica, 9, 185-199.
subdivided in two distinct steps: identification Awda, F. (1980) State variable approach to the
and control. The first is mainly concerned with design of self tuning regulators. Ph.D Thesis,
updating an approximate process model so as to University of Leeds, U.K.
guarantee more reliable predictions of the system Boehm, H. (1982). Adaptive control of a dry etch
behaviour. It involves some form of on-line process by microcomputer. Automatica, 18,
recursive parameter estimator which also provides 665-673.
estimates for the system states. In the second Buchholt, F., K. Clement, and S.B. Jorgensen
step i.e. in the control stage, it is usually (1979) Selftuning control of a fixed-bed
required to recalculate the coefficients of the chemical reactor. Proc.of 5th IFAC Symp. on
controller so that appropriate control action can Identification and System Parameter Estimation,
be derived. However, the determination of new Darmstadt, Germany.
values for such parameters involves the solution of Buchholt, F., and M. Kummel (1979). Self-tuning
a set of recursive matrix equations. For a control of a pH-neutralization process.
multivariable system this becomes a very Automatica, 15, 665-671.
time-consuming procedure, so some improvement is Castro, J.A.A.M. (1983). Aspects of Modelling
highly desirable, particulary when real-time Chemical Processes for Adaptive Control. PhD
control is to be applied. Thesis, University of Leeds, U.K.
Clarke, D., and P.J. Gawthrop (1975) Self-tuning
One of the main attractive features of the adaptive controller. Proc. IEE, 122, 929-934.
control scheme considered in this work is that it El-Sherief, H., and N.K. Sinha (1982) Suboptimal
does not require the solution of such complex control of linear stochastic multivariable
matrix equations. Instead, by making use of the systems with unknown parameters. Automatica,
real process closed-loop configuration, a simple 18_, 101-105.
and effective procedure to update the coefficients Hubbard, M., and T. Da Silva (1982). Estimation of
of the controller has been proposed. This is feedstock concentrations in cement raw material
achieved by extending the identification algorithm blending. Automatica, 18, 595-606.
so as to consider the estimation of those Isermann, R. (1982). Parameter adaptive control
parameters together with the system states but in algorithms: A tutorial. Automatica, 18,
an alternative form, along with the estimation of 513-528.
the model parameters. In this way, it is possible Ko, Κ.Y., B.C. Mclnnis, and G.C. Goodwin. (1982)
to substantially reduce the computational burden Adaptive control and identification of the
characteristic of the optimal self-tuning dissolved oxygen process. Automatica, 18,
regulators when applied to multivariable systems. 727-730.
McGreavy, C , and J.A.A.M. Castro (1981) State
It has been shown that optimal control policies can estimation techniques in the identification of
be excellent. However, when the process models on chemical reactive systems. Proc. Chempor-81,
which they are based are only approximate, the Oporto, Portugal.
performance can be rather poor. In contrast, the McGreavy, C , and A. Vago (1972) Application of
regulator presented here exhibits good global nonlinear filtering techniques to adaptive
performances, particularly in terms of robustness. optimal control. A.I.Ch.E. 65th annual
In fact, it has been shown to be reliable for a meeting, New York, Paper 94-C.
wide range of filter and controller design Morris, A.J., Y.Nazer, and K.Chisholm (1979)
parameters providing a suitable sampling period is Single and multivariable self-tuning
used relative to the dominant time constant or microprocessor based regulators. Proc. of 3rd
where the degree of approximation represented by Int. Conf. on Trends in on-line computer
the model is very crude. control systems, University of Sheffield, U.K.
Morris, A.J., Y. Nazer, and R.K. Wood (1981)
The strategy takes advantage of the closed-loop Multivariable self-tuning control-theory and
nature of the identification procedure, leading to experimental evaluation. Proc. of IEE Conf.,
an equivalent open-loop dynamic representation for No.194, 106-112.
the overall system, as shown in Eq. (13) and Schumann, R. (1982) Digital parameter-adapt ive
Eq. (16). All the information related to the control of an air-conditioning plant.
process model as well as the controller has Automatica, 18, 569-575.
therefore been incorporated in the global Vago, A.G. (1974) Investigation of on-line
transition matrix. As a result, the simultaneous estimators for optimal control of chemical
estimation of both model and controller parameters reactors. Ph.D Thesis, University of Leeds,
ensures that there is some degree of self U.K.
compensation to guarantee the best response of the
process.
ACKNOWLEDGEMENT
Thus, it can also be applied to other systems not
exhibiting overdamped responses where simple first Dr. J.A.A.M. Castro, who is on leave from the
order transfer functions containing time lags give Department of Chemical Engineering, University of
poor performance. Moreover, the multivariable Coimbra, Portugal, wishes to acknowledge the
formulation of the technique can accommodate high financial support provided by the Calouste
order s^steteis without requiring decoupling Gulbenkian Foundation, Lisbon for carrying out this
procedures, so that it is well-suited to deal with research.
 C. McGreavy and J. A. A. de M. e Castro

Plant
PROCESS
Uncontrolled Response
0.02

£ . ISTATE & MODEL

PARAMETER
ESTIMATOR 0.00

OPTIMAL Computer
CONTROL -0.02
1
ALGORITHM
K

Fig. 5. Comparison of the two self-tuning regulators

CONTROLLER
using a sample time of 30 seconds and wrong
initial values for the model parameters
(P, 1.5; = 2.0)
Fig. 1. Structure of the optimal self-tuning q
P
regulator

Uncontrolled Response
PROCESS 0.02
>
OST
0.00
STATE 10 20 30
ESTIMATOR
MST Sample Times
and and
-0.02

CONTROLLER MODEL
PARAMETERS I PARAMETERS

Fig. 6. Comparison of two self-tuning regulators

using a sample time of 6 seconds and wrong
CONTROLLER initial values for the model parameters
P = 3.0 and p^ 4.0)
(P, 2
Fig. 2. Structure of the modified self-
tuning regulator
Uncontrolled Response
cd 0.02

0.00
20 0 4

Sample Times
-0.02

Fig.3. The stirred tank heating system

Fig. 7. Comparison of two self-tuning regulators
using a sample time of 6 seconds and wrong
initial values for the model parameters
(p = p = 6.0 and p = 8 . 0 )
2 3
Uncontrolled Response
0.02. -a
Uncontrolled Response
0 . 02

0.00"

-0.02,
MST
40 80
SAMPLE TIMES
120

0 . 00 /Γ/ > MST

— ^ 20
Sample Times
30

S -0.02
Β

ig. 4. Comparison of the two self-tuning regulators

using a sample time of 30 seconds and exact
initial model parameters (ρ = 1.0) Fig. 8. Performance of the modified self"tuning
regulator for different values of
Q (Q 10-3 I)
9 9
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

AN ADAPTIVE SYSTEM WITH AN

IDENTIFIER (ASI) FOR CONTROL OF THE
TEMPERATURE PROFILE IN A
TUBULAR REACTOR
V. M . Chadeev and Ye. M . Maksimov
Institute of Control Sciences, Moscow, USSR

A b s t r a c t . The pape r i s concerne d w i th c o n t r o l o f a c a t a l y t i c

reacto r use d f o r producin g malei c anhydrid e from benzene . The
objectiv e o f th e c o n t r o l i s th e y i e l d o f th e produc t s y n t h e s i z e d
an d i s obtaine d throug h c o n t r o l o f th e v e r t i c a l d i s t r i b u t i o n o f
th e temperatur e p r o f i l e i n a r e a c t o r . A mathematica l mode l i s use d
whic h d e s c r i b e s th e dynamic s o f therma l an d chemica l processe s i n
a r e a c t o r . The mode l parameter s ar e update d b y adaptiv e i d e n t i f i c a t i o n
a l g o r i t h m s . An algorith m o f adaptiv e i d e n t i f i c a t i o n w i th orthogona -
l i z e d inpu t v a r i a b l e s improve s th e accurac y o f paramete r e s t i m a t i o n
whe n inpu t v a r i a b l e s s t r o n g l y c o r r e l a t e . The computin g loa d i s s o
low tha t a micro-compute r w i ll do th e j o b . An i d e n t i f i e r update s
the proces s parameter s whic h i s a c r u c i a l advantag e s i n c e dynami c
p r o p e r t i e s o f a plan t ma y chang e s u b s t a n t i a l l y du e to d i f f e r e n t
r e a s o n s , l i ke c a t a l y s t a g e i n g , change s i n th e environmen t a i r com -
p o s i t i o n , s t u f f e d hole s i n th e benzen e s p r a y i n g d e v i c e s , e t c . S p e -
c i a l a t t e n t i o n wa s g i v e n t o th e syste m r e l i a b i l i t y . An algorith m
wa s designe d f o r p r o t e c t i n g th e informatio n store d i n RAM i n c l u d i n g
the program s whic h us e e r r o r - c o r r e c t i n g codes .

Keywords . C a t a l y t i c r e a c t o r ; malei c anhydride ; adaptiv e syste m

with a n i d e n t i f i e r ( A S I ) .

PRODUCTIO N OP MALEIC ANHYDRIDE The reacto r c o n s i s t i n g o f s e v e r a l t h o u -

M a l e i c anhydrid e with th e s t r u c t u r a l f o r -
mul a H - i s obtaine d throug h o x i d a -
H - C ^
t i o n o f benzen e b y th e a i r oxyge n ove r a
c a t a l y s t . Oxidatio n i s a n exotherma l reac -
t i o n , consequently , som e hea t mus t be r e -
moved . The reacto r diagra m i s show n i n
Pig .
san d c a t a l y s t - f i l l e d tube s i s house d i n a
v e s s e l w i th a c i r c u l a t i n g mel t o f n i t r i d e -
n i t r i d e s a l t s whic h remove s hea t to th e
wate r b a t h . The s a l t mel t temperatur e v a -
r i e s w i t h in 370° to 420°C. The a i r heate d
to 170°C i s fe d to th e evaporato r formin g
a n air-benzen e mixtur e (ABM). A 35:1 a i r -
-benzen e mixtur e i s f u r t h e r directe d to
th e c a t a l y t i c r e a c t o r wher e malei c anhyd -
r i d e i s formed . The mixtur e i s f i n a l l y fied
to a p a r t i a l condense r wher e th e malei c
anhydrid e vapou r i s condensed .

The reacto r e f f i c i e n c y i s measure d i n

term s o f th e malei c anhydrid e y i e l d . O x i -
d a t i o n proceed s ove r a s o l i d c a t a l y s t i n
a tubula r r e a c t o r . The benzen e o x i d a t i o n
i n c r e a s e s v/ith th e r e a c t i o n temperature ;
th e maxima l temperature , however , i s c o n -
P i g . 1 Reacto r Diagram .

323
324 V. M. Chadeev and Ye. M. Maksimov
strained because at some Tq the reaction
yields byproducts and the catalyst ages
fast* The catalyst has to be replaoed nor-
mally every 2 to 3 years. If T 0 is excee-
ded regularly, the catalyst lifetime redu-
ces sharply. Figure 2 shows the vertical
temperature distribution in a reactor.
Ζ γ
r »| &eaotor~l
* Controller
Ι *
1 Κ
s Choice of
working models
# Identifier
Fig. 2. Diagram of ASI for
Reactor Control.
With catalyst ageing the temperature pro-
file peak shifts downwards as shown by the
dashed line in Fig* 2* Temperature measu-
rements are taken in M points regularly
distributed over the reactor height* Uow
let us proceed to the control system pro-
per.
BASIC CONTROL LOOP
The catalytio reactor control system is
implemented as an adaptive control system
with an identifier (ASI). Its diagram is
given in Fig. 3*
H
Τ
Fig. 3· Vertioal temperature
Profile in the Reactor.
Several models predict the temperature in
each measurement point at different reac-
tor levels. The control objective is to
maintain the maximal temperature in the
reactor as close as possible to the maxi-
mal acceptable T 0 for the specified pro-
bability of exceeding .this temperature.
As seen from Fig. 3. the control loop in-
cludes an identifier in which a spectrum
of dynamic models is generated to predict
temperatures in all measurement points in
the reactor. There is a unit in which
some of them are chosen and used thereaf-
ter as operating ones* These models are
used to determine the oritical points in
1 » which the temperature is the olosest to
that acceptable for the reactor, T 0. Cont-
rol is updated by these working models wi-
thin their accuracy and the associated pro-
bability of their correct prediction* Let
us now consider the functioning of indivi-
dual units of the control system*
CONTROL ALGORITHMS
Control algorithms include identification
algorithms and the control algorithms pro-
per.
It takes several identification algorithms
to obtain mathematical models* To develop
original models the apriori experimental
data are processed using the least-squares
technique* The models are then used as ba-
sic for the adaptive identifier* The main
adaptive identification algorithm has the
followingSÎform:
K + )
N+1 % §N % M •
where % is a vector of the plant pa-
rameter estimation, Xjj +<j is a transfor-
med vector of input variables including
disturbances Ζ and controls U ,
i*e. the vector X consists of two sub-
vectors Ζ and U *
In (1) we use the following denominations:
x
where Y is the plant output, Y is the
predicted output (model output)*
The model structure is specified in ad-
x
vance as follows:
y * F ( X F Y, K)« (3)
Transformation of input data in (1)
amounts to their orthogonalization compa-
red to the earlier input data* The ortho-
gonalization improved the convergence of
the adaptive identification algorithms*
particularly for the dynamic plant models
in the control mode* The orthogonalization
increased but little the computing load
thanks to the way the data is handled in
 Control of the Temperature Profile in a Tubular Reactor 325

the algorithm ( 1 ) . The control algorithm relationship between an increase in the

has a hierarchical structure» At the upper average temperature and a higher yield of
level the real input vector is used to the maleic anhydride is given by the con-
choose from the model spectrum the one in ventional formulas of chemical kinetics*
-which the plant output temperature (need-
less to say, only the predicted one) is the The potential economic effect of applying
highest a control system can "be calculated using
the experimental data obtained for the
reactor working without such a system. A
This is done by substituting the vector coefficient of multiple correlation bet-
X in all models and comparing the predic- ween the plant output and measurable in-
ted temperatures. put disturbances is calculated for this
purpose. The relation between the output
At the next stage of control the chosen mo- variance Q| and the residual variance
del is used for the highest risk cross- Q2 is described as
-section. In this model the substitution 2 2 2
of the predicted output y* by the desi- Q = Q (I - R )·
red T Q yields from (3) an equation for
This residual variance determines the ul-
the control action.
timate capabilities of the control system
TQ » Ρ (Z, U, K ) . (5)
with an identifier.
The components of the vector X are given
here explicitly. Prom (5) using the speci- The main identification algorithm is des-
fied disturbance vector and the desired cribed in (Chadeev, 1967), and the main
value of T Q the control U is calculated control loop is described in (Rajbman
with an acceptable probability Ρ of ex- 1978).
ceeding the temperature.

Control Efficiency REFERENCES

Using a control system one can reach a
narrower function of temperature distri-
bution in the operating points
I P(T)
Chadeev, V.M. (1967). Adaptive Models of
Plants. Preprints of the IPAC Symp.,
(Pig. 4 ) . Prague CSSR, Academia - Prague.
Fundamentals of Process Control (1978).
Ed. by Rajbman, U.S., Moscow, Nauka
Publishers.

Pig. 4. Temperature Distribution

Punction with the Control Loop
Involved.

Hence for the same probability of the tem-

perature in the reactor going beyond the
acceptable limits we can increase the
average temperature in the reactor at
least by (<3j;y - ) , where Σ ^ Γ is the
temperature variance in the operating po-
int without the stabilization system,
(jr is the variance in the operating
point with a control system included. The
 AUTHOR INDEX

Adomaitis, R. A. 295 Kapoor, N. 99

Albujo-Cuadrado, M. 105 Karba, R. 231
Aldren, D. B. 67 Karin, N. 215
Andersen, H. W. 175 King, R. 93
Aral, Y. 159 Kozub, D. J. 147, 285
Atanasijevic, M. 231 Kridiotis, A. C. 249
Kummel, M. 175
Kurt-Erik Haggblom, 243
Bonvin, D. 301
Bremsak, F. 231
Brierley, R. J. P. 25 Laggoune, N. 129
Bujac, P. J. 67

MacGregor, J. F. 31, 147, 285

Canavas, C. 273 Marquardt, W. 123
Carling, G. A. 167 Matsimov, Ye. M. 323
Chadeev, V. M. 323 McAvoy, T. J. 99
Chiu-Nan Wei 117 McGreavy, C. 73, 181, 317
Chylla, R. W. Jr. 295 Meerkov, S. M. 153
Cinar, A. 153, 295 Mijares, G. 255
Clement, K. 43 Minambres, J. J. 313
Cole, J. D. 255 Morris, A. J. 203, 209
Moteki, Y. 159

de Haas, K. 135
de la Sen, M. 313 Naka, Y. 261
de M. e Castro, J. A. A. 317 Nazmul Karim, M. 105
Desages, A. 215 Neyran, B. 129
de Valliere, P. 301
f
0 Shima, E. 261
Edwards, J. B. 189, 195 Odloak, D. 73
Ellingsen, W. R. 37
Epple, U. 279
Papadopoulou, S. Ar. 141
Penlidis, A. 147
Fele, L. 231 Piet-Lahanier, H. 307
Preisig, H. A. 255

Gani, R. 105
Georgakis, C 249 Recelj, T. 231
Gilles, G. 129 Retzbach, B. 225
Goldschmidt, L. 267 Rigopoulos, K. 153
Golob, J. 231 Romagnoli, J. A. 215
Gros, S. 61 Rong, G. 87
Guardabassi, G. Ill Rotstein, H. 215
Guilandoust, M. 195 Ruiz, C. A. 105
Guilandoust, M. T. 203

Scattolini, R. Ill
Hallager, L. 43, 267 Schuler, H. 135
Hamielec, A. E. 147 Shakouri, A. 49
Happel, J. 307 Shu, X. 153
Hashimoto, I. 237 Sourander, M. 61
Hashimoto, Y. 237
Heusinkveld, G. H. 55
Holland, C D . 255 Tabriz!, M. H. N. 189
Hsiao-Ping Huang 81 Takamatsu, T. 237
Tan, G. H. 181
Tham, M. T. 203, 209
Jhy-Shyong Chang 81 Thomasset, D. 129
Jorgensen, S. B. 11, 43, 221, 267 Toftegard, B. 43, 221
Togari, Y. 237

327
328 Author Index

Vagi, F. 209 Wang, J. C. 87

van Heusden, A. R. 55 Wood, R. K. 167, 209
Wright, J. D. 285

Waller, Κ. V. 1, 243
Walter, E. 307 Yung-Cheng Chao 81
 SUBJECT INDEX

Adaptive control, 11, 73, 203, 209, 313, 317 Fixed-bed reactor, 273, 279
Adaptive system with an identifier (ASI), 323 Fluid composition control, 189, 195
Adaptive systems, 203 Foaming in distillation columns, 37
Alarm systems, 93 Frequency domain control design, 49
Application of observers, 225 Frequency response, 175, 189, 195, 215, 301
Approximate models, 99 Fuzzy logic controller, 87
Autothermal reactors, 295
Gel-permeation chromatography, 141
Batch polymerization reactor, 87
Batch reactors, 147 Heat pump, 261
Bayes methods, 93 Heterogeneous catalysis, 307
Bilinear control, 129 Hydrodesulfurization, 81

Calorimetry, 301 Identification, 221

Cascade control, 49 Industrial control, 49
Catalytic reactor, 285, 323 Inferential control, 203
Chemical industry, 25, 111, 123, 167, 273 Integrated control system, 49
Chemical reactor control, 153 Interaction, 255
Chemical reactors, 11, 135 Interaction diagram, 221
Chemical variables control, 87, 189, 195 Interaction measure, 249
Closed loop systems, 167 Internal model control, 285, 295
Closed-loop, 105 Isobutene-isobutane, 307
Column dynamics, 181 Iterative methods, 255
Column instability, 37
Column packings, 25 Kalman filtering, 135, 147
Compartmental modelling, 307 Kalman filters, 93, 273
Computer applications, 117
Computer control, 61, 87, 117, 153, 255 Linear quadratic optimal control, 285
Computer-aided design, 111 Load regulation, 55
Continuous reactors, 147
Control system analysis, 1, 117, 167, 175, 295 Maleic anhydride, 323
Control systems synthesis, 1, 55, 237, 243 Mechanic design, 181
Controllability, 25 Membership set estimation, 307
Model reduction, 279, 295
Decoupling, 1, 61, 237, 249 Model representation, 11
Digital control, 129, 209 Model structure, 221
Discrete time model structure, 267 Modelling, 31, 55, 99, 123, 135, 167, 189, 195,
Distillation, 25, 37, 99, 105, 123, 189, 195, 261 243, 273, 279, 301
Distillation column, 37, 117 Models, 61, 261
Distillation column control, 203, 209 Multicomponent distillation, 181
Distillation columns with side stream, 225 Multiple columns, 105
Distillation control, 1, 37, 175, 243, 249 Multivariable PI control, 225
Distributed parameter systems, 81, 123, 273, 279, Multivariable adaptive control, 267
317 Multivariable control, 249
Dynamic analysis, 301 Multivariable control systems, 1, 49, 215, 243, 255,
Dynamic behavior, 135 317
Dynamic compensator, 49 Multivariable-multirate control, 209
Dynamic model, 105 Multivariable systems, 117, 189, 195, 209
Dynamic modeling, 99 Multivariate control, 285
Dynamic quantity, 221
Dynamic response, 25, 81, 117, 167, 261 Nonlinear behavior, 105
Dynamic simulation, 117 Nonlinear control, 31, 153
Dynamics, 11 Nonlinear control systems, 129
Nonlinear filtering. 93, 141, 273
Emulsion polymerization, 147 Nonlinear model, 73
Energy management, 55 Nonlinear systems, 87, 123, 273, 279
Estimation, 203 Numerical analysis, 111, 313
Expert fuzzy control, 87 Nyquist criterion, 189, 195, 255
Extractive distillation, 225
Oil refinery, 81
Fast adaptive algorithms, 313 Oil refining, 61
Fault detection, 93 On-line operation, 61
Filtering, 273, 295 Open-loop, 105

329
330 Subject Index

Optimal control, 11, 31, 129, 317 Semibatch operation, 135

Optimization, 55, 61 Sensitivity analysis, 175
Optimizing control, 11 Sensors, 31
Singular values, 175
Packed bed reactor, 11 Stability, 175
pH control, 129 Stability criteria, 255
pH measurements, 313 Start-up of distillation columns, 225
PID control, 237 State estimation, 135, 141, 147, 301
PID controllers, 73, 261 Step response, 55
Parameter estimation, 93, 243, 301, 317 System analysis, 243
Parameter uncertainty, 307 System order reduction, 123
Partial differential equations, 189, 195, 273 Systems design, 25
Partial model matching, 237
Pattern recognition, 93 Temperature control, 317
Penalty function, 73 Thermal cracking, 73
Periodic control, 153 Thermal lags, 181
Periodic distillation, 221 Time-varying systems, 129
Petro-chemical control, 261 Transfer functions, 99
Polymerization, 147 Transient response, 105
Polymerization reactors, 31, 141 Transient tracing, 307
Prediction, 93 Tray damage, 37
Pressure profile, 181 Trays, 25
Process control, 1, 175, 209, 215, 221, 237, 249, Trickle-bed reactor, 81
261
Profile position control, 225 Varying structures, 93
Vibrational stabilization of nonlinear systems, 153
Quality control, 49, 141
Quality estimation, 49 Waveform analysis, 279
WIener-Hopf, 285
Reactor dynamics, 267
Robust estimation, 307
Robustness, 175
IFAC Publications, Published and Forthcoming volumes
*ADALI & TUNALI: Microcomputer Application in Process
Control.
AKASHI: Control Science and Technology for the Progress of
Society, 7 Volumes
*ALBERTOS & DE LA PUENTE: Components, Instruments and
Techniques for Low Cost Automation and Applications
*ALONSO-CONCHEIRO: Real Time Digital Control
Applications
ASTROM 8c WITTENMARK: Adaptive Systems in Control and
Signal Processing
ATHERTON: Multivariable Technological Systems
BABARY 8c LE LETTY: Control of Distributed Parameter
Systems (1982)
*BALCHEN: Automation and Data Processing in Aquaculture
BANKS 8c PRITCHARD: Control of Distributed Parameter
Systems (1977)
*BARKER & YOUNG: Identification and System Parameter
Estimation (1985)
*BASANEZ, FERRATE 8c SARIDIS: Robot Control
"SYROCO '85"
*BASAR 8c PAU: Dynamic Modelling and Control of National
Economies (1983)
BAYLIS: Safety of Computer Control Systems (1983)
BEKEY 8c SARIDIS: Identification and System Parameter
Estimation (1982)
BINDER 8c PERRET: Components and Instruments for
Distributed Computer Control Systems
BRODNER: Skill Based Automated Manufacturing
BULL: Real Time Programming (1983)
BULL 8c WILLIAMS: Real Time Programming (1985)
CAMPBELL: Control Aspects of Prosthetics and Orthotics
*Van CAUWENBERGHE: Instrumentation and Automation in
the Paper, Rubber, Plastics and Polymerisation Industries
(1980) (1983)
CHESTNUT: Contributions of Technology to International
Conflict Resolution (SWIIS)
CHESTNUT, GENSER, KOPACEK 8c WIERZBICKI:
Supplemental Ways for Improving International Stability
"CHRETIEN: Automatic Control in Space (1985)
CICHOCKI 8c STRASZAK: Systems Analysis Applications to
Complex Programs
CRONHJORT: Real Time Programming (1978)
CUENOD: Computer Aided Design of Control Systems!
*DA CUNHA: Planning and Operation of Electric Energy
Systems
De GIORGIO 8c ROVEDA: Criteria for Selecting Appropriate
Technologies under Different Cultural, Technical and Social
Conditions
DI PILLO: Control Applications of Nonlinear Programming and
Optimization
DUBUISSON: Information and Systems
ELLIS: Control Problems and Devices in Manufacturing
Technology (1980)
ELZER: Experience with the Management of Software Projects MAHALANABIS: Theory and Application of Digital Control
FERRATE 8c PUENTE: Software for Computer Control (1982) *MANCINI, JOHANNSEN 8c MARTENSSON: Analysis, Design
FLEISSNER: Systems Approach to Appropriate Technology
Transfer
*FLORIAN & HAASE: Software for Computer Control (1986)
*GEERING 8c MANSOUR: Large Scale Systems: Theory and
Applications (1986)
GELLIE 8c TA VAST: Distributed Computer Control Systems
(1982)
*GENSER, ETSCHMAIER, HASEGAWA 8c STROBEL: Control
in Transportation Systems (1986)
*GERTLER 8c KEVICZKY: A Bridge Between Control Science
and Technology, 6 Volumes
GHONAIMY: Systems Approach for Development (1977)
HAASE: Real Time Programming (1980)
HAIMES 8c KINDLER: Water and Related Land Resource
Systems
HALME: Modelling and Control of Biotechnical Processes
HARDT: Information Control Problems in Manufacturing
Technology (1982)
HARRISON: Distributed Computer Control Systems (1979)
HASEGAWA: Real Time Programming (1981)t
HASEGAWA 8c INOUE: Urban, Regional and National
Planning—Environmental Aspects
HERBST: Automatic Control in Power Generation Distribution
and Protection
*HRUZ 8c CICEL: Automatic Measurement and Control in
Woodworking Industry — Lignoautomatica '86
ISERMANN: Identification and System Parameter Estimation
(1979)
*ISERMANN: Automatic Control, 10 Volumes
ISERMANN & KALTENECKER: Digital Computer
Applications to Process Control
JANSEN 8c BOULLART: Reliability of Instrumentation Systems
for Safeguarding and Control
JANSSEN, PAU & STRASZAK: Dynamic Modelling and Control
of National Economies (1980)
JOHANNSEN 8c RIJNSDORP: Analysis, Design, and Evaluation
of Man-Machine Systems
*JOHNSON: Modelling and Control of Biotechnological
Processes
*KAYA 8c WILLIAMS: Instrumentation and Automation in the
Paper, Rubber, Plastics and Polymerization Industries (1986)
*KLAMT 8c LAUBER: Control in Transportation Systems (1984)
*KOPACEK, TROCH 8c DESOYER: Theory of Robots
KOTOB: Automatic Control in Petroleum, Petrochemical and
Desalination Industries
LANDAU, TOMIZUKA 8c AUSLANDER: Adaptive Systems in
Control and Signal Processing
*LARSEN 8c HANSEN: Computer Aided Design in Control and
Engineering Systems
LAUBER: Safety of Computer Control Systems (1979)
LEININGER: Computer Aided Design of Multivariable
Technological Systems
LEON HARD: Control in Power Electronics and Electrical Drives
(1977)
LESKIEWICZ 8c ZAREMBA: Pneumatic and Hydraulic
Components and Instruments in Automatic Controlf
MAFFEZZONI: Modelling and Control of Electric Power Plants
(1984)
and Evaluation of Man-Machine Systems (1985)
MARTIN: Design of Work in Automated Manufacturing
Systems
*MARTOS, PAU, ZIERMANN: Dynamic Modelling and Control
of National Economies (1986)
 McGREAVY: Dynamics and Control of Chemical Reactors
and Distillation Columns
MILLER: Distributed Computer Control Systems (1981)
MUNDAY: Automatic Control in Space (1979)
NAJIM 8c ABDEL-FATTAH: System Approach for
Development (1980)
NIEMI: A Link Between Science and Applications of Automatic
Control, 4 Volumes
*NORRIE & TURNER: Automation for Mineral Resource
Development
NOVAK: Software for Computer Control (1979)
OLLUS: Digital Image Processing in Industrial Applications—
Vision Control
O'SHEA & POLIS: Automation in Mining, Mineral and Metal
Processing (1980)
OSHIMA: Information Control Problems in Manufacturing
Technology (1977)
*PAUL: Digital Computer Applications to Process Control (1985)
*PONOMARYOV: Artificial Intelligence
QUIRK: Safety of Computer Control Systems (1985) (1986)
*RAMAMOORTY: Automation and Instrumentation for Power
Plants
RAUCH: Applications of Nonlinear Programming to
Optimization and Controlf
*RAUCH: Control of Distributed Parameter Systems (1986)
RAUCH: Control Applications of Nonlinear Programming
REMBOLD: Information Control Problems in Manufacturing
Technology (1979)
RIJNSDORP: Case Studies in Automation related to
Humanization of Work
*RIJNSDORP, PLOMP & MÔLLER: Training for T o m o r r o w -
Educational Aspects of Computerized Automation
RODD: Distributed Computer Control Systems (1983)
RODD & MULLER: Distributed Computer Control Systems
(1986)
*ROOS: Economics and Artificial Intelligence
*SANCHEZ: Fuzzy Information, Knowledge Representation and
Decision Analysis
SAWARAGI 8c AKASHI: Environmental Systems Planning,
Design and Control
SINGH 8c TITLI: Control and Management of Integrated
Industrial Complexes
*SINHA 8c TELKSNYS: Stochastic Control
SKELTON 8c OWENS: Model Error Concepts and
Compensation
SMEDEMA: Real Time Programming (1977)t
*STRASZAK: Large Scale Systems: Theory and Applications
(1983)
SUBRAMANYAM: Computer Applications in Large Scale
Power Systems
SUSKI: Distributed Computer Control Systems (1985)
SZLANKO: Real Time Programming (1986)
*TAL': Information Control Problems in Manufacturing
Technology (1986)
TITLI & SINGH: Large Scale Systems: Theory and Applications
(1980)
*TROCH, KOPACEK 8c BREITENECKER: Simulation of
Control Systems
UNBEHAUEN: Adaptive Control of Chemical Processes
*VALADARES TAVARES 8c EVARISTO DA SILVA: Systems
Analysis Applied to Water and Related Land Resources
*WANG PINGYANG: Power Systems and Power Plant Control
*WESTERLUND: Automation in Mining, Mineral and Metal
Processing (1983)
van WOERKOM: Automatic Control in Space (1982)
*YANG JIACHI: Control Science and Technology for
Development
*YOSHITANI: Automation in Mining, Mineral and Metal
Processing (1986)
*ZWICKY: Control in Power Electronics and Electrical Drives
(1983)

*IFAC Proceedings Series

"\Out of stock—microfiche copies available. Detaih of prices sent on request from the IFAC Publisher.

IFAC Related Titles

BROADBENT 8c MASUBUCHI: Multilingual Glossary of Automatic Control Technology
EYKHOFF: Trends and Progress in System Identification
ISERMANN: System Identification Tutorials (Automatica Special Issue)