You are on page 1of 322


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)

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.

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





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

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


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
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


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


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-

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

2 Κ. V. Waller

The interaction can be eliminated or reduced by in- structure as has Rosenbrock's scheme in Fig. 2.
serting compensators, much like feedforward con- Bequette and Edgar state that the pairing for
trollers, between the two primary feedback loops. their column should not be the same as suggested
The approach is called (external) decoupling in by Ryskamp (1982).
the distillation control literature and has been
shown much interest since Luyben's (1970) early pa- Different design methods to obtain non-interacting
per. control were studied by Bequette and Edgar and
they all resulted in essentially the same scheme,
A decoupling scheme is shown in Fig. 6. To decou- i.e. one using the sum and the difference between
ple the two feedback loops, the decouplers have to two temperature measurements as inputs to the con-
be chosen in such a way that they counteract the trollers. The methods studied were (1) column pro-
interaction caused by the two process elements G-j2 file control, (2) implicit decoupling, (3) modal
and G 2 1 · The decouplers in Fig. 6 then become control, (4) output decoupling, and (5) extensive
-G21/G22 and -G12/G1V variable control (Georgakis, 1986). It should be
noted, however, that the study aimed at a control
The decoupling scheme of Fig. 6 has the structure system for a column where the manipulators were
of a multivariable control system where each of chosen to be D and V.
the two outputs are connected to each input. In
standard multivariable control theory the system A multivariable structure where both outputs are
is treated as a whole when the controllers are de- connected to both inputs is obtained also by other
signed. This has the implication that interaction combinations of measured variables from the two
in the system that might be beneficial for disturb- ends of the column. In an attempt to get linear-
ance rejection is taken advantage of in the design. ization in addition to non-interaction, Weber and
In the decoupling approach the basic idea is to Gaitonde (1982) use the conventional manipulators
eliminate all interaction, be it detrimental or reflux flow L and vapor boil up V to control cer-
not, and make the system behave as two isolated tain combinations of top and bottom compositions.
SISO-loops. It has been stated that decoupling con- The variables controlled are called
trol is the opposite to multivariable control.
fractionation = χ
Rosenbrock's and Davison's suggestions in Figs 2
β+ K(1-x )
and 4 can be compared with the decoupling scheme which is controlled by reflux L, and
in Fig. 6. Rosenbrock's scheme can be drawn as in χ χ
Fig. 7. Fig. 7 can be redrawn as in Fig. 8 to show cutpoint = (1- [))/ β
the structural equivalence between Rosenbrock's
scheme and the decoupling scheme of Fig. 6. which is controlled by boil up V.
Davison's scheme of Fig. 4 can analogously be re-
drawn as in Fig. 9. x
Weber and Gaitonde calculate the constant Κ from
1 X
the normal operating point as B ^ " D O ^ *
A comparison between Rosenbrock's and Davison's 0
schemes as plotted in Figs 8 and 9 and the decoup- McAvoy (1983) discusses Weber's and Gai tonde's
ling scheme of Fig. 6 shows that the elements in scheme and states that a "better", i.e. less in-
Rosenbrock's and Davison's schemes that correspond teracting, choice for Κ would be Xp/(1-Xp).
to the decouplers in Fig. 6 are not directly based
on the properties of the process (only indirectly McAvoy (1983) suggests a scheme where a variable
through the controllers and their tuning). It ξ defined as
therefore seems that they cannot generally be ex-
pected to decouple the loops and make the system ? = +X X
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. is controlled by distillate flow D.
6 and to emphasize that all three schemes can be
structurally viewed as general multivariable con- McAvoy state that use of ξ as a controlled varia-
trol systems where all (in this case two) outputs ble results in an almost one-way steady-state de-
are connected to all inputs, as in Fig. 7. coupled system, i.e. changes in V do not signifi-
cantly affect ξ . The relationship between the con-
Also Ryskamp (1982) discusses schemes where one trolled and the manipulative variable is further
temperature above the feed plate and one below the stated to be linear (McDonald and McAvoy, 1983).
feed are used as the two measurements from which
product composition is inferred. Ryskamp suggests Instead of ξ as a controlled variable, resulting
a scheme where heat input is set by the sum of the in an almost one-way steady-state decoupled sys-
temperatures and the difference between them is tem, a similar variable could be used giving per-
used to set reflux. fect one-way steady-state decoupling when the ma-
nipulators are D and V. This variable is (Haggblom
Structurally Ryskamp's scheme resembles Rosen- and Waller, 1986)
brock's and Davison's schemes. It differs, however,
ξ= x
in two respects. Firstly, Rosenbrock and Davison B+ K'x
use distillate flow D as a manipulator in addition
to boil up V, whereas Ryskamp use reflux flow L in where the constant K' can be calculated from
addition to V. Secondly, the temperature difference steady-state compositions or flows as
X x D
is used to set boil up in Rosenbrock's and Davison's
schemes,1 whereas the sum of temperatures is used in F o - B ox _ o
Ryskamp s scheme. Β
Do " F o
Recently Bequette and Edgar (1986) used so called In McAvoy's scheme (McAvoy, 1983, McDonald and
singular value analysis to design a control scheme McAvoy, 1983) the second controlled variable is
for a simulated column. A structurally non-inter- separation factor S, which is controlled by boil up
acting system was obtained when distillate flow D V. The separation factor is defined as
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
Distillation Control System Structures 3

B A list of ratio control schemes suggested in the

χ 1-x
D literature is given in the book by Rademaker,
S = - - Rijnsdorp, and Maarleveld (1975). The book discus-
1-x x
D B ses various ratio-control schemes, among them also
This loop is not decoupled from the other one, two-ratio schemes, in which both the manipulators
changes in D affect the separation factor S. are flow ratios.

Separation factor control is also discussed by One of the schemes recently most discussed where a
Shinskey (1984). It could be added that Boyd (1975) flow ratio is used as manipulator is a scheme sug-
and Ya and Luyben (1984) also have discussed con- gested by Ryskamp (1980). The manipulators for
trol schemes where temperature differences, in composition control in the scheme are D/(L+D) and
their schemes differences between differences, are V.
used as controlled variables.
The rationale for the scheme is expressed by Rys-
It is interesting to note the similarities and dif- kamp (1982) as follows. The scheme "holds reflux
ferences between the schemes discussed above. For ratio constant if the top AC output is constant.
symmetrical separations, i.e. D = B , 1-xrj = Χβο> An increase of heat input from the bottom AC does
0 Q 0
and x p = 0 . 5 , the controlled variables and mani- not make top product as impure as would occur with
pulators become the ones shown in Table 1. reflux constant (conventional control) nor as over-
pure as would occur with distillate flow constant
(material balance control)". Thus, this property
CHOICE OF CONTROL VARIABLES of the scheme results in a certain decoupling ef-
fect and the scheme is often said to result in
In the previous section were reviewed various sug- "implicit decoupling", in contrast to "explicit
gestions for transformation of controlled varia- decoupling" accomplished by external decoupling
bles mainly to make the two control loops noninter- elements, discussed in the previous section and
acting for dual-composition control of distilla- later in this paper treated more in detail.
tion. 1
A modification of Ryskamp s scheme to a scheme
At the other end of the control system there are where the manipulators are D/(L+D) and V/B has
the control variables or manipulators used to con- been suggested by Takamatsu, I. Hashimoto and Y.
trol the process. They, too, can be chosen in var- Hashimoto (1982) (1984) and by Shinskey (1984).
ious ways, and this is also a subject much discus- 1
sed in the recent distillation control literature. An implementation of Ryskamp s scheme can be il-
lustrated as done in Fig. 10 for the top of the
For feedback composition control of a distillation column.
column with two products there are the following
four primary control variables: distillate flow D, An important difference between the manipulators
reflux L, bottoms flow B, and vapor boil up V (the in the scheme of Fig. 10 and the schemes previ-
last one indirectly manipulated through heat input ously discussed with D, L, V, and Β as manipula-
to the reboiler). tors should be noted. In the schemes where two of
the basic manipulators D, L, V, and Β are used to
The manipulators in the so called conventional or control composition, the other two are used for
energy balance scheme are L and V. They are the level control (as illustrated e.g. by Fig. 4 ) .
manipulators in Fig. 1, whereas D and V are the These level control loops have not been considered
manipulators in Rosenbrock"s scheme in Fig. 2 and in the discussion above, in order to simplify the
Davison's scheme in Fig. 4. Schemes where either D treatment. However, strictly speaking, the control
or Β are manipulated to control composition are problem discussed is concerned with control of
usually referred to as material balance control four outputs using four manipulators, i.e. a 4x4
schemes. problem.

McDonald and McAvoy (1983) use the "material bal- In Ryskamp's (1980) scheme in Fig. 10 both the m a -
ance variable, D or B", to control a linear combi- nipulators L and D at the top end of the column
nation of product compositions (see previous sec- are simultaneously used for control of both compo-
tion). The "energy balance variable", V or L, is sition and level, i.e. each output is connected to
used to control the separation factor. each input, as illustrated in Fig. 11, which shows
the structure of Ryskamp's scheme in block diagram
Between the two base cases of energy balance con- form.
trol and material balance control there are a num-
ber of combinations. Many suggestions to use vari- If, in analogy with the case discussed above, the
ous ratios between flows as manipulators can be bottoms composition x and the reboiler level are
found in the literature. An early example is given
controlled by a flow ratio, like V/B, this means
by Rijnsdorp (1965), who suggests the ratio of re- that both these outputs are connected to both the
flux flow and top vapor flow as a manipulator for inputs V and B.
the top loop (L/(L+D) «-x ) (here the arrow point-
ing backwards denotes feedback). Stainthorp and The various schemes using various flow ratios as
Jackson (1974) experimentally studied a scheme in manipulators have as a rule been obtained by heu-
which the top loop manipulator was L/D. For a num- ristic reasoning. Obviously they have been found
ber of simulated columns McAvoy (1977) studied the advantageous in industrial practice. Theoretically
steady-state interaction also for Rijnsdorp's sug- this is not unexpected: they have the structure of
gestion, both in combination with (V-*-x ) and - as an optimal system where each output is connected
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 still better control can be anticipated if other
smallest amount of steady-sxate interaction for algorithms than pure division between flows were
the last scheme, (L/(L+D) X Q ) ( B ^ - X ) . He also used.
extended the idea in search of such functional com-
binations of manipulated variables as would make Furthermore, the four outputs and four inputs have
the loops noninteracting, the final result was de- been split into two groups. Still better control
generacy (Jafarey and McAvoy, 1978). 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 LQ-design results in a scheme where all the
the column. state variables are used for feedback. The size of
the feedback coefficients may, however, differ con-
siderably. By neglecting small feedback coeffi-
EXTERNAL DECOUPLING cients, Oakley and Edgar (1976) concluded for a
specific example analyzed that the resulting
Inserting compensators between interacting loops, scheme could be approximated with good accuracy by
as shown in Fig. 6, is usually called (external) a scheme having the same structure as a one-way
decoupling in the distillation control literature. decoupled scheme, where changes in reboiler heat
This subject has attracted much interest in the duty are fed to the top loop, not only through the
literature ever since Luyben's paper in 1970 column but also through the decoupler. The same
(Luyben, 1970), and it was discussed in connection conclusion is obtained in a study by Tung and
with Rosenbrocks's and Davison's schemes (see Figs Edgar (1978).
6, 8, and 9 ) . Below some complementary aspects are
given. In chemical process control, the number of sensors
is usually much lower than the number of state
The decoupling scheme as drawn in Fig. 6 is actual- variables used to describe the process. In that
ly one of several possible decoupling schemes. case an optimal multivariable (LQ) controller con-
Which one to use is usually determined by realiza- tains an observer or state estimator, by which the
bility aspects (Waller, 1974). unmeasured states are calculated before they are
used for feedback.
Two-way decoupling (where there are two decouplers
between the two primary loops, as in Fig. 6) and An LQ-design can very well start from a system
one-way decoupling (only one decoupler) are com- model consisting of simple transfer functions con-
pared by Fagervik, Waller, and Hammarstrbm (1983). taining dead times. An illustration of the various
Results obtained in the study also indicate that, steps of the design starting from experimentally
in general, the best disturbance rejection is not obtained transfer functions is given in Hammar-
obtained by perfect decoupling, but that the best strbm, Waller, and Fagervik (1982). The paper is
response is obtained by some tuning of both decou- focused on how errors in the process models affect
plers and feedback controllers. the control properties of the multivariable con-
trol system obtained. Connections between model
We are then left with a parametric optimization structure, the performance index, the control
problem of a considerable size. Even if the types quality, and the sensitivity are illustrated.
of the feedback controllers and the form of the de-
couplers are decided upon, the problem of simulta- One further aspect on state estimation may be men-
neous tuning of "decouplers" and feedback control- tioned here, since it has to do with the structure
lers is still considerable. It is further desira- of the state estimator. When there are disturb-
ble that various decoupling structures (such as ances with non-zero mean, such as step disturb-
one-way versus two-way decoupling in a 2x2 system) ances, estimation e.g. by a Kalman-fliter results
can be tried, as well as various forms for the de- in steady-state estimation offset if a model of
couplers. the deterministic disturbances is not included in
the estimator. Also the resulting control quality
Use of the Inverse Nyquist Array (INA) technique may be drastically reduced, compared to the case
for simultaneous feedback controller and "decoup- when a model of the disturbances is included in
ler" tuning is demonstrated by Waller, Wikman, and the estimator. In an distillation example treated
Gustafsson (1985). The paper also shows that mini- by Hammarstrbm and Waller (1974), the control was
mizing interaction at the critical frequency often, improved by several orders of magnitude simply by
but not always, is a good criterion for design of including estimation of an occuring step disturb-
the decouplers. It is also shown that it may be ance in feed composition. It should be noted that
enough to use pure gains as decouplers; adding dy- this improvement was accomplished not by any added
namics need not significantly improve control sensors or disturbance measurements, but by
quality. changing the structure of the estimator so that
the control system was informed about the fact
It should further be emphasized that although the that there might be disturbances of step type.
control quality, as measured e.g. by error inte- Estimation of more deterministic disturbances
grals after step disturbances, is not very differ- (such as steps both in feed composition and flow)
ent for different decoupling schemes, the robust- than actually occured (in feed composition) did
ness of the scheme may significantly speak in fa- not degrade control quality. However, not more
vor of one of the schemes, as is quantitatively il- disturbances can be estimated than there are meas-
lustrated by Fagervik, Waller, and Hammarstrbm urements on the process (Hammarstrbm, 1980).
(1983). Further illustration of LQ-control in distillation
can be found in Hammarstrbm (1980).


In the previous treatment it was found that many
of the schemes for dual-composition control dis- The number and location of sensors are important
cussed in the literature and there treated as mul- issues directly related to the control system
tiloop SISO systems can be viewed as ΜΙΜΟ (multi structures.
input multi output) systems.
Waller, Gustafsson, and Hammarstrbm (1974) illus-
An alternative would then be to start from the trated, through simulation, how drastic an effect
structure of the system (be it 2x2, 3x3, or 4x4) the location of a third sensor (two fixed in the
and use a true multivariable design method and cal- product lines) may have on distillation control
culate the controller by some optimization proce- quality. Also when state estimation (Kalman-fli-
dure, instead of determining part of the control- ter) was used (Hammarstrbm and Waller, 1974) the
ler by more or less heuristic choices. One such location of the third sensor strongly affected
method is the so-called linear quadratic (LQ) de- control quality in LQ-design. This could easily be
sign. A number of simulation studies of LQ-design understood if the measurements were inaccurate,
have been reported, reviews can be found in Edgar but in the study mentioned perfect measurements
and Schwanke (1977) and Waller (1982). were assumed.
Distillation Control System Structures 5

A frequent approach in the literature for sensor compositions to desired values? This last question
location is to use some kind of observability or was recently discussed by Georgakis (1986).
sensitivity consideration. In recent literature
so-called singular value analysis has been popular. Among interesting problems for future research in
distillation control can be mentioned the question
Already in 1974, Hammarstrbm and Waller (1974) of sensor number and location (observability con-
studied the relation between control quality and siderations do not give the whole picture) and
observability properties of a distillation system. their relations to the complexity of the process
No correlation was found between the observability models used for design. Here the combination of
index used and control quality. Observability con- various types of sensors is an interesting and
siderations are probably not enough to solve the important subproblem.
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- NOTATION
lection of measurements to be used in optimal feed-
back control of linear stochastic systems.Applica- Β bottoms flow
tion of the algorithm to a distillation system D distillate flow
gave results which were reported to be "consider- F feed flow
ably different from what would be expected for L reflux flow
pure estimation" (Mellefont and Sargent, 1978). 0 as index denotes steady state
Τ temperature
The location and number of sensors are also direct- V boil up
ly related to the optimal size and complexity of X composition
the process model used for controller design.
Dahlqvist (1980) investigated, both experimentally
and through simulation on an 11-plate pilot column, REFERENCES
LQ-controllers based on various model sizes. The
measured states were obtained thorugh state esti- Bequette, B.W., and T.F. Edgar (1986). The equiv-
mation. Using two sensors, a second order model alence of non-interacting control system de-
gave better results than did a sixth order model. sign methods in distillation. Proc. ACC,
With three sensors, a sixth order model gave Seattle, WA, USA, 31-38.
better control than a 13th order model did, and Boyd, D.M. (1975). Fractionation column control.
also better control than a fourth order model gave. Chem. Eng. Progr. 71 No 6, 55-60.
Dahlqvist, S J \ . (1980). Control of a distillation
The relation between sensor number and location column - application of sampled data control
and process model size and complexity in multivar- to a large pilot plant. Tech. Dr thesis,
iable control is likely to be a function of the Royal Institute of Technology, Stockholm,
quality of the measurements, such as accuracy and Sweden.
dynamics. These relations are also functions of Davison, E.J. (1967). Control of a distillation
the properties (both static and dynamic) of the column with pressure variation. Trans. Instn
process to be controlled, as well as of the prod- Chem. Engrs 45, T229-T250.
uct specifications. Edgar, T.F., and C O . Schwanke (1977). A review
of the application of modern control theory
Another closely related problem, not much studied to distillation columns. Proc. JACC,
so far, is how to combine slow and direct (desir- 1370-1376.
able) measurements (like those of a gas chromato- Fagervik, K.C., K.V. Waller, and L.G. Hammarstrbm
graph) with fast and indirect ones (like tempera- (1983). Two-way or one-way decoupling in dis-
tures). tillation? Chem.Eng.Comm. 21, 235-249.
Georgakis, C. (1986). On the use of extensive
variables in process dynamics and control.
DISCUSSION Chem.Eng.Sci. 41, 1471-1484.
Gustafsson, T.K. (1984). A study of modeling and
Dual-composition control of distillation has been control of pH in fast acid-base reaction proc-
treated emphasizing the structure of various con- esses based on chemical reaction invariance.
trol schemes. Many suggested SISO schemes have Tech. Dr thesis, Abo Akademi, Finland.
been shown to be structurally equal to a general Gustafsson, T.K., and K.V. Waller (1983). Dynamic
multivariable scheme. For the problem studied, it modeling and reaction invariant control of pH.
then seems that the structural question could Chem.Eng.Sci. 38, 389-398.
equally well be viewed as an algorithmic question. Haggblom, K.E., and K.V. Waller (1986). In prepa-
ration. Abo Akademi, Finland.
So far only feedback control strategies have been Hammarstrbm, L.G. (1980). An investigation of
treated. For distillation feedforward control is multivariable control strategies for high-
often used to cope with changes in feed flow rate. order processes with few sensors. Report 80-7,
If feed composition can be measured, feedforward Process Control Lab., Abo Akademi, Finland.
can be used also to compensate disturbances in Hammarstrbm, L.G., and K.V. Waller (1974). State
feed composition. In this way the two feedback estimation in multivariable ΡΙ-control. Report
loops normally used in dual composition control 74-10, Process Control Lab., Abo Akademi,
could be replaced by one feedforward loop and one Finland.
feedback loop, as suggested by Jafarey and McAvoy Hammarstrbm, L.G., K.V. Waller, and K.C. Fagervik
(1980). (1982). On modeling accuracy for multivariable
distillation control. Chem.Eng.Comm. 19,
An issue that seems to be, at least partly, open 77-90.
when controlling various (linear) combinations of Jafarey, Α., and T.J. McAvoy (1978). Degeneracy of
concentrations (or temperatures) instead of the decoupling in distillation columns. I&EC Proc.
product compositions themselves is the following. Pes.Dev. 17, 485-490.
Does good control of a combination of concentra- Jafarey, Α., and T.J. McAvoy (1980). Steady state
tions (or temperatures) also implay good control feedforward control algorithms for reducing
of the product concentrations which usually are energy costs in distillation. ISA Trans. 19,
the important variables to control? What about ro- No. 4, 89-100.
bustness? How should e.g. integral action in the
control system be implemented to bring the product
6 Κ. V. Waller

Lau, H., J . Alvarez, and K.F. Jensen (1985). Syn- Takamatsu, Τ,. I. Hashimoto, and Y. Hashimoto
thesis of control structures by singular value (1982). Multivariable control system design of
analysis: dynamic measures of sensitivity and distillation columns system. Proc. PSE 82,
interaction. AIChE Journal 31, 427-439. Kyoto, Japan, Tech.Session 243-252.
Luyben, W.L. (1970). Distillation decoupling. AIChE Takamatsu, T., I. Hashimoto, and Y. Hashimoto
Journal 16, 198-203. (1984). Dynamic decoupler sensitivity analysis
McAvoy, T.J. (1977). Interacting control systems: and its application in distillation control.
steady state treatment of dual composition con- Preprints 9th IFAC World Congress, Budapest,
trol in distillation columns. ISA trans. 16, Hungary, Vol. Ill, 98-103.
No. 4, 83-90. Tung, L.S., and T.F. Edgar (1978). Analysis of
McAvoy, T.J. (1983). Interacti on an alys i s. ISA control-output interaction in dynamic systems.
McDonald, K.A., and T.J. McAvoy (1983). Decoupling 71th annual meeting AIChE, Miami.
dual composition controllers. 1. Steady state Waller, K.V. (1974). Decoupling in distillation.
results. Proc. ACC, San Francisco, USA, AIChE Journal 20, 592-594.
176-184. Waller. K.V. (1980). Current problems in process
Mellefont, D.J., and R.W.H. Sargent (1978). Selec- control. Paper and Timber 62, 128-132.
tion of measurements for optimal feedback con- Waller, K.V. (1982). University research on dual
trol . I&EC Proc.Pes.Dev. 17, 549-552. composition control of distillation: a review.
Oakley. D.R., and T.F. Edgar (1976). Optimal feed- In Seborg, D.E. and T.F. Edgar (Eds) Chemical
back control of a binary distillation column. process control 2, Engineering Foundation/
Proc. JACC, 551-557. AIChE 395-412.
Rademaker, 0., J.E. Rijnsdorp, and A. Maarleveld Waller, K.V., and P.M. Makilâ (1981). Chemical
(1975). Dynamics and control of continuous dis- reaction invariants and variants and their use
tillation units. Elsevier, Amsterdam_0xford- in reactor modeling, simulation, and control.
New York. I&EC Proc. Pes.Dev. 20, 1-11.
Rijnsdorp, J.E. (1965). Interaction in two-variab- Waller, K.V., S.E. Gustafsson, and L.G. Hammar-
le control systems for distillation columns. strbm (1974). Incomplete state-feedback inte-
Automatica 3, 15-51. gral control of distillation columns. Proc.
Rosenbrock, H.H. (1962). The control of distilla- Kern. Tek 3, Copenhagen, Denmark, Paper C.8.
tion columns. Trans.Instn.Chem.Engrs 40, Waller, K.V., K.E. Wikman, and S.E. Gustafsson
35-53. (1985). Decoupler design and control system
Ryskamp, C.J. (1980). New Strategy improves dual tuning by INA for distillation composition con-
composition column control. Hydrocarbon proc- trol. Chem.Eng.Comm. 23, 149-174.
essing 59, No. 6, 51-59. Weber, R., and N.Y. Gaitonde (1982). Non-inter-
Ryskamp, C.J. (1982). Explicit versus implicit de- active distillation tower control. Proc. ACC,
coupling in distillation control. In Seborg Arlington, VA, 87-90.
D.E. and T.F. Edgar (Eds), Chemical process Ya, C.C., and W.L. Luyben (1984). Use of multiple
control 2, Engineering Foundation/AIChE 361- temperatures for the control of multicomponent
375. distillation columns. I&EC Proc.Pes.Dev. 23,
Shinskey, F.G. (1984). Distillation control. 2nd 590-597.
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.
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
T + T
Ryskamp (1982) V
1 2
Weber and Gaitonde (1982) L
D " +B
McAvoy (1983) x x (= ξ ) D
T D +B T
Bequette and Edgar (1986) D
T1 T2
1 " 2
Haggblom and Waller (1986) D+ χ
Β ( = ξ ' )

D x,

En 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.


-K> c2 >22

Distillation Control System Structures 9

'ci '11

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

Control system Process


Fig. 7. Rosenbrock's scheme.


Control system
4 3



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

Rosenbrock's scheme.

Control system Process

22 -6
10 Κ. V. Waller

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


0.6K ,

Control system


D / (L+D)

Fig. 10. Implementation of control

scheme suggested by Ryskamp

^D. set
Controller 1 χ


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


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.


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

12 S. B. J0rgensen

purpose defined as (MacGre gor et al. 1983): application of control designs for fulfilling the
primary and some additional control objectives
Manual: All but the most fundamental regulatory i.e. designs for stabilization, disturbance rejec-
loops are in manual control. The operator sets tion and set-point tracking. This section pre-
actuators and flow setpoints. This level may sents first linear quadratic control design and
have an advisory computor system. then reviews selection of the number and location
of sensors. The largest part of section 5 contains
Single Loop: The basic measurements are under a review of the applied control designs mainly
single loop control. Such loops may require grouped according to the applied model representa-
advancxed techniques as adaptive or nonlinear tion. Section 6 contains a brief review of optimi-
control. zing control. Finally the main conclusions are
Multi Loop: Coordinated yet simple interaction
occur between two or more loops, such as cascade,
feedforward, interaction override or compensation. 2. FIXED-BED REACTOR CONFIGURATIONS

Supervisory: Control coordination occur between Catalytic gas-phase chemical reactions play an
data from several sources. This coordination in- important role in chemical industry, and has also
cludes handling filtered data, e.g. from process found important applications in pollution abate-
analyzers and executing multivariable control, ment of exhaust gases. Contacting of the reactant
e.g. model based. gas with the catalyst is mainly effected in reac-
tors where the catalyst is fixed and to a much
Optimizing: Plant steady state or dynamic behavior smaller extent in fluid beds. The various types of
is evaluated to optimize a profit criterion. beds with fixed catalyst employed in present day
reactors cover a wide spectrum. A large fraction
Solution of the basic control objective can on of these reactors may be described by one or a
some fixed-bed reactors be possible using a single combination of some generic types:
loop, which however may be non-minimum phase, and
therefore model based control is advantageous. a) A single tubular bed with axial reactant flow.
However often additional requirements are imposed
such as keeping the reaction zone position and/or b) A single annular bed with radial reactant flow.
the maximun temperature. In these cases it may be
necessary or advantageous to use more than one c) A channel with wall deposited catalyst.
input variable. Therefore fixed-bed reactor con-
trol problems tend to be multivariable, and to be The type a bed is the traditional catalytic bed.
approached using modelbased control designs. Con- Type b beds have developed to make it possible to
sequently most of the control designs to be dis- use large amounts of catalyst with small diameter
cussed subsequently have been carried out at the particles without getting excessive pressure
supervisory level. drops. Type c, which clearly is not a packed bed,
have been developed for handling exhaust or emis-
The present paper aims at reviewing the control sion gases, with still lower pressure drops. The
design methods which have been applied on fixed- channel geometry may vary as well as the support
bed reactors to fulfill one or more of the control construction material. Type c is used with many
objectives, at pinpointing limitations of the parallel channels, whereas type a and b beds may
techniques; and thereby indicating possible direc- be applied as single beds. These single beds may
tions for future research. approach adiabaticity if well insulated and/or of
large diameter (type a) or length (type b ) . How-
A procedure for fulfillment of a particular set oj ever in many cases external cooling or heating is
control objectives using a model based approach desirable or necessary to reduce parametric sensi-
may be achieved through the following steps: tivity. In these cases two types of configurations
may be used or possibly combined:
i) Investigation of process statics and dynamics
ii) Selection of model representation. d) Multiple parallel beds (type a) which are ex-
iii) Selection of identification and control ternally cooled (heated) in co- or countercur-
strategy. rent flow.
iv) Selection of number and location of sensors.
v) Process identification either off- or on-line e) Multiple consecutive beds (type a or b) with
vi) Control design for stabilization, disturbance interstage cooling (heating), e.g. using a
rejection and set-point tracking. quench stream.
vii) Optimizing control design.
The feed stream in exothermic processes is often
The following sections contain a discussion of used for cooling purposes through heat exchange in
most of these steps as they have been dealt with type d and between type e beds or as a quench
in the available investigations. First some typi- stream between type e beds. In these cases auto-
cal fixed-bed reactor configurations will be clas- thermal operation may be possible. In some proces-
sified in section 2. Section 3 starts with discus- ses (partial) recycle of effluent gas is used,
sing fixed-bed rector models and summarizing ob- e.g. to obtain full conversion of an expensive
served static behaviors. Then follows a presenta- reactant.
tion of the basis for the following treatment
which is the underlying process dynamics with The beds whose dynamics and control have been
special focus upon experimental investigations and investigated in the laboratory have mainly been of
demonstrated abilities of the applied mathematical type a and d; and in a few cases in configura-
models. Among these abilities has the locally tions with autothermal opeartion or with reactant
linear behavior been extensively exploited. Sec- recycle. Some of the experimental investigations
tion 4 covers process identification and starts are summarized in tables 2 and 3, where the inve-
with a presentation of locally linear model repre- stigations are grouped according to the bed type
sentations, then model order reduction procedures or configuration. The term cooling is used in the
are touched upon since these techniques may be tables mainly to indicate that the cooling tempe-
important in many cases. Thereafter techniques rature possibly could be manipulated. It is clear
applied for parameter estimation,which may use from these tables that many of the reactions are
either off- or on-line data acquisition, are re- only model reactions, which are used to elucidate
viewed briefly. Section 5 contains a review of 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
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
) 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 )
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
# spatia
l dim.
A. Pseudo
one phase
Β. Two phas
8C 1 d C
F R(C,T) (1)
one 1. Plu
g Flo
1. Externa
l resistanc
e onl
y at Pe dz dz
one 2. Axia
l dispersio
2. Externa
l and interna
l res
two 3. Axia
l and radia
l Externa
l heat and 2
n interna
l mas
s resist
* 3T I A Τ
(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- Experimental investigations in fixed bed reactor
tion for the same operating conditions and nonpa- dynamics are summarized in Table 2. There are two
thological rate expressions, can be demonstrated sets of investigations; those concerned with the
theoretically for all models, except A-l, in Table dynamics around steady states (entries 2,3,6-8,10,
I, . This possibility is due to interaction of and 13-16) and those concerned with the dynamic
axial dispersion and reaction. This subject has behavior when transition between steady states
been treated in the above reviews in particular by occur (entries 1,4,5,9, and 12). The latter set of
Hlavacheck and van Rompay (1981) and Jensen and investigators have focussed on CO and ehtane oxi-
Ray (1982). They both review experimental findings dation as model reactions. The main results of the
of multiple stady states. In the cases where mul- latter group of investigators are that transition
tiple steady states may occur, it is possible to between multiple steady states occur as propaga-
have high sensitivity of the reactor conditions ting fronts, where the front shape changes very
towards changes in inlet conditions or in parame- little and the speed of propagation is very low
ter values. Cases of high parametric sensitivity compared to the thermal wave speed. This behavior
occur near bifurcation of steady states i.e. where can in general be simulated reasonably using A-2
multiplicity of steady states shows up. Puszynski models. In some instances it appears nescessary to
et al.(1981) show using an A-3 model that mul- account for radial velocity profiles (Vortmeyer
tiple steady states may occur in nonadiabatic and Winther 1982). Puszynsky et al. 1982 observed
packed beds for strongly exothermic reactions even oscillatory behavior, which could not be simulated
when the Peclet number is large, as in most indu- using an A-2 model.
strial reactors. For the pseudo-one phase axial
dispersion model with a first order irreversible The main results of the first group of investiga-
reaction the sufficient conditions for uniqueness tions concerning the local behavior around a stea-
show that the solution will be unique for suffi- dy state are that the observed local phenomena can
ciently high values of the unequal Peclet numbers, in general be modelled quite well using pseu-
large heat transfer coefficient or small values of do-one phase models with dispersion or two phase
the Damkohler number. It has been shown computa- models. An exception is the study by van Doesburg
tionally that the region of multiplicity is enlar- and de Jong (1976), where however the wall heat
ged and shifted towards lower Damkohler numbers by capacity had a major influence. The basic dynamic
increasing adiabatic temperature rise or the acti- features of a type a bed can bee qualitatively
vation energy. An increase in the reaction order outlined using the DYQUID in figure 1. The signi-
reduces the region of multiplicity. In the adiaba- ficant difference between the mass and thermal
tic reactor a multiplicity of three is possible, residence times has a decisive influence upon the
whereas more steady states can be found in the dynamic behavior. Inverse response, or 'wrong
nonadiabatic fixed-bed reactor. Jensen and way' behavior (Crider and Foss 1966), have been
Ray(1982) investigate theoretically the bifurca- observed upon inlet temperature and flow rate
tion behavior especially for the pseudo-one phase disturbances. The presence of a concentration node
axial dispersion model. Their results show a point (Crider and Foss 1966, 1968) has also been
wealth of bifurcations at low Peclet numbers. observed in several cases. Locally linear models
However at Peclet numbers relevant for industrial give a reasonable description of the local dyna-
packed beds it is concluded that bifurcations to mics and are therefore obvious candidates for use
periodic solutions should not occur. in model based control designs.

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

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
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?
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
given by Chylla et al. (1986) for an autothermal Autotheraal
reactor. 15. Water-gas Wong, Bonvin, MeHichamp and pulse A-2
shift Rinker (1983)

Jergensen, Bortolotto, step T,e, A·2
16. :iidat.
and Kaas· Petersen F
Fixed Bed Reactor Dynamics and Control 15

4. PROCESS IDENTIFICATION be modelled as filtered white noise. In the pre-

sentation above the disturbances have been assumed
Process identification is concerned with obtaining to be stochastic. In some cases are the distur-
process models and the necessary estimates from bances predominantly deterministic. In that case
experimental data, in the present context with the the above model representation also can be used,
purpose of control synthesis. Below follows a but then e(t) represents a sequence of impulses
presentation of locally linear model representa- that occur irregularly but relatively seldom.
tions and of model reduction techniques whereby
the model structures may be obtained. Thereafter One input-output model representation, which by
estimation of model parameters is treated. tradition is popular, is the frequency response
representation; or in the discrete time case the
Locally linear model representations pulse transfer matrices, which may be written:
The linear models to be used for control design y(t) - Giq' ) u(t) + v(t) (12)
represents generally either internal states i.e.
are in state space form or process inputs and where stationary noise may be modelled as:
outputs only, i.e. are in input-output form. One
state space and two input output forms will be v(t) - IKq" ) e(t) (13a)
1 1
presented below.
where Giq" ) and HCq" ) are rational transfer
A state space representation may be obtained from function matrices, which may be expressed as
the partial differential equations (1-7) by appli- polynomials of the general form:
_ 1 1 2
cation of the quasistationarity assumption to the
X(q ) - X + Xjq" + X q " + ... (14)
gas phase component balance, through linearization q 2
around a steady state profile and by use of ortho-
gonal collocation to discretize the axial coordi- where G - 0 and H - I.
ο ο
nate, as shown for the state space model by Mi-
When the transfer function representation is ap-
chelsen et al (1973). The resulting model may be
plied the linear transfer functions may be deter-
mined directly from input-output data using multi-
variate time series analysis techniques, e.g.
Α χ + Β u + Β e (8) spectral analysis as described in Box and Jenkins
dt (1970).
Where χ is the n-dimensional vector of state vari-
ables, u the m-dimensional vector of controls, Another input-output model representation is the
and e a zero mean white noise vector process with ARMAX model:
intensity matrix Q. The measurements may be col-
1 1 l
lected in a ρ-dimensional y-vector, where ν is a Piq" ) y(t) = CKq" ) u(t) + R(q~ ) e(t) (15)
zero mean white noise vector process with inten-
sity matrix : where u and y are input and output vectors as
above, and e is discrete time white noise. P, Q
y = Cx + Du + v (9) an<jl R are matrix polynomials in the delay operator
q , of the general form:
The use of LaGrange interpolation (Villadsen and 1 1 2 Π
Michelsen 1978), makes it possible to interpolate Xiq' ) - X + X^' + X q" + . . . + Χ ς* Χ (16)
from the collocation points to arbitrary axial
q 2 χ
measurement positions with high accuracy. The where Ρ - R - I whereas Q = 0 . The left hand
second term on the right in eq.(9), which repre- side term in eq. (15) is an autoregression of the
sents direct transmission from the inlet to the outputs, the noise term on the rigth hand side
measurements is due to the quasistationarity as- represents a moving average and the input is exo-
sumption and the global LaGrange interpolation. genous .
The above two equations are shown in continuos
time form, whereas practical application usually
occurs at discrete points in time, due to the The three model representations given above can be
discrete time action of the computer. The measure- shown to be input-output equivalent for stable
ments are available at the sampling instants: plants. For unstable plants predictors based upon
the three representations can be shown to be
y(t) - C x(t) + D u(t-l) + v ( t ) (10) equivalent. The transfer function representation
d reflects only the controllable and the observable
where time is normalised with the sampling inter- part of the process. Whereas the state space and
val (=1) for convenience, and the discrete time the ARMAX models may include unobservable and
noise covariance is V^. The corresponding dis- uncontrollable modes. The state space and the
crete time state space model is obtained by inte- ARMAX models may be transformed into nonunique
grating eq.(8) over the sampling interval (Âstrom cannonical forms, i.e. one which is observable,
1970, Sorensen et al 1980), giving: called left matrix fraction description in the
ARMAX case, and one which is controllable, called
x(t+l) - A rigth matrix fraction description in the ARMAX
dx(t) + B
du(t) + e (t+1) (11)
with A
d - A d( l ) _; A1 (t) - exp(At), and
The noise in the above representations has been
B - ( A - I ) A B , e (t+l) -JjA(l-t')B e(t')dt' assumed stationary. Most often the noise in chemi-
Qd= d ,B
t B Ad( t d t n cal processes contains a nonstationary component,
which may be observed as drift in the inputs and
d ioV ) nQ n d '> ' outputs. This nonstationarity may be removed by
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 using the incremental value y(t) - y(t-l). This
model in eq.(ll) is that it is based upon the filtering may however introduce a zero on the unit
continuous time noise description. Thus any subse- circle in the noise polynomial R(q" ) in the ARMAX
quent design is potentially less dependent upon case. An alternative procedure is to model the
sampling time than designs starting with a dis- noise as nonstationary, e.g. by differencing the
crete time noise with a constant covariance. The noise. In this case eq.(13a) may be written:
B^ matrix is included to allow some freedom in the
1 _ 1
disturbance modelling, i.e. the disturbances may ( 1 - q " ) ! v(t) - H ( q ) e(t) (13b)
16 S. B. J0rgensen
model development seems to preclude industrial use
The resulting model representation contains an
of mechanistic models.
integrated mowing average (IMA) of the noise.
MacGregor and Wong (1980) use an empirical trans-
Model order reduction fer function approach, where the input-output
characteristics are described with an autoregres-
Various types of model order reduction techniques sive integrated mowing average (ARIMA) model, i.e.
have been applied on fixed-bed reactors. Some of eq.(12 and 13b). They then determine a minimal
these techniques will be briefly discussed. state space realization. Kozub et al.(1986) also
use an ARIMA model. In both cases the parameters
The number of significant eigenmodes in the axial are determined offline using a multivariate maxi-
dispersion model eq.(l-7), is very high for most mum likelihood method.
practical values of the axial Peclet number. It Harris et al.(1980) use black box ARMAX models and
turns out however by application of orthogonal determine the parameters on-line, with recursive
collocation that a limited number of collocation least squares. A similar procedure is used by
points gives a satisfactory description of the bed McDermott et al.(1986a,b), where the approximate
dynamics. Thus application of orthogonal colloca- model orders are detemined using simulation stu-
tion renders by itself a very useful model order dies. A different approach to using ARMAX models
reduction technique when a single bed is modelled. were taken by Hallager et al. (1981,1983 and 1984)
In several cases, e.g. when wall or environmental and Jergensen et al. (1985). These investigators
dynamics have to be modelled further model order use a procès knowledge based approach to specify
reduction may be desirable. Silva et al.(1979) ap- only the unknown parameters in ARMAX model struc-
plied Davison (1966)'s method to reduce a 14'th tures. This usage of parsimonious model structures
to a 7'th order model. Bonvin and Mellichamp makes it possible to obtain relatively fast con-
(1982b) compare various model order reduction tech- vergence to new parameter values during on-line
niques, and in Bonvin et al.(1983c) they apply operation, when necessary.
Litz (1979a,b)'s method to reduce their 24'th to a
5'th order model. Bonvin and Mellichamp (1982b)
The fixed-bed reactor identification studies de-
demonstrate that information may be obtained
monstrate a clear tendency in that the early stu-
about the structural dominance characteristics.
dies use mainly mechanistic based, state space
This information may be applied advantageously to
models, which were fairly manpower demanding to
compare structural aspects of different reactor
develop and require nonlinear regression methods
designs as used by Chylla et al (1986). Rivera and
for obtaining parameter estimates. Whereas later
Morari (1985) have developed a frequency domain
work has concentrated around developing and iden-
model reduction procedure that yield near optimal
tifying models at a significantly lower manpower
models with respect to user designed closed loop
cost. These models were either transfer function
based models, where the parameters are obtained
off-line using spectral analysis, or ARMAX models
A different approach to model order reduction is which are linear in the parameters which therefore
taken by Hallager and Jergensen (1981,1983) and can be estimated on-line.
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 5. PROCESS CONTROL DESIGN
setup a discrete time model structure of an ARMAX
model. The model structure can include the possi- This section starts with a brief summary of linear
bility of variable flow rate, i.e. of variable qudratic control design including state recon-
process delay. The advantage of this procedure is struction. Thereafter follows a brief discussion
that the resulting model structure is parsimo- of sensor selection and location for control pur-
nious , which speeds up convergence of parameter poses . Then follows a summary of the investigated
estimation and that the model structure is valid control designs. This major part is grouped accor-
over a relatively wide range of operating condi- ding to the applied model representations, star-
tions, when the model parameters are estimated on- ting with the mechanistic model based control
line. designs, followed by the transfer model based and
Epple (1986) uses the interpretation of the finally the adaptive control designs which mainly
reaction zone as a moving wavefront to propose a are ARMAX model based.
nonlinear model reduction technique, where the
reactor response to inlet or cooling temperature Most of the fixed bed reactor control experiments
changes at an arbitrary location may be described listed in Table 3 have been using linear qudratic
by a sixth order model. (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
Parameter estimation is to minimize (for stochastic disturbances: the
expectation of) the objective function:
The early model based fixed bed reactor control
Nv T T T T
studies used mechanistic models wherein the para- 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
meters either were estimated using offline data
from the reactor (Jutan et al.1977b), or were where Ν is the time horizon of the optimal con-
based mainly, but not entirely, upon prior infor- trol, and η can be the tracking error y - y ,
mation (Serensen 1977), or were based upon a com- the process output y or another linear combina-
bination thereof (Wong et al. 1983), and (Silva et tion of states and inputs :
al. 1979). In the latter case a few parameters
were updated prior to the experimental run . Jutan n(t) = C x(t) + D u(t) (18)
η η
et al.(1977b) collected their data under univa-
riate closed loop operation, presumably due to using the state space model and the knowledge that
nonstationary character of the stochastic distur- the controls are constant over a sampling period
bances, using an external dither added to the eq.(17) may be written:
feedback signal. The remaining investigators used N-1A
χρΤA T A
open loop experiments as a basis for parameter J„ =7 ( x+2u F x+u F u)+n (N)F n(N) (19)
C 4V»Q χ ux u oc
estimation. It is noteworthy that in no case were
the model parameters derived entirely from inde- where F , F , and F are integral functions of
pendent experiments. It is a conclusion in several the weighting matrices in eq. (17). These inte-
of these early works (Jutan et al. 1977, Wallmann grals are evaluated over the sampling interval,
et al. 1979) that the effort necessary for the and depends also upon the process dynamics.
Fixed Bed Reactor Dynamics and Control 17

Most often an objective function of the following The sensor selection and location problem differ
form is applied: in principle according to the type of feedback
J N-l( n nT +F 2 nu +T uuF) T +F used. If output feedback is used then the sensor
types are dictated by the control objective and
D -^ nd und ud ^ W F ^ ) (20) their location by the location of the control
This function may also be brougth into the form of zone. If sensors are not available or only can
eq.(19), but the weigthing matrices are then sim- provide measurements less frequent than desired
ple functions of the weigthing matrices in eq. then an estimate may be evaluated based upon se-
(20) and the process dynamics. Differences between condary measurements. The latter application was
the two approaches can thus be expected when there demonstrated by Tremblay and Wrigth (1977). In
are significant variation in the process state their case the predictor parameters were updated
over the sampling interval. Such differences may on-line when the less frequent concentration mea-
be reduced by reducing the sampling interval. surements became available.

The minimization of either form of eq. (19) re-

sults in a control which is a linear function of When a detailed process model is available in the
the states or the state estimates. The control form of a state space model then it is possible to
gain matrix is in general time varying, but often use a state reconstructor to reduce the effects of
the stationary solution obtained for Ν approaching process and measurement noise, and to provide
infinity is used. The calculation of the feedback estimates of unmeasured states for feedback. Here
reveals that it is desirable to have access to the the sensors should be positioned such that the
full state vector, which is rarely possible in model is observable, which usually is not a pro-
fixed bed reactors. Instead state reconstruction blem in fixed bed reactors due to the predominant
can be accomplished. In the case of Gaussian dis- convective behavior, and such that optimal esti-
turbances it is possible to determine the optimal mates are obtained for the control objective. This
estimator, which minimizes the variance of the one problem have been investigate by several groups
step prediction error. This procedure gives the e.g. Mellefront and Sargent (1978) and Harris et
Kalman gain matrix Κ to be used in the Kalman al.(1980). When state reconstruction is used then
filter: in addition unmeasurable disturbances may be esti-
mated. The minimum number of sensors necessary to
observe external disturbances is equal to the
x ( t + y t ) - A x(tyt-l)+B u(t)+K(y(t)-y(tyt-l)) (21)
d d number of linearly independent external disturban-
ces (Morari and Stephanopoulos 1980, and Jcrgensen
where y(t^t-l) - C x ^ t - l ) + Du(t-l)
et al. 1984). In the latter study it is recommen-
ded to locate temperature sensors just upstream of
The filter gain matrix Κ is in general time va-
the control zone for estimation of distributed
rying, but most often the stationary filter gain
acting disturbances. If the disturbance is a boun-
have been used.
dary forcing to a delay ( e.g. inlet temperature )
then it is important to obtain a measurement as
The theoretical basis for using Linear Quadratic
close to the disturbance entry point as possible.
control is the validity of the separation theorem
Generally it is advisable to use redundant infor-
for linear process models. This implies that the
mation for both state and disturbance estimation.
optimal control strategy can be split into two
It was however demonstrated by Joseph and Brosilow
parts: one which estimates the states from the
(1978a,b) that a high number of sensors increases
measured outputs and one which computes the linear
the sensitivity towards modelling error. Alvarez
feedback from either the estimated or the measured
et al.(1981) and Romagnoli et al.(1981) propose to
states. The obtained closed loop is guaranted to
use a variable location of the sensors according
be stable, provided the process is linear. However
to the present operating conditions, and suggest a
fixed-bed reactors are not linear, therefore an
fast algorithm for this purpose. Kuruoglu et al.
important issue is how to tune the linear quadra-
(1981) develope a steady state distributed filter
tic control, i.e. to select the control weigthing
for estimation of concentration and temperature
matrices in eq. 17 or 20, in order to obtain a
profiles, and a dynamic filter for estimation of
control design which is robust against parameter
slow catalytic activity changes. The combined
variations and against possible variations in
filters are simulated using the Clough and Ramirez
disturbance characteristics. Although the specific
(1976) fixed-bed reactor model for styrene produc-
form of the objective functions in eq.'s (18 and
tion by dehydrogenation of ethylbenzene. They
20) does not have a physical background the LQ(G)
demonstrate for the profile estimation that inter-
design approach have been widely applied in fixed-
nal measurements allow for subdivision of the
bed reactor control studies. One reason for this
spatial domain and thus yield improved profile
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. Mechanistic model based control designs

Sensor location Vakil et al.(1973) used the continuous time LQG

approach on a simulated adiabatic fixed bed reac-
In control system design both selection of the tor, with the parameters of the Sinai and Foss
sensor types and their position are design vari- (1970) liquid phase packed bed reactor. The state
ables. In practice however there may not be reli- space model is derived using orthogonal colloca-
able on-line sensors available for measurement of tion in space. The control problem was to keep the
product compositions. Similarly for load distur- reactor close to the steady state operating condi-
bances, whether they are of internal origin e.g. tions using manipulation of the temperature or
catalyst activity variations or of external ori- concentration of a secondary reactant stream in-
gin, they may not be feasibly measured e.g. due to jected partway along the bed to compensate for
inhospitality of the environnement. Therefore upstream disturbances in the feed concentration
variables of secondary interest with regard to the and temperature. A temperature measurement at the
purpose of control may often be measured and used injection point is used as the only output. Thus
in a state reconstructor to estimate the vari- the sensor and the control input locations were
able (-s) of primary interest for the control ob- coupled. This work inspired a number of subsequent
jective. In this section mainly the principal studies in several ways. Cinar (1984) studied the
aspects of sensor location for control will be location of the sensor and control input in a
discussed. A few particular examples will be men- similar two bed gas phase fixed-bed reactor, using
tioned and other are covered in the discussion of the parameters of the Hoiberg et al. (1971) gas
control designs. 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
: Experimenta
l Investigation
s of fixed-be
d reacto
r estimatio
n and co disturbances entering the second bed. A fairly
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
Ο ί trol of the effluent concentration from the second
. Buchholt,Clement.Jergenaen(1979
O c
) e
a off . KF LQ
Ο ί
bed, based upon an estimate of this concentration
Clement and Jergense
n (1981
) aa off . IK
Ο c from measuring four temperatures in the second
Hallager et al.(1981,1983,1984
r bed (Wallman and Foss 1981). In the adjoining
and Jergense
n (1985
, 1986
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
Tremblay and Wrlgth (1977
) a a off
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
O< integral LQG control is achieved using the objec-
Kozub and MacCregor (1986
) F T off
tive function J in eq. 20 with η including incre-
3. Butan
e Lee and Lee (1985
O ο
mental states and the selected states to be inte-
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
Ο fi
Oxldat . Wallman , Silv
a and Foaa (1979
) sa off
Ο s concentration and temperature controller operated
Wallman and Foes (1981
) as off
O fi
satisfactorily when four temperatures were mea-
Autothermal: sured in the second bed. Using only one tempera-
5. Water- Wong, Bonvin
, Mellicham
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
O ο +
ΜΙΜΟ ο + quench flow is avoidance of control input satura-
tion. Quench flow is effectively used for rapid
6. H
o and Jergense
n (1985
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 19

three-input multivariable control system for the Transfer function based design
two-bed reactor system described above with the
objective of regulating the product concentration MacGregor and Wong (1980),table 3.2 use an off-line
and temperature. Dynamic compensators are deve- identified transfer function model developed from
loped such that interaction among variables is production rates which are inferred using a linear
suppressed and the effects of concentration di- regression on current temperatures and flowrates.
sturbances are controlled. The control is effi- The regression coefficients are updated every time
cient, but uses about eigth times the control a gas chromatograph measurement is available
effort needed for concentration control alone due (Wright et al.1977). The transfer function model
to a near singular system matrix. It is concluded with nonstationary noise is transformed into a
by the authors that proper design of dynamic com- minimal order state space realization. The resul-
pensators relies heavily on the designers experi- ting fourth order model is used in a LQG design to
ence. Clement and Jorgensen (1981), Table 3.1 control the production rates of propane and bu-
investigated the direct Nyquist array (DNA) design tane. The objective function is eq.(20) with η
with constant pre and post compensator matrices. containing the two outputs and using incremental
The control objective was to maintain the steady controls of hydrogen and butane to cope with non-
state profile, when the reactor was upset by up- stationary noise. The resulting control perform
stream disturbances. The controls utilized were well in rejection of a coolant temperature load
inlet concentration and temperature. However a disturbance. Drift is observed in the manipulated
relatively poor behavior was obtained for the variables possibly due to catalyst activity chan-
controlled reactor. This seemed to be connected to ges.
the thermal delay, therefore a delay compensation In their recent study Kozub et al. (1986), table
was investigated, however without much success 3.2, use an off-line identified empirical trans-
(Terndrup et al.(1983)). fer function representation similar to that of
MacGregor and Wong (1980). Kozub et al. (1986)
Economu and Morari (1986) develop and investigate transform their model into a right matrix fraction
through simulation a multiloop IMC design ap- description, which is convenient for control de-
proach. The method treats interaction terms as sign. They investigate two different control de-
additive pertubations and uses an interaction signs: LQ output feedback control using eq.(20)
measure, which is closely related to the Gershgo- with η being the tracking errors and incremental
rin bands of DNA, to determine either a reasonable controls; and deterministic internal model control
configuration or to decide for a full multivari- (IMC). IMC may be tuned to achieve satisfactory
able IMC design (Garcia and Morari 1985a and b ) . performance and robustness towards model plant
The multiloop method gives a rather good design mismatch using diagonal filter matrices on the
for the Foss et al. (1980) problem. Whereas a feedback and setpoint signals. The manipulated
multivariable design seems most promising for a variables were setpoints to the hydrogen flowrate
methanation reactor where outlet and hotspot tem- and to a hotspot temperature controller. The hot-
perature is controlled using inlet flow rate and spot temperature is controlled by butane flow
temperature as manipulated variables. rate.The results indicate that both designs were
restricted by model plant mismatch, due to nonli-
nearities and catalyst activity changes. Trial and
The frequency domain design techniques employed
error adjustment of the tuning matrices led to
thus far rely on a trial and error procedure for
robust designs with acceptable regulation and
determining simple compensators. In the multiloop
setpoint change performance over a fairly wide
IMC design approach only the univariate control-
operating region.
lers are detuned using simple filters to account
for interactions and the performance can be di-
rectly evaluated.
Adaptive control design
Modal design has also been investigated in a coup-
le of instances. Georgakis et al.(1977b,c) inve- A number of studies have been concerned with inve-
stigate the stabilization of a cooled tubular stigating various adaptive control algorithms for
reactor at an unstable steady state, using coolant fixed bed reactor control. The main purpose has
flow as manipulated variable and measuring tempe- been to reduce the control sensitivity towards
ratures and concentrations at the collocation model plant mismatch, using input-output data to
points. Modal design is demonstrated computa- provide information for on-line tuning of the
tionally to give a fairly large domain of attrac- control. Two main approaches to adaptive control
tion when the two largest eigenvalues are relo- have been used. In one case are the controller
cated. It is also shown that an observer can be parameters estimated directly using an implicit
used to estimate the control variable directly model, in the other case is model parameters ex-
from temperature measurements only. Wong et plicitly estimated, and then used to calculate the
al.(1983), Table 3.5 use modal design with a re- control. This general area has recently been re-
duced order model to stabilize the autothermal viewed by Seborg et al.(1986).
reactor at an unstable state manipulating feed
stream heat input. The number of states in the Much of the early theoretical development was
reduced order model were measured directly, The carried out by the group in Lund, Sweden, where
control design performed well in disturbance re- the univariate selftuning regulator was described
jection both with a fifth and a third order re- by Âstrom and Wittenmark (1973) and Àstrom et al
duced model. (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-
The problem of robustness of control design is variable case. Clarke and Gawthrop (1979) derived
treated by Mandler et al. (1986) in a computatio- an implicit controller which minimize the univa-
nal study of univariate control of a cooled metha- riate objective function in eq.(20) with a one
nation reactor operating close to ignition. The step time horizon. This controller is often called
inlet temperature is used to control the product a generalised minimum variance controller (GMV). A
temperature. The control design is based upon a multivariable version of the implicit GMV control-
reduced order model. Internal model control (Gar- ler was developed by Koivo (1980) for an equal
cia and Morari 1982a) is used for control design number of inputs and outputs (square plants) and
coupled with Doyle's (1982) structured singular by Grimble (1981) for nonsquare plants. An impli-
value analysis, whereby a tradeoff between robust cit univariate LQG controller has been developed
stability and performance is accomplished. 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 results were satisfactory. In subsequent works has
al.(1979) and Àstrom and Wittenmark (1980). Prager the control design been improved to an an assymp-
and Wellstead (1980) proposed a multivariate ver- totic LQ design, which has been tested on the
sion. These algorithms are mainly of the explicit fixed-bed reactor under more and more demanding
model type. conditions. The control design utilizes the objec-
tive function in eq.(20) with
The first attempts at implementing adaptive con- T T T
trol design used model reference adaptive control n(t) = (y (t-l),..,y (t-n l),u (t-2), .
(MRAC). Here the objective is to make the output T T p T
of an unknown plant assymptotically approach that .,u (t-n -l),z (t-l)) (22)
of a specified reference model. The initial at-
tempts lacked theoretical background. The MRAC where z(t) contains integral outputs, if included.
approach can be shown to be similar to the selftu- The control horizon is set to infinity and the
ning regulator in a number of cases. Tremblay and Riccati equation is iterated by direct substitu-
Wright (1977), Table 3.2, performed an early inve- tion once at each sampling instant. The feedback
stigation of MRAC. They used a state space model gains will converge to the optimal values when
for the extents of the reaction augmented with the the parameters are constant. The feedback gains
production rate of ethane and propane. The slecti- will vary relatively slowly when the model parame-
vity of propane was to be controlled using the ter values change. The weigths in the control
flow rates of hydrogen and butane as manipulated performance criterion are chosen using available
variables. The model reference algorithm was based knowledge about the limitations of the model
upon Liapunov's direct method. The results were structure in order to obtain robustness towards
not satisfactory. However in the course of this unmodelled dynamics. Hallager and Jergensen (1981)
study a linear regression model was developed for investgate this design experimentally using incre-
the effluent concentrations as a function of three mental inputs in eq.(20) to achieve offset elimi-
functionals of the reactor temperature and the nation in a least squares sense, since there are
inlet flow rates. This regression provided predic- maximally two inputs: inlet temperature and con-
tions for the concentrations at every sampling centration and five outputs. Effective rejection
instant. The regression coefficients were updated of upstream pulse disturbances in temperature,
using recursive least squares (RLS) estimation concentration and flowrate is demonstrated when
when actual concentration measurements were avai- the reactor is operated at low inlet concentra-
lable at every 12'th sample (Wright et al.1977). tion. In Hallager and Jergensen (1983) is the
Harris et al.(1980), Table 3.2, apply univariate model structuring expanded to be able to handle
self-tuning control in various forms to control disturbances in flow rate, i.e. the thermal delay;
the hotspot temperature, independent of position, and to include integral action on specific out-
manipulating butane flow rate. A GMV controller puts. The ability to reject an upstream tempera-
with incremental control signal provided conside- ture disturbance is demonstrated to be satisfacto-
rably better control than a conventional PI con- ry and to be nearly independent of sampling time,
troller. In a recent study Underwater et al.(1986) provided the model structure is modified accor-
dingly, when the distributed inlets also are used
have used a GMV controller to tune the same loop
as manipulated variables. Hallager et al.(1984)
using a logarithmic transformation of the hotspot
investigate various identification aspects using
temperature. The input weight in eq.(20) was used
RELS with variable forgetting factor and show in
to provide a trade off between acceptable hotspot
particular that the adaptive controller is able to
stabilization and smooth variation of the butane
reject disturbances in upstream temperature and
flowrate also at medium high oxygen concentration.
The parameter estimation, which is RELS with va-
McDermott et al.(1986a), Table 3.5 apply a uni- riable forgetting, is relatively rapidly able to
variate pole-placement controller with the abili- estimate varying parameters after a flow rate
ty to auto-tune a pole location on-line. They change. Goldschmidt et al.(1986), table 3.1, inve-
control a temperature just upstream of the hotspot stigate the servo behaviour obtained through a
manipulating the heat input to the entering reac- first order filtered change of the integral refe-
tant gas. Good disturbance rejection behavior is rence states. During a sequence of setpoint chan-
demonstrated both at an open loop unstable and at ges the reactor is brought from almost no reaction
an upper stable steady state and similarly for up to a very active reaction, with a high thermal
setpoint changes between the two operating condi- gain. The adaptive control behaves very well and
tions. The good setpoint behavior is obtained by is able to follow the parameter changes even
on-line adjustment - auto-tuning - of a desired though the process in addition is slowly time-
pole-position by minimizing the error variance η varying due to catalyst activity changes.
- y - y (McDermott and Mellichamp (1984)). In
the adjoining paper (McDermott et al.(1986b)) a
multiloop version of the pole-placement controller Bortolotto and Jergensen (1985), investigate di-
is applied to maintain the temperature profile sturbance rejection and set point tracking of a
with two loops : in one the temperature near the reactor with recirculation of unreacted reactants,
hotspot is controlled by the top heater (i.e. at using the same control design as above. The mani-
the turning point of the gas), in the other loop pulated variables are inlet temperature to the
a temperature near the exit is controlled by mani- reactor and limiting reactant mass flow rate to
pulating either - in simulations - the heat input the recirculation loop. The outputs are five
to the entering reactant gas or - in an experi- temperature measurements and the exit reactant
mental test - the gas flow rate . A dynamic de- concentration. In this case the control weights in
coupler is used to minimize loop interactions. The the objective function were selected carefully in
model structures, i.e. the polynomial orders η ,n order to account for model structure limitations.
and n^, were chosen based upon simulation explri- Disturbance rejection is effective and a setpoint
ence. The above configuration behaved well in change which brings the reactor close to a Hopf
simulations and also in an experimental test bifurcation point is also very well handled.
around the upper stable state.
Both the transfer function based and the ARMAX
based adaptive control designs lead relatively
Buchholt et al.(1979), Table 3.1, investigate quickly to on-line implementation. Especially the
disturbance rejection of the adiabatic fixed-bed process knowledge based model structuring proce-
reactor using inlet temperature as manipulated dure seems promising in this regard. Most of the
variable with five equidistant axial temperature early control design approaches include a step,
measurements. The control design was GMV and RLS which often can be of trial and error type, where
was used for parameter estimation. The control the designers artistic abilities often play an
Fixed Bed Reactor Dynamics and Control 21

important role for a succesful result. The on- that linear(-ised) mathematical models can des-
going development of how to apply control design cribe the dynamic behavior in a reasonable neigh-
methods such that they are robust against known borhood arouund an operating point. In cases with
inaccuracies in model and/or disturbance charac- very high sensitivity it may be necessary or ad-
teristics holds promise for reduction if not eli- vantageous to use nonlinear dynamic models.
mination of artistic elements in this design step.
Three somewhat different types of linear model
representations have been applied in identifica-
6. OPTIMIZING CONTROL tion studies on fixed-bed reactors. The most pro-
mising approach for practical application seems to
Long term optimal operating conditions for fixed- be ARMAX models with parsimonious model structures
bed reactors may be difficult to determine due to where the parameter to be estimated are selected
timevarying process characteristics such as cata- based upon knowledge of qualitative process dyna-
lyst deactivation and the associated modelling mics . The advantage of this approach is that the
difficulties, which persist even though sinificant resulting model structures are valid over a wide
progress have been achieved in this area. There- range of operating conditions and that the rela-
fore it is not desirable to apply open-loop opti- tively low number of parameters may be determined
mizing control polices to fixed-bed reactors. Due relatively fast on-line.
to the time-varying physics this is an obvious
area for application of on-line identification
techniques. All three model types and various identification
methods have been successfully applied in control
Hong and Lee (1985) propose on line computation of studies. The sucessful control design methods have
a measure of catalyst activity using measurement predominantly been based upon LQ(G) designs where
of inlet temperature, exit concentration, and for both the primary but also additional control ob-
a cooled bed also of coolant temperatures. They jectives have been investigated. Modal design have
apply their measure of catalyst activity for eva- mainly been used for satisfying the primary design
luation of the desired quasi steady inlet tempera- objective. The mechanistic modelbased control
ture to maintain the product concentration within design approaches have in some cases reported low
a specified range. Thus their control objective robustness towards varying parameters or modelling
could presumably be met with techniques from the errors. The recent design procedures which addres-
previous section which also can address the prima- ses the problem of providing robust performance
ry control objective. Lee and Lee (1985), table and stabilization represent a promising develop-
3.3, study optimization by simulation and expe- ment towards reducing the amount of trial and
rimentally on a non-adiabatic fixed-bed reactor error necessary for multivariable control design.
for the production of maleic anhydride using coo- It is noteworthy that sequences of stepchanges
lant temperature and flow rate as manipulated covering a relatively large operating region have
variables, with a temperature near the hot spot mainly been reported for robust control and for
and the inferred product concentration as out- adaptive control methods. Thus these two methods
puts. The temperature near the hot-spot was con- seem most promising for practical application.
trolled in an adaptive loop by the coolant tempe-
rature with a feedforward from flow rate. Assuming Optimizing control has been demonstrated to work
this control to be perfect a second order Hammer- well in one case, where adaptive control was used
stein model was assumed for maleic anhydride con- to satisfy the primary control objective. This
centration as a function of hot-spot setpoint and result is also a most promising development.
flow rate. This model is a quadratic function of
the input variables, but linear in the parameters. A few attempts have been made towards developing
These parameters are determined using recursive a methodology for carrying out reactor design
instrumental variables. The steady state relation- while accounting for poosible control problems
ship from this model is used in a profit function: 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.
where the first right hand side term is the net
profit per unit time and the second a penalty term In summary the area of fixed-bed reactor control
on high bed temperatures. The maximization of J is have matured significantly and practical applica-
locally constrained to the region where the iden- tions of advanced modelbased control techniques
tified model is valid. The search direction is seem most likely. However several aspects deserve
determined, in the experiments at every 100'th additional theoretical and experimental studies.
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-
A,Β Matrices in state space model
strated to perform well in two experimental opti-
C Concentration
mization runs.
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.
The experimental investigations of fixed-bed reac- M Thermal residence time to heat transfer
tor dynamics and control have mainly been carried time constant ratio,
out with type a beds in various laboratory reactor η ^ vector variables in control objective.
configurations, where the Reynolds number often is
P(q ) Autoregression matrix polynomial.
significantly lower than in industrial practise.
Pe^ Peclet number (axial)
The dynamic behavior of fixed-bed reactors is
q ^ Backward shift operator.
qualitatively well understood. The dynamic beha-
Q(q ^) Input matrix polynomial.
vior can be quantitatively modelled using conser-
R(q ) Noise matrix polynomial.
vation principles with reasonable accuracy when
R(C,T) Reaction rate
the reaction kinetics are known. However a number
Τ Temperature
of theoretical modelling issues remain unsolved.
u vector of inputs (controls),
Theoretical results indicate that axial thermal
ν measurement noise vector,
dispersion also is important for the dynamics of
χ state vector,
industrial fixed-beds. It has been demonstrated
y output vector.
ζ axial distance, integral output.
22 S. B. J0rgensen

subscripts Clement.Κ.;Jergensen,S.Β. and Serensen,J.Ρ.

d discrete time 1980 "Fixed-Bed reactor Kalman Filtering and
h heat Optimnal Control-II: Experimental Investigation
m mass of the Discrete Time Case with Stochastic
η noise or output function Disturbances".Chem.Engng.Sci.35,1231-1236.
u input Clement,K.;Jergensen,S.Β. 1981 :"Experimental
Investigation of a Fixed Bed Chemical Reactor
Greek letters Control System Designed by the Direct Nyquist
β adiabatic temperature rise Array Method" Proc.triennal Congr. IFAC 2743-50
7 activation energy Pergamon Press
Clement,K.;Jergensen,S.B. 1983 "Experimental
REFERENCES Investigation of Axial and Radial Thermal Dis-
persion in a Packed Bed".Chem.Engng.Sci.38,835-
Ahrensberg,A., Goldschmidt,L.; Clement,Κ.; Seberg, 842.
H.; Jergensen,S.B. 1983 :"A Note on Discrete Clough,D.E.; Ramirez,W.F. 1976 : "Mathematical
Time Optimal Control of Continous Time Systems" Modelling and optimization of the Dehydrogena-
Proc.IASTED Symp. Applied Control and Identifi- tion of Ethylbenzene to form Styrene". AIChE
cation ρ. 2.6 - 2.10, Copenhagen Journal 22 1097-1105.
Alvarez,J.; Romagnoli,J., Stephanoupolos,G. 1981 : Crider,J.E.; Foss,A.S. 1966 :"Computational Stu-
"Variable Measurement Structures for the Control dies of Transients in Packed Tubular Chemical
of a Tubular Reactor" Chem.Eng.Science 36 Reactors" AIChE J^ 12 514-522
1695-1712 Crider,J.E.; Foss,A.S. 1968 "An Analytic Solu-
Âstrom,K.J. 1970 Introduction to Stochastic Con- tion for the Dynamics of a Packed Adiabatic
trol Theory, Academic Press Chemical Reactor" AIChE Journal 14, 77-84.
Âstrom,K.J.; Wittenmark.B. 1973 :"On Self-Tuning Danckwerts, 1953 :"Continuous Flow Systems"
Regulators",Automatica 9 185 Chem.Eng.Sci. 2 1-13
Âstrom,K.J.; Wittenmark,B. 1980 :"Self-Tuning Davison,E.J. 1966 "A method for simplifying linear
Controllers Based on Pole-Zero Placement" IEE dynamic systems" IEEE Transac. Aut. Control AC
Proc. 127,D 120 11 93-101
Âstrom,K.J.; Borisson.U.; Ljung,L.; Wittenmark,B. Denn,M.M. 1969 :"Chemical Reaction Engineering -
1977 :"Theory and Applications of Self-Tuning Optimization, Control, and Stability" I.&E.C.
Regulators" Automatica 13 457-476 61no.2, 46-50
Bonvin,D. ; Rinker.R.G.; Mellichamp,D. 1980 .'"Dy- Dixon,A.G. and Cresswell,D.L. 1986 ."Effective
namic Analysis and Control of a Tbular Autother- Heat Transfer Parameters for Transient Packed-
mal Reactor at an Unstable State" Chem.Engng. Bed Models". AIChE Journal 32 ρ 809-819.
Science 35 603-612 Doyle.J.C. 1982 :"Analysis of Feddback Systems
Bonvin,D.; Mellichamp,D. 1982a "A Generalized with Structured Uncertainties" IEE Proc.
Structural Dominance Method for the Analysis of 129,D(6) 242-250
Large-Scale Systerns"Int.J.Control 35 807-827. Economou.C.G.; Morari,M. 1986 : "Internal Model
Bonvin,D.; Mellichamp,D. 1982b "A Unified Deriva- Control. 6. Multiloop Design". Ind.Engng.Chem.
tion and Chritical Review of Modal Approaches to Process Pes.Dev. 25 411-419.
Model Reduction".Int.J.Control 35 829-848. Eigenberger,G.; Butt,J.B. 1976 "A Modified Crank-
Bonvin,D.; Rinker.R.G.; Mellichamp,D. 1983a-d :" Nicolson Technique with Non-Equedistant space
On Controlling an Autothermal Fixed-BED Reactor steps" Chem.Eng. Sci. 31 681-691
at an Unstable State-I: Steady-State and Dynamic Epple.U. 1986 : "Model Reduction for Nonlinear
Modelling"; "-II: Discrimination among Rival Systems with Distributed Parameters". Precedings
Models to Achive Suitable Internal Structure"; " of IFAC Symposium on Dynamics and Control of
-III: Model Reduction and Contro Strategies Chemical Reactors and Destination Columns.
which avoid State Estimation" Chem.Engng.S c ience Brighton.
38 233-244; 245-255; and 607-618. Fieguth.F.; Wicke,E. 1971 :"Der Ubergang vom zund
Borisson.U. 1979 :"Self-Tuning Regulators for a /Losch-Verhalten zu Stabilen Reaktionszustanden
class of Multivariable Systems" Automatica bei einem Adiabatischen Rohrreaktor" Chem.Ing.
15 209-215 Techn. 43 604
Bortolotto,G.E.; Jergensen.S.B. 1985 : "Multiva- Foss,A.S.; Edmunds,J.M.; Kouvaritakis,B. 1980 :
riable Adaptive Control of a Recycle Reactor". "Multivariable Control System for Two-Bed Reac-
Proceedings of IFAC workshop on Adaptive Control tors by the Characteristic Locus Method" I.E.C.
of Chemical Processes. ed. Η.Β.Unbehauen Frank- Fundam. 19, 109-117
furt oct. 21-22. Froment,G.F. 1974 :"Fixed Bed Catalytic Reactors
Box,G.E.P.; Jenkins,G.M. 1970 "Time Series Analy- - Technological and Fundamental Design Aspects"
sis Forecasting and Control" Holden Day Chemie Ing.Techn. 46 374-386
Buchholt,F.; Clement,K.; Jergensen,S.B. 1979 Garcia,C.E.; Morari,M. 1982 :"Internal Model Con-
"Self-tuning Control of a Fixed Bed Reactor" trol 1. A Unifying Review and some New Results"
5'th IFAC symposium on Identification and system Ind.Eng.Chem.Process Pes.Dev. 21 308-323.
Parameter Estimation, Darmstadt, paper C3.5. Garcia,CE.; Morari,M. 1985a,b : "Internal Model
Canavas.C. 1986 : "Estimation of the Dynamic Control 2. Design Procedures for Multivariable
Behavior of a Fixed-Bed Reactor Through Filte- Systems" and "3. Multivariable Control Law Com-
ring" . Precedings of IFAC Symposium on Dynamics putation and Tuning Guidelines" Ind.Eng.Chem.
and Control of Chemical Reactors and Destina- Process Pes.Dev. 24 472-484 and 484-494.
tion Columns. Brighton. Garey,G.F.; Finlayson.B. 1975 "Orthogonal Colloca-
Chadeev,V.M.; NMaksimov,Y.M. 1986 : "An Adaptive tion on Finite Elements"Chem.Eng.Sci. 30 587-596
System with an Identifier for Control of the Georgakis,G.; Aris,R.; Ammundson,N.R. 1977a,b,c :
Temperature Profile in a Tubular Reactor". "Studies in the Control of Tubular Reactors I:
Precedings of IFAC Symposium on Dynamics and General Considerations", "II: Stabilization by
Control of Chemical Reactors and Destination Modal Control", and "III: Stabilization by
Columns. Brighton. Observer Design". Chem.Engng.Science 32 1359-
Chylla,R.W.; Adomatis,R.A.; Cinar,A. 1986 : "De- 1369, 1371-1379 and 1381-1387.
sign and Control of Tubular Autothermal Reac- Gilles,D. 1974 :" Quasi-stationares Verhalten von
tors : An Evolutionary Approach". Precedings of wandernden Brennzonen". Chem.Engng.Science 29
IFAC Symposium on Dynamics and Control of Chemi- 1211-1216.
cal Reactors and Destination Columns. Brighton. Gilles,D.; Eigenberger,G. and Ruppel,W. 1978 ."Re-
Cinar,A. 1984 :"Controller Design for a Tubular laxation Oscillations in Chemical Reactors".
Chemical Reactor". Can.J.Chem.Eng. 62 746-754 AIChE Journal 24 912-920.
Clarke.D.W.; Gawthrop,Ρ.J. 1975 :"Self-Tuning
Controller" IEE Proc^ 122 929-934
Fixed Bed Reactor Dynamics and Control 23

Grimble,M.J. 1981 :"A Control weighted minimum- Jorgensen.S.B.; Hansen,K.W. 1976 ."Dynamic Model-
variance controller for non-minimum phase sy- ling of a Gas Phase Catalytic Fixed-Bed Reactor-
stems" Int.J.Contr. 33 751-762 III: Identification Using Binary Pulse Sequence
Grimble,M.J. 1984 :"Implicit and Explit Self- Test Signals" Chem.Engng.Sci.,31, 473-479.
Tuning Controllers" Automatica 20 661-669 Jorgensen.S.B. ; Clement,K. 1977 .-"Experimental In-
Goldschmidt,L.Β.; Hallager,L.; Jorgensen,S.Β. vestigation of Reaction Zone Control in a Fixed
1985 "Multivariable Fixed-Bed Reactor Adaptive Bed Reactor" Chem.Eng. with Per Soltoft Danish
Setpoint Tracking".Proceedings of IFAC workshop Technical Press, Copenhagen 69-85
on Adaptive Control of Chemical Processes. ed. Jorgensen,S.Β., Goldschmidt,L.G.; Clement,Κ.
H.B.Unbehauen Frankfurt oct. 21-22. 1984 "A Sensorlocation Procedure for Chemical
Goldschmidt,L.B.; Hallager,L.; Jorgensen,S.Β. Processes"Com Chem.Engng.8 195-204.
1986 : "Dynamics and Control of an Adiabatic Jorgensen.S.B., Goldschmidt,L.G.; Hallager,L.
Fixed Bed Reactor with Varying catalyst 1985 -."Sparse Process Modelling for Robust Mul-
Activity". Preceedings of IFAC Symposium on tivariable Adaptive Chemical Process Control"
Dynamics and Control of Chemical Reactors and Proc. 7'th IFAC symposium System Identification
Destination Columns. Brighton. and Parameter Estimation, York England
Hallager,L.; Jorgensen,S.B. 1981 "Experimental Jorgensen,S.Β., Bortolotto,G.Ε.; Kaas-Petersen,C.
Investigation of Self-tuning Control of a Gas 1986 :"Bifurcation behaviour of a Recycle Fi-
Phase Fixed Bed Ctalytic Reactor with Multiple xed-Bed Reactor"in preparation.
Inputs". Proc.triennal Congr. IFAC 2743-50 Jutan,A.; Tremblay,J.P.;MacGregor,J.F. and Wright,
Pergamon Press J.D. 1977a "Multivariable control of a butane
Hallager,L.; Jorgensen.S.B. 1983 "Adaptive hydrogenolysis reactor I: State Space Reactor
Control Of Chemical Engineering Processes" IFAC modelling".AIChE Journal 23 732-742. 742-750,
workshop on Adaptive Systems in Control and 751-758.
Signal Processing" San Francisco. Jutan.A.; MacGregor,J.F. and Wright, J.D. 1977b,c
Hallager,L.; Goldschmidt,L.Β.; Jorgensen,S.Β. "Multivariable control of a butane hydrogenoly-
1984 "Multivariable Adaptive Identification sis reactor II: Data collection, Parameter Esti-
and Control of a Distributed Chemical Reactor". mation, and stochastic diaturbance Identifica-
Large Scale Systems 6 323-336. tion" and "III: On-line Linear Quadratic Stocha-
Hansen,K.W. 1973 "simulation of the transient stic Control Studies" AIChE Journal 23 742-
behaviour of a pilot plant fixed-bed reactor" 750, and 751-758.
Chem.Engng.Sci. 28, 723-734 Jutan,A.; MacGregor,J.F. and
Hansen,K.W.; Jorgensen.S.B. 1974 "Experimental Wright,J.D. 1984 "Multivariable computer control
Investigation of the Dynamics of a Catalytic of a butane hydrogenolysis reactor".AIChE
Fixed Bed Reactor".Adv.Chem.Ser.,133, 505-518 Journal 30 220-226.
Hansen,K.W.; Jorgensen.S.B. 1976a,b "Dynamic Kershenbaum,L.; Lopez-Isunza,F. 1982 "Dynamic
Modelling of a Gas Phase Catalytic Fixed Bed Behavior of an Industrial Scale Fixed-Bed Cata-
Reactor -I: Experimental apparatus and determi- lytic Ractor" 7'th Internat.Sym Chem.React.
nation of reaction kinetics" and "-II Results Engng.Am.Chem.Soc. 196 pl09-120.
from dynamic experiments":Chem.Engng.Sci.,31, 579 Koivo,H. 1980 :"A Multivariable Self-Tuning Con-
586 and 587-598. troller" Automatica 16 351-366.
Harris,T.J.; MacGregor,J.F.; Wright,J.D. 1980a : Kozub,D, .J.'; MacGregor,J.F. ; Wright,J.D. 1986 :
"Self-tuning and Adaptive Controllers: An "Multivariable Control of a Catalytic Tubular
Application to Catalytic Reactor Control". Reactor Using Both Wiener-Hopf Controller Design
Technometrics 22 153-164. and Internal Model Controller Design Approaches"
Harris.T.J.; MacGregor,J.F.; Wright,J.D. 1980b : Precedings of IFAC Symposium on Dynamics and
"Optimal Sensor Location with an Application to Control of Chemical Reactors and Destination
a packed Bed Tubular reactor".AIChE J^ 26 910- Columns. Brighton.
916 Kumar,S.; Seinfeld,J.H. 1978 "Optimal location of
Hlavacek,V.; Votruba.J. 1974 :"Experimental Study Measurements in Tubular Reactors" Chem.Eng.Sci.
of Multiple Steady States in Adiabatic Catalytic 33 1507-16
Systems" Adv.Chem.Ser. 133 545-558 Kuruoglu.N.; Ramirez,W.F.; Clough,D.E. 1981 : "
Hlavacek,V.; Votruba,J. 1977 :"Steady state opera- Distributed Parameter Estimation and Identifica-
tion of Fixed-Bed Reactors and Monolithic Struc- tion for Systems with Fast and Slow Dynamics. A
tures" in Chemical Reactor Theory ^ A Review ed. Tubular, Fixed-Bed Reactor to Form Styrene".
L.Lapidus and N.R.Amundson 315-399 Prentice Chem Engng.Science 36 1357-1364.
Hall Lapidus,L. 1961 :"On the Dynamics of Chemical
Hlavacek,V.; van Rompay,P. 1981 :"Current Problems Reactors". Chemical Progress Symposium Series 36
of Multiplicity and Sensitivity of States in 57 34-39
Chemically Reacting Systems" Chem.Eng.Sci. 36 Lee.K.S.; Lee,W.-K. 1985 :"0n-lLine Optimizing
1587-1597 Control of a Nonadiabatic Fixed Bed Reactor"
Hlavacek,V.; Puszynski.J. van Rompay,P. 1982 : AIChE J^ 31 667-675.
"Multiplicity and Propagating Fronts in Adiaba- Lee.R.S.H.; Agnew,J.B. 1977a,b "Model Studies
tic and Nonadiabatic Fixed-Bed Reactors" Am. for a Vinyl Chloride Tubular Reactor. 1.Steady-
Chem.Soc.Symp.Ser. 196 89-96 State Behavior" and "2. Dynamic Behavior" Ind.
Hofmann,H. 1979 :"Fortschritte bei der Modellie- Engng.Process Pes.Dev. 16 490-494 and 495-
rung von Festbettreaktoren" Chem.Ing.Techn. 51 501.
257-265 Lehr,C.G.; Yurchak,S.; Kabel.R.L. 1968 : Response
Hoiberg,J.A. ; Lyche.B.C: Foss.A.S. 1971 "Ex- of a Tubular Heterogeneous Catalytic Reactor to
perimental Evaluation of Dynamic Models for a a Step Increase in Flow Rate". AIChE Journal 14
Fixed-Bed Catalytic Reactor".AIChE J.,17, 1434- 627-636.
1447 Litz,L. 1979a "Ordnungsreduktion Linearer Zustan-
Hong,J.-C; Lee,H.H. 1985 : "Fixed bed with Cata- dsmodelle durch Beibehaltung der dominanten
lyst Deactivation: On-Line Estimation of Deacti- Einbewegungen" Regelungstechnik 27 80-86
vation, Control and Design".AIChE J^ 31 302- Litz,L. 1979b "Praktische Ergebnisse mit einem
311. neuen modalen Verfahren zur Ordnungsreduktion"
Jensen,K.F.; Ray,W.H. 1982 : "The Bifurcation Re gelungs technik 27 273-280
Behavior of Tubular Reactors" Chem.Engng.Science Lubeck,B. 1974 "Numerische Losungsmethode fur
37 199-222. nichtlineare partielle Modelgleichungen der che-
Joseph,B.; Brosilow,C.Β. 1978 :"Inferential Con- michen Verfahrungstechnik" Chem.Eng.Sci. 29
trol of Processes: Part I. Steady state Analysis 1320-28
and Design " and "Part II. The Structure and MacGregor,J.F. and Wong,A.K.L. 1980 ."Multivariate
Dynamics of Inferential Control Systems" AIChE model Identification and Stochastic Control of a
J. 24 485-492 and -509. Chemical Reactor" .Technometrics 22^ 453-464.

24 S. B. J0rgensen
MacGregor,J,F,; Penlidis,Α.; Hamielec,Α.Ε. 1983 Studies", and "-II: Effect of Pertubations "
"Control of Polymerization Reactors" fifth IFAC Chem. Engng.Science 34 p613-624 and 625-634.
PRP Conference, Amsterdam Silva,J.M.; Wallman,Ρ.H.; Foss,A.S. 1979 "multi-
Mandler,J.A.; Morari.M.; Seinfeld,J.H. 1986 : bed Catalytic reactor Control Systems:
"Control System Design for a Fixed-Bed Configuration Development and Experimental
Methanation Reactor". Chem.Engng.Science 41 Testing".Ind.Engng.Fundam.,18, 383-391.
1577-1597. Sinai,J.; Foss,A.S. 1970 "Experimental and Com-
McDermott,P.E.; Mellichamp,D.A. 1984 :"An Auto- putational Studies of the Dynamics of a Fixed-
Pole -Placement Self-Tuning Controller" Int.J. Bed Chemical Reactor".AIChE J. 16 658-669.
Cont. 40 1131-1147 Serensen,J.P. 1976 "Experimental Investigation
McDermott,P.E.; Mellichamp,D.A.; Rinker,R.G. 1986 of the dynamics of a fixed-bed reactor", chem.
a,b : "Pole-Placement Control of a Fixed-Bed Engng.Sci.,31 719-725
Autothermal Reactor Part I: Single Variable Serensen,J.P. 1976 "Experimental Investigation of
Control" and " Part II: Multivariable Control" the optimal control of a fixed-bed reactor",
AIChE Journal 32 1004-1014, and 1015-1024. Chem.Engng.Sci.,32 763-774
McGreavy C. 1983 :"0n-Line Computer Control System Serensen,J.P.;Jergensen,S.Β.,; Clement,Κ. 1980
for Chemical Reaction Processes"Com Chem.Engng. "Fixed-Bed Kalman Filtering and Optimal Control-
7 529-566 I: Computational comparison of Discrete vs
Mellefront,D.J.; Sargent,R.W.H. 1977 :"Optimal Continous Time Formulation" Chem.Engng.Sci. 35
measurement policies for control purposes"Int.J. 1223-1230.
Control 26 595-602. Stangeland,B.E.; Foss,A.S. 1970 "Control of a
Michelsen,M.L. ,-Vakil,H.B. ,-Foss,A.S. 1973 "State Fixed-Bed Reactor".Ind.Engng.Chem.Fund.,9,38-48
Space Formulation of Fixed-Bed Reactyor Dynamics" Sundaresan,S.; Amundson,Ν.R.; Aris,R. 1980 ^'Ob-
Ind.Engng.Chem.Fund. 12,323-328. servations on Fixed-Bed Dispersion Models: The
Morari,M.; Stephanoupolos,G. 1980 :"Studies in the Role of the Interstitial fluid" AIChE J^ 26
Synthesis of Control Structures for Chemical 529-536
Processes. Part III: Optimal Selection of Secon- Terndrup,H.; Clement,K.; Jergensen,S.Β. 1983
dary Measurements in the Framework of State "Multivariable Fixed Bed Reactor Control using
Estimation in the Prersence of Persistent Un- Frequency Methods" Proc.IASTED Symp. Applied
known Disturbances" AIChE JL 26 247-260 Control and Identification p.4.12-4.17, Copenha-
underwater,D.; MacGregor,J.F.; Wright,J.D. 1986 gen
"Nonlinear Temperature Control of a Catalytic Tremblay,J.P. and Wrigth.J.D. 1974 "Computer Con-
Tubular reactor Using Self-Tuning Reactors". trol of a Butane Hydrogenolysis Reactor". Can.
Submitted to Can.J.Chem.Eng. J.Chem.Engng. 52,845-47.
Padberg,G.; Wicke,E. 1967 :"Stabiles und insta- Tremblay,J.P. and Wrigth.J.D. 1977 "Multivariable
biles Verhalten eines adiabaten Rohrreaktors am model reference adaptive control of a pilot
Beispiel der katalytischen CO-Oxidation".Chem. scale packed bed tubular reactor." Proceedings
Engng.Science 22 1035-1051. IFAC symposium on Digital computer applications
Padmanabhan,L.; Lapidus,L.:"Control of Chemical to process control. North-holland publishing
Reactors" in Chemical Reactor Theory A Review" company.
ed.L.Lapidus and N.R.Amundson.
Prentice-Hall Vakil,H.B.; Michelsen.M.L.; Foss,A.S. 1973 "Fixed
314-399. bed reactor control with state estimation".Ind.
Prager,D.L.; Wellstead,P.Ε. 1980 :"Multivariable Engng.Chem.Fund. 12,328-335.
Van Doesburg,H.; De Jong,W.A. 1976a,b : "Tran-
Pole-Assignment Self-tuning Controllers" Proc.
sient Behaviour of an Adiabatic Fixed-Bed Metha-
IEE 128,D, 9-18
nator I: Experiments with Binary Feeds of CO or
Puszynski,J.; Snita,D.; Hlavacek,V.; Hofmann.H
CO^ in Hydrogen" and "II: Metahanation of mix-
1981 : "A Revision of Multiplicity and Parame-
tures of Carbon Monoxide and Carbon Dioxide"
tric Sensitivity Concepts in Nonisothermal Non-
Chem.Eng.Sci. 31 45-51 and 53-58.
adiabatic Packed Bed Chemical Reactors" Chem.
Villadsen,J.; Michelsen,M.L. 1978 Solution of
Eng.Sci. 36 1605-1609
differential equation models by polynomial ap-
Puszynski,J.; Hlavacek,V. 1984 : "Experimental
proximation. Prentice Hall, New Jersey
Study of Ignition and Extinction Waves and Os-
Vortmeyer,D.; Jahnel,W. 1972 :"Moving Reaction
cillatory Behavior of a Tubular Nonadiabatic
Fixed bed Reactor for the Oxidation of Carbon Zones in Fixed Bed Reactors Under the Influence
Monoxide ". Çhe mJE ngng.Science 39 681-692 of Various Parameters" Chem.Eng.Sci. 27 1485-
Ray.W.H. 1972 : "Fixed-Bed Reactors; Dynamics and 1496
Control". Proc.Second Internat.Sym Chem.Reac. Vortmeyer,D.; Winther R.P. 1982 :"Impact of Poro-
Engng. A 8-1 to A 8-19. Elsevier Amsterdam sity and Velocity Distribution on the Theoreti-
Rivera,D.E.; Morari,M. 1985 :"Internal Model Con- cal prediction of Fixed-Bed Chemical Reactor
trol perspectives on model reduction" Proc.Am. Performance" Am.Chem.Soc. Symp.ser. 196 49-61
Contr.Conf.,Boston MA 1293-1298. Wallman,P.Η., Silva,J.M.; Foss,A.S. 1979:"Multi-
Romagnoli,J.; Alvarez,J.; Stephanopoulos,G. 1981 : variable Integral Controls for Fixed-Bed Reac-
"Variable measurement structures for process tors" .IndJtagngJFunda^ 392-399.
control"Int.J. Control 33 269-289. Wallman,P.Η., Foss,A.S.1981 ."Experiences with
Root.R.B.; Schmitz.R.A. 1969 : "An Experimental Dynamic Estimators for Fixed-Bed Reactors".
Study of Steady State Multiplicity in a Loop Ind.Engng.Chem.Fundam.20 234-239.
Reactor". AIChE JOurnal 15 670-679 Wellstead,P.E.; Prager,D.; Zanker,Ρ.1979 ^Pole-
Root,R.B.; Schmitz,R.A. 1970 : "An Experimental assignment Self-Tuning Regulator" Proc.IEE 126
Study of Unstable States in a Loop Reactor". 781-789
AIChE JOurnal 16 356-358. Wicke,E.; Padberg,G.1968 :"Einfluss von Stoff-
Schmitz,R.A. 1975 : "Multiplicity, Stability, and und Warmetransport bei reaktionen in gas-
Reacting durchstromten Kornerschichten", Chem.Ing.Tech.
Sensitivity of States in Chemically
Systems- A Review" Chemical Rection Engine- 40, 1033-1038.
ering Reviews. ed. H.M.Hulbert. Adv.Chem.Ser. Wicke,E.; 0nken,H.U. 1986 : "Statistical Fluctua-
148 156-211. tions of Conversion and Temperature in an Adia-
Schnelle,P.D.; Richards,J.D. 1986 :"A Review of batic Fixed-Bed Reactor for CO Oxidation". Chem.
Industrial Control: Difficult problems and Wor- .Science 41 1681-1687.
kable Solutions" Procedings Chem.Process Control Wong,C.; Bonvin,D.; Rinker,R.G.; Mellichamp,D.
III 749-802, ed.M.Morari & T.J.McAvoy, Elsevier 1983 " On Controlling an Autothermal Reactor
Seborg.D.; Edgar,T.F.; Shah,S.L. 1986 : "Adaptive at an Unstable State-IV: Model Fitting and
Control Strategies for Process Control: A Sur- Control of the Laborator Reaktor"
vey" AIChE J^ 32 881-913 Chem.Engng.S c i enc e 38 619-633.
Sharma.C.S. and Hughes,R. 1979a,b ."The behaviour Zhao-Ying,Z.; Àstrom,K.J. 1981 :"A microprocessor
of an adiabatic fixed bed reactor for the oxida- implementation of a LQG self-tuner" Department
tion of carbon monoxide -I: General Parametric of Automatic Control, Lund Institute of Techno-
logy, Sweden
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986

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-

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

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
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

26 R. J. P. Brierley

The table shows that in this example, which Moving from the bubbly regime to mixed froth and
is reasonably typical, the liquid inventory then to spray, the liquid inventory on the tray
in a packed column system is little more than decreases - especially between the last two.
half that in a trayed column system. Further- Within a given regime it is possible - though
more, looking specifically at that part of the design correlations rarely predict it - that
system in which mass transfer is occurring (ie liquid inventory varies inversely with rates.
the trays and the packing), the contrast is
even more striking: the trays account for some
60% of the total inventory, whereas the packed TRENDS IN DISTILLATION SYSTEM DESIGN
beds hold only 20% of the total inventory.
These numbers will vary considerably from case Since the oil crisis of 1973-4, much thought has
to case, but the general message will remain. been given to energy saving through the develop-
ment of sophisticated system design and analysis
Liquid inventory is the major capacitance techniques. Whatever happens to oil prices, it
affecting the dynamic performance of distilla- is likely that these techniques will continue to
tion systems. In packed columns, the lags be used and will continue to produce designs
resulting from those parts of the liquid inven- which are more sophisticated and energy effi-
tory which are not directly involved in the cient than those of 15 years ago. These des-
mass transfer process are especially signifi- igns have significant effects on the dynamic be-
cant. This refers to liquid in the column haviour and control requirements of distillation
base, reboiler, reflux drum, pipework, etc. systems, especially when substantial amounts of
As a result, control systems for packed disti- heat integration are used. Control may not nec-
llation columns are often conservative. Typic- cessarily be more difficult - it may be easier.
ally, the heat supply to the column is set at Several areas in which developments have occurred
a fixed rate and such a column must be using or are possible are discussed below.
more energy than necessary in order to ensure
that product specifications are always met.
Sidestream Columns for Chemical Separations
These lags can be reduced but factors other
than the design and performance of the control Sidestream columns have been used in the oil in-
system under normal operation must be consid- dustry for many years, but the separations requ-
ered. For example, is the reflux drum really ired there are relatively sloppy and are not
necessary? In normal operation it almost cer- based upon a separation of identifiable chem-
tainly is not, but problems can arise in upset icals, but on the achievement of effects in the
conditions, and start-up may well be more diff- various products - eg vapour pressure, cloud
icult if that liquid inventory at distillate point, pour point. The outline flowsheet shown
composition is not available. in Fig.l is typical of many chemical processes -
reactor product passing through two columns, the
Hydraulics first to remove light impurities and the second
to remove heavy impurities. Depending on the
Another area where trayed and packed columns detailed process requirements, it is often poss-
perform differently is in their hydrualic be- ible to achieve both separations in one column,
haviour. The term "hydraulics" in this con- by withdrawing the product as a sidestream.
text refers to the behaviour of the vapour and
liquid phases on the trays or in the packed It is not generally possible to get a high rec-
beds. It has already been mentioned that the overy of the byproducts into the top and bottom
pressure drop of packings is much less than streams, but since they will be present in the
that of trays. More relevant is the varia- feed in small concentrations high recoveries are
tion of liquid inventory with variations in often not necessary. In such circumstances,
liquid and vapour rates. To a first approxi- considerable energy savings (as much as 50%) may
mation, in distillation, the liquid/vapour be achieved by using a sidestream column. But
ratio at a given point in a column will be con- how should it be controlled? With three product
stant in normal operation. streams there are in principle three composition
specifications to be met.
In packed beds the process is vapour phase con-
tinuous with liquid streams flowing down the A feed stream which is to be split into two main
packing surface. The liquid inventory falls product streams, but with a small byproduct
with reduced rate; at half design rates the stream of intermediate volatility, presents
liquid inventory is likely to have fallen by another situation where a single sidestream col-
something like 30%. umn can be used. In this case the byproduct
sidestream will not be pure but will contain
On trays, the variation of liquid inventory significant proportions of the two main products.
with rates is much more complicated than in But since the sidestream rate is small the loss
packed beds. At low rates the process is of products may well be justified by the savings
liquid phase continuous - discrete vapour achieved with a single column.
bubbles pass through a relatively quiescent
liquid layer on the tray. At higher rates
the vapour does not pass through as discrete Vapour Recompression
bubbles, but creates a turbulent froth. At
still higher rates little liquid exists on the Vapour recompression (Fig.2) is a technique des-
tray floor; most of it is blown up into the igned to drastically reduce the energy consump-
vapour space in the form of droplets. These tion in distillation by compressing the overhead
characteristic conditions are known as the vapour so that it will condense at a higher temp-
bubbly, mixed froth and spray regimes. Hofhuis erature than the column base. Thus the heat of
and Zuiderweg (1979) described a fourth condensation may be recycled. This system,
regime - emulsion flow. The regime obtained which appears to be highly attractive at first
depends not only on flowrates, but also on sight, has not been extensively used. The main
tray design. At pressures around atmospheric reason for this is that in many cases the
it is normal to have the mixed froth regime pressure ratio required for the compression of
with a spray in the vapour space above it. the overhead vapour would result in an expensive
Developments in Distillation and System Design 27

multi stage machine which makes the scheme is based on consideration of a feed mixture con-
look unattractive. taining three components. For a more detailed
discussion and evaluation, consult Doukas and
Appendix 2 extends the example of Appendix 1, Luyben (1978).
assuming the separation of toluene from meta
xylene, to calculate the compression ratio The simplest complex column is a sidestream
required for both a trayed column and a column, discussed above. As was noted, the
packed column with structured packing. The simple sidestream column is of limited value
calculated values of compression ratio are 3.9 since it is not possible to produce three pure
for the trayed column and 2.8 for the packed products, and the applications described were
column. Neither case looks promising, but cases which do not have this requirement. Three
the benefit of packing is clearly demonstrated. pure products can be achieved by the addition of
Even with no pressure drop and no temperature either a sidestream stripper (Fig,3) or a side-
difference in the condenser/reboiler, a com- stream rectifier (Fig,4). In both cases the
pression ratio of 2.1 is required. Therefore, feed to the sidestream column contains only two
it can be seen that vapour recompression is components. The sidestream column may have its
likely to be viable only with systems of low own condenser and reboiler but it is more usual
relative volatility. Vapour recompression has to omit the condenser (stripper) or the reboiler
been applied to such systems in commercial (rectifier) as shown.
plants, for example ethylene/ethane splitters
on olefine units. An alternative approach is to do a partial sepa-
ration in a prefractionation column - the
distillate being a binary mixture of components
Heat Integration A and B, while the residue is a binary mixture of
components Β and C. These may then be separated
Vapour recompression may be regarded as an ex- in two further columns as shown in Fig.5. Clearly
ample of heat integration in which the integra- these two columns may be combined. The pre-
tion is confined to a single column and its fractionator may have its own reboiler and con-
ancillaries. Another form of heat integration, denser or it may be directly linked to the main
which has been in use for many years and is column as shown in Fig.6. Although this may
similarly confined to the column and its ancill- show substantial benefits on the basis of a
aries, is feed/product heat interchange and the steady-state design, control problems may render
use of condensation heat to preheat the feed. it inoperable. Dynamic simulation seems to be
The most extensive use of these techniques has the only way to solve the control problems or to
been in crude oil distillation. demonstrate conclusively that they are unsolv-
Much attention has been given in recent years
to the heat integration of distillation columns
with the rest of the process (Linnhoff and co- CONCLUSION
workers, 1983). Many chemical processes in-
volve not just one distillation step but sev- This paper has aimed to demonstrate that in at
eral. The heat requirements of distillation least two areas of development in the industrial
are often large compared with those of the application of distillation technology, current
rest of the process and the most obvious scope developments in the simulation of column dynamics
for heat integration is in linking columns and in the design of control systems could prove
together so that the condensation heat from very helpful and timely. Of course this will
one column is used to reboil another. We all hopefully be in addition to progress with the
know that distillation column control is not many long-standing and well-known problems of
easy - putting two columns in series is surely distillation system operation and control.
asking for trouble. This has been the atti-
tude in the past to the linking of columns by
heat integration. Where such systems have REFERENCES
been used, various steps have been taken to
avoid undesirable interactions: Doukas, N. and W.L. Luyben (1978). Economics of
alternative distillation configurations for
a) adjacent columns not linked unless with the separation of ternary mixtures.
intermediate storage, Ind. Eng. Chem. Process. Pes. Dev., 17,
b) either or both columns operated at con- Fair, J.R. (1973). Gas-liquid contacting. In
stant reboil rate, R.H. Perry and C H . Chilton (Ed.), Chemical
Engineers Handbook, 5th ed. McGraw-Hill,
c) a trim condenser installed in parallel Kogakusha, Tokyo. Chap. 18, pp 3-25.
with the linking condenser/reboiler so that Hofhuis, P.A.M. and F.J. Zuiderweg (1979).
disturbances in the higher temperature column Sieve plates: dispersion density and flow
are not passed to the other, etc. regimes. 3rd. International Symposium on
Pistillation. IChemE Symposium Series
The use of linked columns for energy saving is No.56. Pp 2.2/1 - 2.2/26.
an important area which is currently con- Linhoff, Β., H.A. Punford and R. Smith (1983).
strained by an inadequate understanding of the Heat integration of distillation columns
dynamic performance of such systems and how into overall processes. Chemical Engineering
they can best be controlled. Science, 35, 1175-1188.
Meier, W., W.R. Hunkeler and Ρ Stocker (1979).
Performance of the new regular tower packing
Complex Columns "Mellapak". 3rd International Symposium on
Pistillation. IChemE Symposium Series
A simple column is one with only two product No.56. Pp 3.3/1 - 3.3/18.
streams, distillate and residue. For a bi- Strigle, Jr, R.F. and K.E. Porter (1979). Metal
nary feed a simple column will in principle Intalox - a new distillation packing. 3rd.
achieve any desired separation of the two com- International Symposium on Pistillation.
ponents. This discussion of complex columns IChemE Symposium Series No.56. Pp 3.3/19 -
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
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
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 vol%
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)

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

Turbo blower

Product Feed-


^ Distillate
"ends —•-Residue

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

Feed _
A + Β + C
A + Β + C

Fig. 3. Sidestream stripper system

Fig. 4. Sidestream rectifier system

A + Β

Feed ,
A + Β + C •Β A + Β + C

Β + C Β + C

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


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 (emulsion or solid catalyzed Ziegler- Natta) mass transfer rate
limitations between the phases may sometimes be important. In
There are many aspects to the control of polymerization reactors. ionic Ziegler-Natta catalyzed systems the major uncertainty lies
Industrial emphasis has centered around the sequencing and in trying to model the active site complexes of the catalyst system.
automation of batch and s e m i - batch r e a c t o r s , and the For example, in homogeneous Z-N systems such as the solution
temperature control of the reactors. In this paper I shall polymerization of butadiene, there is as yet no agreement on the
concentrate exclusively on problems of controlling polymer number, nature, or role of the active catalyst species. Similarly, in
quality variables, and ignore the automation and temperature heterogeneous systems little knowledge exists about the activity,
control problems except in as far as they affect quality. No or role of the active catalyst sites.
attempt will be made to give a comprehensive review of the
literature, since excellent reviews have ben presented at past Most of the polymerization models reported in the literature are
IFAC conferences on Instrumentation and Automation in the deterministic in nature. They in no way account for the stochastic
Paper, Plastics and Polymerization (PRP) Industries (Amhren, variability that exists in industrial processes. For free radical and
1977; Hoogendoorn and Shaw, 1980; MacGregor et al., 1983). ionic polymerizations the major sources of stochastic disturbances
Rather, reference to papers will be made only to illustrate certain are varying amounts of reactive impurities. Trace amounts of
aspects of specific control problems 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 .
The paper is outlined as follows. Given the importance of Fluctuations in these impurity levels are the major reason for the
fundamental or mechanistic mathematical models in the control batch-to-batch variations in polymerization rate and polymer
of quality variables, a discussion of the nature of these models and quality, and for the drifting conversion and quality observed in
their limitations is first presented. Attention is focussed continuous reactors. Therefore, in order to be able to relate
exclusively on free radical and catalyzed ionic polymerizations. A mechanistic model predictions to observed plant behaviour, (e.g.
discussion of past work on open-loop optimal operating policies, is through state estimation), or to be able to control polymer
followed by a discussion of on-line measurements and state production rate and quality using these models, it is important
estimation. Finally, past work and future directions on the that the effect of these reactive impurities be better understood,
closed-loop control of continuous and semi-batch reactors are and accounted for in some manner in the models. Recently some
discussed. 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 Molecular Weight Development
more of the following set of balances in one form or another:
overall material balances for all the major components; molecular There are a number of different approaches that have been taken
balances for molecular weight development; population balances to develop expressions for the molecular weight development in
for particle size development. 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
Material Balances 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
states the number and weight average molecular weights are
A set of differential equations accounting for the accumulation,
inflow, outflow, and disappearance by reaction of the monomers, easily calculated.
solvent, initiators and catalysts, chain transfer agents, and other
ingredients such as emulsifiers or stabilizers comprises the main For free radical polymerizations in which either termination or
part of any model. The overall balance equations are usually easy transfer reactions dominate the models appear to be reasonably
to write down in the case where spatial mass and heat transfer is well developed (e.g. Hamielec and MacGregor, 1983). The major
not a problem. The uncertainties usually lie in the description of uncertainities lie in obtaining reasonable e s t i m a t e s of the
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 termination rate parameters (including the gel effect parameters)
polymerizations at low conversions are among the easiest to and the rate constraints for the transfer reactions (to modifier, to
model. At higher conversions or lower solvent concentrations polymer, to internal double bonds, etc.) Obtaining good estimates
diffusion controlled macromolecular reactions (Trommsdorf or gel of these parameters is far from straightforward. Large amounts of
effect) usually become important, and bulk mass and heat transfer data on molecular weights are required, and often the type of data
in the vessel lead to spatial variations in concentration and required (GPC, NMR) is very difficult to obtain due to the poor

32 J. F. MacGregor

precision of the measurements or to the insolubility of the co- and ter-polymers, etc., are still in the development stage.
branched polymer, etc.. Polymer structure properties (e.g. tacticity, sequence length
distribution, etc.) require sophisticated NMR techniques. The
For Ziegler-Natta systems the understanding of the molecular analysis of particle structure or morpholgy in heterogeneous
weight development process is much less well understood. Again polymerization requires meticulous sample preparation and the
this involves a more detailed understanding of the nature of the use of electron microscopes. Even with the successful analysis of
active catalyst sites, and of the effect of impurities on them than such fundamental properties, we are still often left with the
currently exists. Diffusional effects in solid catalysed gas phase difficult task of understanding how they relate to the end-use
reactions may also play a role in molecular weight development performance of the product.
(Laurence and Chiovetta, 1983).
Therefore, on-line measurements are only feasible, at this time,
for certain simple fundamental properties, and for certain
Population Balances surrogate variables which relate to some fundamental and
end-use properties. Most of the sophisticated analyses are not
In heterogeneous polymerizations such as emulsion systems feasible for on-line application because of the lack of instrument
where the particle size distribution plays an important role in robustness to the demands of the plant environment, and because
both the final product quality, and the rate of reaction, population of the need for meticulous sample preparation.
balances accounting for the birth, death and growth of polymer
particles is an important part of the model. Min and Ray (1974) On-line conversion measurements using densitometers (both
and Penlidis et al., (1986) have summarized the population vibrating tube, and nuclear) have proven very successful in many
balance and age distribution approaches. Major uncertainties lie applications (Shork and Ray, 1981, Penlidis, 1986). On-line
in modelling the particle nucleation and coagulation phenomena, energy balances around the reactor to track instantaneous
particularly in the presence of non-ionic stabilizers. Usually these reaction rates, monomer conversion, and heat transfer coefficients
phenomena are so unpredictable and occur so rapidly that little (Harris and Rushing, 1982; Wu, 1985; MacGregor, 1986) have also
can be done to actively control them during a polymerization. The become widely used. On-line particle size analyses for emulsion
best that can be done is to try to avoid having to deal with them by polymerization based on turbidity and dynamic light scattering
using seeded emulsions, and an adequate amount of stabilizer. 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
OPEN-LOOP OPTIMAL POLICIES effective means of monitoring many industrial processes.

Given the lack of availability of on-line sensors, and the lack of Given the scarcity of on-line measurements there would appear to
well developed nonlinear feedback control methods, it is not be a great incentive to use state estimation methods to infer the
surprising that most early research on polymerization reactor values of other state variables in mathematical models of the
control centered around the simulation, and the development of process.
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 Two other papers at this conference deal with Kalman Filtering
stirred-tank reactor policies which optimized some measure of methods, one to track the material and population balance states
performance, subject to certain constraints, were u s u a l l y in emulsion polymerization using densitometers, etc. (MacGregor
developed through the use of the Maximum Principle. Common et al., 1986) and the other to estimate the molecular weight
objectives were to minimize batch or start up time subject to distribution using an on-line GPC (Papadopoulou, 1986). The
producing polymer with a given number average molecular former paper also stresses the importance of modelling the effect
weight or polydispersity. In some simple cases, experimental of impurities, and of incorporating nonstationary stochastic states
verification was obtained. For the case of copolymer composition into the model of the process in order to obtain realistic state
control, a necessary condition for an optimal policy could be estimators that are capable of tracking an industrial process.
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 Observability plays an important role in state estimation for
Principle. Final selection among various possible constant polymerization processes. In general, the molecular weight states
composition feed policies would then have to be based on other are unobservable from measurements on the material balance or
considerations such as minimizing branching, minimizing batch the particle population balance states of the model. Since on-line
time, or safety considerations. A review of the literature is measurements for the latter states are more readily available, this
presented in MacGregor et al. (1984). means that tracking the molecular weight development of a
polymerization process is often very difficult.
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 FEEDBACK CONTROL
also been treated. These studies may be useful for pointing to
potential operating problems in reactors. However, the results are Continuous Processes
highly dependent upon the structure and parameters of the model.
The predicted bifurcations often occur well outside of the In high volume continuous polymerization processes, regulation
conditions under which the model has been developed, and are or disturbance rejection is a major priority, although start-up and
difficult to verify experimentally. grade change policies can also be very important. Open-loop
start-up or change-over policies can be developed by the
In general, simulation and optimization based on deterministic optimization procedures discussed earlier, and approaches to
polymerization reactor models are useful mainly for the improved feedback control during these periods would be similar to those
understanding they generate, and for comparing alternative discussed later for semi- batch reactors. Therefore in this section I
operating policies. Because of the uncertainties in these models, shall concentrate on the regulation of continuous processes about
particularly in the molecular weight part, and because the models given set-points.
ignore the effect of reactive impurities, etc., the optimal policy
results are rarely directly applicable. This is discussed again later Importance of Reactor Design: The importance of considering
under the section on feedback control. 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
ON-LINE MEASUREMENTS AND STATE limitations in the process, by the lack of provision for surrogate
ESTIMATION measurements throughout the process, by the inflexibility of the
purification and feed system, and by the poor dynamic behaviour
Even the off-line measurement of many polymer quality variables of the process. Although numerous industrial examples exist to
in the laboratory presents a formidable task. Molecular weight illustrate each of these situations, I will focus here on only one
measurement methods for highly branched homopolymers, and for
Control of Polymerization Reactors 33

examples which illustrates the importance of considering the important at the given sampling interval, and if there is a cost
dynamic behaviour of the process at the design stage. associated with taking a corrective action. In other cases, it is
more appropriate to identify discrete empirical transfer function
In the design of continuous reactors it is usual to consider only and disturbance models (ARMAX models) relating these output
steady state models, and then consider the dynamics and control measurements to the process inputs, and then use minimum
problems later. In the case of emulsion polymerization in CSTR variance of LQG-type stochastic controllers. Examples of this
trains this approach can lead to disastrous consequences. It has approach are treated in MacGregor and Tidwell (1980), Rushing
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 (1986), Kelly et al. (1986).
commercially important monomers in CSTR's leads not to a
constant steady state operation, but rather to a sustained A final point, that is apparent to most people working with
oscillation or limit cycle behaviour in all the latex and polymer polymerization reactors, is that these processes are highly
properties. Using dynamic models, verified through extensive interactive, and usually must be treated as true multivariable
experimentation, Kiparassides et al. (1979, 1980) showed systems. A change in any one of the input feedrates (e.g. initiator,
conclusively that the reason for such behaviour lay in the monomer, etc) will usually affect many of the output quality
discontinous particle nucleation phenomena occurring in the early variables. It is for this reason that so much effort has been spent
reactors. in efforts to develop mechanistic models of polymerization
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 Semi-Batch Reactors
control theory. However, they did show that a reactor train
redesigned to include the splitting of the monomer and water feeds Much of the polymer industry is based around semi-batch reactor
between a very small first reactor and the subsequent large technology. In particular, most low volume and specialty
reactors would eliminate the oscillations. Furthermore, by polymers are made in these reactors They allow the flexibility of
regulating the split of monomer and water between these reactors, producing a wide variety of products within the same equipment.
and the initiator flow to the first reactor, both the monomer
conversion and latex particle size could easily be controlled in a Most of the optimal open-loop control policies discussed earlier
stable, non-oscillatory manner. Figure 1 shows a schematic of this were aimed at developing temperature, and initiator or monomer
split-feed reactor system and Figure 2 shows some conversion feedrate policies versus time which would produce polymers or
results taken from a pilot plant on the continuous emulsion copolymers with specified molecular weight, composition, or
polymerization of vinyl acetate (Penlidis, 1986). In part A a single morphological properties in a minimum batch time. These
CSTR was employed and the oscillatory behaviour is clearly policies are all based on deterministic models, and precisely
evident. At point Β the small reactor and the split feed system known initial conditions. Unfortunately, implementation of these
were switched in, and the oscillations immediately disappeared. policies as a preprogrammed trajectory over time poses serious
A change in the feed splits was then made to effect a conversion problems. First, the deterministic models are often poor,
change at point C. Hence, the major controller problem particularly with respect to MW development, and the policies are
(oscillations) was eliminated at the design stage, and a new, much often very sensitive to mismatch between the process and its
more flexible and stable system resulted. model. Secondly, there are unmodelled, stochastic elements to the
process, such as the concentration of impurities in the initial
Regulation of Cont: AUQUS Processes. The control of continuous charge and in the subsequent semi-batch feeds, the activity or
polymerization processes, where on-line measurements of state or active site distribution of Ziegler-Natta catalysts, and unexpected
surrogate variables is available, or where inferred values are temperature variations that occur during the batch cycle. Such
available from state estimators, parallels closely the control of uncertainties in the initial conditions and the subsequent
most other continuous industrial processes. A large body of linear operation of the reactor would lead to wide variability in the final
univariate and multivariate control theory is available to handle product, unless feedback from on-line measurements is used to
such problems. However, there are a number of things that are correct the trajectories. This area of nonlinear, multivariable
common to many polymerization processes and should be feedback control, aimed at compensating for the effect of
addressed in developing a control scheme. Some of these are unmodelled dynamic and stochastic effects on optimal trajectory
discussed below. 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.
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 A common approach to trajectory control employed in other areas
feed monomers themselves have varying impurity levels, and such as aerospace is to linearize the model about the precomputed
consideration should always be given to the feasibility of trajectory, and apply control adjustments based on linear control
controlling some of this feedstock variability. More important, in theory to try to bring the reactor states back to their open-loop
processes with recycle of recovered monomer, is the buildup of optimal trajectories. In principle this approach is feasible. It
impurities in the recovery and recycle system. In these processes would require local linearizations, extended Kalman Filtering,
it is not uncommon to concentrate the control effort on the and finite horizon LQG controllers, for example. However, even
recovery system. If, in this way, one can achieve control over the the objective of bringing the reactor back to a trajectory that is
impurities entering the reactor, the need for advanced control based on a purely deterministic model or even based on past
around the reactor itself is greatly reduced. experience is questionable, because in the presence of stochastic
variations that have caused the reactor states to deviate from this
Another characteristic of polymerization reactor control is the trajectory, the precomputed trajectory will no longer be optimal.
need to deal with noisy and infrequent measurements. Some of On the other hand, to use general nonlinear programming
these measurements often come from off-line analyses performed methods to recompute optimal control trajectories on-line at each
in a quality control laboratory. Given the difficulty in getting any control interval would appear to be computationally impractical
measurements of polymer quality, it is important that one make given the highly nonlinear, multivariable nature of the models.
use of all the available data. Examples of such off-line Suboptimal approaches to feedback control which make use of the
measurements useful for process control are: Mooney viscosity nonlinear mechanistic models plus a few on-line measurements
and solids measurements in the production of butadiene rubbers; are currently an active area of research.
melt-index and density measurements in polyolefin production;
viscosity and fiber characteristics such as dye uptake, etc. in the Approaches to handling these nonlinear semi-batch reactor
manufacture of synthetic fibers. A common industrial approach to control problems through the use of traditional adaptive control
using such off-line data is to implement statistical quality control with linear models have been suggested (e.g. Kiparissides and
procedures such as Stewart, CUSUM or EWMA charts, and take Shaw, 1983; Houston and Shork, 1986). However, in this author's
corrective action only infrequently whenever a statistically opinion, such approaches could have serious problems. The
significant deviation from target is detected. Such procedures, parameters of these linear mdoels often have to adapt rapidly to
however, are appropriate only if process dynamic effects are not account for reaction nonlinearities. Such rapid parameter

34 J. F. MacGregor

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

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

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

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

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.

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 %






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


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

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


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

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 trays, doubling the vapor rate will increase the
pressure drop by a factor of four. This is
The industrial tray design standard for all sizes because the flow rate is approximately equal to a
of distillation columns is for a minimum tray constant times the square root of the pressure
differential pressure of 0.24 psi (6.6 in. of drop.
water). Because of the high competition among
the tray manufacturers, most of them design for For the valve trays, the resistance to flow
the 0.24 psi loading. (A minimum tray spacing of remains nearly constant until all valves are
18 in. is also an industrial standard. This is completely open. Doubling the vapor flow might
partly for ease of maintenance.) For small dia- only slightly increase the pressure drop. In
meter columns, the cost to increase the tray either case, if the flow is increased by the
strength is usually proportionally less; in fact, square root of two (for the case of fixed opening
the common materials used usually result in tray in the trays), the pressure drop will double from
strength greater than the minimum standard of say 3 to 6 in., which is still within the design
0.24 psi. maximum load. For valve trays this is even more
conservative and, as mentioned, the flow can be
The differential pressure standard is for ver- close to double. The vertical pressure drop will
tical stressing in both directions. When the remain close to 3 in., or 50% the maximum
column is operating, the vertical differential designed vertical load.
pressure caused by the vapor flow partly counter-
acts the downward force of the tray liquid Flooded Trays
At a given nominal gas rate providing an upward
Another general industry standard is for the gas force of 3 in. across the tray and, with the tray
pressure drop across a tray to be on the order of flooded to 9 in., the net downward force is
1 to 3 in. of water. Therefore, if the vapor 6 in., which is near the maximum design load.
rate is providing 3 in. of upward force gas drop Since spacing between the trays is 18 in., if a
across each tray and the liquid level on each tray were to flood to 18 in., then there would be
tray is 3 in. of water providing a downward a downward force of 15 in., which exceeds the
force, the net force on the tray would be zero. design load by more than double.
Now consider the following two cases where the
trays are dry or flooded. The importance of the above scenario is that no
matter what the cause of flooding or of a tempo-
Dry Trays rary draining of trays, especially in the case of
flooding, it is relatively easy to double the
If the vapor rate provides 3 in. of dry gas pres- designed force on a tray. Often when flooding
sure drop, then the upward force will be 3 in. of occurs and the column is generally upset, an
water, which is half the design force. For trays operator might greatly reduce the steam to the
with fixed openings such as sieve and bubble cap column and, therefore, reduce the boilup rate.
Diagnosing and Preventing Tray Damage 39

This would increase the downward hydraulic force Items 5 and 6 will cause instability but probably
by the amount of reduction in gas pressure drop exhibit adequate average separation.
due to the reduction in boilup. Thus, for a
column with 18 in. tray spacing, flooded 18 in., Poor Separation. If the separation becomes
and the vapor rate reduced, the downward hydrau- poorer during normal operating conditions and
lic force could be close to three times greater if the column is stable, then tray damage or
than the maximum design load for the tray. an error in one or more measurements could be
the cause. For example, if the steam flow-
What might be concluded from the above analysis meter erroneously reads high, then the boilup
is that preventing flooding of the trays should is lower than it appears. Also, if the reflux
prevent most of the tray damage. This, however, flowmeter erroneously reads high, then there
does not seem to be true since most of the dam- would be less reflux. (Note that an unknown
age, especially when related to having a high change in feed composition would also appear
liquid level in the bottom of the column, is for as a change in separation under normal column
the trays to rupture in the upward direction. operating conditions.) Poor separation can
However, not all of the column damage is being also be caused by fouling or plugging. Foam-
examined carefully and, if done so, it might ing may cause a poor separation because of
reveal that some of the damage is being caused by entrainment. However, usually instability is
downward rupturing of the trays. Only about 3 of also apparent. Instability, if minor cycling
the 17 incidents surveyed appeared to be asso- is occurring in the key variables, usually
ciated with downward rupturing of the trays. It will not result in poor separation. On the
seems prudent, however, in any case, to prevent other hand, large instability such as dumping
flooding of the trays. (Note that in the above of the column affects the separation.
analysis, the hydraulic pressure would be less in
proportion to the degree of gasification on a By maintaining an operating performance
tray. For example, if 50% of the volume were history, any suspicion of poor separation
bubbles, then the liquid would be 50% less and so could be verified by comparing current
the hydraulic loading would be 50% less.) operating parameters with a similiar or
identical set of earlier operating conditions.
Cause of Tray Damage When the Bottom of the
Column Is Flooded With Liquid A mathematical model previously matched to a
normal operating column could also be used in
In most cases of column damage, where the steam comparing with current operating conditions,
is added to the column with the bottom trays to determine if the separation has degener-
flooded with liquid, it cannot be reasoned that ated.
the damage was due to the high vapor rate even if
it were two-phase flow. Therefore, there prob- Instability. Minor instability, such as con-
ably are vertical hydraulic forces, such as the tinuous cycling of one or more variables, is
water hammer effect, that cause the damage. It most likely due to improper controller tuning.
appears that these damages would be especially Continuous cycling can also be caused by a
dominant during the very initial introduction of continuous cycle or a disturbance in a primary
steam to the column. Therefore, boiling up very variable such as the column feed rate, pres-
slowly to begin with, until vapor equilibrium is sure, and others. Severe cycling, especially
established in the column (in other words until when the period of the cycle is greater than
vapor begins condensing at the top of the the natural period of any of the key column
column), should reduce the chance of damage variables, can be due to tray damage, fouling
occurring. or plugging of the trays, or foaming.

Foaming. Isolating foaming as the cause of

DIAGNOSING COLUMN TRAY DAMAGE column misbehavior from other causes is very
important since the solution to the problems
Some Causes of Column Misbehavior are quite different. It is very costly and
frustrating to shut down a column and examine
1. Tray damage. it for tray damage and find none.

2. Fouling or plugging. It can be very difficult to isolate column

behavior caused by foaming from those caused
3. Foaming. by plugging or tray damage. However, there
are some guides to the cause of foaming which
4. Measurement errors in key column variables, are listed as follows:
such as pressure, temperature, level, flows, and
concentration. 1. Particulates in liquid on trays can cause
foaming. (This was observed in a solvent
5. Instability caused by improper controller column where water and solvent were being
tuning and/or poor control strategy (such as not separated. A small amount of sodium chloride
providing sufficient control rangeability and/or impurity was precipitating and causing foam-
decoupling of variables that interact). ing. An antifoam agent was added, and it
eliminated the problem.)
6. Instability caused by uncontrolled external
variables such as feed rate changes to the 2. Two liquid phases where it becomes a single
column, feed composition changes, and variations phase. Foaming may occur in the region of the
in steam quality and/or steam supply pressure column where this occurs.
(with no steam flow controller), and others,
depending on the nature of the separation and 3. Cannot have foaming with two phases always
equipment being used. present. When two phases are present, the
phase with the lower surface tension acts as
Determining Cause of Column Misbehavior an antifoaming agent.

Items 1 to 4 are manifested by poor separation 4. When separating two components, where the
and can cause instability. component being removed from the bottom of the

40 W. R. Ellingsen

column has a lower surface tension than the liquid is up in the column (in other words, veri-
component removed from the top, foaming will fying that the calandria level measurement is
not occur unless (1) or (2) exist. This is" correct), but aid him in deciding how to elimi-
called the Marangoni effect. An example of nate the liquid. If the liquid level is above
this is the separation of Dimethylacetimide one or two trays, the operator might elect to
(DMAc) and water. This cannot foam because very slowly boil that liquid out. Whereas, if
water being removed from the top of the column the liquid level is up 25% of the column, then
has a higher surface tension than DMAc being the operator would elect to pump the liquid out
removed from the bottom and vice versa. before introducing steam to the column.

Foaming can usually be eliminated by adding Obtaining a Column Density Profile

antifoaming agents in the 10-100 parts per
million range, which is usually acceptable. Scanning a column empty and during normal and
abnormal operating conditions with a nuclear
Effect of Foaming in Columns radiation scanning system can isolate column
problems. This approach cannot be used in
1. Sometimes foaming causes a reduction in preventing tray damage, but is very useful in
separation due to reduced tray efficiency. diagnosing tray damage and other column prob-
This is caused by foam entrainment from tray lems by showing the liquid mass profile in the
to tray. column. However, installing the instrumenta-
tion mentioned above would, in most cases,
2. Foaming can cause tray flooding by reducing help in preventing the tray damage, as well as
the density of liquid in the downcomer by slow in diagnosing the cause of tray damage and
disengagement of bubbling in the downcomer. other column problems.

3. Foaming on a tray will produce a dam or a

partial blocking of the entrance to the down- PREVENTING COLUMN TRAY DAMAGE
corner. (This has been observed in test
installations.) The incentive to preventing distillation column
tray damage has to be placed in an economic
prospective. If we want to spend the money, we
GENERAL APPROACH TO TROUBLESHOOTING can design distillation columns that will never
be damaged under the most severe operating condi-
Although troubleshooting a distillation column tions. Thus, a tradeoff exists among reducing
has to be tailored somewhat to each column, there capital investment (which increases the proba-
are three general levels of problem isolation bility of tray damage), the attendant cost of
that can be used on any column. These are: repair, and loss of production due to downtime.
A related question to be asked is, "Can we use
1. Assure that the measurement instrumentation the industry minimum design standard column con-
and control loops are functioning properly. struction, but eliminate tray damage by improved
instrumentation and control?" The cost tradeoff
2. In the case of column instability, determine becomes minimizing capital investment in column
if disturbances are being introduced into the construction, but adding some capital for addi-
column externally such as by variations and feed tional instrumentation and control. (Note that
rate, reflux rate, steam rate, reflux temperature enhanced instrumentation and control could lead
variations, and other external variables. to more economical operation of a column under
normal operating conditions.)
3. If the items mentioned in 1 and 2 are not the
cause of the column problem (instability or poor The answer to simultaneously minimizing capital
separation), then the source of the problem is investment cost and operating cost is not simple.
internal. The troubleshooting can now focus on The answer for an existing column versus a new
whether the instability or poor separation is due column would be different. Based on the analysis
to plugging, foaming, or tray damage. of the 17 incidents of column damage and the
apparent causes of the damage, some recommenda-
Aids in Diagnosing Column Tray Damage tions that can be used as a guide in reducing the
and Other Problems probability of column tray damage are given
Aids used in diagnosing column problems can also
be used in preventing tray damage. Probably the Liquid Above the Bottom Tray
most significant is to measure differential
pressure and temperature across sections of the With a column that has liquid above the bottom
column. The ideal situation would be to measure tray, the best procedure is to pump the liquid
the pressure and temperature and composition on out until the calandria level is normal; then
each tray. A compromise, in order of priority, proceed to turn on the steam to bring the column
would be to measure the pressure drop across to normal operating conditions. This would
sections of the column, say 10% or at least 25% eliminate what appears to be over 70% of the
sections. This will help locate the region in source of distillation column tray damage. Since
the column where a problem has developed. A this will not always be practical, a compromise
temperature profile consisting of at least five procedure would be to:
temperature measurements up the column would
greatly enhance the analysis of the state of the 1. Ramp the steam flow rate to the column to
column. The ability to draw liquid (or vapor) approximately 25% of the design rate over 30 min-
samples from sections in the column for analysis utes to one hour.
could greatly enhance diagnosing problems because
the degree of separation would be known from an 2. Hold at that rate until it is obvious that
analysis of the samples. vapor is being condensed at the top of the
Providing a liquid level measurement for the
bottom 25% of the column (differential pressure) 3. Ramp to the minimum design rate over a period
would aid the operator not only in verifying that of 30 minutes to one hour.
Diagnosing and Preventing Tray Damage 41

4. Hold at that rate until the liquid has to the maximum vapor flow design rate should be
reached a normal calandria level. During the limited directly on the control valve. In
second phase, the reflux should be set to the addition, the maximum flow set point on the steam
minimum acceptable rate until the liquid in the flow controller should be limited.
column reaches the normal calandria level.
It should be noted that tray damage, in most
Bottom Tray Liquid Seal incidents in the survey, occurred when the steam
flow to the column was within normal or reduced
Construct the bottom tray seal to be especially rate. This observation shows that limiting the
strong. This is a low-cost item, but if the steam flow rate to the column is not sufficient
bottom seal is broken, especially on trays with- protection in preventing tray damage in a column
out inlet weirs, vapor may bypass a large number with liquid up in the lower tray section.
of trays by going up the downcomers. A partial
bypass of vapor up a damaged bottom seal may Inlet Weir
cause periodic flooding and dumping of the above
tray which might propagate up a number of trays. An inlet weir should be provided on the bottom
25% of the trays to prevent vapor bypass up
Construction of Bottom Trays downcomers and possible flooding and dumping
propagation should one of the bottom trays be
For columns prone to tray damage, construct the damaged. Inlet weir are necessary even for valve
bottom 25% of trays to be stronger than the above trays since they do not completely seal. It is
trays. The design standard is approximately especially important for sieve trays. The bubble
6 in. of water differential pressure, which, if cap trays do not need inlet weirs since they
it were increased to 18 in., would be consider- retain liquid on the tray.
ably more capable of handling stresses produced
by high vapor velocity or the liquid flooding Fatigue Failure Due to Tray Vibration
load. Since the usual trace spacing is 18 in., a
completely flooded tray with no counter vapor Fatigue failure due to tray vibration usually
force would not rupture the tray in the downward occurs at low vapor rates (near the dumping rate)
direction. at pulses of 1 to 50 Hz. This does not seem to
be a significant problem. Educating the oper-
Measuring Liquid Level in the Column ating personnel as to the nature of the problem
should be adequate. If needed, instrumentation
A liquid level differential pressure measurement can be installed for analyzing frequencies of
should be provided for the bottom 25% of the vibration.
column. This would give the operator critical
information needed as to the degree of flooding Additional Measurements for Analyzing
in the column and help him in deciding the best Column Problems
method to use in reducing the liquid level. For
example, if the liquid were only above the bottom Install five or more sample taps and thermowells
tray, a reasonable approach would be to ramp the at various locations on the column. The loca-
steam to the column as the means for reducing the tions should be determined by the temperature or
liquid level. On the other hand, if the liquid composition profile. The sample taps should be
is up 25% in the column, the course of action capable of being used to both measure pressure
would be to turn off the steam to the column (if and to sample the liquid on the tray. The tem-
not already off) and to pump the liquid out of perature profile should be measured and at least
the column to the normal base level. be available on demand.

Pumping Out Excessive Liquid Depending on the particular situation, the

sectional differential pressures can be measured
Provision should be made for easy diversion of continuously and recorded in the control room; or
bottom liquid to either the feed supply tank or when a problem occurs, that a portable differen-
to other temporary storage, so that the operator tial pressure instrument can be used on a one-
will freely use this method of reducing the time basis to measure a sectional differential
liquid in the column to its normal level. This pressure. Also, a liquid sample could be
diversion or pump-out system flow should be obtained and analyzed for use in diagnosing an
actuated from the control room. apparent column problem.

Flow Control of Steam to Column

Nuclear Radiation Scan
The steam to a column should be automatically
flow controlled. The pressure of steam upstream On critical columns, especially the large-
(of the flow measuring device) should be pressure diameter columns (greater than 5 ft), a nuclear
controlled so that a constant pounds-per-hour of radiation scan to obtain the density profile up
steam will be sent to the column. The set point the column would be useful in diagnosing problems
to the steam flow controller should have a ramp such as tray damage, pluggage, or foaming.
function so that when the operator increases the Ideally the column should be scanned empty and
steam rate to say 10%, it will increase to 10% again scanned when operating under normal condi-
over a given period of time. The ramp rate tions. Then, if and when a problem occurs, an
should be adjustable. A set point rate limiter additional scan can be run to help isolate the
should also be provided if the controller is used nature and region of the problem. It is
in a cascade control loop. This will prevent an important that the problem be defined because of
automatic controller, such as a temperature the large cost of a shutdown.
controller or level controller, from rapidly
increasing the steam rate. If the equipment is The most frustrating problem is that of foaming
available, the rate of change of steam flow in the column. This can easily be interpreted as
should be measured and alarmed. A bypass should tray damage or plugging which would call for a
be available on the ramping function so that the shutdown. However, the solution to foaming is to
operator can initiate an emergency shutdown of add an antifoaming agent or eliminate the cause
steam. The maximum valve opening to correspond of the foaming.
42 W. R. Ellingsen


Frank E. Rush, Du Pont Company, Wilmington,

The survey indicated that about 12% of the inci- Delaware, was especially helpful in providing
dents of tray damage that are operational based cause and effect information on foaming in
(as distinguished from mechanical failure or columns. He also reviewed the paper and made
corrosion), are related to having calandria valuable suggestions. Noel G. O'Brien, Du Pont
liquid level that extends above the bottom trays Company, Wilmington, Delaware, provided infor-
in the columns. High liquid level in the bottom mation on column construction and references on
of a column is an easily recognized condition, tray failure due to vibration and other problems.
and can be remedied if dealt with carefully.

A careful diagnosis must be made to determine the O'Brien, N. G., and Curran, B. F., Du Pont
cause of column misbehavior since some causes, Company, Wilmington, Delaware, Engineering
such as control loop instability, foaming, and Department Report No. 15576, January 1978,
others, may exhibit the same phenomenon as caused "Mechanical Failures of Seive Trays - Survey
by tray damage. Installing some additional of Tray Vendor Experience".
column instrumentation will aid in diagnosing
problems but, in addition, will provide useful Morehead, P., "Pulsation Studies", Glitsch, Inc.,
operating information. November 1980, Report No. 1-80.
Copyright © IFAC Control of Distillation
Columns and Chemical Reactors,
Bournemouth, UK 1986


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

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

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

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 (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).

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

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.


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

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-
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.

Single control loops. <

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
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
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
4 FM 3 CV4 MAN. 0 stage a digital PID contro lier was developed. At
5 TS2 - CV5 AUTO 2390 f,700
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
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 47

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 ad
1 1
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.


The plant is built to satisfy the requirements of

the fire and labor inspection authorities. Due to Fig. 6. Safety valves in the heat pump section,
the magnitude of the system considerable efforts
have gone into obtaining a relatively safe system.

ο The basic security requirements must be full-

Passive security measures. filled before anything may be done. On this
level pressures, concentration of explosive
The plant is physically spread over several loca- gasses, ventilation of the compressor enclosure
tions. The column section ( corresponding to fig. and cooling water to the compressor system are
1) is placed in a 16 m high tower building ap- supervised.
proved as a fire hazardous area, using complemen-
tary equipment. The compressors, which do not have ο If the basic requirements are ok, most of the
explosion proof motors are placed in a separately 'passive' equipment along with the heat pump
ventilated room, serving as a sound and fire bar- system may be operated. The 'passive' equipment
rier . consist in valves opening the different pas-
sageways of the system, but not allowing heat
Security valves are placed in the column section into the system. Tank levels are monitored and
(see Fig. 5 ) : in three positions around the top of valves leading to a tank may not open if the
the column system and in one position at the bot- level is too high.
tom. The outlets of these valves are connected to
a common venting and draining system, which vents ο The 'active' parts (pumps, steam valve) of the
to the environment above roof and drains to the equipment only may be operated if not blocked
extra tank. Furthermore the drain system of the downstream. This means that a pump is not al-
floor is connected to this system, in order to lowed to run if the downstream tubing is not
prevent liquid losses into the public sewage sys- open to a larger vessel, and that the steam
tem. The heat pump section is equipped with secu- valve only may open if one of the feed valves
rity valves to the environment (see Fig. 6) from of the column (BV40-BV44) is open.
both the high and the low pressure side of the
compressors. 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.


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
Fig. 5. Safety valves and venting system in the initial tests have been performed with only the
column section. simple PID controllers working, and with a limited
amount of automatic data storage.

Active security measures. The plant has been started on several occations by
means of the preheater (HEFS) and kept running at
The plant operation is monitored by a commercial total reflux only by means of the heat pump. How-
Programmable Logic Controller. The strategy behind ever many of the simple loops are highly nonline-
the active security system contains two lower ar, necessitating tuning upon change of operating
levels of requirements to be satisfied before point. Consequently they are obvious choices for
allowing entering energy or material into the some type of adaptive or automatically tuned con-
process : 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

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 ) .


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.


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.

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.

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


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


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

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

50 A. Shakouri

methods (e.g. linear quadratic control: Kwakernaak The diagonal dominance can be graphically inter-
and Sivan, 1972; Ten Hacken and Van Wijk, 1984). preted via an Array of the Nyquist plots of the
diagonal elements Qi^(s) with, in each frequency
This paper is organized as follows. Firstly the DNA point, a circle with a radius equal to dcj(Q(s))
design method is summarized, after which the or drj(Q(s)) for column and row dominance respec-
decoupling algorithm for the compensator design is tively; these bands of circles are called the
presented. The process under consideration is des- Gershgorin bands. If the point (-1,0) does not lie
cribed briefly and the off-line simulation results within the Gershgorin band, diagonal dominance is
are presented. Finally, the performance improvement obtained. In case R(s) is (either row or column)
achieved is evaluated and a number of conclusions diagonal-dominant the Nyquist stability criterion
are drawn. can be used for evaluation of the stability of the
closed-loop system by application of the following

The notation used in this section corresponds to η, - η (5)

h q
that adopted previously (Rosenbrock, 1969, 1974). i=l
The control configuration considered is given by
Fig. 1. The open-loop transfer function matrix Q(s) with n-j_: the number of encirclements of the
is defined as Gershgorin band of Q ^ ( s ) of the point
(-1,0) in the Nyquist plot
Q(s) = P(s) C(s) = L(s) G(s) K(s) C(s) (1) n^: number of unstable poles of H(s)
nqi number of unstable poles of Q(s)
with G(s): η χ η process transfer function matrix
L(s): η χ η post-compensator transfer function If Q(s) is not diagonal-dominant, the above-men-
matrix tioned rule cannot be applied. The Gershgorin bands
K(s): η χ η pre-compensator transfer function and the Ostrowski bands can be plotted in the
matrix frequency plots to evaluate the system interaction,
C(s): η χ η diagonal controller transfer func- for the open-loop and closed-loop process respecti-
tion matrix. vely. For the design of the controllers use is made
of classical criteria like gain and/or phase


-J^S)-^ C(S) - { - » . K(S) L(S)
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-
. the ability of the designer to manipulate the
Fig. 1. Feedback structure of multivariable compensator parameters with respect to some
control system using DNA method interaction criterion.

The design of a controller by the DNA method is The order of the compensator is a compromise: on
split into two phases. Firstly, the interaction in the one hand the order should be kept limited for
the process model G(s) is reduced over a pre-speci- implementation reasons; a high compensation order,
fied frequency region with the use of a pre-compen- on the other hand, gives a better reduction of the
sator and/or post-compensator. In the next section interaction and allows better control performance.
this compensator design is discussed in more This emphasizes the need of a method for automatic
detail. The decoupled process P(s) can then be generation of compensator matrices, resulting in a
controlled by a diagonal controller C(s). The diagonal-dominant system for a given compensator
stability of the closed-loop transfer function structure. Thus, different compensator structures
1 _1 can be evaluated in a fast and efficient way,
H(s) = (I + Q(s))" Q(s) = R(s) Q(s) (2) enabling the above compromise to be found. The
algorithm proposed therefore utilizes a numerical
with R(s): η χ η return difference matrix optimization method to minimize the system inter-
action, defined as a function of the compensator
can be investigated via the following procedure. parameters.
Use is made of the property of diagonal dominance,
defined by 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
|l+Q ..(s)| > E | Q - - ( s ) | -dc.(Q(e)) (3)
1 IJ J
I = . parameter optimization
. evaluation of the interaction reduction
for all columns and over the pre-specified . design and implementation considerations.
frequency range.
In the first step use is made of an automatic pro-
Equation (2) considers "column diagonal dominance"; cedure to detect the maximum frequency of interest
"row diagonal dominance" is defined in a similar (Leininger, 1979b). A good initial pre-compensator
way 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
|l + Q..(s)| > Σ |Q ..(e)| = dr.(Q(s)) (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
again over the pre-specified range. 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
) = max
Ω i=l
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,
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 Given a compensator structure for a column, the
consider the case of the design of a pre-compensa- parameters which minimize the criterion are deter-
tor K(s) in order to reduce the column interaction mined. If the interaction reduction is not suffi-
of P(s). L(s) is taken as the unity matrix. Reduc- cient the order is increased until a satisfactory
tion of interaction in the columns, as opposed to reduction is found. The procedure is then repeated
reduction of interaction of the rows, is less for all columns, after which one can commence with
complex in that it limits the number of parameters the controller design. The success of any numerical
in the optimization procedure. The columns of the optimization method depends on the shape of the
compensator can be designed independently. Other contours of the performance function and the
cases, like post-compensator design, follow concep- convergence properties of the algorithm employed.
tually the same procedure. In an attempt to reduce the impact of the contour
irregularities the variable-matrix algorithm is
The following parametrization is used for each used (Powell, 1981). By using this method the
compensator element Kjj(s) to satisfy the above constraints on the parameters are translated into
requirements some analytical function to eliminate the need to
use constraint optimization.
a . . (1 + a . .s)
n 0 CO)
) (1 +ea. . . s I +L aL -. . s
- _ o y

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
with a
kj i> 0, k = 1, 2, 3.





Fig. 2. Simplified scheme of the HVU


> T
From a number of cases of application in practice
it has been found that a general structure for a Here the relations between manipulatable variables
distillation columns exists, which can be expressed F (for instance reflux flows, reboiler duties,
as follows: 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) =
G(s) ( ' 2 6 ( e ) )
3 4
This structure proves to have a number of charac-
3( S7) / \F 7(s)/
teristics influencing both the modelling procedure /t o(S)\
and the final control scheme. The first part has
H(s) 3
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 The interaction in the system can be found by
reduced by making use of a pre-compensator, thus inspection of the column interaction indices Fj_
allowing the use of single-input/single-output PID (i = 1, 2, 3) as defined by equation (8). Figure 4
controllers. To overcome the adverse effect of time shows a decoupled first column; Fj close to 0 over
delays on the control performance Smith predictors the entire frequency region of interest. The second
or quality estimators can be used (Smith, 1957; and third loops, however, have a high interaction
Alevisakis and Seborg, 1973); intermediate index.
variables are used to predict the actual qualities.

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



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

O I(NR) E L . are necessary to enable the compensator to be
implemented in an integrated control system; these
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.
0 0 2 40 36 2 04 0 36
F I. Ga6
Alevisakis, G., and D.E Seborg (1973). An extension
of the Smith predictor method to multivariable
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
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-
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.
O I (NR) E L . Powell, M.J.D. (1981). Some Properties of the
) 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
/ 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
Lo SAE M P. No

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
T30. 2°c 6.5
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
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
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^-)

IL 4 S

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

Master controllers

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


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

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


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-

lowing examples will be illustrated below:
There is an ongoing trend in the process indus- i. periodical changement of electricity costs
try towards generating utilities for one*s own ii. steam-boilers in parallel
use. Even cogeneration of steam and electricity iii. cogeneration of steam and electricity.
becomes more and more a normal situation. Usu-
ally several parallel units have been installed i. Periodical changement of electricity
(sometimes historically determined). When opti- costs.
mizing such powerhouses, it is quite obvious to If the process is connected with the national
choose the demands (dictated by the process- grid, there may be periods of increased rate
management) as independent variables. Optimiza- e.g. day/night or week/weekend. Lowering the
tion of the load-distribution between the par- consumption during increased rate and visa
allel units is then possible, and 2 to 5 per- versa leads to optimality, see Fig. 1. The mean
cent saving in primary energy is attainable. value (F) of the feedrate of the processes has
For relatively small utility plants, non-linear to meet the planned amount. Restrictions are
programming is suitable (Cho and Blevins, 1981 twofold (this also goes for ii and iii).
or Draga, 1983). If the structure of the utili- a. There must be a certain overcapacity
ties is more complicated, linear programming <F > F ) .
(LP) techniques are more adequate (Boenink X
b. Storage capacity of products and intermedi- 1984), in order to reduce the amount of ates has to be adequate.
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.


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-


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. DESCRIPTION OF THE APPARATUS
If the Kth boiler (see Fig. 2) is only partly
loaded, the best operation is to use K-l and Κ In order to study continuous distillation, a
boilers alternately. 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, 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-
. PRODUCTION OF STEAM rates (pumpsteering), reflux-ratio (time-modu-
lation of the solenoid valve) and three heat-
FIG. 2 Fuel costs as a function of the amount of full-loaded
inputs (electronic regulation of the current to
the electric heaters).
boilers (K) in use.
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.
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- COOLING WATER
ables) .

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.




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

at constant demand of electricity.


Fig. 4 Lay-out of the distillation column.

Distillation Column with Time-varying Energy Costs 57

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

TABLE 1 The operating points of the column temp, feed/power feed 0.675 180 70
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

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.


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

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 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
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

Some remarks: The correction consists of a time shift:

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

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-
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

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

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

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

utilization in total chemical plant operation.
Journal A, 25_ 168 - 176.
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).


Fig. 7 Switch-over behaviour from low to high

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

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


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

Keywords. Computer control

On-line operation
Oil refining

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

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

62 M. L. Sourander and S. Gros

product finishing. This means difficulties in of both crude columns and thus propagate dis-
the control mechanisms but also possibilities turbances in the overhead system and the
for additional benefits from interunit subsequent fractionation columns back to each
manipulations. crude column mid section.The lowest side draw
Hence also a refinery wide, closed loop optimi- product in any column, heavy gasoil or cutter
zation system, developed within Neste OY, was stock is maximized against a number of
installed in the Naantali computer, but is not constraints using special logic in the constraint
described in this paper. controller. This is because of the sensitive
contamination of the product by entrained residue
The control objectives for the refinery section when overflash approaches nil. Maximization is
which is described here are as follows: economically justified because of the substantial
value difference between sidedraw and residue.
1. Maximize recovery (or yield) of every
distillate product against its most Pumparound control
critical quality property
Since vac. overhead is returned to crude unit 1
2. Allocate energy optimally to emphasis pumparound as an additional feed, also
separation vs. energy interaction disturbances in each crude unit sidedraws and in
vacuum unit are propagated back to the crude unit
3. Operate within the equipment 1 midsection. Similar disturbances enter from
limitations and drive against the other units to crude unit nr 2. The above
constraint which represents the biggest underlines the importance of stabilization and
profit optimization of the pumparound section in the
column. The special features of pumparound loops
4. Maximize control responsiveness by make the conditions of stabilization difficult.
use of proper control structure, The stabilization of the pumparound duty, for
implicit decoupling and dynamic element example, does not stabilize the internal
conditions of the column. Therefore the
5. Retain control robustness to emphasis calculated internal flow rates at some trays
the predicted process condition above the pumparound return are controlled to
variations targets by manipulating a pumparound flow
controller setpoint and a bypass flow
simultaneously. The total pumparound duty is
CONTROL STRATEGY stabilized by adjusting the heat returned to
crude preheat when changes in other pumparound
Product quality control heat exchangers (light naphtha fractionators
reboiler) or elsewhere occur.
In fractionation of petroleum, different column
products represent distinctive cuts of the In order to assure operation inside feasible
feedstock which are measured by (standard range of the operating window, internal flow
laboratory testing) distillation curves. The rate min/max constraint control features are
split between two adjacent cuts is determined by added to internal flow controllers.The
the cutpoint between them. The cutpoints can be constraints are, besides the physical equipment
calculated by using column direct measurements limits, heat exchanger maximum duties, valve
and they are used for every petroleum fraction positions, temperature limits etc, also the
as a primary quality property to control the column flooding and downcomer back-up limits.
quality of the lighter product heavy end.

The quality control consists of yield and Pressure control

balancing control, cutpoint control and analyzed
property control as cascade type logic as shown Pressure minimizing constraint controller is
in the Fig. 2. The sidedraw product yields (in employed to ride the constraints affecting
percentage of column feed) are stabilized using pressure (and internal flow rates). During
dynamically compensated yield controllers and constrained operation, pressure setpoints are
feedforward controllers to decouple the relaxed dynamically in the fastest way, while
propagation of disturbances further than to the the internal flow rate controller reacts slowly
adjacent product below. Cutpoints are controlled to the constraints that affect both the vapor
to targets by manipulating the yield controller rate and the pressure. The longer term economic
setpoints in order to maintain constant product impacts that need interunit calculations is then
fraction width. The cutpoints in their turn are performed in the energy optimizer, which
adjusted to keep constant analyzed product calculates the optimum internal flow rates and
qualities (Fig. 2 ) . These controls are signifi- column feed temperature simultaneously. Observ-
cantly slower than the cutpoint controls because ations show that in our cases the flow rates
of analyser and process delay times. The two have been maximized within other limitations,
crude units are identical in control structure. while sacrificing the pressure. Our steady state
simulations, however, show that this somewhat
The overhead product of the atmospheric crude unique observation is no more necessarily valid
units is further distilled to a number of if a vacuum unit operating at minimum floating
fractions in downstream units and has no pressure is used instead. Then the optimum
distillation range requirement.This fact, the operation would also be much more sensitive
overhead system in two stages and the overhead to the relative product values and the value of
noncondensible gas being charged to the vapor fuel gas. This underlines the need to use energy
recovery compressor gives the oportunity to optimizer, capable of taking into account the
minimize the overhead light naphtha for the refinery structure and economics.
benefit of heavy naphtha without the ordinary
drawbacks of light overhead product. Therefore Energy versus yields
the overhead temperature minimizing constraint
control is used. Note that light naphtha is fed In a crude distillation column the separation
to the debutanizers, light naphtha splitter, must be understood as a function of the
deisopentanizer and depentanizer in this order. distillation curves. This is illustrated in
The reboilers of these use the pumparound heat Fig. 3. The figure shows the crude oil true
Integrated Fractionation Processes in a Petroleum Refinery 63

boiling point distillation curve and the is the absorbed heat in crude heater fired with
corresponding product TBP-curves. If the fuel gas or oil. The feed flashes and generates
separation was complete the product distillation internal vapor flow, the volume of which expands
curves should coincide with the crude oil curve while meeting internal reflux and becoming
for the distillation product in question. A low lighter at every tray. The vapor rate can
separation gives distillation "tails" containing decrease only by removing heat from the column
high boiling point fraction (which for the product at the pumparounds. The column geometry requires
for instance LGO will result in a low cloud some heat to be removed and also the overhead
point). A higher separation gives a smaller condensers are sized to match this.
"tail"(and a better cloud point) which in turn
enables a better yield on account of the product
yield below. Optimization of the pumparound duty

Generally the pumparound duty can widely vary

PARAMETERS AFFECTING THE SEPARATION without exeeding column constraints. A decreased
duty causes increased vapor flow in the upper
It is important to know which parameters are part of the column, which in turn improves
possible to manipulate in order to affect the separation and the yields of the lighter products
separation efficiency. These parameters are the above the pumparound at the expense of the lower
variables which are manipulated when the objective products. The pressure effect will cause an
function is maximized. In the following a crude increased long residue yield and the preheat
distillation system is analyzed. changes of the heater feed will cause an
increased fuel gas usage. Interaction to vacuum
The column pressure affects the separation, due pumparound exchangers will increase their duty
to the increased relative volatilities at affecting favourably the heavy vacuum gasoil
decreased pressures. In our case the optimal yield. Steady state modelling technique with
operation is minimization of pressure against feedback information from the unit to adjust
constraints as for instance the maximum control the models for successively increased accuracy
valve position, flooding etc. is used. The final feedback on long term is the
product laboratory distillation curves which
The stripping stream decreases the hydrocarbon are used on daily basis.
partial pressure and thus has the same effect
as the above discussed column pressure. The Optimization of column feed temperature
optimum stripping steam ratios in crude columns
and side strippers have been investigated through In these columns the other possibility to affect
steady state simulations. The hydrocarbon to energy usage is to adjust the column feed
steam ratios are maintained through ratio temperature. This optimization is even more
controllers. complex because it is dependent on the factors
described earlier, overflash recovery in crude
units and in vacuum unit and vacuum unit heater
The internal flowrates are important parameters operation credit and the optimum value of the
when trying to affect the separation. Increased pumparound duty. The optimum column feed
internal flows leads to a better separation. temperature is calculated and implemented via
The crude columns have one pumparound and one feed temperature constraint controllers.
top reflux with both warm and cold reflux, see
Fig. 2. If the pumparound duty is decreased
the top reflux duty will automatically increase. Vacuum unit optimization
This will lead to increased internal flows in
the upper part of the column and subsequently The vacuum unit has total draw-off trays for
also to better separation in the upper part. products and the unit is operated against a
The separation below the pumparound is not fixed minimum overhead product temperature.
affected. This shows that if higher yields of Light and heavy vacuum gasoil are blended and the
lighter (and in our case more expensive) products heater operates at its maximum metal
are desired the column should be cooled mainly temperature.Therefore, in our case, the energy
by top reflux. A high feed temperature also optimization is straightforwardly maximization
leads to higher internal flows, and better of heavy gasoil pumparound duty (at the expense
separation. of light gasoil pumparound duty) against
operating constraints. Note that Hy.VGO is
maximized against its colour and other
OPTIMIZATION OF ENERGY RECOVERY AND YIELDS constraints by manipulating the washflow back
to the column, whereas during normal operation
The objective function to be optimized is the the overflash is returned to the feed.
sum of product income and net energy cost.

High feed temperature results in an increased Light ends fractionation optimization

energy consumption. Cooling in the upper part of
the column using cold reflux means that energy In light ends fractionation chemical compounds
is transferred to water instead of preheating can be analyzed and defined as light and heavy
crude, having the same effect. key components and treat the column as pseudo
binary distillation. For example, in debutani-
To be able to calculate how much extra energy is zer i-pentane and in deisopentanizer n-pentane
needed, rigorous simulations of the heat exchanger is the heavy key component. In all these columns
network have been carried out (see Fig. 4 ) . composition control hard target is the distillate
Also rigorous distillation simulations have been product heavy key composition, which is
used to adjust the functions describing the controlled according to specification, by
yields change with changing pumparounds, top indirectly shifting the column material balance.
reflux and feed temperature. The soft composition target at bottoms product
is allowed to have more variations and is
In crude and vacuum units (with no reboiler) all indirectly controlled to a long term optimum
the primary energy for fractionation is entered value by using energy optimization which
into the column as feed enthalpy and its source 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.
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


1 11




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.


Ο». Ο



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


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

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.


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

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

68 D. B. Aldren and P. J. Bujac

losses to the surroundings. The heat balance for

DYNAMIC MODEL the gas flow assumes a well mixed CSTR. The
MACRO has been configured to allow additional
The dynamic simulation language used to model the N2 and 'B' flows at specified inlet tempera-
plant is the 'Advanced Continuous Simulation tures. A heat flowrate term has also been
Language , ACSL (ref 1 ) . The language is a high included to allow for catalyst carry over, say
level, FORTRAN based, user friendly package which from the fluidised bed into headspace. In
may be used interactively to give on-line control practice this facility was not used.
of the simulation. The models are prepared from an
equation or block diagram description. Libraries 3. TRANI0: models a temperature transmitter as a
of commonly used model elements have been created first order lag with a selectable time constant
within ICI enabling complete models to be easily and This was used for temperature measurements
quickly assembled using a building block approach. throughout.
The "lumped modelling technique, which assumes
homogeneous elements, has been used extensively; with 4. C0NTB: models a three term controller with
distributed systems being modelled by dividing into selectable tuning parameters. This was used
a finite number of lumped elements in series. The for temperature control of the heat tracing
dynamic elements of the model include the reactor on the transfer pipe and secondary reactor.
headspace, transfer pipe and the secondary reactor.
5. MOTOR: models a valve actuator with selectable
Main Assumptions stroke time for the Nitrogen control, valve.

As with all models, there is a compromise between too 6. GVALV: models a gas valve with selectable
simple a system description and too much complexity characteristics. This was used for the addi-
which leads to excessive computation time and over- tional Nitrogen flow control to the fluidised
specification of parameters. To achieve a time and bed. The valve sizing gives a maximum
cost effective model, certain assumptions have to Nitrogen flowrate of 100 mass units/hour
be made. In this model the following assumptions
were made Additional features included in the model are:

a A simple non-dynamic and empirical model was a Continuous calculation of the component distri-
chosen to describe the fluidised bed. This model bution from the fluidised bed allowing the bed
was based on previous laboratory and kinetic studies and feed flowrates to be changed during
and describes the distribution of major reaction simulations.
components leaving, the fluidised bed. The model
related the concentration of 'X' to the primary b The inital section of the model calculates,
parameters, bed temperature, residence time, and the Arrenhius constants, the initial concentra-
'Β/Α' feed ratio. For a given 'X' concentration, tions for each cell, the average specific heat
the distribution of Ύ ' and 'Ζ' were calculated. of the gas, and converts flowrates from mass
to molar.
b Reactions 2 and 3 are described as first order
Arrhenius type expressions The main inputs to the model include

d(Y) = - djX) = A2 exp (-E/RT) (X) (B) Flowrates for A, B, C and Nitrogen
dt dt Bed temperature
Wall set point temperatures
d(Z) = - d(Y) = A exp (-E/RT) (Y) (B)
dt dt
Typical outputs include for each cell

where (Β), (Χ), (Y) and (Z) are molar concentrations Product distributions
Ε is activation energy, assumed the same for
both reactions Gas temperatures
Τ is absolute temperature
R is the universal gas constant Wall temperatures

c The reactor sytem was described as a series of SIMULATION STUDIES

continuous stirred tank reactors (CSTRs) or well
mixed vessels. The flow pattern in the reactor The model has been used for both steady state
disengagement headspace will be complex; a simple and dynamic simulation studies. Steady state
2 cell model was however considered to be adequate simulations may be carried out by effectively
on the basis of previous plant data. Although the ignoring the large time constants of the reactor
transfer pipe and secondary reactor are more properly walls and insulation eg by setting wall and insu-
described by a plug flow model, most of the reaction lation densities to small values. Steady state
is complete before the gases reach this section and results have already been used to determine stable
no significant loss in accuracy should occur if a operating points for the plant to give improved
CSTR model is used for each unit. This simplification product yields.
allows the definition of a basic model element (the
cell) to describe the various reaction stages. This report is concerned with dynamic simulation.
Dynamic simulation runs were carried out
The basic cell
a on the existing plant control system to obtain
The basic cell with typical input and output is shown a better understanding of the response of the
in Fig 3. The model consists of the following reactor to changes in the main control
specific elements (MACROS) and standard elements from variables,
the ICI libraries.
b on alternative control systems and potential
1. REAC: describes the reaction kinetics for a CSTR plant developments
for reactions 2 and 3. In addition the MACRO
calculates the heats of reaction and component For these runs, the model was used to describe the
concentrations. system response, from steady state, to a step
2. ΗΡΑΝ: describes the heat and mass balance for change in one of the main operating parameters.
the gas flow reactor wall, insulation and heat
Complex Gas Phase Reactor System 69

The base case operating point used in the majority a 'Β' flowrate of 3 3 0 units/hour the transter
of the simulation runs is described by the following pipe wall temperature set point was increased
from 3 5 0 to 3 9 0 degC. Fig 9 shows the response
Flows to the fluidised bed A 1 0 0 Mass units/hour of the gas temperatures in the transfer pipe and
Β 3 5 0 Mass units/hour secondary reactor increasing by 2 0 degC in 0.4 hrs
C 1 0 0 Mass units/hour Fig 1 0 shows the effect on the product concentra-
N2 1 5 0 Mass units/hour tions in the transfer pipe and exit the secondary
reactor. The transfer pipe wall temperature
Fluidised bed temperature 3 5 0 degC takes approximately 0.3 hours to reach the new
setpoint. The wall temperatures must be optimised
Transfer pipe wall temp set point 3 5 0 degC as part of the overall plant operating strategy.

Secondary reactor wall temp set point 3 5 0 degC Alternative Control Systems

This base case corresponds to a final reactor exit Sensitivity trials using the model indicated that
gas composition of the headspace gas temperature could be controlled
by several methods, with improvements in the over-
X 3% molar all control of reaction. The most practical
Y 50% molàr and simple method is to control the flow of
Ζ 20% molar diluent Nitrogen to the bed. The advantages of
Others 27% molar controlling the nitrogen are; it does not affect
the stochiometry of the reactions, it is readily
Extensive sensitivity trials have been undertaken available a reasonable cost and presents no post-
to examine both over and under reaction. The processing problems. The model has been used to
simulations included in this paper give an indica- examine the effects of controlling the upper head-
tion of the plant responses and the main interactions space gas temperature by adjusting the nitrogen
flowrate between 0 - 1 0 0 units/hour above a base
Gas Temperature Control flowrate of 1 2 5 units/hour, required to fluidise
the bed. The controller set-point has been
The headspace gas temperature has no direct closed initiated to approximately .the value found in
loop control system on the existing plant. The previous results for the same 'Β' flowrates. A
temperature is manipulated by adjusting the flow simulation run with 'Β' increased from 3 3 0 - 3 5 0
controller setpoints. units/hour has been undertaken. The controller
tuning parameters used were
The transfer pipe and secondary reactor control
systems were originally designed to control the Proportional gain = 1 0 . 0
gas temperatures by manipulating the heat input to Integral action = 4 0 seconds
the respective walls by means of electrical heat Derivative action = 4 0 0 seconds
tracing. The model was used to confirm that the
control systems were impractical due to the large Fig 1 2 shows the exit product concentration and
difference between the gas residence times and the the main waste by-product concentration increasing
thermal time constants of the walls eg for the by 0.5 and 1% respectively. The simple closed
transfer pipe the gas residence time is typically loop control system is difficult to tune to give
4 seconds whereas the thermal time constant of the improved response. Some of the factors limiting
wall is approximately 1 2 0 0 seconds. The model was the performance of the control system are; the
used to investigate the effect of reconfiguring measurement time constant of 1 5 sees, the valve
the control sytems to maintain constant wall stroke time of 1 0 sees, the lag of the nitrogen
temperatures. Installation on the plant has pre- flow through the fluidised bed (not modelled) and
vented damage to the heat tracing from high the limited heat capacity of the nitrogen entering
temperatures and stabilised the plant by eliminating the headspace at the bed temperature of 3 5 0 degC.
oscillations in the wall temperatures. This The lags are significant when compared to the gas
modification is considered part of the existing residence time of the headspace which is typically
control sytems throughout the rest of the study. 4 0 sees. The simulation demonstrates the head-
space temperature can be controlled to improve
Response to 'Β' flowrate changes product yield. Further developments such as
introducing nitrogen at a reduced temperature
As both these simulation runs and plant experience directly to the headspace have been simulated
indicate that the system is particularly sensitive with improved control. However, no plant data on
to changes in the 'Β' flow, this response will be the effects of the mixing in the headspace are
examined in some detail. The 'Β' flowrate was available. The simple control system would
reduced from 3 5 0 to 3 3 0 units/hour at 0.5 hours require minimum plant modifications, whereas the
simulation time. Figs 4 and 5 shows the rapid more complex schemes would require significant
reduction in the gas temperatures throughout the and costly changes.
plant eg headspace temperatures drop by 3 0 deg C
in a few minutes. The gas temperature continues CONCLUSIONS
to fall at a reduced rate after 8 hours. Fig 6
shows the headspace wall temperature responses The model explains the observations from the
confirming the long time constants. Fig 7 and 8 plant that the reactor system is difficult to
show the gas component concentrations exit the control ie it is not easy to maintain optimum
secondary reactor. The initial rise in product product concentrations in the off gas. The
concentration is followed by the slow reduction as simulation runs indicate the highly sensitive
the simulation continues. The results demonstrate and non-linear response of the *B' flow, the
the difficulties experienced by plant operators rapid response of the reacting gas system and
in attempting to maximise product make at accept- the long lags of the reactor walls. With
able by-product concentrations. variations in feed rates, general "noise", and
limited off gas analyses (only every few hours)
Responses to reactor wall temperature changes it is not surprising that good steady control is
difficult to achieve.
The transfer pipe has been chosen to demonstrate
the effect of changing the wall temperature, Plant experience and the model results have
achieved by changing the controller setpoint. With 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%

Fig. 1. The Reactor System


Fig. 2. The Product Distribution

Complex Gas Phase Reactor System

CELL ( i )


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




QC .




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

IS8L_. Tgft BEL - - -

τ450.ο ο 450.
•450. G
Ό αι αι
43C. S 430, .30. Ξ
OJ 00
-· 410. -η 410. •410. ^
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

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.

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 -
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.


Ν ·

<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 -
53.0- Χ
>- 52.5 -
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


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

INTRODUCTION effective (White, 1976), although the computational

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

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:


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
= 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.

- V ° (2)
An Adaptive Control Algorithm

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

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

and Δ Τ _ = 3 Δ Τ
Τ 2 ρ 0
03 Η 9 3 3 3
4Bj A T ° a
- «°

2+ 4 4ASS(l,2) T ° a
2+ 4Σ3Τ£ 0 ] Δ Τ
2 Ρ
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 +


δΤ = C. δΤ + d, διι
g 1 Ρ 1

where _PF _
H ( T dt
3 δ Τ ρ+ d 3δη - ^
t d
["F -g>L
/ ΓΤ,ο ο 3H
4a, a T - u
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°
Κ+ J
ôu(k+l) = ou(k) - {rT°(k) - T°(k-l)1
+ 4GS(1) ΣΤ°3 ΔΤ - ΔΤ = 0
g g °Τ ρ

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
ΔΤ (K) + Κ
ρ c
4 [SS(1,1) - Α ^ ] Σ Τ
Γ °°
Τ ( ) kT k T
3 Q
- &p - set< ^ i " T?6T (k) -
P ΔΤ = 0
3Τ~ Ρ
- [Τ°0Ο - 2T°(K-L) + T°(K-2)] Δ Τ
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
Α Ε ) ΣΤ° - ^ ΔΤ
+ 4GS(1) Σ Τ ° d, δη
g ι

Κ - C
0 Δ Τρ
+ d
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
+ ΓΚ° + 2Τ°]δΤ (k-1) - Τ° δΤ (k-2) = 6u(k) yo)
c d-i ρ d p

- ÔK
cr- iν26Ίί - r ôT
3 d- SjôT (k)
- 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

and becomes
W(k) is a matrix that can be determined by the
relat ion :
d T (k) T a+ +U d +
~ 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.
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)
- x s e] t <a[i(k) - x s e] t
{ d r 6 K +Ôr T +
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)
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

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)

An Adaptive Control Algorithm 77

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


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

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

Gill, P.J. (1974). Ph.D. Thesis, University of
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


Cracking! radiantl ITERATION

Fip;. 3 PI controller optimal parameters for

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


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

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.


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

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
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 ,
emj . el
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)"
C - ( 8 . 9 1 10-6) Ρ + (4.16 ΙΟ" ) eH U
(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 (1-jS Ψ -jS Ψ )
d _c
- = D 3 C d
' V V ' c c 2 " l -
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

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
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
k 1.22xl0 1xe x0p ( - 3 1 . 3 x1 0 3/ R T ) k *χ exp(-30 103/RT)
x x A
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)
0.261(31.4wt7o) 7.4X10" X(3.48X 10 -1/P )X(0.32-C )
^ max ν A
« 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

*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
-p' O
r -A - E · € -K -C.-C (2)
1 cat
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-
cients of liquid ans gas phase respectively; r
is the reaction rate of desulfurization shown in
= 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
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 ,
E.. are the effectiveness factors for sulfur and
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.
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.
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
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
- l _ ( p -c · ν ·τ) (δ)
g pg g t_ c a c
a χ = lOOr / p (17)
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)
Hg Q (x,0) f (x)
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;
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-
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
(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

8 1.833xl0 kcal/kg T.(x,0) 658°K 2
0.7493 kcal/kg °K Ρ 140 kg/cm . G
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
PEAT 0.266xl04 kg/m 3 0.12x10" kg/kg
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
5.056 kg/m 3 system Ρ & Τ)
Ps 3.99xl0 kg/m
R 1.987 cal/g mole °K


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

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
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. -


- 2 -

0.2 - 6
5 43""""

680 h
< ^ /
/ ——•


640 I 1 1 1 1 1 1 1 1
0.2 0.4 0.6 0.8 1.0

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


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


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

88 G. Rong and J. C. Wang

A/D—D/A converter has been used not only to imple- heating voltage (V).
ment various control modes "but also to automate
batch reactor's operations such as process monit- Fuzzy selector is derived from fuzzy linguistic
oring and experimental data handling. The plant is rules ( 1 - 7 ) » given in Appendix. These rules are a
a pilot scale ( 2 0 0 0 ml) sealed vessel of stainless summary of human operator's decision making
steel, equipped with agitator, pressure and temp- strategies adopted to decide which control modes
erature sensors,cooling coil and electric heater. will be needed during different stage of polymer-
The major control variables are voltage of elec- ization. If we treated these rules by using the
tric heater (V) and cold water flowrate to cooling compositional rule of inference as described by
coil (F). The output variables, which can be mea- Zadeh ( I 9 6 5 ) , a huge decision making look-up table
sured and controlled directly, are pressure (P) would be obtained. It is not efficient to imple-
and temperature (T) of the reactor. Negative ion ment the table in a microcomputer control system
polymerization of propylene oxide (sometimes, due to large number of data involved. So we simply
ethylene oxide) produces polyether product which , but effactually, implement the selecting rules
is a kind of useful chemical auxiliary. The pilot by translating them into normal logic decision
batch reactor is designed for laboratory scale process. The translation is based on look-up table
test of development of new polyether products and relating fuzzy subsets representing fuzzy value of
for experimental investigation of kinetics of inputs to quantised input values. The look-up
polymerization processes. table has the same form as described by King and
Mamdani ( 1 9 7 7 ) . Fig. 3 shows how the fuzzy selec-
tor works.
Pressure-temperature Control of the Reactor
Fuzzy linguistic rules (20-28) in Appendix are
Usually the rate of polymerization can be contro- used to design pressure fuzzy controller in the
lled indirectly by changing temperature of the form of look-up table. Human operator uses EP and
reaction when neither concentration of monomer nor CP to affect flowrate of cold water. Only absolute
conversion rate of monomer can be measured in real value of flowrate is considered in the controller
-time enviornment. In the reactor, pressure of because the flowrate is changed in such a way that
vapor phase is directly proportional to tempera- the normally-closed valve of cooling coil's entry
ture and concentration of the volatile components. is opened in a part of the time interval of samp-
Hence changing pressure is another way to control ling and control. Positive EP is eliminated by
reaction rate. Volatile monomer propylene oxide cooling. Negative EP will be eliminated by the
will evaporate and pressure of the reactor will exothermic reaction itself when EP is small, or by
rise as the temperature increases during stage of selected heating action if EP is negative-big.
heating initiation of batch polymerization. When The fuzzy controller takes similar structure as
the pressure reachs desired value (pressure set described by King and Mamdani ( 1 9 7 7 ) . The number
point), monomer reacts to produce polyether and of elements in the universe of discourse about EP,
heat is released at a rate of 2 0 kcal/mol. Press- CP and F are 14-, 1 3 and 7^espectively. For example
ure should be maintained at its set point as long , 7 elements in the universe of discourse about F
as possible if the temperature does not exceed represent 7 levels of control action, and these
prescribed high limit. In latter stage of the 7 values are assigned to the grade of membership
polymerization, the temperature will be higher to get the definition of four fuzzy subsets desc-
than its limit in order to prevent pressure from ribing control action, as shown in Table 1 .
decline. Temperature control should be introduced Similar definitions of fuzzy subsets about EP and
and take the place of pressure control as soon as CP can be written out.
the temperature is higher than its limit.

TABLE 1 Look-up Table Relating Fuzzy Subsets

Isothermal Operation and Temperature Control to quantised Control Values

Polymerization of propylene oxide is carried out 0 1 2 k 6

3 5 7
under isothermal condition for investigation of
kinetics of the reaction. Depolymerization of PB 0 0 0 0 0.1 0Ά 0.8 1.0
polymer product has been observed in isothermal
polymerization process. This phenomenon is inde- PM 0 0 0.2 0.7 1.0 0.7 0.2 0
cated by the anomalous restoration of pressure and PS 0 Λ 1.0 0.8 0 Λ 0.1 0 0 0
concentration of monomer. Thus the task of temper-
ature control includes removal of heat when reac- 0E 1.0 0.5 0 0 0 0 0 0
tion releas heat and supplement of heat during
stage of initiation and stage of depolymerization. note ι PM (positive medium^ P3 (positive small)
PB (positive big) and 0C (nil) are fuzzy
subsets about control action.
FOR PRESSURE-TEMPERATURE CONTROL According to the procedures of design (King and
Mamdani,1977; Liu, 1 9 8 5 ) , a look-up table, Table 2
Early investigators (King and Mamdani, 1 9 7 7 ? Liu, , has been established, which represents a map
I985) reported their applications of fuzzy logic from levels of EP and CP to levels of control (F).
controller to temperature tracking in batch react-
or process and achieved successful control results. Similarly, fuzzy linguistic rules (9)-(iç;) given
in Appendix, are used to build up fuzzy controller
Motivated \sy their work, a fuzzy control system
for tracking reaction conditions like pressure and of temperature, which is another look-up table.
temperature has been considered. The closed-loop
system is shown in Fig. 2 . Three parts of the
algorithm are fuzzy selector, fuzzy temperature EXPERT FUZZY CONTROLLER
controller and fuzzy pressure controller. The
algorithm's inputs are error of pressure (EP), General Idea of Expert Fuzzy Controller
rate of its change (CP), error of temperature (ET)
and rate of ET's change (CT). The outputs of the Tong ( 1 9 8 ^ ) pointed out that expert fuzzy contro-
algorithm are flowrate to cooling coil (F) and 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- Results of Pressure- temperature Control
ssure-temperature regulation mentioned in previous
section can be considered as a kind of EFG with The implement of the fuzzy control algorithm for
hierachical structure. The high-level rules are regulation of pressure and temperature in the
linguistic rules for fuzzy selector and the low- pilot batch plant, where a poorly-understood ion
level rules include on-off control policies for polymerization of propylene oxide takes place,
heating and fuzzy linguistic rules for pressure gives pressure-temperature tracking profiles as
and temperature control. 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
Expert Fuzzy Control of Reactor's Temperature algorithm has met the technical needs, and it is
better than the performance of simple on-off
During isothermal operation of the batch polymer- control and manual control. The fuzzy control
ization reactor, it is a hard work to keep temper- algorithm can also obtain satisfactory results, as
ature at its desired value because of the non-lin- shown in Fig. 5» when initial conditions (recipe
ear and time-variant characteristics of the batch of reactants) are changed. Robustness of the
process. Fuzzy logic controller (King and Mamdani, algorithm is not difficult to find according to
1977? Liu, 1985) has been applied to eliminate Fig. 5 and Fig. 4 . Because of on-off mode of cool-
error of temperature of the polymerization reactor ing conventional PID algorithm cannot be imple-
by Rong (1986) and found that fuzzy logic control- mented; and by lack of mathematical model of the
ler (FLG) with constant parameters cannot give 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

Simulation Results of EFC PI: Kp=0.5

0.28 1.24 -I.29
Ti=1.0 min
A mathematical model of the pilot batch reactor
has been built up by Hong ( 1 9 8 6 ) , which includes EFC: use rule ζ-Λ
kinetic behaviou of the polymerization of propy- Rl if té [0, 41 ]
lene oxide. Continuous regulation of the flowrate R2 if té [41,386 ] 0.24 0.87 -O.61
of cold water is available in the model so that R3 if té [386,442]
the fuzzy logic controller (FLC) designed by King R4 if té 1422,600]
and Mamdani (1977) can be directly used to elimi-
nate temperature error of the reactor under iso- note:
thermal operation. Well turned digital PI contro- t—reaction time in minute
ller is also applied to control of temperature of Overshoot— (*C)
the plant. Expert fuzzy controller based on fuzzy Error-M—the worst negative ex*ror after
stage recognition and FLCs with different control 2 reaction
initial stage of the
parameters has been applied to the model of pilot
Variance^—- ^ [T(i) -Tsp]
batch polymerization reactor. Table 3 lists the
simulation results of FLC, PI controller and EFC.
In all runs the sampling time is 0.5 min.
PI controller gives less overshoot tnan FLC, but
pessimistic negative error of temperature because
of the nonlinear and time-variant characteristics 26 (in Chinese).
of the process. FLC has been turned to ensure the
Maiers, J. and Y. S. Sherif ( 1 9 8 5 ) · Applications
stability of the system and decrease overshoot.
of fuzzy theory. IEEE Trans. Syst. Man Cybem.
When the FLC works in the closed-loop system, t t
negative error of reactor's temperature is not . 1 5 . i» 175-189.
avoidable after initial stage of the polymeriza- Mamdani, Ε. H. and S. Assilian ( 1 9 7 5 ) ·
tion. EFC gives the best tracking of the reactor's An experiment in linguistic synthesis with a
temperature. The reason for EFC eliminating nega- fuzzy logic controller. Int. J. Man-Machine
tive temperature error effectively is that EFC Stud. , 7 , 1 - 1 3 ·
adopts differently turned FLCs and necessary Maroquin, G. and W. L. Luijben ( 1 9 7 3 ) · Practical
heating policies in different stages of the batch control studies of batch reactors using reali-
polymerization process. 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,
CONCLUSION 1 2 - 1 6 and CEER, 1 9 7 0 , 4 , 27-32" ~
Ray, K. S. and D.~Dutta Majumder ( 1 9 8 5 ) ·
Particularly designed fuzzy control algorithm for Fuzzy logic control of a nonlinear multi-
regulation of pressure and temperature of the variable steam generating unit using decoup-
pilot batch reactor has obtained desired pressure- ling theory. IEEE Trans. Syfit.>Man, Cybern.,
temperature tracking performance and shown its 1 5 . 4, 539-558.
robustness. Using different control rules as the Rong, G. ( I 9 8 6 ) . Microcomputer control of a pilot
polymerization progressing is a way to overcome batch polymerization reactor and investigation
time-variant characteristics of the process. of fuzzy control algorithms. Master Thesis,
Zhejiang University. (in Chinese)
Tong", R-M-.M^BocFand A. Latten (1980). Fuzzy
REFERENCES control of the activated sludge waste water
treatment process. Automatica, 1 6 , 695-697·
Amerehn, H. (1977). Computer control in polymeri- Tong, R. M. (1977).A control engineering review of
zation industry. Automatica. 13» 533-5^5· fuzzy system. Automatica, Γ3, 559-569·
Ham, A. A. and J. H. Liemburg (1975;. Comparison Tong, R. M. ( 1 9 8 4 ) . A retrospective view of fuzzy
of five temperature control systems for a control systems. Fuzzy Sets & Systems, 14,
exothermic batch reactor. JournalA, 16, 73- 199-210.
78. ' " Zadeh, L. A, (1965). Fuzzy sets. Inform. Control,
Hoogendroorn, Κ. (I98O). Control of polymerization 8, 338-353.
processes. Proceedings of the 4th IFAC Conf-
erence on PRP~Automation. ICC, Ghent, Belgium.
Kickërt, W. J. M. and""!. R.~Van Nauta Lemke (1976). APPENDIX
Application of a fuzzy controller to a warm
water plant. Automatica, 12, 301-308. Linguistic Rules for Pressure-temperature Control
King, P. J. and Ε. H. "Mamdani "(1977).
The application of fuzzy control system to The universe of discourse about decision has three
industry processes. Automatica, Γ), 235-242. values, i.e. ms=(1,2,3 } , and three fuzzy subsets
Kiparissides, C. and S. L~. Shan (1983)· Heat (heating)
Self-turning and adaptive control of a batch MS=< CTRL-P (control pressure)
polymerization reactor. Automatica, I 9 , . CTRL-T (control temperature)
Liu, M. (I985). A case design of fuzzy temperature Let ms = l means MS=Heat; ms = 2 does MS = CTRL-P;
control for chemical reactor. Control and and m s = 3 does MS = CTRL-T.
Instruments in Chemical Industry, 12, 2.* 20- 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
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
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)
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
if EP=PS and CT«(C0 or PS) then F=PS
if EP=PS and GT=(NS or NM or NB) then F=0G
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*)
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


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

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

Physical properties, easily measured, such as tem-

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

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
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)
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-
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
formed out of one disulphonic acid A and one aro- The principle procedure can be explained by means
matic compound A in reaction (R3). Therefore, the of a sulphonation reactor (King, Gilles, 1984).
yield of A will be small as long as (Rl) is not Sulphonation of aromatic compounds is usually
4 performed in a cooled semi batch-reactor. Some of
Τ the possible operating modes are given in Table 1.
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
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.
of the last innovation y is c a l c u l a t e d ^ â
Kalman filter based on model M where the innova- Penetrating water leads to a rapid increase in
tion sequence r is supposed to be a zero mean temperature. At high temperatures the decomposi-
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:
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 ..
! (a.. pJ. ( t ) ) , J7i
(5) and time constant τ.. . of Eq. (8)
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)
In on-line experiments the transition rate a
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
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- seen for experiment n , it is run with a great ex-
lar to this subsystem the remaining transition cess of A . Therefore, the increasing temperature
rates a-j_j are found. The transition rate to model
can be explained for a long time with the forma-
M tion of polysulphonic acids (R2). Hence, with nor-
2 will be proportional to the estimated amount
of penetrating water. mal operation mode. Only at higher temperatures
the normal operation mode model is not able to ex-
Bayes rule in the example. When measurements are plain the temperature development. Its probability
processed an update of the model probabilities is will decrease and an alarm will be given. This
possible. To use the statistical informations of example shows that the multiple filter algorithm
the three filters in the new disjunctive classes, is able to identify different modes in real time
Bayes rule must be rewritten. The normalizing experiments as wel1.
factor Ν in Eq. (4) is the total probability of
the innovation y , see Fig. 6.
2 r M r
Ν I p( ^ V k-1> P( il k-i> REFERENCES

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

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-
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
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


Process ^^^^^

normat operation

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

1. fault mode

Filter 2

2.fault mode

Fig. 3. Multiple filter method for fault


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

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.


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
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

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
Gain Θ r Gain θ τ

1 .281 10.0 8.0 1.51 8.0 8.0










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

T N G7
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

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
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

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
" ) (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)
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

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-
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 101


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
Fig. 2a. K K vs χ ρ of Gt's, H e and H s in eqn. 1 the moments of eqn. 1 are
ES equated to the moments of a first order transfer function
with dead time, the form of which is given as:

K -T {x )s
Fxp e(x )s+1
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).|
d{ x f )
The final expressions are:

K = 7 K9 RR+1

1 — ΚχηΚι (6)
( l - K K )
θ 50 θ 60
Loop l(Enr. Sect.) Loop 2

1 — Λ10Λ11

Fig. 2b. Bvsxp

KeKs — K\K>j = l\ (8)

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 t d { * f ) = ( t
d) 9 (9)

obtained approximately half way between the design and

perturbed conditions. For example, if the tower models are The expressions for the gains uTi's, TVs and t d q are given in
forced from xp = . 4 to .5 then the time constants should be Table 3. Equation 7 shows that the time constant of the x d
estimated at xp =.45. The results presented in this paper response due to χ ρ changes is comprised of three parts. The
are obtained by linearizing tower models at conditions half first term, labeled Loop 1, is due to the recycle structure
way between the initial and final steady state values. formed by integration of the enriching and the stripping
sections. The Loop 1 time constant refers to the response
To this point this paper has presented a review of the nu- time of the composition of the vapor entering the enriching
merical technique of Kapoor et al. (1986). Our objective section due to a change in xp. The Loop 1 time constant
is to develop analytical expressions for all the composition can be extremely large if the variable Ιχ which is the product
tranfer functions of a tower for all forcing variables. Further of the enriching and stripping section gains, K e K s i is very
we want to reduce the final complicated transfer functions close to 1.0. I n Fig. 2a it is shown that at design conditions
(for example eqns. 1 and 2 for xp changes) to a first order of high purity towers li — • 1.0 and therefore the Loop 1
with dead time model by the method of moments(Gibilaro time constant in eqn. 7 is extremely large. The Loop 1
and Lees(1969)). A reduction in order of final transfer func- time constant is the major contributor to the time constant
tions will provide simple analytical relationships between a of x d in the small neighborhood near steady state. As can
tower's dynamic model and its physical variables. The next be seen from Fig. 2a, the loop gain l\ drops off drastically
section presents analytical expressions for x d respones for from 1. at perturbed conditions, implying that the Loop 1
forcing in xp . Additional analytical transfer function ex- time constant is not necessarily the major contributor to the
pressions for other forcing variables are given in Kapoor time constant of the x d response at perturbed conditions
and McAvoy (1986). 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++
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
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 =
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
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-
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
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.


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.)
& x
b( x f ) = — —
+T +


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-
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

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

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

Hr = Reboiler
He = Condenser
Holdup Holdup


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.
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.







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 .


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


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 .



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 .

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


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.


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

106 R. Gani et al.

fer. For multiple column processes it is possible columns in the BTX process than in the butanes/
to simulate the different columns with different pentane separation process ( see Table 4 ).
degrees of complexity. The dynamic model has built-
in thermodynamic submodels based on equations of Figure 1 shows some of the composition profiles
state ( Soave, 1972 ) as well as group contribut- of the three processes. The changées of the compo-
ion methods ( Fredenslund and coworkers, 1975 ). tion profiles due to the disturbances are illusta-
Thus, a very wide variety of mixtures can be stud- ted in Fig. 2. From these steady state composition
ied by the dynamic model. Also, the dynamic model profiles, the location of the sensitive regions of
has a builtin numerical methods submodel which pro- the columns were determined by applying the method
vides robust, reliable and efficient numerical te- of Brignole and coworkers (1985). It was found
chniques ( Cameron and coworkers, 1986 ). that column 2 always had a greater sensitivity
than column 1. Table 5 shows the location of the
Details of the Problems Studied sensitive points and the corresponding values of
the sensitivity indices ( Brignole and coworkers,
Three multiple column distillation processes have 1985 ) for all the columns. The locations of the
been studied ( Murphy, 1976, Doukas and Luyben, sensitive points were found to vary ( in some cas-
1981, Hess and Holland, 1976 ). These processes es ) with respect to the dynamic response ( that
were earlier employed by Gani and coworkers (1985) is, these locations varied with time before sett-
to test the dynamic model and, for a preliminary ling down to a steady value ).
study of their dynamic behavior.
Figures 3 and 4 illustrate some of the dynamic
The first process involves a butanes/pentane sepa- responses for the butanes/pentane separation pro-
ration scheme. The distillate of column 1 is the cess and the BTX process respectively. In all ca-
feed to column 2, where, i-butane and η-butane are ses, for the butanes/pentane process, the respon-
separated. Pentane is removed through the bottoms ses from both the columns can be represented by
of column 1. The second process involves a benzene first order transfer functions ( in some cases
toluene and xylene ( BTX ) separation scheme. The with delay ). In the BTX process, while column 1
distillate and bottoms from column 1 enter as feed behaved like any other single column ( and having
into column 2. Besides the distillate and the bot- first order transfer functions ) , column2 in some
toms, column 2 also has a side product. The third cases showed nonlinear behavior ( that is requir-
process is a styrene separation unit. The feed ing second order or higher transfer functions ).
mixture ( entering column 1 ) contains toluene, Figure 5 shows the dynamic responses of column 2
ethylbenzene and styrene. The distillate of column for two different sizes of disturbances in the
2 is recycled back to column 1. Table 1 lists some same variable. The size of the disturbance causing
of the necessary details of the three processes. the nonlinear behavior depends on the type of var-
iable being disturbed, configuration of the colu-
It can be seen that the first two configurations mns and the type of variable being measured. Thus
do not contain any recycle streams while the third design and control problems can be formulated for
process has a single recycle stream. Of the first which the nonlinear behavior will not exist.
two processes, the butanes/pentane process has a
much simpler configuration. 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
ANALYSIS OF THE OPENLOOP DYNAMIC RESPONSE configuration and the location of the feed. These
factors are therefore important both for control
As mentioned above, the analysis of the openloop as well as for design of the process.
dynamic responses is made by perturbing one var-
iable at a time and observing the resultant res- Configuration With Recycle
ponse with respect to the reference steady state*
The variables that are perturbed are the variables Comparison of the steady state profiles show simi-
that are usually manipulated for control purposes lar disturbance effects on either columns and sim-
or those variables which are designated as input ilar sensitivity indices on either columns. This
( or design ) variables. In case of configurations is reasonable since a recycle stream is coming ba-
without recycle, only variables of column 1 is ck to the upstream column from the downstream col-
disturbed. For the styrene separation process, umn. Table 6 shows the reference steady state and
variables of both the columns are disturbed ( only the steady states obtained due to disturbances in
one variable at a time ). the reboiler heat duty and the feed composition.
Some of the dynamic responses corresponding to the
Configurations Without Recycle results of Table 6 are illustrated in Fig. 6. It
can be seen that column 1 behaves in a manner sim-
Table 2 shows the reference steady state and the ilar to the column 1 of the butanes/pentane process
steady states obtained from simulations tor tne while column 2 has a tendency for nonlinear respo-
butanes/ pentane separation process. These results nse ( note also the comparitively larger time del-1
correspond to disturbances in the reboiler heat ays in column 2 ). Again, by increasing the distu-
duty and the feed composition. Table 3 shows the rbances, it was possible to obtain the nonlinear
same results for the BTX separation process but behavior. But in this process, it was possible to
for disturbances in the reflux rate and feed com- obtain nonlinear behavior for both columns. In the
position. Several observations can be made from results of Fig. 6, the effect of the recycle str-
the results of Tables 2 and 3. The disturbance eam on column 1 is almost negligible. On the other
has greater effect on the downstream column than hand, the continuous change from column 1 is imp-
in the upstream column. However, the effect on the ortant for column 2. By changing the operating
BTX process is greater than on the butanes/pentane conditions however, it should be possible to incr-
process. The reason for this is the following : in ease the effect of the recycle stream on column 1.
column 1, only one variable is perturbed while in
column 2, the feed that enters, may undergo chan- One interesting aspect of configurations with re-
ges in flowrate, temperature and composition. The cycle streams is that the process might not come
BTX process undergoes greater changes because both to a steady state ( in the openloop ) if the dis-
the feeds are products of column 1. As will be turbance is too large or, if the initial state is
shown later, this is also the cause for nonlinear too far from the desired steady state. One way the
responses in this process. The steady state gains desired steady state can be reached in this case,
also show a larger interaction between the two is to control ( at least ) the flow of the recycle
Analysis of the Dynamic Behavior of Distillation Trains 107

stream ( shown in Fig. 7 ). 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
IMPLICATION ON PROCESS DESIGN AND CONTROL steady state gains and the values of the sensitiv-
ity indices, the most sensitive variables can be
From the results of the openloop dynamic responses identified.
several typical behavior can be noted :
- The downstream columns undergo greater changes If nonlinear behavior is present and by changing
than the upstream columns and because of this the design variables, it cannot be completely eli-
reason they are more sensitive. minated, then control action on the stream linking
- Interaction between the columns depend on the the two columns can be considered. This is shown
process configuration and conditions of oper- by a simple example. For the disturbance for which
ation. nonlinear response was obtained in the BTX process
- The sensitive location of the downstream col- ( see Fig. 5 ) , a simple flow control was used for
umns can vary dynamically with time. the stream linking the two columns. The resulting
- For some configurations, columns can give non- response is also shown in Fig. 5. It can be clearly
linear responses ( openloop ) due to disturb- seen that significant improvement is possible.
ances in the upstream column. The amount of
disturbance necessary to cause nonlinear res- Therefore, the nonlinear behavior can be reduced
ponses depends on the process configuration ( if not eliminated ) by controlling the product
and design variables ( related to conditions of the upstream column ( which is the feed for the
of operation ). downstream column ). The control strategy for the
- The dynamic responses in the downstream show upstream column therfore should be ( in addition
time delays and the length of these depend on to satisfying any product requirement ) to control
the process configuration, type of variable stream that is connected to the downstream column.
being disturbed and on the location of the This can easily be done by applying the modern co-
feed. ntrol techniques once the dynamic analysis has been
made and proper changes in the design have been
The effect of plate hydraulics on the dynamic beh- made.
avior has also been studied in this work. It has
been found that the plate hydraulics are only imp- If the nonlinearity of the downstream column is
ortant when highly nonlinear dynamic behavior is eliminated by the control action of the upstream
present. The dynamic response of the height of the column, then the design of any control strategy is
liquid on the downcomer for the styrene process is significantly simplified. Control techniques taking
also shown in Fig. 7. Besides the above mentioned into account time delays can easily be applied for
conditions, plate hydraulics have been found to be this column.
important during startup and shutdown operations
of distillation columns ( Gani and coworkers, 1986). Closedloop response for the BTX process. For the
BTX process, two PI controllers controlling the
Important Factors For Design And Control Problems top and bottom temperatures of column1 by manipul-
ating the reflux rate and reboiler heat duty were
Based on the observations of above, several 'acti- designed. For column 2, the temperature of the most
ons' that can result in better design of the pro- plate was controlled with a PI controller by mani-
cess and at the same time simplify the control st- pulating the reboiler heat duty. From the openloop
rategy are now proposed. The objective for perfor- responses of Fig. 4, it can be seen that the time
ming these 'actions' are to a) reduce the sensiti- delays are almost negligible. Figure 8 shows the
vity of the downstream columns to changes in the openloop and closedloop responses of the top and
corresponding upstream column, and b) eliminate the bottom temperatures of column 1 for +1% change in
the possibility of any nonlinear response. the feed enthalpy ( that is changing the feed tem-
erature ). Figure 9 shows the openloop and closed-
Reduction of sensitivity will also reduce the int- loop responses of the top and bottom temperatures
eraction between the columns. Even though the int- of column 2 for the same disturbance. The openloop
eractions can be 'handled' by modern control tech- response of the top temperature of column 1 is also
niques, the control problem will be simpler if they plotted in Fig. 9 to illustrate the difference in
are minimized. The elimination of the nonlinear the magnitude of the changes in the two columns.
responses will also eliminate the requirement of
more complex control strategies. Also, the effect
of plate hydraulics need not be considered. Finally CONCLUSIONS
the time delays should be correctly predicted so
that the control techniques can be efficiently app- The dynamic model has been succesfully applied to
lied. The question of the elimination of the time simulate and analyze the openloop transient respo-
delays or the length of this period need to be stu- nces of three multiple column processes. Important
died togather with the control technique ( or stra- factors related to process configuration, column
tegy ) being employed. design and conditions of operation have been iden-
tified as those influencing the final design and
The most sensitive variable with respect to column control of the process. Numerical results showing
sensitivity and nonlinear behavior should be first how by dynamic simulations important interrelated
identified ( a dynamic simulation can be used for problems in process design and control can be sol-
the purpose ). Depending on which is possible, cha- ved or simplified have presented. And, finally,
nges in process configuration, conditions of opera- although the closedloop responses for the BTX pro-
tion or individual column details can be considered cess show promise, more work is necessary before
in order to reduce the importance of the most sen- any general guidelines can be established for these
sitive variable ( or variables ). By dynamic simu- problems.
lations the effects of these changes can be studied.
Development Of A Control Strategy
Brignole, E.A., R. Gani and J.A. Romagnoli (1985),
A simple control problem was defined for the purp- A simple algorithm for sensitivity and oper-
ose. The objective is to keep the product specifi- ability analysis of separation process, Ind.
cation ( temperatures of the products ) at the de- Eng. Chem. Process Pes. Dev., 24, 42-48.
sired conditions. From the openloop dynamic respo- Cameron, I.T, C.A. Ruiz and R. Gani (1986) , A
nse the sensitive points of the columns and the
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
9 Table 5 Steady State Gains with respect to
Type* fresh fresh
1 °Γ 1
2 1 Reflux Rate of Column 1 and Composi-
tion of Compnent i in Plate ρ
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/
D p Distillate from column i
B.: Bottoms from column i Column 1 2 1 •0071
Column 2 3 16 •0525
Column 1 1 20 •0150
Column 2 1 22 •0139
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 ).

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.

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

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

/ 0.02 H
.609 H
i column 1
.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
γ C1 C4



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)
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 /
* (
. . . . ... _
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
_ _ _ _ _ _
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


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


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
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

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

E2 Cond.
D Vi(t) = V ( t ) + μ (Vi - V )
x x
V ( t ) = V" + k
1 x w(t)
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
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
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-
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
χ ί (+
A := I - Θ f (x(x),u(x)) ôt
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
in the implicit equation for δχ , thus obtaining
Yi + φ 1(τ) ô y i1+= Γ ^ χ ) , i = 1,2,...,η-1 ,
the linear equation: ôy
n= Γ (χ)
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
+ 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 ) )
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+
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] , The so obtained step size updating rule can be
± ±
then, if ôt > otmin, set ôt = ôtmin and further refined by currently controlling the
recompute δχ ; else, send a message and wait maximum relative increment δχ~(χ) . As long as it
for instructions (abort, reduce ôtmin). ranges between 0.01 and 0.1, we leave the updating
rule unchanged, whereas we increase the growing
The second rule helps facing a slightly less rate up to 5 times its "natural" value if δχ~(χ)
catastrophic event. To express it, we need some is less than 0.01 , while we decrease it down even
preliminary definition. Let 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)


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:
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
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 ;

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~)

ôt(x) = e^/T := ôt(0) e /
; ,
Fig. 3 : The function β (.)
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:
ôt(x+ôt(x)) = ( ^ ( Ô t ( x ) / T
[ôt(x+ôt(x)) - ôt(x)]/ôt(x) = ôt(x)/T ; m ,aô xx~ ) ) ôt(x) ,

3(v,w) := β (ν) £ M
ôt(x+ôt(x)) = (l+ôt(x)/T) ôt(x) χ 2
and the functions β (.),β2(.) are shown in
On the basis of this analysis, we tentatively set 1
Fig.2 and in Fig.3 , respectively. Otherwise,
Ôt(x+ôt(x)) = ôtmin .
ôt(0) = ôtmin
In connection with Rule #3, it is worth noticing
ôt(x+ôt(x)) = (l+3 (ôt(x)/T )) ôt(x) , that the sampling period of the output (and input)
1 max
variables is usually much larger than ôtmin . Of
where the function β^(.) is given the shape of course, the rule is in itself most effective in
Fig.2 . The deviations of β^(.) from the identity computing responses to input functions which are
are justified, on one side, by the requirement constant over relatively long time intervals (e.g.
that, in order to keep the total number of steps step functions).
within reasonable bounds, the growing rate should
not be less than 10% , even for very small values Due to high stiffness ratio, the values of ôt
of ôt(x)/T . On the other side by the constraint: produced by the updating formula of Rule #3 may
readily become quite large relative to T i . As a
result, numerical instability may show up in some
ôt(x) < ôtmax = T
xa state variable, with undamped or poorly damped
Dynamic Behaviour of Binary Distillation Columns 115

oscillations of period roughly equal to 2 ôt . An structure of the package,the memory occupation is

effective remedy is provided the next rule. 30KB at most.

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


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.


[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,
[2] R.G.E.Franks, Mathematical Modeling in Chemi-
cal Engineering, John Wiley & Sons, New York,
[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


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

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

118 Chiu-Nan Wei

Test 10 is given as an example to illustrate the Modification of the flooded condenser:

study procedures. Figure 2 shows the process The condenser would be operated at
transient when the concentration of F, the approximately 80 percent of capacity
heaviest component in the heavy feed, was rather than 50 percent as described in the
decreased from 62 Wt% to 40 Wt%. As depicted in vendor's design specification because the
the figure, the concentration of F component in operational tower pressure would be lower
the bottom product stream decreased from 99.99 than that designed.
percent to 91 percent, and the impurity of
component F in the middle sidestream increased Relocation of tray temperature control
from 80 ppm to 550 ppm. Both of the stream point (T-14) from tray No. 36 to tray
products are off-specification. Fractionation is No. 33.
inadequate. Two corrections are required to
effectively operate the tower under this upset: Changes of pipe size, control valve, and
flow transmitter range for the middle
1. Increase the firing rate to provide enough sidestream.
heat for fractionation.
Installation of one analyzer in each feed
2. Regulate properly the liquid flow from C-2 to stream: Component A analyzer in the light
C-l to provide enough reflux. feed stream and component F in the heavy
feed stream.
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 CONTROL SYSTEM DESIGN
intercolumn flow was regulated manually. As
shown in Fig. 3, the concentration of component F Process responses from the upset tests were used
from the middle sidestream decreased signifi- to guide the design of the control scheme. This
cantly, but the impurity of C-l bottom increased is probably the easiest and the most effective
(i.e., F concentration decreased). On the other way to test alternative control schemes prior to
hand, if the firing rate was increased by 20 plant start-up. Many experiments can be per-
percent of the design rate and T-14 was put on formed in a short time and the results are not
automatic, the bottom product was on- distorted by noise or other disturbances. Not
specification, but the loss of component F from only the basic regulatory control system (i.e.,
the middle sidestream increased during the properly pairing of the controlled variables and
transient period of four hours by approximately the manipulated variables), but also advanced
1.2 barrel more than that under normal operating control systems such as feedforward-feedback
conditions. This is illustrated in Fig. 4. As control system, inferential model-based control
shown in Fig. 5, all products were on- system, etc., can be designed and tested.
specification and the loss of component F reduced Furthermore, any design changes can easily be
significantly when the firing rate was increased implemented, less expensively, during the design
by 20 percent of the design rate and the stage. As discussed in the previous section, we
intercolumn liquid flow was manually regulated. have changed the location of tray temperature
This implies that the present process design control point (T-14). Figure 6 shows the actual
configuration can handle this upset if it is plant data; it indicates that the sensitivity of
provided with a better control scheme. The the two temperatures is quite different even
firing control scheme should respond to feed rate though only three trays apart. Temperature which
and composition changes, and an analyzer had a shorter deadtime corresponded to feed
controller should regulate the temperature changes, also improved process controllability.
setpoint of T-14 due to the shift of tray
temperature profile. Also a more sensitive The overall control scheme as illustrated in
temperature control point is required. Fig. 7 was designed and implemented as follows:

The other tests were carried out in the same way 1. Regulatory controls - Material balance
as the example described above. Since the control (Cheung and Marlin, 1982):
objective of the study at this stage is a process
evaluation rather than control system design, an Control Variables Regulated Variables
overall supervisory control system has not been
pursued. However, for some upsets, the columns Pressure Reflux flow
could not maintain smooth operation or could not Top sidestream spec. Top sidestream flow
produce on-specification products; these problem T-l bottom level Middle sidestream flow
areas were investigated, identified, and cor- Middle sidestream Intercolumn transfer
rected during the design phase. Based on the spec. flow
results of this study, the following conclusion T-2 bottom level Bottom product flow
was made: Bottom produce spec. Firing rate (furnace
The process design configuration as shown
in Fig. 1 is capable of producing on- Advanced controls - Model-based (dynamic
specif icat ion products. No additional reconciliation) control (Bartman, 1981):
intercolumn heat exchanger is required.
However, a better control system is needed A cascaded scheme (analyzer to temperature to
to improve the tower performance. feed forward ratio to flow) ultimately sets
the sidestream flow to control the sidestream
In addition, several shortcomings were revealed specification as shown in Fig. 8. The
in the study and hence, the following modifi- analyzer feedback controller uses a model-
cations were made to improve the process based control. The technique uses the
operability and system controllability. following relation between the heavier
impurity and C-2 tray No. 77 temperature.
Distillation Columns for a New Manufacturing Plant 119

C = a + b * Τ (1) intent to provide operator training, the program

was designed and executed in a reasonable amount
where C = impurity analysis of computer time, 0.25 sec/cycle. The model was
Τ = tray temperature then integrated with the system executive
a = constant software, the plant process control computer
b - steady state process gain system (Honeywell 4500 system) , and the Foxboro
Spec 200/Videospec instrumentation system to
Feedback control is achieved by updating the perform an interactive simulation. As indicated
constant in equation ( 1 ) . Dynamics are taken in the simulator configuration in Fig. 11,
into account by delaying and lagging the tempera- operators were trained on the same operation
ture by the dead time and time constant between consoles as those they would use to operate the
analysis and temperature. Delaying and lagging actual plant. To provide thorough training, two
the temperature synchronizes it with the current separate simulators were constructed. They are
analysis. Using the updated constant in the tower simulator and the "plant" simulator,
equation ( 1 ) , a new temperature setpoint is which integrated the tower model with other
calculated. The models required were obtained process models such as reactors, adsorbers, heat
from the dynamic tests on the simulator. Those exchangers, steam system, etc.
models were also used for the plant start-up.
The simulator allows dynamic transients from a
Using the simulator, all the control schemes were cold start-up condition to a normal (design)
developed and tested in both the same process operating condition and vice versa. A number of
control computer (Honeywell 4500 system), and possible plant upsets and malfunctions were
Foxboro Spec 200/Videospec as would be used for included and could be entered from the instructor
the actual plant. The software developed was station. Thus, operators were given training in
used for the plant start-up. It is preferred troubleshooting. The simulator provided more
that a test or evaluation uses the actual than ten initial conditions for different
hardware of the system, rather than a mathe- operating conditions and was capable of making
matical representation of the hardware. Thus, it "snapshot" recordings of the simulated plant
avoided some of the problems that are related to conditions at any time. The snapshot contained
the specific implementation or to characteristics sufficient information to allow restart of the
of the process control computer system. simulator from that point. The simulator also
provided selection of fast-mode operation (up to
We did not put all the supervisory controls in ten times of real-time) for certain slow
service when the plant started up even though all processes.
the control software was ready; all the instru-
ment validity and analyzer reliability needed to
be checked first. Therefore, the flow rate was CONCLUSION
initially on DDC (direct digital control) and its
setpoint was changed manually. As shown in Dynamic simulation can be used as a powerful tool
Fig. 9, when the heavy feed rate increased, the for operational analyses, process and control
heavy component (F) increased in the middle design assessment, and operator training in the
sidestream due to the increase of heat input and manufacturing environment. This paper presented
the failure to increase the intercolumn transfer a case study of a dynamic simulation of distil-
flow. As the operator noticed the problem, he lation columns for a new manufacturing plant to
increased the flow, but increased it a little too illustrate such applications.
much, causing the lighter components to go down
to the bottom product. This oscillation As discussed in the paper, the model has been
continued as operators had difficulty regulating used for wide ranging "what if" type of studies
the flow correctly. After two days we added to find the process operability. An expensive
temperature control on top of the flow control to intercolumn heat exchanger system which costs
regulate the flow setpoint. The temperature millions of dollars was eliminated. In addition,
controller detected the heavy component going up several design flaws were found and corrected
before it showed up in the middle product stream. prior to start-up. The simulation was also used
The operation would have been fine if the to develop and test the control systems. To
operator kept alert and correctly changed the implement model-based controls, several dynamic
setpoint of the temperature controller when the tests were conducted on the simulator to obtain
feed rate changed. When the analyzer reliability simplified process models. It has not only saved
increased two days later, we added the analyzer enormous time for plant testing, but also
controller to regulate the setpoint for the determined the best controls for initial start-
temperature controller. As depicted in Fig. 10, up. Finally, the simulator was used to train
the tower ran smoothly even when the heavy feed operators in the same environment as they would
rate changed. operate the actual plant. The start-up team
acquired many hours of realistic hands-on
The control loop discussed above is given as an training, including cold starts and upsets. This
example of commissioning advanced control loops created a feeling of confidence and eliminated
in the plant start-up. The same procedures many surprises for the plant start-up.
applied to the other critical control loops. The
plant was started up successfully with advanced All the activities described above contributed
controls and produced on-specification products greatly to the successful plant start-up. It is
within ten days. 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
OPERATOR TRAINING ensure that we are doing the right things and we
are doing them in the most cost-effective way.
Since the model was also developed with the

120 Chiu-Nan Wei


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

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

B N E (T-14
9. Increases from 62 Wt% to 80 Wt%. -.A COMPONEN
10. Decreases from 62 Wt% to 40 Wt%.
E. Fuel Gas Heating Value -.
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

Fig. 2 Response of process to a step change in

concentration of component F in heavy

Firing rate: unchanged

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




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

E (T-14

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

Firing rate: increased by 20%

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


E (T-
D Fig. 7 Column supervisory controls
0 0 20

Fig. 5 Response of process to a step change in

concentration of component F in heavy

Firing rate: increased by 20%

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

8 S8



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



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


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-

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

124 W. Marquardt

conditions. In both columns, after a step change eigenfunction, which is exactly the first deriva-
in the reflux rate, the profiles are propagating tive of the steady-state profile plus a constant.
in the direction of the reboiler in both the strip- The convergence of any initial profile to the con-
ping and the rectifying sections. Their principal stant travelling wave solution is not yet proven.
shape is conserved during the propagation. An ob- However, simulation results gives one to believe
server moving with a coordinate system linked to the existence of this property, defined here as
the front would hardly recognize any variations in shape stability.
the profiles. The main difference in both cases
is the column's sensitivity to disturbances. In Wave front shape. The shape of the spatial pro-
contrast to column H, the propagation velocity of file can be determined analytically only for a
column Li concentration profiles rapidly decreases particular choice of the equilibrium line. It is a
to zero. Complex multicomponent distillation structural property of the autonomous system and
plants behave in qualitatively similar way (Retz- therefore independent of initial and boundary con-
bach, 1986). ditions.

Wave propagation velocity. It is shown, that the

MODEL REDUCTION TECHNIQUE velocity is only a function of the concentrations
in the streams entering and leaving the column
A Rigorous Dynamic Model section. The determining equation is interpreted
to be a material balance of the infinite long
The principal dynamics of distillation are given column section.
by a simple model, which ignores hydrodynamics and
energy balance. In the case of a packed column Model Reduction Principles
balancing of an infinitesimal volume element
yields the dimensionless hyperbolic equations The wave propagation properties are used to deve-
lop a reduced model for one column section. It is
|aft - | zf = - B-[y*(x)-yl
a postulated, that the dynamics of a finite extended
, (la) system is comparable to that of the infinite long
ζ G [0,1] section discussed above.
t e [0,T]
& * Î £ - B V ( x > - y i (lb)
Due to propagation of the profiles, the model
Eqs.(l) are transformed first in a ξ-t coordinate
for the liquid and the vapour concentration χ and y system moving with a velocity w. Ignoring the va-
respectively. Β is a dimensionless mass transfer pour phase holdup, we obtain
coefficient, A is the ratio of the liquid and va- î V
pour streams in the column section and ε stands
|X- || -B.[y
for the holdup ratio of both phases. Frequently, ( 1 += w ) (x ) - y ] (2a)
ε is taken to be zero due to the small holdup in
the vapour phase compared to that in the liquid | f = A»B[y*(x)-y] (2b)
phase. The equilibrium relation is given by y"(x).
Each column section is modelled by a set of equa- The wave coordinate ξ is defined by
tions of type (1). Its boundary conditions are de-
termined by material balances of the column bound- ξ = ζ - ζ' = ζ - / w dt , (3)
ary systems, i.e. the feed tray, reboiler and con-
densor. A model similar to the one stated was used where ζ = ζ' is the ξ-t coordinate system origin.
by several investigators, such as von Rosenberg The boundary conditions of (2) are formulated in
and Hadi (1980). The distributed parameter formu- this moving coordinate system as
lation is suitable not only for the behaviour of
packed columns but also for staged columns. This x(l-z\t) = ( t ) , y(-z',t) = y ( t ) (4)
is true even where only a small number of stages X l Q
are concerned, if the mass transfer coefficient is
x^(t) and y ( t ) are the concentrations in the
adjusted. This model forms the starting point for Q
liquid and the vapour streams respectively enter-
the following model reduction considerations.
ing the column section. The essential problem now,
Wave Propagation Phenomena 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
The depicted dynamic behaviour of distillation col-
system. This assumption is close to reality, be-
umns is termed as nonlinear wave propagation. In
cause the main part of the dynamic behaviour is
mathematical physics, a nonlinear wave is defined
covered by the movement of the profile with w.
to be a spatially structure moving with constant
Hence, the concentration profiles may be calcula-
propagation velocity and constant shape along a
ted using the stationary Eq.(2). Indefinite inte-
spatial coordinate/This phenomenon can be studied
gration leads to
analytically in an idealized, spatially infinitely
extended column sectign using Eq.(l). The investi-
y = A(l+w) χ - C , (5)
gations include the development of general exist- x
ence conditions of a wave front solution, its sta-
bility, shape and propagation velocity. (1+W)dx + Co (6)
Existence. A wave front solution exists for all Βί^+ν'Ίχ) " A(l+w)x]
physical meaningful system parameters, if the non-
C and C are integration constants to be determi-
linear equilibrium line y*(x) is continuously dif- x 2
ned with Eq.(4). Hence, using Eq.(6) χ can be cal-
ferent! able and either strictly monotonically in-
creasing or decreasing in the concentration inter- culated independently of y, which itself is deter-
val covered by the wave. 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
Stability. Wave front solutions are invariant
against shifts along the spatial coordinate. This concentration profile χ(ζ) is given by the follo-
property is established by analyzing the eigen- wing implicit equation:
value problem derived from Eq.(l) after lineariza-
tion around a steady-state spatial profile. One 0 = ξ - F[xU),p] , (7)
solution is found for an eigenvalue λ = 0 and its
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
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.
are called shape parameters of the profile. To achieve this requirement the only nonlinearity
in the integrand is approximated by another non-
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-
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
maximum slope of the profile; p , which is identi-
cal to the integration constant C , determines the
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
^P^W (p +p ).ln(x-p )] + p
[(P5+Pl) ln(PrX
-χ -

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
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

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 ) =
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 W. Marquardt

long column section are determined by Eq.(15) to for w.

suit the system boundary conditions. Additionally, y ( P J - y (p
they are assumed to be quasi-steady, i.e. algebra- 9 - 1 (18)
ically coupled to the remaining state variables.
The method of weighted residuals (Villadsen and
Michel sen, 1978) is used to deduce the remaining In addition to these physically based methods, it
shape parameter state equations. This technique is is possible to take a purely mathematical approach.
illustrated by those trial functions explicitly The location ζ' is considered as a special shape
solvable for χ(ξ) as is the case for a quadratic parameter and is determined by another MWR equa-
equilibrium line approximation. The spatial trial tion of type (16).
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 The Reduced Model
equated to zero. There remains a residual func-
tion R showing a nonlinear dependence on £ and ξ The application of the propositions stated above
and a linear one on jp, which is due to the lineari- will lead us to a linear differential-algebraic
ty of Eq.(13) with respect to 9x/3t. system of equations

K(s) · s = q(s) (19)

The method of weighted residuals (MWR) provides
differential equations to calculate the remaining
with a singular matrix K. The vector s_ consists of
shape parameters p^ as follows
the shape parameters £ , the front location ζ' and
1-z' sometimes of liquid concentrations at the outlet
of a column section. The algebraic equations are
j R ( p p ^ ) ν^(ρ,ρ,ξ) dç = 0; j=3,...P . (16)
9 the result of the countercurrent flow boundary con-
-ζ' ditions. This structure is the same for column sec-
tions and complete distillation columns. Powerful
These equations are simplified by the choice of numerical techniques for the direct solution of
weighting functions Wj not dependent on the deriva- this kind of system are available (Petzold, 1982).
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
A Distillation Column Section
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 The quality of the reduced model is investigated
of order P-2 for the dynamic shape parameters. In first with a column section. The equilibrium is de-
those cases, where the integral cannot be directly fined by Eq.(10) with α = 5 . The system parameters
solved, it is approximated analytically by a suit- are A = 1.34, Β = 10 and ε = 0. The inlet concen-
trations are chosen to be χ = 0.6 and y = 0.1.
able numerical integration method, for example by χ 0
Three models are studied: model R, based on the
the trapezoidal rule.
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-
Although the method is presented for a trial func-
tion explicit in χ(ξ), it is not limited to this
les ( z , p , p ) whereas model Q uses only two dyna-
type of function. An extension to implicit trial 3 5
mic states ( z ' , p ) . It is found by simulation,
functions (for example Eq.(12)) can be easily 3
that the integration constant type ρ may be taken
achieved. The derivatives in Eq,(13) are determi-
identically as zero, without altering the model
ned by differentiation of the implicit trial
behaviour in any case. In all simulations Eq.(9)
function (7). They result in functions of χ(ξ) on-
and (18) are used for the determination of z' and
ly. Therefore the residual is also a function of
the Galerkin method for the determination of diffe-
χ(ξ) and the integration in (16) must be carried
rential equations for the dynamic shape parameters.
out oyer χ(ξ) rather than over ξ.
Fig. 4 shows the steady-state behaviour of the
Determination of the Propagation Velocity reduced models in comparison to a finite difference
solution (FD) of Eq.(l) with k equidistant grid
The last question is that of the location of the points. Fig. 4a shows, that the FD solution conver-
concentration front ζ'. Here, several approaches ges to the model R solution for increasing grid
are possible. It is known from theoretical consi- point number k. In this case, the reduced model may
derations, that the propagation velocity of the be termed better than any FD solution of Eq.(l), In
profile is determined by a global material balance Fig. 4b the models R, C and Q are compared to an
Of a column section., Using a trial function χ = FD solution with k = 250. The result corresponds
f(ç,£) this balance equation can be written as to the curve fitting analysis of Fig„ 3. Model Q
1-z' does not compare sufficiently to the FD solution,

AW**' 1
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
model C (cubic equilibrium line approximation) is
= (L^CX^TJ-XI-Z^.TJL+^Y^TJ-YIL-Z ,!)] , satisfactory. In the case of stronger boundary in-
fluences in order to compute ζ', Eq.(7), (18)
a linear-implicit differential equation for z'. should be replaced by the more complex Eq.(17) or
The sensitivity function integrals can sometimes by a MWR equation.
be solved analytically. If this is not possible,
an approximation using a finite sum may be used. The structural properties of the reduced model R
are investigated in a wide range of operating con-
Another approach is the use of the simple diffe- ditions. As an example, the control variable A is
rential equation (9) and a physically based defi- varied to move the concentration front from one
nition for w, given by the global material balance section boundary to the other. The steady-state
of a column section infinitely extended over its gains are shown in Fig. 6. p and p show an almost
boundaries. With this assumption an explicit func-
3 5
linear dependence on A in the steady state. The
tion of the limit concentrations pi and p results system nonlinearity is lumped in the front loca-
Nonlinear Model Reduction for Binary Distillation 127

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

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?
ly on the phase equilibrium of the mixture to be
(). χ 1

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


1\ FL equilibrium line 1,38 A

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

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 ) .


FD solutions
Model R S S ^ Y ^

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

Jy ® x 1,0
0.0 0,5 „ z[-] 1,0

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


A: 1,34-1,4


solution (250)