You are on page 1of 11

Journal of the Chinese Institute of Engineers. Vol. 16, No. 1, pp.

101- 1 1 1 (1 993) 101

TWO-DEGREE-OF-FREEDOM DYNAMIC MATRIX CONTROL


FOR DISTILLATION COLUMNS-
23
Pang-Yen Ho and Cheng-Ching Yu*
Department of Chemical Engineering
National Taiwan Institute of Technology
Taipei, Taiwan 10672, R. 0.C.

T.P. Chiang
Union Chemical Laboratories
Industrial Technology Research Institute
Hsinchu, Taiwan 30042, R. 0.C.

Key Words: two-degree-of-freedom controller, disturbance rejection,


dynamic matrix control, distillation control.

ABSTRACT
Generally, most long-range predictive control (LRPC) gives good set
point responses in multivariable process control. However, on several oc-
casions, little improvement in load rejection was observed when comparing
LRPC with well-tuned multiloop SISO controllers. In this study, the dynamic
matrix control (DMC) algorithm is modified to improve disturbance rejec-
tion capability. A two-degree-of-freedom DMC is proposed by incorporating
the step-response models for the major load disturbance. A load tuning fac-
tor is used to adjust the degree of load prediction. A non-linear distillation
column is used to test the effectiveness of the proposed algorithm. Simula-
tion results show that, while maintaining good set point responses, the pro-
posed algorithm outperforms the standard DMC algorithm for both the model-
ed and unmodeled load disturbances.

*Correspondence addressee
Journal of the Chinese Institute of Engineers, Vol. 16, No. 1 (1993)

INTRODUCTION

The development of two computer control


algorithms, dynamic matrix control (DblC; Cutler and
Ramaker, 1979) and model algorithmic control (MAC)
[17,21], lead to an extensive research effort in long-
range predictive control (LRPC) [5,6,8,12-14,201.
Characteristics of LRPC include: (i) use truncated step -
(or impulse -) response model to represent the process,
(ii) prediction of future outputs using the process model,
(iii) improvement of the output prediction (xOl) by Fig. 1. The open-loop system.
estimating load effect (dk) from current measurement
(x,,~) and (iv) computation of the movement of
manipulated variables by minimizing the closed-loop er- variable (e.g., L in Fig. 1). For example, a good approx-
rors [lo, 13,16,22]. Clarke and co-workers [I-31 refor- imation for the feed composition disturbance facing a
mulated the LRPC and termed it as GPC (Generalized distillation column is a steplike change in the feed com-
Predictive Control) using a low-order ARMA model, e.g., position. Stephanopoulos and Huang [24,25] proposed
CARIMA. For a truly low-order process, the GPC is two-port control to improve the disturbance rejection
equivalent to the LRPC based on a truncated step- (or capability. Wellons and Edgar [27] generalized the ex-
impulse-) response model, e.g., DMC or MAC. tended version of the analytical predictor [28] by incor-
However, for high-order chemical processes such as porating a parametric disturbance predictor to improve
distillation columns, e.g., of the order (NT+2) . NC regulatory capability. Under the state space formulation,
where NT is the number of trays and NC is the number Li et al. [9] and Ricker [19] employed state estimation
of components, the high-order step-response model gives to achieve better disturbance rejection. In distillation con-
a better process description. Simulation and experirnen- trol, the major load disturbance is the feed composition
tal results [7,8] show that, generally spealung, LRPC changes (the feed flow disturbance can be compensated
shows better set point responses than multlloop SISO con- with feedforward control). Therefore, specific load
trollers in distillation control. Moreover, little improve- models can be employed in LRPC to achieve better distur-
ment in load rejection (e.g., feed composition changes) bance rejection.
is observed for LRPC when compared to well-tuned The objective of this work is to improve the load
multiloop SISO controllers. rejection capability of Dynamic Matrix Control (DMC)
Disturbance rejection is the major concern in using the concept of twodegree-of-freedom (or two-port)
chemical process control. Under conventional feedback control. A truncated step-response model is proposed to
structure, many attempts have been made to improve the model the load change. This is a straightforward exten-
disturbance rejection capability of control systems. Mantz sion of DMC which requires only a minor modification
and Tacconi [l 11 realized that Ziegler-Nichols tuned PID of the original algorithm. When applied to distillation
controllers exhibit poor tracking, but good regulatory control, the load model for a feed composition is used.
behaviors. A separate controller is designed, in addition The degree of load prediction can be adjusted via a tun-
to the Ziegler-Nichols tuned PID controller, to provide ing factor kL. It provides full load prediction to one end
nearly overdamped set point responses. Shunta and (kL= 1) and reduces to the standard DMC on the other
Luyben [23] proposed the dual algo@thmfor sample-data end (kL=O). Therefore, the proposed algorithm is a
controllers to treat the tracking and regulatory problem generalization of DMC. In this work unconstrained DMC
separately. These types of controllers are generally term- is considered. Although the derivation is based on DMC,
ed as two-degree-of-freedom controllers [15,26]. the results are generally applicable to some other
As pointed out earlier, LRPC generally exhibits algorithms, e.g., MAC etc. This paper is organized as
good set point responses, but tends to provide less im- follows. Section 2 gives a brief review on DMC. The
provement in disturbance rejection. An obvious reason two-degree-of-freedom DMC is derived in section 3 for
is that the typical LRPC is designed for a steplike change both SISO and MIMO systems. A binary distillation col-
in the load effect (e. g ., d in Fig. 1). This is rarely a good umn is used to illustrate the effectiveness of the proposed
assumption in chemical process control, since most algorithm in section 4, followed by the conclusion.
chemical processes face step types of changes in the load Simulation results show that, while maintaining good
P. Y. Ho et al.: Two-Degree-of-Freedom Dynamic Matrix Control for Distillation Columns

servo responses, the proposed algorithm gives better load


responses than the standard DMC for the modeled (e.g.,
feed composition change) as well as the unmodeled (e.g.,
feed flow rate changes) load changes.

DYNAMIC MATRIX CONTROL


for h = 1, 2, ... p (5)
A brief description of dynamic matrix control is
given in this section. For details, please refer to the Here, p is the prediction horizon (or the optimization
following books: Prett and Garcia [16], Seborg et al. [22] horizon as defined by Garcia and Morari [6]). The open-
and Luyben [lo]. A discrete step-response model can be loop error is defined as:
written as:

over the prediction horizon h = 1, 2, ... p. The closed-


loop response can be expressed as:

where x ~ is+the~ output at the k+ lth sampling time, ai


is the ith step-response coefficient, Amk+ - i = mk+ -i
- mkdi is the change in the manipulated variable and
dk+] is the load effect at the k+ lthsampling instance. and the closed-loop error is defined similarly:
Since it is not possible to store all of the step-response
. . coeff~cient,a truncated step-response model is used in
DMC, i.e., n step-response coefficients are employed:

ai = j aifor i 5 n
a, for i > n
(2)
In a matrix form, the closed-loop relationship becomes:

where n is the model horizon which typically falls be-


tween 20-50. Here, n is equivalent to the order of the
model employed. For a truly low-order system, this in-
deed, is an overdetermined model. However, for a high- T
order process such as a distillation column, this gives an where ~ C = I [ek+ 1 CL,ek+2,clt ... ek+p,cll, ~ O =
I [ek+~,~l~
1"
adequate process description. In the DMC approach, cur- e k + ~ , ~...l , ek+p,oll , Am = [am,, amk+l, ..., 4 + , - l l T
rent load effect, dk (Fig. l), can be estimated using the and A is a p x 1 dynamic matrix
process measurement x,,~

Provided with process measurement, xrnjk,Eq. (3) is


simply a rearrangement of Eq. (1). Without further
knowledge about load effect, the following assumption
is made where I corresponds to the number of future (including
the current one) changes which are allowed in the
manipulated variable, i.e., Amk+j = 0 for j 1 1. Here,
I is called the control horizon. For example, 1 = 1means
a single move in the manipulated variable is calculated.
Therefore, the open-loop response (assuming no further This is the well-known "open-loop'' feedback policy [2].
changes in the manipulated variable, i.e., Amk=Amk+ For a given dimension of A, the control moves Am can
= Amk+, = . .. = 0) can be predicted by p-step in the be calculated by minimizing the following objective func-
future tion
104 Journal of the Chinese Institute of Engineers, Vol. 16, No. 1 (1993)

P 1
where Ad is the dynamic matrix which relates the im con-
J = C
i=l
ei+i,cl +p C
j= 1
~m$+~- ( 11) trolled variable to the jth manipulated variable. The
manipulated variable and open-loop error vectors become:

From the objective function, it becomes obvious that


DMC can be viewed as a specific case of LQ control. The Am = [ A ~ I , Aml,k+~
,, , ... Aml,k+~-l;
input suppression parameter k (or weighting factors in the T
LQ control) is introduced to avoid excessive large changes Am2,k2Am2,k+1, .. . Am2,k+1-1 1
in the manipulated variables. An alternative expression -
for the control problems is: , el,k+2,01
- [el,k+l,~l 7 ... el,k+p,o~
;

min
am
( II ec111 22 + PI1 Am 1122 ) ( 12) Therefore, the multivariable DMC algorithm is almost the
same as the SISO version except that an augmented
dynamic matrix is employed.
and the solution is simply
TWO-DEGREE-OF-FREEDOM
DYNAMIC MATRIX CONTROL
+
Am = ( A ~ A k21)-'-4~e,~ (13)
Generally, DMC provides less improvement in load
rejection for the obvious reason that a steplike change in
Notice that only the first element of Am, Amk, is im-
the load effect d is assumed. Unfortunately, most
plemented in each execution of DMC.
chemical processes face steplike changes in the load
Several parameters associated with the design of a
variable. The concept of the truncated step-response
DMC are: (1) sampling time (T,), (2) model horizon
model of DMC can be extended to predict the load ef-
( n ) , (3) prediction horizon (or the optimization horizon)
fects of the p-step in the future.
p , (4) control horizon ( I ) and (5) input suppression
parameter ( k ) . Maurath et al. [14] give a good discus-
1. Load prediction
sion on these five design parameters. Once the design
parameters are specified, the DMC algorithm has the
The two-degree-of-freedom DMC algorithm is
following steps at each point in time:
derived for SISO systems here. For the load prediction,
a step-response model, e.g., a step-response model for
1. Calculate dk from Eq. (3) and assume that dk = GL in Fig. 1, can be used to relate load effect, d , and
dk+ = dk+2 = ... = dk+p. The calculation changes in the load variable, AL.
depends on the process measurement, x , , ~ , and the
past values of the manipulated variables (Amk-l,
Amk-2, ... ).
2. Calculate e,l, [ek+ 1.01, ek+2,ol, . . . ek+p- l Z o ~Tl, from
Eqs. (5) and (6). This calculation depends on the
predicted load effects and the past values of the
manipulated variables. where a l i is the ith step-response coefficient for a load
3. Calculate Am, [Amk, Amk+ .. . , Amk+,- ,I, from change and ALk is the change in the load variable be-
Eq. (13). tween sampling instances (i.e., ALk = Lk - L k - l ) .
4. Implement Amk. Again, the step-response model for the load is truncated
up to n terms.
One of the real contributions of LRPC is that it can
be extended to multivariable systems without any difficul- a l i for i In
ali = (15)
ty. Multivariable DMC is a straightforward extension of aln for i > n
SISO systems. Consider a 2 x 2 system, the dynamic
matrix has the form:
An alternative expression for the step-response model is:
P. Y. Ho et al.: Two-Degree-of-Freedom Dynamic Matrix Control for Distillation Columns 105

where do is the accumulation of load effects for


ALk-, - ALk-n-2 , .... Since at the kth sampling in-
stance dk can be calculated from the process measure-
ment and values of past manipulated variables [Eq. (3)],
Eq. (17) can be used to backcalculate ALk-

Note that a l l is at the denominator of Eq. (17). It is im-


mediately clear that dead-time (greater than the sampling 0.00 c
0.00 20.00 40.00
1
60.00
time) should be removed from the step-response model. TIME (MIN)
Furthermore, for systems with a small value in all, some
type of modification (tuning) is necessary to avoid violent Fig. 2. Load step-response models employed in disturbance prediction
shifts in the manipulated variables. This will be dis- for different values of kL's.
cussed by the introduction of a load tuning constant kL.
Generally, step-response coefficients are obtained from
dead-time free and apart from the load transfer function where El, is the modified step-response coefficient used
( G L ) Once ALk-l, u k - 2 , ... are known, we can in the load prediction [Eqs. (18) and (20)l. Fig. 2 shows
predict the load effects p-step in the future according to how kL changes the step-response model used in the load
"
the load model [Eq. (16)l. Instead of assuming dk = prediction. When kL = 1, the original step-response
dk+ = dk+2 = ... dk+p,the load prediction for the pro- model, GL, is employed in the load prediction. It can be
< . posed algorithm becomes: shown that when kL = 0, the proposed algorithm reduces
to DMC (e.g., dk = dk+i = dk+2 = ... = dk+p; Fig.
2). Furthermore, the disturbance prediction is affected
only by the dynamics (e.g., the time constants) of the
load step-response model (i.e., errors in the steady-state
for h = 1, 2, ..., p gain will not affect the prediction) as also pointed out by
Wellons and Edgar [27]. Note that for systems with a
Note that this is equivalent to assuming that ALk = non-zero kL, it is likely to be less apt to plantlmodel
ALk+ = ... = ALk+p = 0 (i.e., a steplike change in mismatches. This can be understood from the IMC
L). In a matrix form, the load prediction becomes: framework [15]. The closed-loop characteristic equation
for a two-degree-of-freedom IMC is det[Z+ Gc.
GF(G-G)] = 0 wh'ere is the process model, Gc is
the standard IMC controller and GF is the controller for
where d = [dk+l, dk+2, ... dk+plT, AL = [ALk-1, the load prediction. For kL = 0, GF is an identity
ALk-2, . .. Uk-,] and do = [do, do, . .. do]T. AL is the matrix. For kL = 1, GF contains the inversion of GL in
load dynamic matrix with the dimension of p x n: the diagonal elements. Since, in general, GL is strictly
proper, the inversion of the elements of GLhas improper
elements. From the small gain theorem, this pushes the
nyquist plot closer to the limit of stability. Therefore,
a tradeoff between disturbance rejection and robustness
has to be made in selecting an appropriate kk The selec-
tion principle is: the closeness of kL to 1 depends a great
deal on the accuracy of the step response model.
Nonetheless, the two-degree-of-freedom DMC offers this
Similar to DMC, the open-loop response of the proposed degree of freedom for improving load rejection. It is a
algorithm, xkfhSol,can be calculated according to Eq. generalized version of DMC which requires a little extra
(5), except that dk+h is predicted differently. A tuning computing effort [Eqs. (18) and (20)l. Notice that the
factor, kL, is introduced to adjust the degree of load two-degree-of-freedom algorithm only changes the open-
prediction. This is a useful adjustable parameter, loop error, eel, in the general DMC structure. The
especially when plantlmodel mismatches exist. kL is dynamic matrix and the controller matrix remain the same.
defined as: Once the design parameters (including kL) are
chosen, the proposed algorithm has the following steps:
la. Calculate dk from Eq. (3). The calculation
106 Journal of the Chinese Institute of Engineers, Vol. 16, No. 1 (1993)

depends on process measurement, xmIk,and past


values of the manipulated variables (Amk-
Amk-2, ... ). -
lb. Predict future load effects, dk+hfor h = 1, 2, ...
p, from Eqs. (17) and (19). This is a straightforward extension to MIMO systems. The
2. Calculate e,~, [ek+~,~l, T
.
ek+2,ol, ... ek+p- ~ , ~ lfrom
Eqs. (5) and (6). This calculation depends on the
l computational steps for multivariable processes are the
same as the SISO cases.
predicted load effects and the past values of the
manipulated variables. R-V DISTILLATION CONTROL
3. Calculate Am, [Amk,h k + l , ..., h k + l - l ] T , from
Eq. (13). 1. Process
4. Implement Amk.
A binary distillation column is used to test the ef-
The only difference between these two algorithms is that fectiveness of the proposed algorithm. This is an 18-tray
the proposed algorithm made further load prediction by column with the feed tray located at Tray 9. An ideal
backcalculating aLk-l. Load effects, dk+hfor h = 1, vapor liquid equilibrium with relative volatility of two is
2, ... p , are calculated based on ALk-l, U k - 2 ... . assumed (Fig. 3). Product purities on the top and
Therefore, the implementation of the proposed algorithm bottoms are 0.95 and 0.05 (for the light component),
requires only a minor modification of the original DMC respectively [4]. The reflux ratio at design condition is
algorithm. 2.03. Table 1 summarizes typical operating condition.

2. Multivariable system

Similar to DMC, the extension of load prediction


to MIMO systems requires little conceptual effort. Con-
sider a 2 x 2 system, ALi,k-l can be backcalculated from
di,kfor the ih controlled variable.

for i = 1, 2

It may look strange that two changes of the load variable


at the k-lh sampling instance (ALl,k-l and l) are
computed for a single load variable L, since under perfect
model assumption = 1 when the model-
ed load comes into the system. However, if a load distur-
bance other than the modeled one enters the system, this
approach still can handle the disturbance and the dif- Fig. 3. The distillation column under R-V control.
ference between a L l , k - and 6 , k - l can be used as an
indication of the quality of the load models employed. Table 1. Steady-state operating conditions for the
Once a ~ [ ,~ l , k - i ,u 1 , k - 2 J ... aLl,k-n ! 6.k-1, binary distillation column
... 6 , k - ~ T is, available, ~ q (19) . can be
used for load prediction. The load dynamic matrix is a parameter Value
2p x 2n matrix of the form:
Number of Trays
Feed Tray
Top Comp. (mol percent)
Bottoms Comp.
Feed Comp.
where ALi is the load prediction matrix for the ih con- Feed Flow Rate (Kg-mollmin)
trolled variable [Eq. (20)l. The predicted load-effect vec- Reflux Ratio
tor becomes: Relative Volatility (a)
P.Y. Ho et a/.: Two-Degree-of-Freedom Dynamic Matrix Control for Distillation Columns

In order to perform non-linear simulations, the follow- o 97,


ing assumptions are made: (1) saturated liquid feed and ......
--__ _-_
saturated reflux flow, (2) equal molar overflow, (3)
perfect level control, (4) 100% tray efficiency and (5) the
x D 095
0.94
tray hydraulic time constant is 6 sec. The control objec-
tive is to maintain top and bottom compositions by 0.93
0.060
manipulating reflux flow (R) and vapor boilup (V). The
, .---.---.
reflux drum level and the bottoms level are controlled by 0.055 -
__
distillate and the bottoms flow rate. This is the R-V con- XB
0.050.+
trol structure [29]. Four minutes of analyzer dead time 0.045- '.-,,~'- ,__--.._--------
I

is assumed for the composition measurements.


0.040-

2. Control

Three types of controllers are evaluated: (1)


multiloop PI controllers, (2) DMC and (3) the two-degree-
of-freedom DMC. In order to fmd the controller
parameters for PI controllers, ATV tests are performed
on R and V and the process transfer functions are iden-
tified accordingly (Luyben [lo], p. 519).

0.248 e-& -0.213e-4.7S ' 50.0


1 , (42.4~+1)(4.1~+1)(.SO.IS+~)(~.SS+~J O 50 100 150 ZOO
TlME (min)
G,,, =
0.248 e-6S -0.0283 e-& Fig. 4. Set point responses for the proposed algorithm (solid), DMC
1
( 2 8 . 3 ~ +1) ( 5 . 9 ~ +1) ( 3 6 . 8 ~ +1) (1.6~+1). (dash) and PI controllers (dot) with A X D = f 0.1.

The PI controllers are tuned using the BLT method


...
(Luyben [lo], p. 599). For the DMC algorithm, the
model horizon n is chosen such that nT, covers at least 0.952
--_
95 % of the step responses for the slowest step-response. XD
Choosing such a long model horizon can prevent unplea-
sant ringing in the closed-loop responses as a result of 0.948
truncation errors (since we deliberately truncate the step-
response models). In this example n = 40 and T, is the
.... -
same as the analyzer dead time (4 min). The model _ -- _-
XB 0.055
horizon is chosen to be the same as the prediction horizon,
i.e., n=p. The control horizon I is set to half of the
prediction horizon. For this column, an I of 20 is used. 0.045
Here, the ratio of the control horizon to the prediction
horizon is reduced from 315 [7] to 112, since we fmd that
the control horizon does not affect the closed-loop _ _ --
R 36
responses too much if such a large number is used. The
tuning constants, k, and k2, are adjusted empirically such
that the maximum overshoot in the manipulated variables 34 1
56
is about 100% for step set point changes. In this exam-
ple, k, = k2 = 0.0015. In the proposed algorithm, the
feed composition change is considered to be the major . .. __
load disturbance. The step-responsecoefficients for a step 53
feed composition change is used in disturbance predic-
52
tion. The model horizon for the load model is set to be 0 50 100 150 200
the same as that of the process model, since the time con- TIME (min)
stants for the process and load transfer functions are of
the same order of magnitude. The load tuning factors are Fig. 5 . Set point responses for the proposed algorithm (solid),
DMC (dash) and PI controllers (dot) with AXB = 0.1.
also adjusted empirically such that the overshoots in the
108 Journal of the Chinese Institute of Engineers, Vol. 16, No. 1 (1993)

manipulated variables do not exceed 100% or the over- that the feed composition goes through a first-order, lag
shoots of the BLT tuned PI controllers. For this column, type of change with the time constants equal to the load
we have kL1 = kL2 = 0.8. transfer functions.
As expected, the DMC performs better than
multiloop SISO controllers for set point changes in X, A. T w o - D e g r e e - o f - F r e e d o m DMC
(Fig. 4). The BLT tuned PI controllers show significant
interaction for set point changes in X,. Furthermore,
the set point responses for DMC and the proposed
algorithm are effectively the same as shown in Fig. 4.
Similar results can be observed for a set point change in
X, (Fig. 5). Simulation results simply confirm the
observations that LRPC gives better set point responses
than conventional multiloop PI controllers. Since distur-
bance rejection is of primary importnace in distillation
control, the three control systems are tested against'feed B . DMC
composition changes. Fig. 6 shows that the closed-loop
responses are almost the same for both the DMC and the
BLT tuned PI controllers for a step feed composition
change. The results indicate that, as far as the major load
disturbance (feed composition changes) is concerned, the
DMC offers little advantage over multiloop SISO con-
trollers. However, the two-degree-of-freedom DMC
gives much better load responses as shown in Fig. 6. -0.064
Thereason is that the proposed algorithm gives a better 0 50 100 150 200
estimaiiqn in the feed composition changes (AZ) as shown Time, MIN

in Fig. 7 : Fig. 7 shows that the estimated feed composi-


Fig. 7. Estimated feed composition changes form XD (solid) and
tion chang'es (either from XD or from XB) are quite close XB (dash) for the proposed algorithm (top) and DMC
to the true load changes for the proposed algorithm. The (bottoms) when AZ = -0.05.
standard DMC underestimates AZ. Actually, it predicts

..- ____
0.950
...
xD ...
..
........
0.945 - ...... .........
.'....I'
. ,
.2
0.945
0.940 -
0.055 0.06 1
_ -
0.050

XB 0.045
.........................
-- _ -
-
-' ... --
XB
0.04
...... ........

.... _..
......................

/-
_
........
0.03 ........
0.035

-........ - - . ........... ... ..


R 37.5

36.5
56
54 .....
52
....... -.. _ ---- ---_ .......
v 54 "' ....... ............ - . .

53 48 -
46 7
52 0 50 100 150 200
0 50 loo 15.2 200 TIME (min)
TIME (rnin)
Fig. 6. Load responses for the proposed algorithm (solid), DMC Fig. 8. Load responses for the proposed algorithm (solid), DMC
(dash) and PI controllers (dot) with AZ = -0.05. (dash) and PI controllers (dot) with AF = - 10%.
P.Y. Ho et al.: Two-Degree-of-Freedom Dynamic Matrix Control for Distillation Columns 109

controllers. However, for feed composition changes, the


original DMC offers little improvement over the mdtdoop
SISO controllers. The two-degree-of-freedom DMC
shows much better load responses for feed composition
changes. For the unmodeled load disturbance, feed flow
change, the proposed algorithm still performs better than
the original DMC algorithm. Simulation results show that
the two-degree-of-fieedom DMC offers an attractive alter-
-0 081 native for distillation control
0 50 1 GO 150 2CO
Time, MIN

Fig. 9. Estimated feed composition changes form XD (solid) and


XB (dash) for the proposed algorithm when AF = -10%. We thank M. C. Chuang of Union Chemical Lab.
for translating the FORTRAN based algorithms into C.
Next, the proposed algorithm is tested against an PYH wishes to thank Union Chemical Lab., ITRI for
unmodeled load disturbance. Consider a step change in financial support.
feed flow rate. The multiloop PI controllers perform
poorly, as shown in Fig. 8. Since the interaction effects
are not accounted for in the multiloop controllers, the fast
effect of vapor boilup canceled out the action of reflux im step-response coefficient
flow which, subsequently, leads to sluggish responses in im load step-response coefficient
XD. Again, the proposed algorithm still gives better ith load step-response coefficient for the jth
closed-loop responses than DMC even for this unmodel- controlled variable
ed load disturbance (Fig. 8). The estimated load changes modified im load step-response coefficient
1 ' (AZ) also reveal that the load change is different than the CEq. (2211
I modeled one, since AZ estimated from XD is significant- dynamic matrix
ly different from the one estimated from XB (Fig. 9). load dynamic matrix
I
D
A slow change in the load estimated from XD also fits load dynamic matrix for the jm controlled
the description of feed flow rate changes, since the distur- variable
bance hits XB first and affect XD via the changes in vapor bottoms flow rate
boilup. Therefore, the estimated load changes (e.g ., Figs. load effect
7 and 9) can be used to evaluate the quality of the load distillate flow rate
models. If the load changes estimated from XD and XB load effect at km sampling instance
differ significantly, this may imply: (i) the load modeled load effect at km sampling instance from j,,,
is not correct or (ii) the load modeled is not the major controlled variable
load disturbance. DMC dynamic matrix control
eel vector of open-loop error
CONCLUSIONS e0l.k open-loop error at kth sampling instance
ec1 vector of closed-loop error
A two-degree-of-freedom DMC algorithm is pro- closed-loop error at kth sampling instance
posed to improve load rejection capability. Incorporated feed flow rate
with load step-response models for the major load distur- process transfer function
bance, the load effects can be predicted p-step in the load transfer function
future. This type of prediction provides better handling input suppression parameter for DMC
of the steplike load changes which are typical in chemical load tuning factor for the proposed algorithm
processes. A load tuning factor, kL, which adjusts the load variable
degree of disturbance prediction is introduced. It pro- vector of load changes
vides full load disturbance prediction (kL = 1) on the load change between sampling instances
one hand and reduces to DMC (kL = 0) on the other load change between sampling instances
hand. More importantly, the proposed algorithm requires estimated from the jmcontrolled variable
only a minor modification of the original DMC algorithm control horizon
(e.g., a modification in eOl which is the input to the manipulated input at kth sampling instance
DMC controller). A distillation column under R-V con- vector of Ami
trol is used to test the usefulness of the proposed mk - mk-1
algorithm. Simulation results show that, as expected, model horizon
DMC gives better set point responses then multiloop SISO number of components
110 Journal of the Chinese institute of Engineers, Vol. 16, No. 1 (1993)

number of trays J., Vol. 35, p. 241 (1989).


prediction (optimization) horizon 10. Luyben, W .L., Process Modeling, Simulation and
reflux flow rate Control for Chemical Engineers, 2nd Ed, McGraw-
Laplace transform variable Hill, New York, NY (1990).
time 11. Mantz, R.J. and E.J. Tacconi, "Complementary
sampling time Rules to Ziegler and Nichols' Rules for a Regulating
vapor boilup and Tracking Controller, " Znt. J. Control, Vol. 49,
controlled variable at the k* sampling p. 1465 (1989).
instance 12. Marchetti, J.L., D.A. Mellichamp and D.E. Seborg,
process measurement at the sampling "Predictive Control Based on Discrete Convolution
instance Models," Znd. Eng. Chem. Process Des. Dev., Vol.
predicted open-loop response at the kth 22, p. 488 (1983).
sampling instance 13. Martin, G.D., "Long-Rang? Predictive Control,"
predicted closed-loop response AIChE J., Vol. 27, p. 748 (1981).
bottoms composition 14. Maurath, P.R., D.A. &ellicharnp and D. E. Seborg,
top composition "Predictive Controller Design for Single - In-
feed composition changes putlsingle - Output (SISO) Systems," Ind. Eng.
Chem. Res., Vol. 27, p. 956 (1988).
Superscripts 15. Morari, M and E. Zafiou, Robust Process Control,
Prentice-Hall, Englewood-Cliff, NJ (1989).
set set point 16, Prett, D.M. and C.E. Garcia, Fundamental of Pro-
T transpose cess Control, Butterworths, Boston, MA (1988).
-1 inverse 17. Richalet, J., A. Rault, J.L. Testud and J. Papon,
"Model Predictive Heuristic Control: Applications
REFERENCES to Industrial Processes," Autowxztica, Vol. 14, p. 413
(1978).
1. Clarke, D.W., Mohtadi, C. and P.S. Tuffs, 18. Ricker, N. L., "Use of Quadratic Programming for
"Generalized Predictive Control - I. The Basic Constrained Internal Model Control, " Znd. Eng.
Algorithm," Automutica, Vol. 23, p. 137 (1987). Chem. Process Des. Dev., Vol. 24, p. 925 (1985).
2. Clarke, D.W., Mohtadi, C. and P.S. Tuffs, 19. Ricker, N.L., "Model Predictive Control with State
"Generalized Predictive Control - 11. Extensions Estimation," Ind. Eng. Chem. Res., Vol. 29, p. 374
and Interpretation," Automutica, Vol. 23, p. 149 (1990).
(1987). 20. Rotea, M.A. and J.L. Marchetti, "Internal Model
3. Clarke, D.W. and C. Mohtadi, "Properties of Control Using the Linear Quadratic Regulatory
Generalized Predictive Control," Automatica, Vol. Theory," Znd. Eng. Chem. Res., Vol. 26, p. 577
25, p. 859 (1989). (1987).
4. Fuentes, C. and W.L. Luyben, "Control of High- 21. Rouhani, R. and R.K. Mehra, "Model Algorithmic
Purity Distillation Columns, '' Ind. Eng. Chem. Pro- Control (MAC): Basic Theoretical Properties,"
cess Des. Dev., Vol. 22, p. 361 (1983). Automutica, Vol. 18, p. 401 (1982).
5. Garcia, C. E., ' 'QuadraticJDynarnic Matrix Control 22. Seborg, D.E. T.F. Edgar and D.A. Mellichamp,
of Nonlinear Processes: An Application to a Batch Process Dynamics and Control, Wiley, New York,
Reaction Process," paper presented at AIChE An- NY (1989).
nual Meeting, San Francisco, CA (1984). 23. Shunta, J.P. and W.L. Luyben, "Sample - Data
6. Garcia, C.E. and M. Morari, "Internal Model Con- Noninteracting Control for Binary Distillation Col-
trol - 1. A Unifying Review and Some New umns," Chem. Eng. Sci., Vol. 27, p. 1325 (1972).
Results," Ind. Eng. Chem. Process Des. Dev., Vol. 24. Stephanopoulos, G. and H.P. Huang, "The 2-Port
21, p. 308 (1982). Control Systems," Chem. Eng. Sci., Vol. 41, p.
7. Georgiou, A., C. Georgakis and W.L. Luyben, 1161 (1986).
"Nonlinear Dynamic Matrix Control for High-Purity 25. Stephanopoulos, G. and H. P. Huang, "Two-Port
Distillation Columns," AlChE J., Vol. 34, p. 1287 Control Systems: A Generalized Predictive Control,"
(1988). Znt. J. Control, Vol. 45, p. 617 (1987).
8. Ho, B.Y ., "Applications of Model Predictive Con- 26. Vidyasagar, M., Control System Synthesis: A Fac-
trol to Distillation Column,'' Master Thesis, National torization Approach, MIT Press, Cambridge, MA
Taiwan Institute of Technology (1990). (1985).
9. Li, S., Lim, K.Y. and D.G. Fisher, "A State Space 27. Wellons, M. C. and T.F. Edgar, "The Generalized
Formulation for Model predictive Control, " AIChE Analytical Predictor," Znd. Eng. Chem. Res., Vol.
P. Y. Ho et al.: Two-Degree-of-Freedom Dynamic Matrix Control for Distillation Columns 111

26, p. 1523 (1987). Discussions of this paper may appear in the discussion
28. Wong, K.P. and D.E. Seborg, "A Theoretical section of a future issue. All discussions should be
Analysis of Smith and Analytical Predictors," AIChE submitted to the Editor-in-Chief.
J., Vol. 32, p. 1597 (1986).
29. Yu, C. C. and W .L. Luyben, "Design of Multiloop Manuscript Received: Oct. 15, 1991
SISO controllers for Multivariable Processes, " Ind. Revision Received: Dec. 2 7, 1991
Eng. Chem. Process Des. Dev., Vol. 25, p. 498 and Accepted: Dec. 30, 1991
(1986).

You might also like