e Pergamon

Computers chem. Engng. Yol. 19, No.4, pp. 409-421 1995

0098-1354(94)00063-8
Copyright © 1995 Elsevier Science Ltd
Printed in Great Britam . All rights reserved
0098-1354195 $9.50 + 0.00

LIMITATIONS OF DYNAMIC MATRIX CONTROL

P. LUNDSTROM,lt J. H. LEE!':j:, M. MORARI!. § and S. SKOGESTAD'
1 Chemical Engineering 210-41, California Institute of Technology, Pasadena, CA 91125, U.S.A.

2 Chemical Engineering, University of Trondheim. NTH. N-7034 Trondheim, Norway

(Received 5 June 1992; final revision received 13 May 1994; received [or publication 7 June 1994)

Abstract-Dynamic matrix control (DMC) is based on two assumptions which limit the feedback
performance of the algorithm. The first assumption is that a stable step response model can be used to
represent the plant. The second assumption is that the difference between the measured and the
predicted output can be modeled as a step disturbance acting on the output.
These assumptions lead to the following limitations:

I. Good performance may require an excessive number of step response coefficients.

2. Poor performance may be observed for disturbances affecting the plant inputs.
3. Poor robust performance may be observed for rnultivariablc plants with strong interactions

Limitations 1 and 2 apply when the plant's open-loop time constant is much larger than the desired
closed-loop time constant. Limitation 3 is caused by gain uncertianty on the inputs.
In this paper we separate the DMC algorithm into a predictor and an optimizer. This enables us to
highlight the DMC limitations and to suggest how they can be avoided. We demonstrate that a new
model predictive control (MPC) algorithm, which includes an observer. does not suffer from the listed
limitations.

I. INTRODUCTION DMC. Obviously, there are other algorithms m the

literature which do not suffer from these deficien-
Dynamic matrix control (DMC) has been success- cies. In this paper, however. we wanted to adhere to
fully used in industry for more than a decade. the DMC structure because of its popularity and its
Several authors have reported improved control
capability to handle constraints. We demonstrate
performance by use of DMC as compared to "tradi- that the limitations can be avoided by use of a new
tional" control algorithms (Cutler and Ramaker, observer based algorithm, by Lee et al. (1994),
1980; Prett and Gillette, 1980; Garcia and which is a direct extension of DMC.
Morshedi, 1986). DMC has the ability to deal with DMC belongs to the family of model predictive
constraints, which probably is one of the major control (MPC) algorithms, The main idea behind
reasons for its popularity. It also allows set point these algorithms is to use an explicit model of the
changes to be "announced" in advance and it facili- plant to predict the open-loop future behavior of the
tates feedforward control. However, the feedback controlled outputs over a finite time horizon . The
properties of a DMC controller are limited by two predicted behavior is then used to find a finite
restrictive assumptions which are implicit in the sequence of control moves which minimizes a parti-
algorithm: cular objective function without violating prespeci-
At. A stable step response model can be used to fied constraints. Usually only the first input move is
represent the plant. implemented and the procedure is repeated at the
A2. The difference between the measured and next sampling instant.
the predicted output can be modeled as a This algorithm can be separated into two parts, a
step disturbance acting on the output. predictor and an optimizer. By splitting up the
algorithm in this manner. similarities with state-
The objective of this paper is to clearly point out observer state-feedback controllers become appar-
these assumptions and to illustrate in which situa- ent. In fact, Lee et al. (1994) show that uncon-
tions they may limit the feedback properties of
strained DMC is equivalent to an optimal state
t Present address: Chemical Engineering, University of observer (Kalman filter) and linear quadratic feed-
Trondheim, NTH, N-734 Trondheim, Norway. back, using a receding horizon approach and special
t Present
address: Department of Chemical Engineering,
Auburn University, Auburn, AL 36849-5127, U,S.A. assumptions abou t disturbances and measurement
§ To whom all correspondence should be addressed. noise.
409
410 P.

In this paper we use the predictor-optimizer rep-

resentation of DMC. In this framework the limi-
tations of DMC can be traced to the predictor part
of the algorithm. We only consider unconstrained
LUNDSTROM

S;=
et al.

[
SI.I.i

52. I. i
.
Su.;

52.2. i
Sl.n".1
S2. nil' i
:
l •
1=1, ... ,n. (2)

DMC, but the results carryover to the general case Sn.' t.. Sn,.. 1.1 Snv,n ll • i

with constraints, since the issue of constraints only The step response model can be represented in
affects the optimizer. the following state space form, which is equivalent
The limitations we want to illustrate are: to that presented by Li et al. (1989):
Ll. Good performance may require an excessive Y(k + 1) = MY(k) + SAuCk), (3)
number step response coefficients.
y(k) = NY(k), (4)
L2. Poor performance may be observed for
"ramp-like" disturbances acting on the plant where
outputs. In particular, this occurs for input Au(k)=u(k)-u(k-l), (5)
disturbances for plants with large time con-
stants. Y(k) = [y(kfy(k + If ... y(k + n - W]T (6)
L3. Poor robust performance, due to input gain o In, 0 () 0
uncertainty (which is always present in prac- o 0 In, () 0
tice), may be observed for multivariable
plants with strong interactions. M= nXny;
0 0 U In, 0
0 0 0 0 111\
In addition, there is the obvious limitation that the
0 0 0 0 If/v
plant has to be stable.
This paper is organized as follows. In Section 2 we SI
present the algorithms we will use. The purpose is to S2
give a coherent overview of the algorithms and to
S=
point out the implicit assumption made in DMC. We Sn-2 (7)
also use this section to define the nomenclature. Sn-l
Readers not familiar with MPC are referred to Sn
Garcia et al. (1989). In Section 3 we use a simple
single-input-single-output (SISO) plant and a multi-
and
input-multi-output (MIMO) distillation column to
illustrate the limitations of DMC and demonstrate
,
that the algorithm by Lee et al. (1994) can be used to N=[In , 0 0 ... 0 0]. (8)
avoid these limitations. Section 4 contains a
Discussion and Section 5, Conclusions. Au(k) is a vector of changes in the manipulated
inputs at time k. y(k) is the output vector at time k.
The vector Y( k + 1) represents the dynamic states of
2. MODEL PREDICTIVE CONTROL the system. Each state, y(k + l), has a special inter-
pretation: it is the future output vector at time k + I
assuming constant inputs [i.e. Au(k + j) = 0 for
2.1. Dynamic matrix control j ~ 0]. The new state vector Y(k + 1) is the old vector
Y(k) shifted up n, elements plus the contribution
2.1.1. Modeling the plant. In the original DMC made by the latest input change Au(k).
formulation (Cutler and Ramaker, 1980) a step 2.1.2. The predictor. The DMC algorithm is illus-
response model of the plant is used to predict the trated in Fig. 1. The objective of the predictor is to
future behavior of the controlled variables. generate a vector, rw(k + 11k), of predicted open-
Let the step response of a SISO system be repre- loop outputs over a horizon of p future time steps,
sented by the sequence: the prediction horizon. This prediction vector is
then used as an input to the optimizer.
(1) The DMC predictor is described by the following
where the kth element is the output at time k caused equations:
by a unit step input at time O. For a stable plant this Y(k) = MY(k -1) + SAuCk -1), (9)
sequence will asymptotically reach a constant value,
y(k) =NY(k), (10)
i.e. Sn""'Sn+I' For a MIMO system with n; inputs and
ny outputs we get: "Y(k + 11k) = Mp Y(k) + .'1{9(k)- Y(k)] (11)
 Limitations of dynamic matrix control 411

6.u(k) y(k)
Plant

Y(k + 11k)

.
I~ __ ~~,
.
' • • • • • • • • • • • • • • • • • • • • • • • • • • • _ - - • • _ - - - . _ . __ • ._ •• _._ - _ •••• _. •••• _ •• •

Optimizer Predictor
Fig. 1. DMC controller separated into a predictor and an optimizer.

where M; is the first p x ny rows of M and: 0 1

p x n,
T
51
52 51 °
0 I!
,

s = [/ny In,
A

... In, InJ

,

(12) ';j'm=
I
p ( 18)
5n> 5m~1 51
We use to denote that the output is from the
model and not from the true plant. y(k) is a vector s, 51'_1 S,. m
of measured outputs at time k. Y( k) and y(k) are
2lt(k + 11k) is a vector describing the desired output
discontinuous at L while u(k) is discontinuous at
trajectory (set points) over p future time steps r
k+, i.e. y is measured slightly before time k and u is
and A are weighting matrices and arc usually chosen
adjusted slightly after time k.
to be diagonal.
2.1.3. The optimizer. We use the QDMC objec-
The least squares solution to this problem I'
tive function from Garda and Morshedi (1986):
6.OU(klk) = W:f;')T[Tr:J';" + AI1\] I(:j;~'t Ir' r
J= min {ilr[i!Y m (k + 1)[k) - 2lt(k + l jk)lW
tl.'U(klk) x [?Jl(k + Ilk) _o)j(k + Ik)] (19)

+ IIA~OU,(klk)112}, (13) Only the first input move is implemented. and the
resulting optimizer is a constant gain matrix, K,1I'(
where
6.u(k) = [I () .. .O]L'l.jlL(klk)
~OU(klk) = [6.u(klk)T ~u(k + 11k)T
=K M Pd?Jl(k+lik)-O)J(k+ Ilk)!. (20)
... 6.u(k+m-1Ik)ly, (14) KM PC = f1 0" .O]f(:f;?[Tr:J';"+A] I

i!Ym(k + 11k) = [Ym(k + llk)lYm(k + 21k? x (:J;")T[T[ (21)

... Ym(k + plk)TjT, (15) 2.1.4. DMC assumptions, The DMC controller
can only be used with stable plants, There are two
and
reasons for this: (1) The internal model [equations
2lt(k + 11k) = [r(k + llk)Tr(k + 21k? ... (9-10)] can only describe a stable plant: (2)\'(kt
y(k) can grow unbounded for unstable svsterns lead-
r(k + plk)lY. (16)
ing to internal instability.
~OU(klk) is the optimal control sequence computed The internal model of the DMC predictor [equa-
at time k for m future input moves, where m is the tions (9-10)1 does not yield an estimate of the true
input horizon. "Ym(k + 11k) is a vector of outputs plant output. It computes the open-loop model out-
predicted at time k, over a horizon of p future time put, Y(k), for previous input moves. but does not
steps, including the effect of the m optimal input account for the effect of disturbances and model-
moves: plant mismatch, This means that generally v (k) -
y(k) is not zero when there is no steady-state offset
and y(k) = O. Rather. -y( k) equals the accumulated
where effect of disturbances and model-plant mismatch
412 P. LUNDSTROM et al.

Equation (11) gives the predicted open-loop out- outputs. It also allows modeling of integrating
put vector, "Y( k + 11 k). It is the predicted effect of systems:
previous input moves, M, Y(k), plus a simple bias
Y(k+l)=MY(k)+S~u(k)+T~w(k), (25)
adjustment given by the mismatch between the mea-
sured output, y(k), and the output form the internal y(k) = NY(k), (26)
model, y(k). y(k) = y(k) + v(k), (27)
To achieve good control performance, "Y(k+ 11k)
~w(k)=w(k)-w(k-l) is a vector of changes in
should be close to the true open-loop output. This
disturbances and v(k) is a vector of measurement
requires that n, the number of coefficient matrices in
noise:
S, is chosen such that S; = Sn+ I' otherwise M; Y(k)
will be in error. It also requires that y(k) - y(k) 0 In, 0 0 0 0 0
stays approximately constant. 0 0 In, 0 0 0 0
We formulate these requirements as two implicit
assumptions made in the DMC algorithm: M= 0 0 0 t; 0 0 0
0 0 0 0 In, 0 0
AI. A stable step response model with Sn=Sn+1 0 0 0 0 In, Cu Cw
can be used to represent the plant.
0 0 0 0 0 Au 0
A2. The difference between the measured and 0 0 0 0 0 0 Aw
the predicted output can be modeled as a
step disturbance acting on the output. n X ny +- dim{xu}+ dim{xw} , (28)
SI 0
2.2. DMC with general state space model
S2 0
The DMC algorithm can also be derived for a
general discrete state space model (Prett and S= Sn-2 T= 0 (29)
Garcia, 1988) instead of the step response model Sn-l 0
[equations (3-4)] used in the previous section. We Sn 0
will denote this algorithm DMCss. The only differ- Bu 0
ence between DMC and DMCss is the represen- 0 Bw
tation of the internal model. We include DMCss in
N = [In y 0 0 ... 0 0 0 0], (30)
this paper because it allows us to study DMC with-
out the effects of truncation errors caused by S, '* Y(k) = [y(k)T y(k + 1?... y(k + n -1)T
Sn+l' X.(k)Txw(k)T]T. (31)
Let the plant model be defined by the following
Au, B; and C, constitute a state space description
equations:
of the residual plant dynamics after n sampling
x(k+l)=Ax(k)+Bu(k), (22) intervals. A w , B; and C; describe the dynamics of
the disturbances. Xu and X w are state vectors for
y(k) = Cx(k). (23)
residual plant dynamics and disturbance dynamics,
Using this model the DMC algorithm can be de- respectively.
scribed by the block diagram in Fig. 1 by making the This representation allows very general modeling
following substitutions, Y(k) = M(k) = /H(k)- of plant and disturbances. However, we will approx-
i(k-l), M=A, S=B, N=[O .. .0] and imate the residual dynamics with ny x n u first-order
systems, each describing the slow response from one

[C~:AcA:]
input to one output. This approximation gives:

M=p (24) (32)

L;~ICA

(33)
2.3. Observer based model predictive control
This algorithm is from Lee et al. (1994), we will
denote it "OBMPC". n;»; n u
Lee et al. use the following extended version of ~

the step response model in equations (3-4). The Cu = [In" In" ... In..], (34)
extension is made in order to include measurement We also restrict measurement noise and distur-
noise and general disturbances acting on the plant bances to the following special case:
 Limitations of dynamic matrix control

~u(k) y(k)
Plant

Y(k + 11k)
' ~ - -" -" -" _.

Optimizer Predictor

Fig. 2. Observer based MPC controller separated into a predictor and an o ptimizer

1. The measurement noise at each output is equation (38) is parametrized as follows (Lee ,'I al.,
uncorrelated white noise. 1(94):
2. The disturbances at the outputs are integrated
white noise filtered through first-order dyna-
mics.
rIIIIn
For this special case we get the following diagonal
covariance matrices;
K= 1"-
I I"
lfJ' u,) 1j
E{~w(k)~w(k)"'1=W= [WI ... ,.J' (35)
ll.1
II
" ' !II

() l
E{v(k)v(k)T}= V= [Vt...] (36)
V,,\ I".. ,
+
t; + sa (1,1 I
!
if:A.1 (UII"J
n
s': I

1I.,(fJ?
(/;,), ~ --------: - I "" i -s n (4J)
I + 11.,' o{t,),

The adjustable parameters. Ct')i' are detcrrnmed by

the disturbance -\C, -noise ratio for the rth output,
W/V,:
Y(klk) = Y(klk -1) + K{Y(k) - y(klk - I)}. (38)

Y(k+1Ik)=MY(klk)+S~u(k). (39) (/J .1) fo r W.it "D_ (44)

where U,l ,I for v. / t ~ o:

Y(klk -1) = [y(klk -1)Ty(k+ 11k -1)T. . .

Hence, wc may implement the Kalman filter It uhout
Y(k + n -11k -1)Txu(klk - WxAklk -1)11T. (40) solving a Riccati equation. and (f,), and a, mav be
y(k + 11k) is the estimate of y(k + I) based on meas- used as on-line tuning parameters
urements up to and including time k. The predicted 2.3.2. Suue-obserner statr-Ieedbact. uuerpre-
output vector (the input to the optimizer) is: tation. The unconstrained OBMPC described above
is a state-observer state-feedback controller using a
o.!I(k+llk)=M"Y(klk). (41)
receding horizon approach The optimal stare -
For the special noise and disturbance case defined observer is defined by equations (38~l)) and the
in the previous section, the optimal filter gain K in linear quadratic state feedback gam is A\IP,M Fhe
414 P. LUNDSTROM et al.

closed-loop dynamics is determined by the following

!iEfH
state transition matrix:

Y(~)
[ Y(k) - Y(klk)
] = [M- SKMPcMp
0 o 20 40 60 80 100 120 140 160 180 200

:83
Time [min]
Y(k-l) ]

J:[ ; ~
x [ Y(k--l)-Y(k-llk-l) .
(46)
:A"
The eigenvalues of M - SKMPcMp are the regulator
o 20 40 60 80 100 120 140 160 180 200
poles and the eigenvalues of M - KNM are the Time [min]
observer poles.
Fig. 3. Effect of truncation. Response for the SISO plant
If the measurements are noise-free and the distur-
[equation (48») with DMC controller A (Table 1). A unit
bances are random steps acting on the plant outputs, step disturbance acts on the plant output at t = 10.
then an unconstrained DMC controller where S; =
Sn+ 1 is equivalent to the unconstrained OBMPC
plant output at time t= 10. The simulation is per-
controller. That is, for this special case:
formed with a dead beat DMC controller (controller
Mp Y(klk) = M; Y(k) + J[y(k) - y(k)], (47) A, Table 1). The truncated step response causes an
erroneous prediction (a "jump") n -1 sampling
and DMC is an optimal state-observer state-
intervals after the disturbance occurred. The error
feedback controller.
here is so large that it leads to instability.
From this example we conclude that heavily trun-
3. LIMITATIONS OF DYNAMIC MATRIX CONTROL
cated models cannot be used, and thereby the
computer hardware may restrict the choice of sam-
3.1. Limitation 1: Good performance may require pling interval and the achievable closed-loop band-
an excessive number of step response coefficients width. This is especially important for plants with a
In the previous section we stated that the DMC large open-loop time constant.
step response model requires S; = Sn+ I' In this
section we demonstrate the consequence of sacrific- 3.2. Avoiding limitation 1
ing this requirement. Limitation 1 may avoided by using a state space
Assume that a high closed-loop bandwidth is model which has no truncation error. For example,
desired for the plant described by the following the DMCss algorithm requires only 2 states to repre-
model: sent equation (48) for ~ T = 1, one state for the
100
first-order transfer function and one for the delay.
P(s) = e ", (48) More states are needed if ~ T is less than the delay.
100s + 1 The OBMPC controller can also be used to avoid
In order to achieve the desired bandwidth a short limitation 1. Instead of truncating the response after
sampling interval is required. (A common rule is to
use ~ T'-S:2nllOwB' where W B is the closed-loop 80r-~~-~~~-~~~-~--'

bandwidth, e.g. Middleton, 1991.) We select ~ T=

1 min. According to common practice (e.g. Cutler
70 --Exact model
........Truncated model (DMC) ... .'
.'

60 .-_. --Integrating approximation...··

and Ramaker, 1980), the truncation error should (OBMPC) ./
not be larger than about 5%, which in our case 50
s .'
~~~ ,

yields 300 step coefficients. However, there is ~ 40 "

o
always a practical limit on the number of coefficients 30
(state) that can be used in the internal model, since a 20
large number of coefficients leads to an excessive 10
use of computer memory and a high computational
load. 00 10 20 30 40 50 60 70 80 90 100
Consider the case of selecting n = 3D, which is a Time [min]
typical industrial choice (e.g. Cutler and Ramaker, Fig. 4. Open-loop response to unit step in u at [ = 0 for
1980). The effect of this truncation on feedback different models of l00/(I00s+l) e ". The DMC step
response model is truncated at n = 30, ""T= I [equations
control is demonstrated in Fig. 3 which shows the (3-4)]. The OBMPC model uses A"=I, 8"=s»+,-s»,
response to a unit step disturbance acting on the n = 30 and A T= 1 [equations (25-26»).
 Limitations of dynamic matrix control -tIS

103
- - Exact model
·····-Truncated model (DMC)
'. .-----Integrating approx imation
~ ~L1!;····=·OutPutdisturbar.lce
o . '_... -Input disturbance
"
]102~----"':":"" (OBMPC)
'§, .-... :. .
o : ---------i
,.....-;:c:c--:'::---;:;';=--~_:.~-~-'
o 5 10 15 20 25 30 35 40 45 so
::E 10' Time [min]
..".

O['!I:
'

~
10°

lOio" 10.3 10.2 10"

Frequency [rad/min]
10° 10'
l;~ II __mmuu
a 5 ----'1:-'::0----':-':15:--2::':0::--~25::--""30:---=:3;:-5
1

4()-4~5 -~()
Fig. 5. Frequency response for different models of 100/ Time [min]
(lOOs + 1) e -'. The DMC step response model is truncated
at n = 30, ~ T= 1 [equations (3-4)]. The OBMPC model Fig. 7. Responses for the SISO plant [equation (411)1 with
uses Au=l, Bu=S"+I-S,,, n=30 and ~T=I [equations DMCss controller B Crable 11 Solid curves: urut step
(25-26)]. disturbance acting on the plant output at 1 10 Dashed
curves: unit step disturbance acting on the plant input at
1= 10

n time steps (as would be the case with a DMC step

response model) we may use Au and B; in equation high and low frequencies. The integrating OBMPC
(32-33) to mode] the slow dynamics. This way we model on the other hand, yields excellent agreement
can reduce the number of states required to repre- with the exact model at high frequencies . hut
sent the plant and thereby allow a short sampling displays large dev iation at low frequencies.
interval. Au and B; allow us to use any first-order Simulations with DMC and OBMPC (controllers
model of the slow dynamics. We could obtain an A and D) arc shown in Fig. 6. The disturbance rs a
exact model of the plant in equation (48) by using unit step on the plant output. We conclude that the
Au = e -II.TIlOO and B; = S,,+ 1- s., However, in this rough integrating approximation of the res idual
example we will approximate the slow dynamics dynamics is better than the truncated model Not«
with an integrator, and select Au = 1, B; = S,,+ 1 - SrI that controller D is tuned for ramp drsturb.mcc«.
andn=30. a = 1 (to take care of the low-frequencx mismatch)
In Fig. 4 we compare the open-loop model re- and some measurement norse. f, ,~U.S ( to adul'\e
sponse to a unit step in u at time 0 for the exact high-frequency robustness]
model [equation (48)] (solid curve) with the trun-
cated DMC model [equations (3-4)] (dash-dot 3.3. Limitation 2: Poor response for ramp-ltkr dis-
curve) and the OBMPC model [equations (25-26)1 turbances
with the residual dynamic approximation given The DMC performance may be very poor for
above (dashed curve). The last model gives a large disturbances which do not act as steps on the output.
error as time increases, but does not have the abrupt Figure 7 shows the responses for the plant in equa-
change at t=n!1T which is characteristic for the tion (48) to a unit step acting on the plant output and
truncated DMC model. In the frequency domain input, respectively. A DMCss controller is used In
(Fig. 5), the truncated DMC model is poor both at hoth simulations in order to avoid truncation effects

!:D\=-=
--DMC controller A
:; .5 ·-----OBMPC controller D
~ 01-------' !.---.."...--------,
o
-.5
L-~~_~~~,........,~__::":""---:"=---:'::--;!
o 5 10 15 20 25 30 35 40 45 50 o 5 10 15 20 25 30 35 40 45 so
Time [min] Time [min]

!~5;':u-; ;;;B
01-----, --DMCss controller B
:; -.5 ...... OBMPC controller [)
0- . PI controller
..::; -1

o 5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 4(1 45 ,n
Time [min] Time [min]

Fig. 6. Responses for the SISO plant [equation (48)] with Fig. H. Responses for the 51S0 plant [equation (4~,)i With
different controllers (Table 1). A unit step disturbance acts different controllers (Table I) A unit step disturbance acts
on the plant output at t= 10. on the plant input at 1= 10
416 P. LUNDSTROM et al.

--DMCss controller B [ I~XlW'~. - - Without uncertainty

10" ------With uncertainty
····--OBMPCcontroller D
.-.-.-.-. PI controller :lo~l..,:\~.> -I
" 10-'
o \::::,'=, _ .

1[£
"0 ·5
a
'2 10- 0z---;':1O~2~0,...-;;:3':::'0-4-:':0:--:5-!';::0----:60::--:7,..,,0-8~0~90--.J100
~
:a 10- 3
i T,mO[m;,): : j
10

10' 10" 10° 10'

Frequency [rad/min) o 10 20 30 40 50 60 70 80 90 100
Time (min]
Fig. 9. Sensitivity function vs frequency for the SISO plant
Fig. 11. Responses for the distillation column [equation
[equation (48)] with different controllers (Table 1).
(49») with DMCss controller C (Table 1). A 0.001 step
disturbance acts on YD at t= 10. Uncertainty as defined in
equation (51).
The controller is tuned for dead beat control (con-
troller B) and the output disturbance (solid curve) is
rejected in one sampling interval, since the distur- other controllers perform well. Actually , the PI
bance is in accordance with assumption A2. The controller is almost as good as OBMPC for this
response to the input disturbance (dashed curve) is simple plant.
extremely sluggish. The reason is that a step distur- The difference between the controllers is also
bance on the input results in a slow, ramp-like illustrated in Fig. 9, showing the sensitivity function
disturbance on the output. In this case assumption vs frequency . The DMCss controller yields a sensi-
A2 does not hold and the output prediction used by tivity function with slope 1 for frequencies below the
the algorithm is incorrect which results in poor bandwidth . This shape of the sensitivity function is
performance. The response cannot be improved by optimal for step disturbances and is a consequence
a different tuning since a dead beat controller gives of assumption A2. However, for "ramp-like" distur-
the highest feedback gain of any choice of I', 1\, p bances we need a stronger disturbance suppression
and m for a given ~ T. at low frequencies . With a DMC controller this can
only be achieved by increasing the bandwidth of the
3.4. Avoiding limitation 2 closed-loop system , since the shape of the sensitivity
An observer based MPC algorithm makes it poss- function is fixed (due to A2) . The maximum band-
ible to avoid the output step disturbance assumption width for a given ~ T is obtained by using A = 0
A2 which causes limitation 2. To demonstrate this (dead-beat) and if the resulting suppression of low-
we compare the dead beat DMCss response (con- frequency disturbances is not enough. then a smaller
troller B) with the OBMPC response (controller D). ~ T has to be used. With an OBMPC controller we

We also included a PI controller in this comparison may use a to adjust the disturbance suppression.
to demonstrate the performance of a very simple Figure 9 also shows that the sensitivity function
controller. The PI controller is tuned according to for DMCss controller B goes to zero at (V = 1(/~ T
Ziegler-Nichols rules taking into account an extra rad/min. This is due to the dead-beat tuning and
delay of half the sampling time (Table 1). makes the controller very sensitive to high fre-
Responses to a unit input disturbance are shown quency uncertainty , e.g . dead time uncertainty.
in Fig. 8. The DMCss response is sluggish, while the
3.5. Limitation 3: Poor response for interactive
MIMO plants

(~,I : : : s;:d
In this section we will show that there are cases
with model-plant mismatch when a DMC controller
does not perform well even when the disturbance
actually is a step acting on the output.
There is always a certain mismatch between a real
10-4 10') 10-2 10-' 10° 10'
Frequency (rad/min] process and a model. The mismatch can have vari-
ous sources: uncertainty in the model parameters
Fig. 10 Magnitude vs frequency plot of the (I,I)-RGA and the model structure. inaccuracies of the actua-
element of the distillation column [equation (49») . The
interactions are large at low frequencies (Au = 35.1), but tors and measurement devices , etc . Multivariable
not at high frequencies (Au = 1.0). systems introduce a special problem here because
 Limitations uf dynamic matr ix control ~1 1

Table I. Tun ing param et er s lor co ntroller s

t1T
T ype r A ( min) m p n IL ,

A DlC I II 3 .\ 30
B DM Css I 1I 3 .\
C DM Css I> : 1l.1l2 J,, : 5 10
D OBMPC 1 II 3 .\ 30 I '\ '1 - ' -.'I" 1 1.1 ..'
E OBMPC l : «: ll . 12S /, . , 5 III 3ll ll'195 ." , - '. " " Il.YY5 OY
F OBMPC li « : 0> : 5 111 30 0995 , . • " '-', ,." 11.995 0.22

' t1 ~
K ~)
PI Cr ,(z) = ('Z(_T; ) + K,; K, = llAS. r, =s.nmin . t1 T= 1 min

the "gain" of a multi varia ble process varies not onl y tr an sfer funct ion. We do this only because know n
with frequency, but also with "dire ctio n" , Skogestad del ays fit better into the MP C framew or k
et al. (1988) show that if a plant is ill-conditioned Skogestad et at. (l990) dem on strat e that" fre -
irrespective of scaling, then the co ntro l performance quency dependent relative gain arra y (RG A ) (Bris-
is strongly affected by input unc ertainty, in particu- tol, 1966) is a useful tool to check how se nsitive a
lar, when the controller is tr ying to invert the plant. plant is to inp ut uncertainty. Figure 10 shows the
The DMC controller is such a co ntro ller. especially , ( 1,1) RGA eleme nt. A. l l . of the dist illa tion col umn.
if the penalt y weight on the inp ut moves is low . as a function of freque ncy . ;. i t is high (:'is. l) »t low
Since there is alw ays some input uncertaint y, it frequencies but falls to one at higher freq uen cies .
should be clear that a DMC co ntro lle r is potentially T his sho ws that a D MC co ntrolle r ma y have pr ob-
bad when used for an ill-conditioned plan t. lem s with low-frequency inp ut uncertaint y
3.5.1. M1MO example. We use a distillation col- 3.5 .2. Eff ect of input uncertainty . We assume tha t
umn as an example process. The model is from th er e is 2()°j" unce rtaint y in the input move s. Fro m a
Skogestad and Morari ( 1988) and is den oted "co l- singular valu e ana lysis, one ca n deter mine that th e
umn A " in their paper. T he co lumn is described by wors t steady sta te effec t is obt aine d whe n the u nce r-
the follo wing equation s: tainties in A I. an d A VB act in opposite di re ction s

)' l (Skogesta d et al .. 1988). In the simulation s Wf use :

l
k ll ll 11 ll
- - - - - - eO'
( k) ++1'k:!S 1+k r,s
dYD ] = 1 + ,1'15 e- o, A L..d".,= I.2A 1. , ,,rnpu,cJ and
[ d XB k~ ,
- - gL(S) e - Os
(
+
k1 1 k 11
- -- - - -
k!l )' . C 17.\
l\ \/ Hactu at = 0.8/1 \!Jh " IUPutl:tl ' (5 1)
1 +1' ,5 I+r ~ I + r, s Respon ses for co ntro ller C (T able 1) to a n.OO I
ste p disturbance acting on I' D are sho wn in Fig 11
(49) E rro rs in th e input gains lead to very sluggish distu r-
ban ce rej ect ion although the dis turba nce IS In
where gL(S) expresses the liquid flow dynamics: accordan ce with the DMC disturbanc e assu mp tio n,
I Th e reaso n for this slow settling is tha t t he effect of
(50) the er ro rs in the manipulated inp uts is similar l O the
ef fect of inp ut step di sturba nces . Not e that this
OL is the overall liquid lag from the top to the bottom sluggish be havior ca nno t be Improved by a lterna tive
of the col umn. nT in equa tion (50) sho uld be equal choices of t uni ng par ame te rs
to the number of tra ys in the co lumn, bu t we use Th is ca n also be dem on strat ed in a plot 0 1 the
nT= 5 to avoid a mod el of unn ecessar y high order. sing ular values of the se nsitivity funct ion (Fig 12) .
Reflux Land bo ilup VB are man ipul at ed inputs , top Bo th the so lid cu rves (no unce rtaint y) have slope 1
co mpos ition, YD and bottom composition XH ar e which is a co nse q ue nce of the dis turbance asvurnp-
controlled outputs. We use th e following par ameter tion . Th ey also lie close to each other. which shows
values ; k ll == 0.878, k ,z = - O.RM, k 21 = 1.082, k 22 = th at the sens itivity function is we ll-cond itio ned .
1.096. T,=194min , Tz=1 5min. 0 = 1 min , 8 L = Since the plant itse lf is ill-con d ition ed we can con-
2.46 min and nT= 5. Skoges tad and Morar i do not elude that the co ntroller IS co mpe nsat ing h)l the
include any specified del ays in their model , inste ad dir ect ion ality o f the plant. Such a co nt ro lle r IS basi -
the y use a norm bounded uncertaint y de script ion to ca lly inve rti ng the plant and the syste m shou td be
cover the effect of de lays and ot he r un modeled high se nsitive to input uncertaint y. Indeed , th i-, I , the
frequency dynamics. In equa tion (49) we assume the case as see n both from the simulation in Fig. 1! a nd
delays to be known and eq ual to 1 min for each from the large difference bet ween so lid and dot te d
418 P. LUNDSTROM et al.

- - Without uncertainty
10° ...... With uncertainty

10 20 30 40 50 60 70 80 90 100
Time [min]

J~~:: ;: :<1
--Without uncertainty
10- "
"
.... --With uncertainty
.' .... -. Reference
IO"''--;--'~~"--:;-~~'"'''''~~'''-;-~~'''-;;-~~:'''
10- 4 IO. J 10 2 10.1 10° 10
1

Frequency [rad/min] 80 10 20 30 40 50 60 70 80 90 100

Time [min]
Fig. 12. Maximum and minimum singular values of the
sensitivity function for the distillation column [equation Fig. 13. Responses for the distillation column [equation
(49)J with DMCss controller C (Table 1). Uncertainty as (49)J with OBMPC controller E (Table 1). A 0.001 step
defined in equation (51). disturbance acts on YD at t= to. Uncertainty as defined in
curves in Fig. 12. The dash-dot curve in Fig. 12 is equation (51).
included as a reference. It is an upper bound on
sensitivity functions which achieve about 20 min both cases the residual plant dynamics for each
closed-loop time constant and a maximum sensi- input-output pair is approximated by b,/(z - a,J
tivity peak of 2. where blj=s'.i.n+I-Si.i.n and a'J=0.995, i.e. a first-
By comparing Figs 10 and 12 we see that there is order response with a time constant approximately
an excellent agreement between the predicted effect equal to 200 min and a gain determined at the
of uncertainty, based on the RGA-plot, and the "truncation" step. The disturbance is assumed to
actual effect seen in the sensitivity plot. [However, have the same dynamics as the plant, i.e. a, = 0.995.
this sensitivity plot is only showing how the control Figures 13 and 14 show responses for OBMPC
performance deteriorate for this specific input error controllers E and F, respectively, for the same
(+20% in Land -20% in VB), and there may be an output disturbance as used in Fig. 11. The OBMPC
even larger effect for other error combinations of controllers perform well despite the uncertainty
the same magnitude, i.e. the plot is not necessarily [equation (51)1 and suppress the disturbance much
showing the "worst case" of a norm-bounded uncer- faster than DMCss does (Fig. 11). The response of
tainty.] controller E is nearly unaffected by the uncertainty.
Controller F yields an almost perfectly decoupled
3.6. Avoiding limitation 3 response when there is no uncertainty (i.e. XB is not
There are two different ways to deal with the affected by the disturbance in YD), while the re-
problem caused by input uncertainty demonstrated sponse with input error clearly demonstrates inter-
in Fig. 12: action between the two loops.
In the case of an input disturbance the difference
1. Use a controller that does not correct for the
between the OBMPC controllers and the DMCss
directionality of the plant.
2. Increase the gain at those frequencies where
the disturbance suppression is poor. xlO-4
10

~
--Without uncertainty
The first method is suggested in Skogestad et al. ...... With uncertainty
"a. 5
~--::::=.:..
(1988). It gives a controller with somewhat sacri- o" 0 '\~
ficed nominal performance, but the performance is ·5 ",\,:,~,:::"
much less sensitive to uncertainty because the con- o 10 20 30 40 50 60 70 80 90 100
Time [min]
troller does not correct for directionality.

J~nv:: :: :::
The second approach will work if the uncertainty
only causes problem at low frequencies. With this
approach the controller is still sensitive to uncer-
tainty, but this is counteracted by increasing the 1
controller gain at low frequencies to make the .80 10 20 30 40 50 60 70 80 90 100
Time [min]
nominal response much better than what is nomi-
nally needed. Fig. 14. Responses for the distillation column [equation
(49)J with OBMPC controller F (Table 1). A 0.001 step
We will now demonstrate the two approaches, disturbance acts on YD at t= 10. Uncertainty as defined in
using OBMPC controllers E and F (Table 1). In equation (51).
 Limitations of dynamic matrix control 419

high that even with uncertainty the performance IS

satisfactory, except over a short frequency range.
10°
Although controller F yields satisfactory perfor-
mance, we may conclude that the plant is rather
sensitive to input uncertainty also at frequencies
above the bandwidth. Thus, the best tuning
approach for this plant is approach I above, used for
- - Without uncertainty
tuning controller E .
.----- With uncertainty
..... Reference
4. mSCUSSION
1~ 1~ 1~ 1~
We have studied feedback limitations of uncon-
Frequency [rad/min]
strained DMC with a quadratic objective function
Fig. 15. Maximum and mini~um s.ingular values of the [equation (13) 1. There are several variants of DMC:
sensitivity function for the distillation column [eq~atlOn
"original" DMC (Cutler and Ramaker, 1980), DMC
(49») with OBMPC controller E (Table I). Uncertamty as
defined in equation (51). with least squares satisfaction of input constraints
(Prett and Gillette. 1980). DMC with constrained
linear programming optimization (LDMe) (Mor-
controller would be even larger, because of limi-
shedi et al.. 1985). DMC with constrained quadratic
tation 2.
programming optimization (QDMC) (Garda and
The sensitivity plot for controller E is shown in
Morshedi, 1986). These variants use different opti-
Fig. 15. This controller is using a high input weight
mizers but the predictor is the same for all of them.
and a large disturbance-to-noise ratio. The plot
Both the limiting assumptions (A I and A2), which
shows that this controller does not try to invert the
we have studied, are implicit in the predictor and
plant; the solid curves (no uncertacnty) do not lie
will not be avoided by modifying the optimizer, so
close to each other. We can also conclude that it is
the results in this paper hold for all these algorithms.
insensitive to uncertainty since the dotted curves
The results also carryover to the general case with
[uncertainty defined in equation (51)] lie close to the
constraints, since the issue of constraints only affects
solid curves.
the optimizer.
Figure 16 shows the sensitivity plot for controller
The requirement S; = S" , i of assumption J, can be
F. This controller has no weight on the inputs (dead
avoided within the DMC framework (as defined by
beat KMPc) but is tuned for substantial measurement
Fig. 1) by using a general state space model instead
noise (fa = 0.22). From the plot we see that this
of a step response model. However, the plant still
controller is inverting the plant for frequencies
has to be stable to ensure internal stability.
above oi =0.01 rad/min. Below this frequency it is
Assumption A2 cannot be avoided unless the con-
still trying to invert the plant, but it does not suc-
stant matrix .J is exchanged with a transfer function
ceed, since the true plant and the model with
states. This is most clearly seen in the DMCss algor-
approximated residual dynamics are slightly differ-
ithm where the input to .9 IS l(k)v(k) cannot be
ent. This controller is sensitive to input uncertainty
"filtered" bv j' since it has no states, no knowledge
in the sense that low frequency controller gain is so
of previous- measurements. Using a gain different
from 1 in .9 would not filter v(k) but rather intro-
duce a steady state offset.
.
-~'-':'''-'''.; ' . ' ' ' ' ' ' ' ' ' '
In this paper we have excluded the feedforward
10°
. . :»> .- part of the algorithms. although feedforward control
.' is a standard feature of MPC. It is simple to include
feedforward in the algorithms by introducing mea-
sured disturbances as inputs to the predictor.
However, this does not affect our results. the feed-
/--Without uncertainty
..' ." ... With uncertainty back limitations arc still present
10'
.' .' ..... Reference In the following we discuss some of the results and
our choice of example processes and controller tun-
10" 10.2 10.1 10°
Frequency [rad/min]
ings.
The SISO example [Equation (48)1 has a time
Fig. 16. Maximum and mini~um singular values of the constant much larger than the time delay. This
sensitivity function for the distillation column [equation
(49») with OPMPC controller F (Table I). Uncertamty as parameter choice is made on purpose to demon-
defined in equation (51). strate limitations Ll and 1.2. since thev are
420 P. L UNDSTROM et al.

espec ially important when the time constant of the NOMENCLATURE

open-loop plant is large. compared to the desired
closed-loop time constant. In the SISO simulations A == Sta te matrix
.<Jl == D isturbance stat e ma trix . equa tio n (37)
we use controllers whe re the optimizer is tuned for a = Elem ent in state mat rix
dead-beat control. The reason for this tuning is th at B = Input matrix
we want to produce clear illustr ative simulations b = Eleme nt input matrix
C = Output matri x
wher e the controller action is easy to understand . In fa= T un ing par ameter in filter ga in K
real applications one should always use a nonzero f~ = T uning par ame ter in filter ga in K
input weight A to achieve some robu stness to noise .'1' == Eq uati on ( 12)
K == Kalm al filter gain
and high-frequency mod el-plant mismatch . K,== Pl -con tro ller gai n
All the simulations present ed in this paper are K MP<" == MP C fee d back gain
without measurement noise. Again the reason is that k ;i = G ain . equa tio n (49)
L=Rcflu x
we want to show clear illustr ative simulations. (We M == Matrix in step response mod el , eq uat ion (3)
have also performed the simulations with noise and m = Inpu t horizon
it does not change our results. ) N = Mat rix in step response mod el , equa tion (4)
/I = ModeJ horizon
We have demonstrated that a truncated step re- /I ll = Number of inputs
spon se model may cause severe model-plant mis- /I , = Number of outputs

match . both at high and low frequencies. p= Prediction horizon

0l = Set point vector
Low-frequency mismatch is not critical as long as the S == Ste p response coefficient mat rix, equation (3)
"sign" of the proce ss is correct , but high-frequency s == Step respon se coefficient
mismatch may yield an unstable syste m (Fig. 3) . T== D isturbance input matrix, eq uati o n (25)
!'1 T == Sam pling time
With the OBMPC controller trun cation is avoided. t ==Time (min)
In controller D we use a rough approximation of the "lI. == O ptima l contro l sequence
residual dynamics (a n integrator) to show that eve n II = man ipul ated input
V= Noise covariance matrix
this approxi mation is better th an truncation. VB ==Boilup
The sensitivity plots for the distillation column v = Measurem en t noise
(Figs 12, 15 and 16) show singular values of the W = Disturban ce covariance matrix
w = D isturban ce
sensitivity function for no uncer tainty and for one x = Sta te variable
specific case of uncert ainty. However , we have used XR == Bottom composition (kmol/ kmo l)
the structural singular value , u, to chec k that the Y == Ou tput vec to r
~ == Dynam ic sta tes of DM C pr ed ictor
controllers will perform well also for othe r cases of y ,= D ynami c states o f OBMPC predic to r
uncer taint y (see Skogestad et al., 1988, for '\lI== Pred icted output vecto r
instance) . y == Contro lled ou tput
We have not put any effort into finding the best
y = Me asur ed outut
YD = Top composition (kmol/kmol)
approximation of the residual dynamics, although z '= Shift ope ra to r
that may have improved the OBMPC responses
Greek symbols
further , instead the simple st choices of Au and B;
are used throughout the paper. Hovd et ai. (1991) a == Par am eter in disturbance mod el, eq uatio n (37)
discuss how to choo se Au and b ll in an optimal r == O utput weighting matrix , equation (13)
fashion . !'1== !'1 u(k )==II(k )-u (k - l )
8==Time del ay (min)
1\ = Input weighti ng matri x. equa tion ( 13)
5. CON CLUSION S
;'11 =( 1,1) RG A eleme nt
We have shown that there are situations where TI = Pi -cont roller integ ra l tim e co nsta nt (m in)
r i o T2'=Timc co nsta nt (min) . equa tio n (49)
the feedback performance of DM C is poor irrespec- W '= Fr equen cy (rad/ min)
tive of tun ing. Th is poor prform ance is due to the We '= Closed-loo p bandwidth (rad/ min)
two assumptions (A I and A2) made in the pr edictor
Abbreviations
part of the algorithm . Thi s explains why the perfor-
mance cannot be improved by different tuning- DM C'= Dynamic matrix control
different tuning onl y affec ts the optimize r part of the D MCss == D ynamic mat rix control with sta te space
algor ithm. model
MIM O == Mult i-input-multi-ou tput
Th e OBMPC algor ithm by Lee et al. (1994) allows MPC = Model pr ed ict ive co ntrol
us to avoid the limitati ons of DM C and still preserve O BMPC = O bser ver based model pr ed ictive control
all the attractive properti es of the DMC algorithm. Q D MC == qu ad rati c programming dynam ic matrix
co ntrol
A ckn owledgement-Partial suppo rt fro m th e National RGA = Relat ive gain array
Science Foundation is grat efull y acknowledged. S[SO = Single input-single output
 Limitations of dynamic matrix control 421

REFERENCES Li S., K. Y. Lim and D . G . Fisher , A state space formula-

tion for model predictive control. AIChE Il 35, 241- 249
(1989).
Bristol E. H. On a new measure of interactions for multi- Middleton R. H. Trade-offs in linear control system
variable process control. IEEE Automat. Control design. Automatica 27, 281-292 (1991).
xc-n. 133-134 (1966). Morshedi A. M.• C. R. Cutler and T. A. Skrovanek ,
Cutler C. R. and B. L. Ramaker , Dynamic matrix Optimal solution of dynamic matrix control with linear
control-a computer control algorithm. Proc. Joint progr amming techniques (IDMC) . Proc. Am. Control
Automatic Control Conj., San Francisco , CA, Paper Conj ., Boston , MA , pp. 199-208 (1985).
WP5-B (1980). Prett D . M. and C. E. Garda, Fundam ental Process
Garda C. E . and A. M. Morshedi , Qu adr atic program - Control . Butterworths, Stoneham, MA (1988).
ming solution of dynamic matrix control (QDMC). Prett D. M. and R. D. Gillette , Proc. Joint Automatic
Chem. Engng Comm/tn. 46, 73-fn (1986). Control Conj., San Francisco , CA (1980).
Garda C. E., D. M. Prell and M. Morari , Model predic- Skogestad S. and M. Morari, Understanding the Dynamic
tive control: theory and practice-a survey. Automatica Behavior of Distillation Columns . A1ChE il 33. 162()-
25,335-348 (1989). 1635 ( 1988).
Hovd M. , J. H. Lee and M. Morari, Model requirements Skogestad S., M. Morari and J. C. Doyle , Robust control
for model predictive control. Europ ean Control Conj., of ill-conditioned plants: high-purity distillation . IEEE
Grenoble, France (1991). Automat. Control 33, 1092-ll0S (1988).
Lee J. H., M. Morari and C. E. Garcia, State-space Skogestad S.. 1'. Lundstrom and E . W. Jacobsen , Selecting
interpretation of model predictive control. Automatica the best distillation control configuration. AIChE JI 36.
30,707-717 (1994). 753-764 (1990).

CACE 19-4-0