Copyright ~ IFAC New Trends in Design of Control Systems.

Smolenice. Slovak RepUblic. 1994

STABLE CONSTRAINED PREDICTIVE

CONTROL
M. FIKAR and J. MIKLES
Department of Process Control, CHTF STU, Radlznskiho 9, 812 37 Bratislava ,
Slovakia , E-mail : jikar@cvtstu.cvt .stuba .sk

Abstract. A multi variable linear controller based upon pole placement design is devel-
oped . This apriori assures stability of the closed-loop system. The generalized controller
from the family of all stabilizing controllers satisfying given constraints on inputs and/or
outputs is then searched . In addition an objective function based on future prediction of
signals is minimized.
Keywords. Pole placement, Multivariable systems , Predictive control, Constraints ,
Quadratic programming

1 INTRODUCTION constraints are achieved, quadratic function min-

imized and closed-loop system has poles in the
prespecified locations. The first results show
Constraints on inputs and recommended con-
that the approach is very promising (Fikar and
straints on outputs are very common in prac-
JI!lrgensen, 1993, Fikar 1994)
tice . If these are not taken into account while
designing the control algorithm, the closed-loop
behaviour may deteriorate. The successful meth-
ods dealing with the presence of the constraints 2 POLE PLACEMENT ALGORITHM
are usually based on the predictive control and
receding-horizon strategy. These methods pre- The following algorithm has been described in
dict possible future control and output sequences Mikles (1990). We consider MIMO system with
according to present state of the controlled sys- r inputs and r outputs illustrated in Fig. 1
tem. The best known representative is Gen-
eralized Predictive Control (GPC) (Clarke et.
al., 1987) . The control sequence minimizing
quadratic performance criterion is calculated and
only its first element is applied . As such the
methods are only open-loop in design and only
a very limited understanding of their stability
properties exists (see for example Zafiriou and
Marchal , 1991) .

In this paper a new type of predictive control is

presented . The algorithm can handle constraints
and minimizes quadratic objective function. In
addition, the pole placement design guarantees Fig . 1: Block diagram of the closed-loop system
stability of closed-loop system in the presence of
the constraints. The resulting generalized con-
The controllable and observable process is as-
troller (GC) is linear, but with time-varying co-
sumed to be modelled by the discrete-time equa-
efficients even in the non-adaptive case. The idea
tion
behind the algorithm is to search for the control-
ler from the family of all stabilizing controllers
and to choose its parameters in such a way that

405
where y, u are process output and control- M is achieved and the pole placement tracking
ler output r - vector sequences respectively. problem has the solution if and only if ALF, BL
AL, BLE n",,[z-l], HyE n"dz- l ] are polynomial are r.l.p. The feedback part of the controller (6)
matrices. The argument z-l will be omitted IS gIven as a solution of the Diophantine equa-
in the text for simplicity. AL(O) is invertible , tion :
BL(O) = 0, and Hy represents the influence of
process initial conditions . AL and BL are rel-
atively left prime (r.l.p.) and describe input - (8 )
output properties of the controlled system .

The reference signal w is r - vector sequence and and the feedforward part RL as a solution of an-
is modelled by the equation other Diophantine equation:

Fw=G (2) (9)

where FEn",,[z-l], GEnrdz-l], the pair F,G where SE n",,[z-l] is a polynomial matrix.
=
is r .l.p ., F(O) is invertible , F diag(f(z-l )) .
The monic polynomial f depends on the type of There are infinity solutions of eqs. (8), (9) . Al-
the reference vector, e. g. f = 1 - Z-l for step though usually a minimum degree solution is
changes . used, it is possible to search among general solu-
tions if any additional requirements on the con-
The controller is described by the equations troller are laid . Thus, the minimum degree con-
PLu = RLW-QLy+H" (3) troller Po, Qo, Ro serves only as a basis for find-
Fu = u+Hu (4) ing the expression for the set of all stabilizing
controllers. For our case, such controllers for
PL, QL, RLE n",,[z-l], H", HuE n"dz-l] . The
specified closed-loop poles can be in left matrix
pairs PL, QL and PL, RL are r.l.p ., PL(O) is in-
fraction representation parametrized as
vertible and the terms H" and Hu represent the
initial conditions of the controller. The second h PLO + ZBL (10)
equation assures that the controller has integra- QL QLO - ZALF (11)
tion properties and gives zero tracking error for
given class of reference vectors w.
RL = RLO + FX (12)
S = Sp - BRX (13)
Let us further define right matrix fraction de- where X, ZE n",,[z-l] are polynomial matrices.
scriptions of the process and controller
This general controller will be used in the fol-
ALl BL = BRARl (5)
l lowing section where the matrices X, Z will be
PilQL QRP R (6) chosen to satisfy additional demands on control-
and assume that these descriptions are right rel- ler.
atively prime (r.r.p.)

Finally, to specify the closed loop poles we define 3 CONSTRAINED PREDICTIVE

diagonal polynomial matrix ME n",,[z-l], which
is assumed to be stable and can be expressed in
CONTROL
the form
In this section the design of the predictive con-
troller will be very briefly outlined. The control-
M=mI (7)
ler is optimal for all cases of the initial conditions
and the design enables to incorporate the knowl-
where m is a stable monic polynomial in the op- edge about the constraints on inputs and/or out-
erator Z-l. We assume that the poles are the puts.
same for the each loop . However, this condition
is not very restrictive as will be shown later . The objective function to be minimized is the
following finite horizon quadratic criterion
If the matrix M is chosen a priori, then the con- N
troller is said to be based on pole-placement de- F(t) L e(t + j f R.e(t + j)
sign. j=N,
N.
Theorem The internal stability of the closed- + L u(t + j - If R,.u(t + j -1l14)
loop system with the poles given by the matrix j=l

406
where NI, N, Nu are minimum, maximum, and
subject to A [ : ] >b
control horizon respectively, Ru = diag( rui),
Re = diag(rei) are positive definite matrices used
as weights on ii and tracking error e . We assume where
that f = 1 - :-1 so that the signal ii is the se-
A (s:r;U ) szU) ) (23)
quence of the control increments. -Sr U) -Sz U)
After some manipulations on the equations (1)- b ( SminU) - soU) ) (24)
soU) - sma:r;U)
(9) the sequences ii and e can be written as
for sU) ET, j = N1 . .. N, j = 1 . .. Nu
ii M-I ARRLGL
+ M-I[F H.:;) - QR(BLH u + F Hy») (15)
e M-ISGL 3.1 User specifications
M-I[PR(F Hy + BLHu)BRH.:;) (16)
If the general controller is searched, the controller There are several variables which user must spec-
polynomial matrices must obey eqs . (10)-(13). ify in the implementation of the algorithm. func-
Moreover, the initial conditions H.:; depend on tion of some among them is similar to other
X, Z and can be written as H.:; = Ho + HJrx + approaches (GPC) . There are following tuning
HJzz where x=coef(vec(X», z=coef(vec(Z)) (see knobs: degrees nx, nz of the design polyno-
Boekhoudt, 1990 for definitions polynomial to mial matrices X,Z, horizons N , N 1 ,Nu , weight-
vector transformations), x, z are column vectors ing matrices Re, Ru, and the closed-loop poles
containing the unknown coefficients of X, Z. The M.
subscript 0 denotes free responses that do not
depend on X, Z and the subscript f forced re- Degree of matrices X, Z . These are the most im-
sponses. portant parameters for the user. The higher the
degrees, the more degrees of freedom exist and
The equations (15), (16) can also be decomposed the less value of the objective function can be
into free and forced responses (see Fikar, 1994) : obtained. However, the lower degrees of results
in fewer computations, so the values should be
deg(ii) = Nu (17) chosen carefully and depend on the speed of the
computer as well as on the time constants of the
process . The matrix X influences mainly set-
e=Eo+Erx+E.z deg(e) = N (18) point changes and the value of nx should be cho-
sen according to their magnitude. The matrix Z
applies to disturbances. As their magnitude is
The objective function (14) can be transformed
usually lower as the setpoint, the value of nz = 0
to the form
is usually satisfactorily.

Minimum Horizon NI. The usual value of mini-

mum horizon is NI = d + 1 where d is minimum
where g is the gradient and H the Hess matrix. known time delay of the process.

The predicted signals sET = {ii, u, e, ... } are Maximum Horizon N. Unlike other predictive
of the form (cf. eq (18) ) control approaches where the value of N depends
on time constants of the process and should be
(20) chosen to the rise - time of the plant, in our
case dependency between Nand n exists. If
deat-beat controller is assumed, then the influ-
and the constraints considered ence of forced responses is finite and given as
N = nx + deg(AL) + d. This value is also recom-
Smin(t + j) ~ s(t + j) ~ smar(t + j) (21) mended if M :f. I

Thus the original constrained control problem is Control Horizon Nu . Control horizon has the im-
finally rewritten as the optimization problem-in pact on the complexity of the computations. Its
this case as Quadratic Programming task: value is usually chosen equal to N. Nevertheless
it is possible to reduce it, so only first few samples
are optimized . Unlike GPC, there cannot be in-
min c + gT [ x ] + [xT, zT)H [ x ] (22)
x ,z z z stability if even Nu = 0 with nonminimum-phase

407
system to be controlled. Thus infeasibility of the QP problem becomes an
excellent tool to diagnose the status of the pro-
Weighting Matrices. The weighting matrices de- cess. If the model of the process is guaranteed
termine the activity of the controller and have to be perfect , then if it is not possible with given
the same role as in GPC. If the norm of Ru is constraints to reach given setpoints then the set-
larger compared to the norm of Re more cautious points should be changed.
controller results. In the opposite situation, the
control error vanishes with faster rate , but the
control actions can be very excessive . 3.4 Control Algorithm

The control algorithm given below summarizes

Closed-loop poles. The choice of the closed-loop
the steps necessary for the predictive constrained
poles is not very crucial. It has been observed,
control design.
that their influence weakens with increasing nx .
This is probably due to the fact, that also the
objective function minimized influences the be- For given controlled system in left matrix fraction
haviour of GC. description AL, BL and desired closed-loop poly-
nomial matrix M choose parameters of the ob-
jective function (14) (nx , nz, N, NI, Nu, Re, Ru)
3.2 Complexity of the Optimization and repeat :

The complexity of the Quadratic Programming

measured by the number of the unknown vari- 1. find minimum degree controller PLO, QLO,
ables nl and the number of the constraints n2 RLo from Diophantine equations (8), (9),
increases with the horizon N, the degrees nx, nz
2. construct predictions U, e - eqs. (17), (18),
and number of outputs r. The unknown variables
are the coefficients of the polynomial matrices 3. form Hessian matrix H and gradient vector
X(z-I), Z(Z-I) . Thus the general formulae are g given in modified objective function (19),
nl = r2(nx+nz+2)
4. if no constraints are specified find unknown
n2 2r(N - NI + Nu)
vectors x, z ; goto step 9,
However, the number of unknown variables can 5. form matrix A, vector b (eqs. (23), (24)),
be reduced if either reference or controlled out-
put vector remain constant over some interval. 6. find zero columns in A and delete them in
The influence of the matrices X, Z is cancelled A, H, g and delete corresponding rows of H,
and the gradient vector and Hessian matrix con-
tain zero columns. These can be removed and nl 7. find zero rows in A and delete them in both
reduced. A,b,

8. solve Quadratic Programming problem (22),

It is also recommended to search for the con-
straints numerically independent on X, Z con- 9. construct controller PL , QL, RL from equa-
taining either zeros or small numbers (for exam- tions (10)-(12),
ple 0 ~ 0) and to remove them from the set of
constraints. These are usually responsible for the 10 . implement control law from equations (3),
incorrect results of the QP. (4),
11 . wait until next sample; goto step 1.
3.3 Feasibility of the solutIon

There are two possibilities, for the solution of 4 SIMULATION RESULTS

the QP problem to become unfeasible. In one
case there may be a large set point change or dis- The aim of the simulations was to demonstrate
turbance, for which the degrees nx, nz are too the behaviour of the non-adaptive generalized
small. However, the sizes of these changes are controller and to show the influence of its param-
usually known in advance and the degrees can eters to the performance of the closed-loop sys-
be chosen sufficiently large to prevent such a sit- · tem . Further, to compare GC with its ancestor-
uation . The second reason for infeasibility is in- basic pole-placement controller (BC) and to dis-
consistency of the constraints with the model. cuss the advantages and drawbacks of the both
This case is usually caused by a wrong model. approaches.

408
The system controlled was discrete system of the N..c-" •• N.O· ······ ItJ.,..- ,..0- ..
form
AL 1+ Alz- l + A2Z-2 2

EL

~~6 ) A2 = ( :~ .2 )
.3
~~--------~.~.~
.. ~
..----------~----~~

:~ ) E2
Unless otherwise stated , the closed-loop poles
= ( ~~2 ~ )
were chosen M =
I (deat-beat response) and the
horizons N = Nu = 7, Nl = 1.
T",.

The first simulation shows the dependency of GC

Fig. 3: The role of weighting matrices
on the most important design parameter nx . In
Fig. 2 the dependency of the objective function
F(t) (see eq. (14)) and maximum control incre-
ments are shown for the reference vector (1, -1 f that the weighting matrices can be the tools for
and weighting matrices Ho = R • . It follows from fine tuning of the algorithm. The extreme cases
the theoretical assumptions, that the objective Ho =
0 or R. =
0 also indicate the maximum
function should be decreasing . The discrepancy ability of GC for given nx if constraints are to be
found for larger nx when F(t) is increasing is introduced .
probably due to the numerical properties of the
algorithm. Thus, if our main interest was in the The influence of closed-loop pole locations to the
minimization of the objective function, in our algorithm is shown in Fig. 4. Case A corresponds
case it would be satisfactory to choose nx = 2. to aperiodic behaviour with the pole z = 0.4, case
B to periodic pole z = -0.4, with other param-
It might be suggested that the use of larger nx eters nx = 0, Ho = R. . Although the objective
is not only computationaly demanding, but also function (14) is minimized, the poles also deter-
not necessary as well. However, the right graph mine the closed-loop performance.
shows the maximum control increments occurred 2 ·· .... A:- u" ....... ,,'2
during the response of the controlled system to · r 8.- ···· o2r - u22
reference change and confirms the assumption, ~ .n i n nl __ . ' _________ _

that for larger nx, the control actions are much o

smoother . This graph also gives a hint in choos- ·r !i ( ... _ _..;i

ing the minimum nx if limits on ii are assumed.

I__
12.5
r.•
- ---- ~rl
~
.:
0

~
.
8.-----
y
y

~-\
-w
• r.a
12.0 ·r
• i r.• o 2 ro

" .5
I r1
=-
iL I r.O
Fig . 4: Choice of the closed-loop poles
11 .0 iD'
L .a
-"-
---.
The last two simulations deal with constrained
rO.5 0 .•
control. In the first case the tracking error for
01
2 3 • 2 3
• 5 variable Y2 was specified to be less than 10% of
the corresponding setpoint-in other words the
overshoot less than 10% was demanded . Fig. 5
Fig. 2: The influence of nx shows that this demand was fulfilled with the
corresponding deterioration of the unconstrained
The second simulation compares BC with GC output .
and also demonstrates the influence of weight-
ing matrices for nx = 3. Figure 3 shows, that In the second case, the constraints on inputs were
the deat-beat response is 5 steps for GC and 2 imposed as lul 1 and liil :s
0.6. The results :s
steps for BC , which is in agreement with theory. in Fig. 6 show slower response to the setpoint
Only one input and output are shown. It is clear change, but both constraints were satisfied .

409
~
5

0
I,f
r'
'Jtf', , ,
:

:
..................
flr' , ~'
rlr.............:::::
:'oo :
---- u:z If the parameters of the process are unknown
or time-varying and adaptive control is re-
quired, the enhancement of the algorithm is
• I ~.
straightfowward by including the appropriate
t _ _'.._:_ _ _ _ _..........---'----''--_ _ _......-...JJ
·5 L RLS method for th e identification of the process
model.

6 REFERENCES
Boekhoudt , P . (1990). Solution of polynomial
Time
equations in H 00 optimal control. Control
Theory and Advanced Technology , 6, 321 -
Fig. 5: Constrained control. e2 ~ W2
337.

'I
l
g j
Clarke, D. W., C . Mohtadi, and P . S. Tuffs
(1987). Generalized Predictive Control -
part I. The basic algorithm. A utomatica ,
~ .:1
"---1- -..i
• •-
• • •- ~ ••_• • • • • • • •
'•-
• •-
• •-~..1
•• 23, 137 - 148.

Fikar, M. (1994) . Control of multivariable pro-

=<: ••
cesses . Parametrization of aI/ stabilizing

~~r~"~~
control/ers, PhD thesis , Dep . of Process

d Control CHTF STU, Bratislava.

Fikar, M., and S. B. J!1Srgensen (1993). Mul-

tivariable constrained adaptive predictive
o 2 W
control based on pole placement. In:
K. Warwick, M. Karny, editors, Mutual Im-
Fig. 6: Constrained control. lul ~ I, liil ~ 0.6 pact of Computing Power and Control The-
ory, Plenum Pub . Corp ., London, 301 - 310.

MikleS, J. (1990). A multivariable self-tuning

5 CONCLUSIONS controller based on pole-placement design.
Automatica, 26 , 293 - 302.
This paper contains derivation of the generalized
Zafiriou, E., and A. L. Marchal (1991) . Stability
controller within the predictive control frame-
of SISO Quadratic Dynamic Matrix Control
work . The controller is able to handle constraints
with hard output constraints. AIChE Jour-
on inputs and/or outputs while satisfying the sta-
nal, 37, 1550 - 1560.
bility of the closed-loop system . With this prop-
erty is GC superior to other model-based predic-
tive control approaches (GPC, DMC, ... ) where
even in the unconstrained case is stability defined
very vaguely. The price paid for this superior-
ity is computational load and time-varying coef-
ficients of the controller even in the nonadaptive
case .

The second advantage of GC lies paradoxly in

the case when it fails . If the degrees nx, nz are
chosen large enough to cope with given reference
changes and disturbances, the only reason for in-
feasibility of the controller is inconsistency of ei-
ther model of the process (in the adaptive con-
trol) or constraints (if any) . Thus the failure of
QP algorithm becomes an excellent tool to diag-
nose the status of the process. It is then up to
control designer to choose the appropriate strat-
egy, usually to reformulate the constraints or ref-
erence vector.

410