Documente Academic
Documente Profesional
Documente Cultură
www.elsevier.com/locate/jprocont
Abstract
In this paper, robust model predictive control (MPC) is studied for a class of uncertain linear systems with structured timevarying uncertainties. This general class of uncertain systems is useful for nonlinear plant modeling in many circumstances. The
controller design is characterizing as an optimization problem of the ``worst-case'' objective function over innite moving horizon,
subject to input and output constraints. A sucient state-feedback synthesis condition is provided in the form of linear matrix
inequality (LMI) optimizations, and will be solved on-line. The stability of such a control scheme is determined by the feasibility of
the optimization problem. To demonstrate its usefulness, this robust MPC technique is applied to an industrial continuous stirred
tank reactor (CSTR) problem with explicit input and output constraints. Its relative merits to conventional MPC approaches are
also discussed. # 2001 Elsevier Science Ltd. All rights reserved.
Keywords: Model predictive control; Uncertain linear system; Structured uncertainty; Linear matrix inequality; Industrial application
1. Introduction
Model predictive control (MPC) techniques are
widely used in industrial process control practice
[19,10]. MPC solves an on-line optimization problem at
each step to compute an optimal control prole over
nite horizon of future time. Typically a sequence of
predicted control moves will be calculated, but only the
rst one is implemented. At the next sampling time, the
optimization problem is solved again with new measurements, and control input is updated.
The key advantage of such a methodology compared
with many other control techniques is that it can handle
input/output constraints directly. Although it is popular
in industrial applications, most MPC techniques lack a
guarantee of stability except for some special cases. The
reason is that MPC is in principle a computational
approach, and an analytic expression for the controller
is generally not available. This limits further study of
closed-loop stability properties which is based on such
information. Standard MPC schemes virtually have no
0959-1524/01/$ - see front matter # 2001 Elsevier Science Ltd. All rights reserved.
PII: S0959-1524(00)00052-4
650
i 1; . . . ; S; j 1; . . . ; Fg;
whereP
the operator
PF norm on i,j is the induced l2 norm
S
and
r
i
i1
j1 mj nw . This general uncertainty
description includes linear time-invariant (LTI) uncertainty, and ane structured uncertainties discussed in
[12] as special cases. Consequently, a scaling set D is
considered
D : diagfD1 ; . . . ; DS ; dS1 Im1 ; . . . ; dSF ImF :
Di 2 Sri ri ; Di > 0; dSj 2 R; dSj > 0; i 1; . . . ; S;
j 1; . . . ; Fg:
X
0
S
1
"
A B2 FT
C0 D02 FT
BT0
A B2 F
C0 D02 F
B0
D00
DT00
>0
1 n
X
xT k ijkQxk ijk
i0
o
u k ijkRuk ijk
651
J1 k
max
u k ijk
4u;max 1; . . . ; p
i50
max
y k ijk
4y;max 1; . . . ; q;
ukijk;i0;1;...
subj: to
i50
is solvable.
Note that both component-wise peak-bounded and
Euclidean norm-bounded input and output constraints
are included, which must be satised over innite time
horizon.
2.2. State-feedback robust MPC synthesis
Fig. 1. Uncertain system y=Tuu.
652
D00 S
3
5
1
xT k
50;
xk
X
u2;max I
YT
Y
50 1; 2; ; p;
X
0
y2;max X
6
6
0
T
6
6C X D Y D Y
4 2;
22;
20;
C0 X D02 Y
D00 T
T
T T 3
XC0 Y D02
7
T DT00
7
750;
5
0
9
wk ijk k izk ijk;
10
wk ijk
C0 D02 FT 1 C0 D02 F
C0 D02 FT 1 D00
DT00 1 D00
11
ijk
D
F
D
;
4
5 4 0
02
00 5
wk ijk
yk ijk
C2 D22 F D20
DT00 1 C0 D02 F
xk ijk
50:
wk ijk
XCT0 YT DT02
7
SDT00
7
7
7
0
7 > 0;
7
0
7
7
5
0
XCT2; YT DT22;
T DT20;
Im
0
T
where the rows of Y, and the rows of C2, D20, D22 are
partitioned conformably to the p-tuple and q-tuple
xT k i 1jkX 1 xk i 1jk
< xT k ijkX 1 xk ijk 8i50:
13
T
1
x k ijkX 1 xk ijk i0 is a strictly decreasing
sequence and bounded below by 0. So limi ! 1 xk
ijk 0 and the uncertain system is robustly stabilized.
Multiply (12) from the left by [xT(k+i|k) wT(k+i|k)]
and its transpose from the right side, and observe (11),
one has
xT k ijkQxk ijk uT k ijkRuk ijk
< Vxk ijk
Vxk i 1jk;
653
vT X 1 v41
4lmax Y X 1 YT ;
which is less than u2;max by (7).
&
xT k 1Xk 1 xk 141;
given xT(k)Xk 1x(k)41 and x(k+1) as one instance of
x(k+1|k). So the optimal solution at step k is also a
feasible solution of (5)(8) for all steps afterwards.
Secondly, because Xk+1,gk+1 is the optimal solution
but Xk, gk only a feasible solution at step (k+1), it holds
1
xk 14
k xT k 1Xk 1 xk 1
k1 xT k 1Xk1
3. Industrial problem
In this section, a nonlinear, continuous-time stirred
tank reactor (CSTR) problem will be considered. The
654
control problem of CSTR has been studied using various control techniques. To exploit its input and output
constraints explicitly, the robust MPC technique is
applied here for constrained controller design. The application to the CSTR model will demonstrate usefulness
of robust MPC to industrial problems, and serve as a
benchmark to evaluate its advantages and drawbacks to
conventional MPC techniques.
3.1. Modeling of CSTR
Stirred tank reactors are commonly encountered in
process industries. The CSTR problem discussed here
represents a rst-order, irreversible, exothermic kinetics
reaction (A!B). The mass and energy balances of the
reaction are described by the following nonlinear dierential equations [22].
dCA Qf
CAf
dt
V
dT Qf
Tf
dt
V
CA
T
k0 CA e
Ea
RI T;
k0 CA
He
Cp
14
Ea
RI T
UAh
T
VCp
Tc :
15
Two state variables of the model are reactant concentration CA and reactor temperature T, and the
manipulated variable is cooling water temperature Tc.
The meanings of all other symbols are listed in Table 1.
To reveal the relevant dynamics, Eqs. (14) and (15) are
normalized using dimensionless variables x1, x2, u and d,
:
x1
x2
x1 Da 1
x2 BDa 1
x1 e
x2
1x2 =
;
x1 e
x2
1x2 =
16
x2 u d;
where
CAf CA
T Tf 0
Tc Tf 0
x2 :
u :
Tf 0
Tf 0
CAf
Tf Tf 0
d :
Tf 0
HCAf
V
UAh
B :
Da : k0 e
:
:
Qf
Cp Tf 0
VCp
Ea
:
RI Tf 0
x1 :
Due to its nonlinear characteristics, previous theoretical interest in control of CSTRs was largely based on
nonlinear geometric control theory, such as feedback
linearization techniques. In practice, PID controls are
widely used and robustness has been achieved through
somewhat ad-hoc controller tuning. The major shortcoming of these techniques is lack of robustness guarantees.
y2
3
x~ 1
6 x~ 2 7
6 7
6 7
6 w1 7
6 7
6w 7
6 27
6 7
4 u~ 5
2
18
d
w1
w2
where
C11
1
0
h
0
2
z1
z2
141 ; 2 41; LTV;
19
i
x20 x2r
Da e1x20 x2r =
=2;
i
x20 x2r
Da e1x20 x2r =
=2;
#
x x
x
x
x20 x2r
Da e1x20 x2r =
h
x20 x2r
R11 Da e1x20 x2r =
"
20 2r
Da 1 x10 1x2020 x2r2r =
1x
x =
e
e 20 2r ;
C12
2x2r
x20 x2r
x20 i
2r
Da 1 x10 h 1xx20 x
e 20 x2r =
e1x20 x2r =
2e1x20 =
:
R12
2x2r
Furthermore, an additional integral action is introduced to eliminate the steady state error in the rst
state, that is
:
x~ 3 x~ 1 :
20
The state x~ 3 is articial and available for state-feedback
use. It will be incorporated into the optimization cost
function to penalize the tracking error of the rst state.
No hard constraint will be put on this articial state.
Another approach for nonlinear system modeling is
through global linearization [3,14]. Consider a nonlinear
system with the equilibrium (0,0,0)
Ah
CAf
Cp
Ea
H
k0
Qf
RI
Tf
Tf0
U
V
:
xt fxt; ut; dt;
21
22
e 20 2r =2;
1 x20 x2r =
2
R12
2r
Da 1 x10 x1r 1xx20 x
e 20 x2r =
2
1 x20 x2r =
Da 1 x10 x1r 1xx20 xx2r =
20 2r
e
=2:
1 x20 x2r =
2
655
Table 2
Parameter settings for the case study
B
Da
Equilibrium (x10, x20,u0)
Operating window
[x10 x1r,x10+x1r][x20 x2r,x20+x2r]
Input constraint umax
Output constraint ymax
1.0
0.3
20.0
1.0
0.072
(0.3,1.96,7.5)
[0,0.6][0,3.92]
10
[0.3,1.96,1]
656
23
The tracking performance from these initial conditions were simulated in Fig. 3. Clearly, states x1,x2
approached their set point quickly (less than 2.5 time
units) while constraints on input and states were satised over the entire horizon. Since the optimization
index is dened over innite horizon and its upper
bound
is always nite during the simulation, this
implies both states converge asymptotically to the equilibrium. On the other hand, the constrained robust
state-feedback controller constructed from initial optimization was implemented, and the convergence rate of
states x1,x2 was found much slower in this case. So the
strategy of on-line optimization is well justied.
For the CSTR problem, the proposed synthesis condition will be the same as Kothare's results because the
block size of uncertainty 1, 2 is 1. However, if we
assume 1=2, then it is possible to reduce the conservatism associated with the synthesis condition in [12]
utilizing full block diagonal scaling matrix. Given initial
condition (0.57, 0.3), the one-step optimization with full
block S produces
=434.15 compared with
=2150.0
resulting from Kothare's synthesis condition. This
clearly demonstrates the advantage of full block diagonal scaling over diagonal matrix.
Secondly, the disturbance rejection ability of the
robust MPC is examined. Although robust MPC was
Fig. 2. Feasibility regions of both uncertain models: 1. Conic sector approach (solid), 2. global linearization approach (dash).
dt
8
<
:
1
0:5 t
1
657
04t < 3
3 34t < 7
t57
Using the same parameters R,Q as (23) rst, the control performance of robust MPC was found quite sluggish. Then Q, R were changed to Q=diag{106, 106,
106}, R=10 2 respectively, reasonable tracking performance was obtained. Simulation results for both cases
were shown in Fig. 4. Clearly, the disturbance was lard
gely cancelled by control input, which is about
.
Both states x1, x2 moved away from their initial values
rst, then went back gradually. The deviation of the
second state from its initial value is relatively large, and
it can be reduced using another integrator.
3.3. Comparison
Despite of their superuous similarity, robust MPC
dierentiates from conventional MPC techniques in
several aspects.
Fig. 3. Set point tracking from various initial conditions; 1. (0.02, 1.1)
(solid), 2. (0.1, 0.4) (dash dot), 3. (0.5, 0.1) (dot), 4. (0.57, 0.3) (dash).
(a) normalized reactant concentration x1; (b) normalized reactor temperature x2; (c) normalized cooling water temperature u.
658
Fig. 4. Disturbance rejection with dierent parameters: 1. Q=diag{1,1,1}, R=1 (solid), 2. Q=diag(106,106,106}, R=10
reactant concentration x1; (b) normalized reactor temperature x2; (c) normalized cooling water temperature u.
4. Concluding remarks
Motivated by the work of [1214], the robust MPC
technique was generalized to an important class of
uncertain linear systems with LFT-type perturbations.
sucient synthesis condition was derived and formulated as LMI optimization, from which the control
action can be calculated. The stability of robust MPC is
guaranteed as long as the optimization problem is solvable at the initial step. Then a constrained CSTR problem was considered using robust MPC design
technique. The simulation results support the applicability of this control technique to industrial problems.
It was further shown that the performance of robust
MPC is closely related to the uncertain model derived
from original nonlinear plant. Finally, its relative merits
to conventional MPC were carefully evaluated. Needless to say, the main advantage of robust MPC technique is its capability to deal with both model mismatch
and constraints as well as stability guarantee. Hopefully
such a discussion will reveal current limitations of
robust MPC and possible directions for its future
development.
References
[1] N. Amann, F. Allgower, m-Suboptimal design of a robustly performing controller for a chemical reactor, Int. J. Control 59 (3)
(1994) 665687.
[2] J.C. Allwight, G.C. Papavasiliou, On linear programming and
robust model predictive control using impulse-responses, Syst.
Contr. Lett. 18 (1992) 11591164.
[3] S. Boyd, L. El Ghaoui, E. Feron, V. Balakrishnan, Linear Matrix
Inequalities in Systems and Control Theory, SIAM, Philadelphia,
PA, 1994.
[4] P. Campo, M. Morari, Robust model predictive control, Proc.
Amer. Contr. Conf., (1987) 10211026.
[5] N. de Oliveira, L. Biegler, Constraint handling and stability
properties of model predictive control, A.I.Ch.E. Journal 40 (7)
(1994) 11381155.
[6] F. Doyle III, A. Packard, M. Morari, Robust controller design
for a nonlinear CSTR, Chem. Eng. Sci. 44 (9) (1989) 19291947.
[7] L. El Ghaoui, J. Folcher, Multiobjective robust control of LTI
systems subject to unstructured perturbations, Syst. Contr. Lett.
28 (1) (1996) 123130.
[8] M. Fan, A. Tits, J. Doyle, Robustness in the presence of mixed
parametric uncertainty and unmodelled dynamics, IEEE Trans.
Automat. Contr. AC-36 (1) (1991) 125138.
[9] P. Gahinet, A. Nemirovskii, A. Laub, M. Chilali, LMI Control
Toolbox, Mathworks, Natick, MA, 1995.
659