Disturbance Models for Offset-Free

Model-Predictive Control
Gabriele Pannocchia
Dept. of Chemical Engineering, University of Pisa, 56126 Pisa, Italy
James B. Rawlings
Dept. of Chemical Engineering, University of Wisconsin, Madison, WI 53706

Model predicti®e control algorithms achie®e offset-free control objecti®es by adding

integrating disturbances to the process model. The purpose of these additional distur-
bances is to lump the plant-model mismatch andror unmodeled disturbances. Its effec-
ti®eness has been pro®en for particular square cases only. For systems with a number of
measured ®ariables (p) greater than the number of manipulated ®ariables (m), it is clear
that any controller can track without offset at most m controlled ®ariables. One may
think that m integrating disturbances are sufficient to guarantee offset-free control in the
m controlled ®ariables. We show this idea is incorrect and present general conditions
that allow zero steady-state offset. In particular, a number of integrating disturbances
equal to the number of measured ®ariables are shown to be sufficient to guarantee zero
offset in the controlled ®ariables. These results apply to square and nonsquare, open-loop
stable, integrating and unstable systems.

Introduction
Model predictive control ŽMPC. arose from the industrial stable and unstable systems. As shown in Lee et al. Ž1994.,
applications called Identification and Command ŽIDCOM. step or impulse response models are particular cases of the
ŽRichalet et al., 1978., and Dynamic Matrix Control ŽDMC. state-space models.
ŽCutler and Ramaker, 1979.. These control algorithms used In these formulations, offset-free objectives are obtained
finite-impulse or step-response models to predict the future by augmenting the system state with integrating disturbances.
process behavior. In order to obtain offset-free control, the In the original formulations of MPC no analysis was given
model is updated with feedback information. Comparing the regarding steady-state offset. In Rawlings et al. Ž1994. it was
current measured process output and the current predicted shown that a constant output disturbance guarantees offset-
output, a constant bias term is added to the future model free performance for square systems without integrating
forecasts. These convolution models cannot be used with modes. The constant output disturbance model cannot be
open-loop integrating or unstable systems. Generalized pre- used in the presence of integrating modes, because the ef-
dictive control ŽGPC. ŽClarke et al., 1987a,b., instead, used fects of the plant integrating mode and of the additional dis-
autoregressive models. In order to cover unmodeled distur- turbance cannot be distinguished. For such cases, double in-
bances andror plant-model mismatch, a disturbance model is tegrated disturbances have been proposed ŽLundstrom ¨ et al.,
designed by choosing the so-called ‘‘T polynomial’’ Žsee, e.g., 1995.. Alternatively, the system can be augmented with inte-
Clarke and Mohtadi, 1989., which allows the designer to grating modes that affect the system states, as shown in Ep-
specify dynamics other than a constant step as in IDCOM or pler Ž1997.. When the number of measured variables is dif-
DMC. State-space formulations of MPC were proposed later ferent from the number of manipulated variables, a natural
ŽMarquis and Broustail, 1988; Kwon and Byun, 1989; Lee et question arises concerning the number of additional integrat-
al., 1992; Rawlings and Muske, 1993., which handle open-loop ing disturbances required to obtain offset-free control. The
purpose of this work is to understand steady-state offset with
MPC algorithms and provide general design criteria to obtain
Correspondence concerning this article should be addressed to J. B. Rawlings. offset-free performance.

426 February 2003 Vol. 49, No. 2 AIChE Journal

 An alternative or complement to this ‘‘design approach’’
for offset-free performance would be a ‘‘disturbance identifi-
cation’’ approach, in which one uses plant identification test
data to identify which disturbances are nonstationary Žin-
tegrating. and design internal models for just those distur-
bances. However, even after such a nonstationary distur-
bance identification procedure, one would still like to design
the additional disturbances that guarantee offset-free perfor-
mance in case the process disturbances change or the
plant᎐model mismatch changes over time. So the questions
addressed in this article are relevant even under a distur-
bance-identification approach to controller design.
In the following section, we present an example of a sys-
tem with three measured variables, two of which are con-
trolled by using two manipulated variables. One might think
that a reasonable way to remove steady-state offset is to add
one integrator to each controlled variable. We show that this
choice leads to steady-state offset. Motivated by this and other
similar examples, we derive conditions that guarantee zero
steady-state offset of the controlled variables when the num-
ber of measurements is greater than the number of inputs.
Arbitrary linear model dynamics are covered in this work and
these results also apply to square systems and nonsquare sys- Figure 1. CSTR.
tems with a number of measurements less than the number
of manipulated variables. The rest of the article is organized
as follows. In the third section, the MPC formulation with The controlled variables are the level of the tank, h, and the
infinite horizon is reviewed. In the fourth section, conditions molar concentration of A, c. The third state variable is the
on the augmented system are derived in order to obtain a reactor temperature, T, while the manipulated variables are
stable estimator first, and then offset-free control. Next, in the outlet flow rate, F, and the coolant liquid temperature,
the fifth section, the motivating example is revisited and sev- Tc . Moreover, it is assumed that the inlet flow rate acts as an
eral controllers are designed according to the results of the unmeasured disturbance. The model parameters in nominal
previous section. A second example of an ill-conditioned dis- conditions are reported in Table 1. The open-loop stable
tillation column is also presented to show that the demand of steady-state operating conditions are the following
offset-free performance in as many outputs as possible may
lead to closed-loop instability in the presence of errors. Fi- h s s 0.659 m, c s s 0.877 molrL, T s s 324.5 K,
nally, in the sixth section, the main accomplishments of this
work are summarized. Proofs of the results of the fourth sec- F s s100 Lrmin, Tcs s 300 K.
tion are reported in the Appendix.
Using a sampling time of 1 min, a linearized discrete
Motivating Example state-space model is obtained and, assuming that all the states
are measured, the state-space variables are
Plant model
A continuous stirred-tank reactor as shown in Figure 1 is
considered. An irreversible, first-order reaction, A ™ B, oc- cy c s Tc yTcs cy c s
curs in the liquid phase, and the reactor temperature is regu- x s T yT s , us , y s T yT s ,
hy h s FyFs hy h s
lated with external cooling. This example is taken from Hen-
son and Seborg Ž1997. with the assumption that the level is ps F0 y F0s .
not constant. Mass and energy balances lead to the following
nonlinear state-space model

dh F0 y F Table 1. Parameters of the CSTR

s Ž 1a .
dt ␲r 2 Parameter Nominal Value
dc F0 Ž c 0 y c . E F0 100 Lrmin

dt
s
␲ r 2h ž
y k 0 c exp y
RT / Ž 1b . T0
c0
r
350 K
1 molrL
0.219 m
dT F0 Ž T0 yT . y⌬ H E k0 7.2=10 10 miny1
dt
s
␲r h 2
q
␳ Cp ž
k 0 c exp y
RT / ErR
U
8,750 K
915.6 Wrm2 ⴢ K
2Uh ␳ 1 kgrL
q Ž Tc yT . . Ž 1c . Cp 0.239 Jrg ⴢ K
r␳ Cp ⌬H y5=10 4 Jrmol

AIChE Journal February 2003 Vol. 49, No. 2 427

The corresponding linear model is in which d k g⺢ 2 is the integrating disturbance, w k g⺢ 3 is
the state noise, ␰ k g⺢ 2 is the disturbance noise, and ®k g⺢ 3
x kq1 s Ax k q Bu k q Bp p Ž 2a . is the output noise. A steady-state Kalman filter is used to
estimate the state, x k , and the disturbance, d k , given the plant
y k sCx k , Ž 2b . measurements, y k . The estimator is tuned assuming zero
in which noise for the state ŽEq. 5a. and the measurement equation
ŽEq. 5c., which is the standard DMC-like tuning. Thus, the
filtering equations are
0.2511 y3.368=10y3 y7.056=10y4
As 11.06 0.3296 y2.545 ,
0 0 1 0 1.734=10y2 0
ˆx k N k ˆx k N ky1 0 0.9934 0
y5.426=10y3 1.530=10y5 s q 0 y8.770=10y2 0
Bs 1.297 0.1218 ˆ
dk N k dˆk N ky1
1 y1.734=10y2 0
0 y6.592=10y2 0 8.770=10y2 1
y1.762=10y5 1 0 0 1 0
Bp s 7.784=10y2 ,
6.592=10y2
Cs 0
0
1
0
0 .
1 ž
= y k yCxˆk N ky1 y 0
0
0 dˆk N ky1 .
1 / Ž6.

In the following, both the linear and the nonlinear models Using the model ŽEq. 5. to replace the controlled variables
are used to generate the plant response in the presence of with the system state, the solution of the unconstrained opti-
disturbances. mization problem ŽEq. 3. is

Estimator and regulator uk s K Ž ˆ

x k N k y x t . q ut , Ž7.
The controller is chosen as the solution of the following
unconstrained infinite-horizon quadratic optimization prob- in which K g⺢ 2=3 is the optimal gain matrix computed from
lem the discrete algebraic Riccati equation, and the state and in-
put targets are solutions of the following system:
⬁
min ⌽ s Ý y kT Qyk q uTk Ru k , Ž3. x t s Ax t q Bu t Ž 8a .
u 0 ,u 1 , . . . ks0
0 1 0 0 1 0 ˆ
in which s x q d . Ž 8b .
0 0 0 1 t 0 1 kN k

1 0 0
0.1 0 It is clear from Eqs. 8 that x t and u t are the targets that
Qs 0 0 0 , Rs .
0 0.1 drive the controlled variables to their setpoint Žthat is, zero..
0 0 1

It is clear from the choice of the tuning matrix, Q, that only Disturbance rejection
the composition and the level are required to be at the set- We assume that at time t s10 min, a disturbance ps10
point Žfirst and third outputs.. enters the plant, which is an increment of 10 Lrmin on the
Since it is assumed that the disturbance, p, is not mea- inlet flow rate. The results of the simulation, reported in Fig-
sured, the model used by the regulator is ure 2, show offset in the controlled variables. In this work we
clarify why offset occurs in cases like this, and we present a
x kq1 s Ax k q Bu k general methodology for designing offset-free model predic-
y k sCx k . Ž4. tive controllers.

However, it is well known that a disturbance model is re- MPC Algorithm

quired to reject unmeasured nonzero disturbances and ob- Preliminary definitions and notations
tain offset-free control in the controlled variables. A common
choice is the so-called output disturbance model, in which an In this work, we consider linear, time-invariant discrete
integrated disturbance is added to each controlled variable. systems
The state-space formulation of this model is
x kq1 s Ax k q Bu k Ž 9a .
x kq1 s Ax k q Bu k q w k Ž 5a . y k sCx k , Ž 9b .
d kq1 s d k q ␰ k Ž 5b .
in which x k g ⺢ n, u k g ⺢ m, y k g⺢ p, Ag⺢ n=n , Bg⺢ n=m ,
1 0 C g⺢ p=n , with the assumption that Ž A, B . is stabilizable and
y k sCx k q 0 0 d k q ®k , Ž 5c . Ž C, A. detectable. If p) m, that is, the number of measure-
0 1 ments is greater than the number of manipulated variables,

428 February 2003 Vol. 49, No. 2 AIChE Journal

 Disturbance model and estimator
In order to achieve offset-free performance we augment
the system state with an additional integrating disturbance
vector. The roots of this idea are in the works of Davison and
coworkers ŽDavison and Smith, 1971; Davison and Smith,
1974; Qiu and Davison, 1993. and in the internal model prin-
ciple ŽFrancis and Wonham, 1976.. In order to remove offset,
one designs a control system that can remove asymptotically
constant, nonzero disturbances ŽDavison and Smith, 1971;
Kwakernaak and Sivan, 1972, p. 278.. To accomplish this end,
the original system is augmented with a replicate of the con-
stant, nonzero disturbance model. Thus the states of the orig-
inal system are moved onto the manifold that cancels the ef-
fect of the disturbance in the controlled variables.
The state and the additional integrating disturbance are
estimated from the plant measurement by using a steady-state
Kalman filter designed for the following augmented system

x kq1 A Bd xk B
s q u q wk Ž 12a .
d kq1 0 I dk 0 k

xk
yk s w C Cd x q ®k , Ž 12b .
dk

in which d k g⺢ n d , Bd g⺢ n=n d , C d g⺢ p=n d . The vectors w k

Figure 2. Linearized CSTR. g⺢ nqn d and ®k g R p are zero-mean white-nose disturbances
Disturbance on the inlet flow rate that causes offset. for the augmented state equation and for the output equa-
tion, respectively. Thus, the state and the disturbance are es-
timated as follows
one cannot attempt to control without offset all the mea-
sured variables Žsee, for example, Corollary 4 of Davison and
ˆx k N k ˆx k N ky1 Lx
Smith, 1971.. However, one can choose, as controlled vari- dˆk N k
s
dˆk N ky1
q
Ld
Ž yk yCxˆk N ky1 yCd dˆk N ky1 . ,
ables, m linear combinations of the measured outputs. If pF
m instead, all the measured variables can be controlled with- Ž 13a .
out offset. We treat both cases within the same framework by
defining the controlled variable as and the prediction of the future augmented state is obtained
by
z s Hy, Ž 10.
ˆx kq1 N k A Bd ˆx k N k B
s q u , Ž 13b .
in which z g⺢ n c , H g⺢ n c=p . We assume the following re- dˆkq1 N k 0 I dˆk N k 0 k
striction
in which L x g⺢ n=p and L d g⺢ n d=p are the filter gain ma-
Iy A yB trices for the state and the disturbance, respectively. Equa-
rank s nq n c . Ž 11. tions 13a and 13b can be condensed into one step, known as
HC 0
Kalman predictor
See Corollary 3 of Davison and Smith Ž1971. for a continu-
ous-time condition equivalent to this one. As we discuss in ˆx kq1 N k A Bd ˆx k N ky1 B
s q u
the target calculation, this restriction on the original system dˆkq1 N k 0 I dˆk N ky1 0 k
implies we are able to compensate for the steady-state distur-
bances in the controlled variables of interest. This condition L1
implies that the number of controlled variables Ž n c . cannot q
L2
Ž yk yCxˆk N ky1 yCd dˆk N ky1 . , Ž 14.
exceed either the number of manipulated variables Ž m. or
the number of measurements Ž p .. It also implies that H has in which L1 g⺢ n=p and L2 g⺢ n d=p are the predictor gain
to be full row rank, that is, the controlled variables must be matrices for the state and the disturbance, respectively.
independent of each other. Clearly when m- p, H defines, Straightforward algebraic calculations show
as controlled variable, a linear combination of the measure-
ments, while if pF m, H can be chosen as the identity matrix L1 Lx
A Bd
and all the measured variables can be controlled without off- s . Ž 15.
L2 0 I Ld
set.

AIChE Journal February 2003 Vol. 49, No. 2 429

For convenience, we introduce the following notation gram that minimizes the offset of the controlled variables in
a least-squares sense ŽMuske and Rawlings, 1993..
A Bd L1 Next, the input sequence is computed as the solution of
˜
As , C˜s w C C d x , ˜
Ls . the following infinite-horizon optimization problem
0 I L2
⬁
T
Thus, the predictor gain matrix is computed from the follow- min ⌽ s Ý Ž z k y ˜z . Q Ž z k y z .
u 0 ,u 1 , . . . ks0
ing
T
y1 qŽ uk y ut . R Ž uk y ut . , Ž 19a .
˜ AÝC
Ls ˜ ˜T CÝC
˜ ˜T q R ®
Ž . , Ž 16.
subject to the model ŽEq. 12. and to input and output con-
in which R ® g⺢ p=p is a symmetric positive definite matrix straints
associated with the output noise, ®k , and the matrix Ýg
⺢ Ž nqn d .=Ž nqn d . is the unigue symmetric positive semidefinite Eu k F e k s 0, 1, . . . Ž 19b .
solution of the discrete algebraic Riccati equation Fy k F f k s 0, 1, . . . . Ž 19c .
y1
˜ A˜T q Q˜w y AÝC
Ýs AÝ ˜ ˜T CÝC
˜ ˜T q R ®
Ž . ˜ A˜T . Ž 17.
CÝ The penalty matrices Rg⺢ m=m and Qg⺢ n c=n c are as-
sumed to be symmetric and positive definite.
The matrix Q ˜w g⺢ Ž nqn d .=Ž nqn d . is the symmetric positive Using the targets computed from Eq. 18, we introduce the
semidefinite associated with the augmented state noise w k . following deviation variables
It is important to notice that the additional disturbances,
d, are not controllable by the inputs u. However, since they wj s ˆ
x kqj N k y x t , ®j s u kqj y u t . Ž 20.
are observable, we use their estimates to remove their influ-
ence from the controlled variables, as discussed in the next Thus the regulator optimization problem ŽEq. 19. becomes
section.
⬁
min ⌽ s Ý wjTQwj q ®jTR®j Ž 21a .
Target calculation and regulator ® 0 , ®1 , . . . js 0

Given the current estimate of the disturbance, dˆk N k , the

state and input target are computed by solving the following subject to
quadratic program ŽMuske and Rawlings, 1993.
wjq1 s Awj q B®j k s 0, 1, . . . Ž 21b .
T
min Ž u t y u . R˜Ž u t y u . , Ž 18a . E®j F ey Eu t k s 0, 1, . . . Ž 21c .
xt , ut

FCwj F f y F Ž Cx t qC d dˆk N k . k s 0, 1, . . . , Ž 21d .

subject to
in which Q s C T H T QHC g ⺢ n=n is symmetric positive
Iy A yB xt Bd dˆk N k semidefinite. We assume that
s Ž 18b .
HC 0 ut y HC d dˆk N k q z
Ž Q1r2 , A . Ž 22.
Eu t F e Ž 18b .
is detectable. It can be shown that this condition is satisfied if
F Ž Cx t qC d dˆk N k . F f , Ž 18d . and only if Ž HC,A. is detectable. This condition forces the
unstable modes of A to be ‘‘seen’’ in the regulator objective
in which x t g⺢ n, u t g⺢ m, and Rg⺢ ˜ m=m
are symmetric function.
positive definite, and z g⺢ and ug⺢ m are the set points
nc
The solution of the infinite-horizon constrained problem
for the controlled and manipulated variables, respectively. ŽEq. 21. is discussed in Chmielewski and Manousiouthakis
Notice that, often, the input set point is not specified and it Ž1996. and Scokaert and Rawlings Ž1998., in which Eq. 21 is
can be assumed zero in order to use, if possible, the optimal reparameterized in terms of a finite number of decision vari-
input in a least-squares sense. E g ⺢ ␯=m and eg⺢ ␯ de- ables and constraints by appending the optimal LQR uncon-
scribe linear input constraints, while F g⺢ ␰=p and f g⺢ ␰ strained control law, ®j s Kwj , at the end of a finite horizon,
describe linear output constraints, in which ␯ and ␰ are the N.
number of input and output constraints, respectively. It is as-
sumed that the feasible region of Eq. 18 is nonempty. Notice Main Results
that Eq. 11 implies that the unconstrained target calculation
Žthat is, Eq. 18 subject to Eq. 18b only. has a solution for all Restrictions on the augmented system
dˆk N k and z. If the feasible region of Eq. 18 is empty, input The disturbance model is defined by choosing the matrices
and output constraints are too stringent, and the n c-con- Bd and C d. More complex disturbance models requiring
trolled variables cannot be tracked to their setpoint without identification from data can be considered, but Ž Bd ,C d . allow
offset. In this case, one can solve a different quadratic pro- enough flexibility to meet most process control objectives

430 February 2003 Vol. 49, No. 2 AIChE Journal

without requiring disturbance model identification proce- in which K g⺢ m=n is the optimal controller gain computed
dures. Since the additional modes introduced by the distur- from the Riccati equation. If
bance are unstable, it is necessary to check the detectability
of the augmented system. Detectability of the augmented sys- null Ž L d . : null Ž K e . , Ž 25.
tem is a necessary and sufficient condition for a stable esti-
˜ LC
mator to exist, that is, such that Ay ˜ ˜ is stable wsee, e.g., there is zero offset in the controlled variable, that is,
Sontag Ž1998.x. We have the following results.
Lemma 1 (Detectability of the Augmented System). The Hys s z. Ž 26.
augmented system ŽEq. 12. is detectable if and only if the
nonaugmented system ŽEq. 9. is detectable, and the following Proof. See the Appendix for proof of Lemma 2.
condition holds When null Ž L d . ­ null Ž K e ., it is easy to construct a plant
such that the closed-loop system is stable and there is
Iy A y Bd steady-state offset. Therefore, the condition stated in the
rank s nq n d . Ž 23. previous lemma is necessary to guarantee zero offset in the
C Cd
presence of arbitrary plant᎐model mismatch. If L d is square
and full rank, the assumption, nullŽ L d . : nullŽ K e ., is satis-
Corollary 1 (Dimension of the Disturbance). The maximum fied independently of K e . When the number of integrating
dimension of the disturbance d in Eq. 12 such that the aug- disturbances is less than the number of measurements Ži.e.,
mented system is detectable is equal to the number of mea- n d - p ., L d is not square, and Eq. 25 must be checked to
surements, that is, guarantee zero offset in the presence of plant᎐model mis-
match. A second problem related to choosing n d - p is that,
n d F p. Ž 24. in general, the structure of K e also depends on the regulator
tuning, and therefore Eq. 25 may not hold if the tuning pa-
A pair of matrices Ž Bd , C d . such that Eq. 23 is satisfied rameter of the controller change.
always exists. In fact, since Ž C, A. is detectable, the submatrix There are particular cases in which offset-free perfor-
Iy A mance can be guaranteed by adding a number of distur-
g⺢ Ž pqn.= n has rank n. Thus, we can choose any
C bances equal to the number of controlled variables Ž n d s n c
Iy A - p .. A class of examples can be described as follows. Let
n d F p columns in ⺢ pqn independent of and build
C the system state be partitioned as x s w x 1T x T2 xT, the output
Bd and C d accordingly. Clearly, this methodology is not in- as y s w y 1T y 2T xT, and let the controlled variable be z s y 1.
tended for design, but just to show that a detectable aug- Let the model and disturbance model matrices be as follows
mented system can always be constructed. Design methods Žeach term is a matrix of appropriate dimensions.
for disturbance modeling are currently under investigation
ŽPannocchia, 2002..
A11 0 C11 0
The continuous-time versions of Lemma 1 and Corollary 1 As , Cs ,
were first stated and proved by Smith and Davison Ž1972, p. A 21 A 22 C21 C22
1214.. Morari and Stephanopoulos Ž1980. provide a nice,
0 I
compact proof of the continuous-time version using the Hau- Bd s , Cd s ,
0 0
tus lemma. Similar conditions have been found useful in other
contexts as well ŽKurtz and Henson, 1998..
in which A11 and A 22 are strictly stable square matrices. The
Q 0
Offset-free disturbance models regular for tuning matrix is Qs 1 and stability of A11
0 0
1r2
The results of the previous section provide a tool for and A 22 implies that Ž Q , A. is detectable. Given any posi-
checking whether or not the augmented system is detectable tive definite matrix, R, the optimal controller gain matrix has
and therefore whether a stable estimator exists. Once the dis- the following structure
turbance model matrices Ž Bd , C d . are chosen such that Eq.
23 holds, the augmented system is detectable. In the motivat- K s w= 0x ,
ing example, it is easy to verify that this condition holds. Nev-
ertheless, the closed-loop response shows offset in the con- which implies that Aq BK is an upper-block triangular ma-
trolled variable. The next results provide sufficient conditions trix. If there is zero noise related to x 1 in the state equation
for zero offset in the controlled variables. and A11 is stable, it can be shown that the filter structure is
Lemma 2 (Zero Offset). Consider a system controlled by
the MPC algorithm, as described in the third section. The 0 0
Lx s , Ld s w I 0x .
target problem ŽEq. 18. is assumed feasible. Assume that the 0 =
closed-loop system reaches a steady state with output ys and
input u s , and suppose that input and output constraints are Given the structure of H, C, Aq BK, and L x , straightfor-
not active at this steady state. Let K e g⺢ n c=p be the follow- ward block matrix calculations show that K e s w I 0x s L d ,
ing and therefore Eq. 25 holds.
However, we are mainly interested in conditions for zero
y1 offset that hold for generic linear systems and disturbance
K e s H C Ž I y Ay BK . Ž Aq BK . L x q I , models. We have the following results.

AIChE Journal February 2003 Vol. 49, No. 2 431

 Lemma 3 (Full Rank of L d ). Given disturbance model in which pk g⺢ n p , Bp g⺢ n=n p , C p g⺢ p=n p , A p g⺢ n p=n p ,
matrices Bd g⺢ n=n d , C d g⺢ p=n d such that the assumptions Bd g⺢ n p=n d . Assuming that the disturbance dynamic matrix,
of Lemma 1 are satisfied, and A p , is strictly stable, all the results developed can be ex-
tended with minor changes. In particular, in Lemma 1 the
n d s p, Ž 27. condition to check in place of Eq. 23 becomes

the disturbance filter gain matrix in Eq. 13a, satisfies Iy A y Bp 0

rank 0 Iy Ap y Bd s nq n p q n d . Ž 32.
L d g⺢ p=p Ž 28.
C Cp 0
rank L d s p. Ž 29.
Proof. See the Appendix for proof of Lemma 3. As a corollary, we again obtain that the maximum number of
Theorem 1 (Main Result). Consider a system controlled by integrated disturbances that can be added without losing de-
the MPC algorithm as described in the third section. The tectability is equal to the number of measurements, that is,
target problem ŽEq. 18. is assumed feasible. Augment the n d F p. If we choose n d s p and the closed-loop system is
system model with a number of integrating disturbances equal stable, there is zero steady-state offset in the controlled vari-
to the number of measurements Ž n d s p ., choose any Bd g able, that is, Hys s z.
⺢ n=p , C d g⺢ p=p , such that The results shown here can be extended to finite Žcontrol
andror prediction . -horizon MPC, provided that the corre-
Iy A y Bd sponding unconstrained control law is nominally stabilizing,
rank s nq p. that is, it corresponds to a linear feedback gain matrix, K,
C Cd
that makes the matrix Aq BK strictly stable. This is typically
accomplished in finite-horizon MPC by choosing the output
If the closed-loop system is stable and constraints are not
and input penalty matrices and the horizon length.
active at steady state, there is zero offset in the controlled
variables, that is,
Case Studies
Hys s z, Ž 30. Re©isiting the moti©ating example
in which ys is the plant output. By applying Lemma 4.1 to the motivating example, we can
Proof. See the Appendix for proof of Theorem 1. verify that the augmented system in Eq. 5 is detectable, and
therefore the Kalman filter used in Eq. 6 results in a stable
estimator. However, the filter gain matrix corresponding to
Remarks the integrated disturbance is
It is important to emphasize that, in proving the previous
results, no assumptions have been made on the plant dynam- 1 y1.734=10y2 0
Ld s ,
ics. That is, these results apply even when the plant does not 0 8.770=10y2 1
satisfy the model ŽEq. 9.. In other words, if the system reaches
a steady state without active constraints, there is no offset in which has rank equal to two and, therefore, has a null space.
the controlled variables regardless of the plant dynamics. If The term e k s y k yCx k yC d d k does not go to zero, while
constraints were active at steady state, for example, one or the product L d e k does. We can verify that nullŽ L d .­
more inputs saturate at steady state, in general it is impossi- nullŽ K e . and that e s s lim k ™ ⬁ e k is not in the null space of
ble to achieve offset-free control. This inability, which be- K e . Therefore, there is a steady-state offset in the controlled
longs to any kind of control strategy and not specifically to variables.
MPC, is due to the fact that input constraints are too strin- The results of the previous section provide conditions for
gent for the controlled variables to reach their setpoint. How- obtaining zero offset in the controlled variables, c and h.
ever, the results presented do not depend on the possibility Lemma 1 states a detectability condition on the augmented
of transient constraint saturation, as shown in the fifth sec- system, given the disturbance model matrices, Bd and C d ,
tion. while Corollary 1 shows that the maximum number of inte-
In this work the disturbance model chosen ŽEq. 12. is an grated disturbances that can be added is three. Finally, Theo-
integrated disturbance that enters through the model dy- rem 1 states that if three integrated disturbances are added
namic matrix, A. If desired, one can choose different dynam- in such a way that the augmented system is detectable, there
ics for the disturbance, as in the following augmented system: is a guarantee of offset-free control.
In the motivating example we added two integrated distur-
x kq1 A Bp 0 xk B bances to two of the three measured variables. One can try to
pkq1 s 0 Ap Bd pk q 0 u k Ž 31a . add three disturbances to all three measured variables, but in
d kq1 dk 0 this case the augmented system is not detectable due to the
0 0 I
presence of the integrating mode in the process model Žtank
xk level.. Other possibilities are adding two integrated distur-
yk s C C p 0 pk , Ž 31b . bances to the manipulated variables and one to a measured
variables, and vice versa. In six cases the augmented system is
dk
detectable; in other cases it is not, as shown in Table 2. It is

432 February 2003 Vol. 49, No. 2 AIChE Journal

 Table 2. Detectability of Disturbance Models
ID Dist. on Inputs Dist. on Outputs Detectable
0 ᎏ c, T, h No
1 Tc , F c Yes
2 Tc , F T Yes
3 Tc , F h Yes
4 Tc Any No
5 F c, T Yes
6 F c, h Yes
7 F T, h Yes

interesting to note that, if the integrated disturbance is added

to the manipulated variable Tc , but not to F, the augmented
system is not detectable. This property comes from the fact
that Tc does not influence the integrating mode associated
with h.
The results of simulation, obtained with all detectable dis-
turbance models reported in Table 2, are shown in Figures 3
and 4. For each augmented system, a steady-state Kalman
filter has been designed assuming zero noise on the state
equation and arbitrarily small noise on the measurement
equation. For each disturbance model, an unconstrained con-
troller has been designed with the same tuning specified in
the second section. As expected, all controllers achieve off-
set-free control of the composition and the level. However,
the robust performance of these offset-free controllers is quite
Figure 4. Linearized CSTR.
different. For example, disturbance models 噛1, 噛2, and 噛5
Rejection of the disturbance on the inlet flow rate: inputs.
show better disturbance rejection and require smaller varia-
tions of the inputs than the other ones. This behavior can be
associated with the fact that, in the other disturbance mod-
els, an integrator is added to the third measurements, that is,
the level, which already has one integrating mode. These sim-
ulations have been repeated using the nonlinear model ŽEq.
1. as the plant, and the results are shown in Figures 5 and 6.
In these nonlinear plant simulations we included the follow-
ing input constraints,

299 K FTc F 301 K, 85 Lrmin F F F115 Lrmin, Ž 33 .

in order to show that transient constraint saturations do not

eliminate the offset-free properties of the proposed design.
As stated in Theorem 1, these disturbance models lead to
zero steady-state offset even if constraints Žnot active at steady
state. are present andror the plant is nonlinear. However,
the difference in robustness of these disturbance models is
emphasized by the nonlinear behavior of the plant.

Offset-free control: The other side of the coin

In this section we emphasize some risks of offset-free con-
trol. We consider the following transfer-function model of an
ill-conditioned distillation column ŽMorari and Zafiriou, 1989.

1 0.878 y0.864
GŽ s. s . Ž 34.
75sq1 1.082 y1.096

We assume that the plant is given by the following transfer

function

1 0.878 y0.880
Figure 3. Linearized CSTR. Gp Ž s . s . Ž 35.
Rejection of the disturbance on the inlet flow rate: outputs. 75sq1 1.100 y1.096

AIChE Journal February 2003 Vol. 49, No. 2 433

Figure 5. Nonlinear CSTR constrained. Figure 6. Nonlinear CSTR constrained.
Rejection of the disturbance on the inlet flow rate: outputs. Rejection of the disturbance on the inlet flow rate: inputs.

䢇 MPC 2. In this controller, two integrating disturbances

The determinants of the gain matrices of G Ž s . and GpŽ s .
have opposite signs, and it is expected that any feedback con- are added to the process outputs:
troller with integral action designed for G Ž s . would lead to
instability if the plant is GpŽ s . wsee, for example, Skogestad 0 0 1 0
Bd s , Cd s ,
and Postlethwaite Ž1996.x. 0 0 0 1
Assuming a sampling time of 5 min, a discrete state-space
minimal realization of G Ž s . has two states, two inputs, and and a steady-state Kalman filter is designed for the aug-
two outputs. The state-space matrices are mented system, assuming nonzero noise for the disturbance
equation, zero noise for the state equation, and arbitrarily
small noise for the output equation. The controller tuning
␾ 0 0.878 y0.864 matrices are the same as in MPC 1.
As , Bs Ž 1y ␾ . ,
0 ␾ 1.082 y1.096 䢇 MPC 3. In this controller, two integrating disturbances

are still added to the process outputs and the same Kalman
1 0 filter is used. However, we require that only the first output
Cs ,
0 1 reaches its setpoint by choosing the following controller tun-
ing matrices,

in which ␾ sexp Žy5r75.. 1 0 1 0

Three controllers are compared: Qs , Rs ,
0 0 0 1
䢇 MPC 1. In this controller, no integrating disturbances are

added and a steady-state Kalman filter is designed for the or, in other words, the controlled variable is z s w1 0x y.
nonaugmented system, assuming zero noise of the state equa- Simulations of a setpoint change from the origin to y s
tion and arbitrarily small noise for the output equation. The w1 0xT using the three controllers are reported in Figures 7
controller tuning matrices as in Eq. 21 are and 8. As expected, MPC 2 leads to an unstable closed-loop
system, because it is designed to remove offset in both out-
1 0 1 0 puts. MPC 1 is not designed for zero steady-state offset in
Qs , Rs , the presence of modeling error and leads to a stable closed-
0 1 0 1
loop system, but with offset in the controlled variables. MPC
3 is designed for removing offset in the first output and leads
which means that both outputs are required to reach set- to a stable closed-loop system. From Theorem 1, we know
point. that if the closed-loop system is stable, there is zero offset in

434 February 2003 Vol. 49, No. 2 AIChE Journal

 Figure 7. Ill-conditioned column outputs. Figure 8. Ill-conditioned column inputs.
Setpoint change in the presence of gain errors. Setpoint change in the presence of gain errors.

sen disturbance model. Next, it was proven that the designed

the controlled variable. It is interesting to note that MPC 3 controllers guarantee zero steady-state offset. These results
uses two integrating disturbances as in MPC 2, which was apply to linear MPC for square and nonsquare, open-loop
unstable. Therefore, closed-loop stability is not related to the stable, integrating, and unstable systems. In the example, sev-
number of integrators added, but to the fact that both out- eral offset-free disturbance models that add integrating dis-
puts were requested to be at setpoint. turbances to input andror output variables were compared. It
For ill-conditioned processes, the demand of offset-free was shown that all the admissible disturbance models, that is,
performance in all outputs can lead to closed-loop instability those with detectable augmented systems, guarantee offset-
if the model identification procedure is not sufficiently accu- free control. The result holds if the plant follows the nonlin-
rate. In such cases, one should relax the offset-free perfor- ear model as well as the linearized model used by the con-
mance in the least important outputs to maintain closed-loop troller.
stability if the sign of the determinant of the gain matrices These disturbance models show quite different behavior in
may be different between the model and the plant. terms of input and output transient variations. The natural
and interesting question concerning the robustness proper-
ties of offset-free disturbance models is the subject of current
Conclusions research ŽPannocchia, 2002..
In this work the problem of designing offset-free model A second example of an ill-conditioned distillation column
predictive controllers was addressed. MPC algorithms achieve was also presented to show some possible risks of demanding
offset-free performance by adding integrating disturbances to offset-free performance in as many variables as possible. The
the process model Žoften to the controlled variables.. How- case of incorrect sign of the determinant of the gain matrix
ever, the effectiveness of this procedure was proven for par- was considered, and it was shown that a request for zero off-
ticular square cases only ŽRawlings et al., 1994.. An example set in all the output variables leads to closed-loop instability.
was presented, in which a common and apparently reason- In such cases, one can relax the offset-free performance in
able choice of disturbance model leads to steady-state offset the least important variables to maintain closed-loop stabil-
without violating detectability restrictions. ity.
Motivated by this example, general conditions have been After this article was submitted, we became aware of a
derived, which allow design of a proven offset-free MPC al- manuscript submitted to another journal by Muske and
gorithm. In particular, a number of integrated disturbances Badgwell ŽMB. that treats a similar problem ŽMuske and
equal to the number of measurements is shown to be suffi- Badgwell, 2002.. We summarize the differences between the
cient to guarantee offset-free performance. A simple rank test two articles here. Although both articles use integrating dis-
is necessary and sufficient to check the suitability of the cho- turbances, the disturbance models are different. MB use a

AIChE Journal February 2003 Vol. 49, No. 2 435

block-diagonal structure for their disturbance model; we use for Nonlinear Systems Affine in the Unmeasured Variables,’’ Com-
an unstructured disturbance model. Each model choice has put. Chem. Eng., 22, 1441 Ž1998..
Kwakernaak, H., and R. Sivan, Linear Optimal Control Systems, Wi-
advantages and disadvantages, and the two are not equiva- ley, New York Ž1972..
lent. Kwon, W. H., and D. G. Byun, ‘‘Receding Horizon Tracking Control
We define a controlled variable, z, which may be different as a Predictive Control and its Stability Properties,’’ Int. J. Control,
than the measurement, y, and determine conditions for which 50, 1807 Ž1989..
Lee, J. H., M. S. Gelormino, and M. Morari, ‘‘Model Predictive Con-
z can be controlled without offset. MB use y as the con-
trol of Multi-Rate Sampled-Data Systems: A State-Space Ap-
trolled variable as well as the measurement, and determine proach,’’ Int. J. Control, 55, 153 Ž1992..
conditions for which y can be controlled without offset. This Lee, J. H., M. Morari, and C. E. Garcia, ‘‘State-Space Interpretation
difference changes the focus of the two articles. MB explore of Model Predictive Control,’’ Automatica, 30, 707 Ž1994..
the issue of how current industrial MPC algorithms can be Lundstrom, P., J. H. Lee, M. Morari, and S. Skogestad, ‘‘Limitations
of Dynamic Matrix Control,’’ Comput. Chem. Eng., 19, 409 Ž1995..
interpreted in their article’s disturbance model framework. Marquis, P., and J. P. Broustail, ‘‘SMOC, A Bridge Between State
They can therefore critique the current industrial practice of Space and Model Predictive Controllers: Application to the Au-
using rotation factors for integrating models, and slow rejec- tomation of a Hydrotreating Unit,’’ Proceedings of the 1988 IFAC
tion of disturbances when using output disturbance models. Workshop on Model Based Process Control, T. J. McAvoy, Y. Arkun,
and E. Zafiriou, eds., Pergamon Press, Oxford, p. 37 Ž1988..
We do not discuss these issues. We show the somewhat coun-
Morari, M., and G. Stephanopoulos, ‘‘Minimizing Unobservability in
terintuitive result that the dimension of the disturbance model Inferential Control Schemes,’’ Int. J. Control, 31, 367 Ž1980..
has to be as large as the measurement vector to obtain Morari, M., and E. Zafiriou, Robust Process Control, Prentice Hall,
offset-free control, regardless of the number of controlled Englewood Cliffs, NJ Ž1989..
variables. Muske, K. R., and T. A. Badgwell, ‘‘Disturbance Modeling for Off-
set-Free Linear Model Predictive Control,’’ J. Process Control, 12,
MB prove the existence of their disturbance model. The 617 Ž2002..
question of existence is simple in the unstructured model Muske, K. R., and J. B. Rawlings, ‘‘Model Predictive Control with
treated here. Existence is not as simple in the block-diagonal Linear Models,’’ AIChE J., 39, 262 Ž1993..
structures model, and MB’s existence result shows that they Pannocchia, G., ‘‘Robust Offset-Free Model Predictive Control,’’
IFAC 2002 World Congress, Barcelona, Spain Ž2002..
do not lose the ability to treat any systems by choosing the
Qiu, L., and E. Davison, ‘‘Performance Limitations of Non-Mini-
block-structured model. mum Phase Systems in the Servomechanism Problem,’’ Automat-
ica, 29, 337 Ž1993..
Rawlings, J. B., E. S. Meadows, and K. R. Muske, ‘‘Nonlinear Model
Acknowledgments Predictive Control: A Tutorial and Survey,’’ ADCHEM Proc., Ky-
The authors gratefully acknowledge the financial support of the oto, Japan, p. 185 Ž1994..
industrial members of the Texas-Wisconsin Modeling and Control Rawlings, J. B., and K. R. Muske, ‘‘Stability of Constrained Receding
Consortium, NSF through Grant CTS-0105360, and the Department Horizon Control,’’ IEEE Trans. Autom. Control, AC–38, 1512
of Chemical Engineering of the University of Pisa ŽItaly.. All simula- Ž1993..
tions were performed using Octave Žhttp:rrwww.octave.org.. Octave Richalet, J., A. Rault, J. L. Testud, and J. Papon, ‘‘Model Predictive
is freely distributed under the terms of the GNU General Public Heuristic Control: Applications to Industrial Processes,’’ Automat-
License. ica, 14, 413 Ž1978..
Scokaert, P. O. M., and J. B. Rawlings, ‘‘Constrained Linear
Quadratic Regulation,’’ IEEE Trans. Autom. Control, AC-43, 1163
Literature Cited Ž1998..
Skogestad, S., and I. Postlethwaite, Multi®ariable Feedback Control,
Chmielewski, D., and V. Manousiouthakis, ‘‘On Constrained Infi-
Analysis and Design, Wiley, New York Ž1996..
nite-Time Linear Quadratic Optimal Control,’’ Syst. Control Lett.,
Smith, H. W., and E. J. Davison, ‘‘Design of Industrial Regulators.
29, 121 Ž1996..
Integral Feedback and Feedforward Control,’’ Proc. IEE, 119, 1210
Clarke, D. W., and C. Mohtadi, ‘‘Properties of Generalized Predic- Ž1972..
tive Control,’’ Automatica, 25, 859 Ž1989..
Sontag, E. D., Mathematical Control Theory, 2nd ed., Springer-Verlag,
Clarke, D. W., C. Mohtadi, and P. S. Tuffs, ‘‘Generalized Predictive
New York Ž1998..
Control: I. The Basic Algorithm,’’ Automatica, 23, 137 Ž1987a..
Clarke, D. W., C. Mohtadi, and P. S. Tuffs, ‘‘Generalized Predictive
Control: II. Extensions and Interpretations,’’ Automatica, 23, 149 Appendix: Proofs
Ž1987b..
Cutler, C. R., and B. L. Ramaker, ‘‘Dynamic Matrix ControlᎏA Proof of Lemma 2
Computer Control Algorithm,’’ AIChE Meeting, Houston, TX Since the estimator matrix Ay ˜ LC˜ ˜ is stable and the input
Ž1979..
Davison, E. J., and H. W. Smith, ‘‘Pole Assignment in Linear Time- and the plant have reached steady values, we have from Eq.
Invariant Multivariable Systems with Constant Disturbances,’’ Au- 14 that the state and the disturbance estimates also reach
tomatica, 7, 489 Ž1971.. steady values. We denote with xy y
s and d s the steady-state
Davison, E. J., and H. W. Smith, ‘‘A Note on the Design of Indus- values of the model state and disturbance, respectively, prior
trial Regulators: Integral Feedback and Feedforward Controllers,’’
Automatica, 10, 329 Ž1974..
to filtering. We also denote with xq q
s and d s the correspond-
Eppler, S., ‘‘Steady Offset and Integral Control in Model Predictive ing steady-state values after filtering Žin general, they can be
Control,’’ Ad®anced Independent Study, University of Wisconsin- different .. Let e s be the steady-state output reconstruction
Madison, 1997. error defined as
Francis, B. A., and W. M. Wonham, ‘‘The Internal Model Principle
of Control Theory,’’ Automatica, 12, 457 Ž1976.. e s s ys yCxy y
s yC d d s .
Golub, G. H., and C. F. Van Loan, Matrix Computations, The Johns
Hopkins University Press, Baltimore, MD Ž1996..
Henson, M. A., and D. E. Seborg, Nonlinear Process Control, Pren- From Eqs. 14 and 15 we have
tice Hall PTR, Upper Saddle River, NJ Ž1997..
Kurtz, M. J., and M. A. Henson, ‘‘State and Disturbance Estimation xy y y
s s Ax s q Bu s q Bd d s q AL x e s Ž A1a.

436 February 2003 Vol. 49, No. 2 AIChE Journal

 0 s Ld es . Ž A1b. tains the eigenvalues ␭ i of the matrix L2 . Moreover, ⌫ can
be chosen in such a way that the eigenvalues appear in any
From the target calculation Eq. 18b, we have order along the diagonal.
Suppose L2 is singular Žor equivalently not full rank., we
x t s Ax t q Bu t q Bd dy can choose ⌫ such that the last eigenvalue ␭ p is zero. Thus,
s , Ž A2.
the elements of the last row of ⌳ are all equal to zero. Then,
in which we used the fact that dq y y we have
s s d s q L d e s s d s . More-
over, since constraints are not active at steady state and infi-
nite horizon is used, the steady-state input, u s , is given by Ay L1C B d y L1 C d
˜ LC
Ay ˜ ˜s
y ⌫⌳⌫ H C I y ⌫⌳⌫ H C d
u s s u t q K Ž xq y
s y x t . s u t q K Ž x s q L x e s y x t . . Ž A3 .

I 0 Ay L1C Ž B d y L1 C d . ⌫ I 0
Combining Eqs. A1a, A2, and A3, we obtain s .
0 ⌫ y ⌳⌫ C H
I y ⌳⌫ H C d ⌫ 0 ⌫H
y1
xy
s y x t s Ž I y Ay BK . Ž Aq BK . L x e s , Ž A4. I 0 ˜ LC
˜ ˜ are equal
Since is unitary, the eigenvalues of Ay
0 ⌫
in which Ž I y Ay BK .y1 exists since A q BK is a strictly to the eigenvalues of
stable matrix Žsee, e.g., Kwakernaak and Sivan, 1972.. Using
Eqs. 18b and A4, we obtain
Ay L1C Ž B d y L1 C d . ⌫
Âs .
Hys y z s H Ž e s qCxy y y y ⌳⌫ C H
I y ⌳⌫ H C d ⌫
s qC d d s yCx t yC d d s .

y1
s H I qC Ž I y Ay BK . Ž Aq BK . L x e s Since the last row of ⌳ has only null elements, the matrix Aˆ
s K e es . has the following form:

Since Eq. A1b we have that e s g nullŽ L d ., and from Eq. 25 = ... ... =
. .
we have that K e e s s 0, and therefore Eq. 26. . .
Âs . . . Ž A6.
= ... ... =
Proof of Lemma 3 0 ... ... 1
The first statement ŽEq. 28. follows immediately from Eq.
27. From Eq. 15 we have that L d s L2 , and we prove that Therefore, 1 is an eigenvalue of Aˆ and of Ay
˜ LC,
˜ ˜ which
˜ LC.
contradicts the stability of Ay ˜ ˜ Hence, L d s L2 cannot
rank L2 s p by contradiction. Under the assumptions of
Lemma 1, the augmented system is detectable. Hence, a fil- be rank deficient.
ter gain, L,˜ exists such that the estimator characteristic
closed-loop matrix Ay ˜ LC˜ ˜ is stable. L2 g⺢ p=p can be Proof of Theorem 1
rewritten using the Schur decomposition wsee, e.g., Golub and Since the assumptions of Lemma 3 are satisfied, we have
Van Loan Ž1996, p. 313.x that L d is square and full rank. Therefore, the null space of
L d is only the zero vector. Thus, the assumptions of Lemma 2
L2 s ⌫⌳⌫ H , Ž A5. are satisfied and there is zero offset in the controlled vari-
ables.
in which ⌫ g⺓ p=p is a unitary matrix Ž ⌫⌫ H s ⌫ H ⌫ s I . and
⌳ g⺓ p=p is an upper triangular matrix whose diagonal con- Manuscript recei®ed Apr. 3, 2001, and re®ision recei®ed July 11, 2002.

AIChE Journal February 2003 Vol. 49, No. 2 437