Chapter 9

Optimal Windup and Directionality

Compensation in Input-Constrained
Nonlinear Systems

M. Soroush
Drexel University, Philadelphia, Pennsylvania
P. Daoutidis
University of Minnesota, Minneapolis, Minnesota

9.1. Introduction

When a plant with actuator saturation nonlinearities is controlled with a

dynamic controller, the closed-loop stability and performance may degrade
significantly due to directionality and/or windup effects. This chapter sur-
veys definitions of these two phenomena, presents an optimal directionality
compensator, and develops a dynamic nonlinear controller which compen-
sates for windup and allows for optimal performance in the presence of
saturation.
Specifically, the notion of directionality in input-constrained systems is
defined, and the class of plants that do not exhibit the directionality are
characterized. The performance of the optimal directionality compensator
is shown and compared with those of clipping and direction preservation, by
linear and nonlinear examples. Given a controller output, the directionality

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 228 Soroush and Daoutidis

Controller
w ? u
> r
Directionality
Compensator

Figure 1: Directionality compensation.

compensator calculates an optimal feasible (constrained) plant input that

results in a plant response as close as possible to the response of the same
plant to the controller output. The compensator can be used for both linear
and nonlinear plants, irrespective of the type of controller being used.
A nonlinear dynamic controller which includes windup compensation is
then presented. In the absence of input constraints, the proposed controller
is input-output linearizing, whereas in the presence of input constraints it
provides the flexibility of achieving performance optimality and a desired
region of closed-loop asymptotic stability. The application and performance
of the controller are demonstrated by a chemical reactor example.

9.2. Directionality and Windup

When a plant with actuator saturation nonlinearities is controlled by an
analytical dynamic controller, the closed-loop response may be considerably
poorer than an integral-of-squared-error (ISE) optimal response. Examples
of analytical controllers are PID controllers, internal model controllers and
input-output linearizing controllers. This poorer closed-loop performance
can be due to windup and/or directionality. If the directionality is not
compensated for, then the plant actuators naturally render the controller
output feasible by clipping (limiting) the controller output components.
The problem of directionality compensation is that of calculating a feasible
plant input on the basis of a given unconstrained controller output (see
Figure 1). It is worth noting that these two problems are not present in
model predictive control, in which constraints are explicitly accounted for
and the controller action is solution to a constrained optimization problem.

9.2.1. Directionality
The phenomenon of directionality usually occurs in multiple-input
multiple-output (MIMO) plants. In single-input single-output (SISO) plants

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 Optimal Windup and Directionality Compensation 229

with actuator constraints, the boundary of the plant-input feasible set

(which is naturally closed and convex) consists of only two isolated points,
and when the controller output is infeasible, one of the two points that
is closest to the unconstrained controller output usually yields an optimal
response, i.e. one that is closest to the response of the same plant to the
unconstrained controller output. In other words, in SISO plants clipping
(limiting) an unconstrained controller output usually leads to an optimal
feasible plant input. In MIMO plants with actuator constraints, however,
the boundary of the plant-input feasible set consists of an infinite number of
points, and when the controller output is infeasible, the feasible point that
is closest (in the plant input space) to the unconstrained controller out-
put may not yield an optimal (in the sense described above) response. In
other words, in MIMO plants clipping the components of an unconstrained
controller output may not lead to an optimal feasible plant input.
Compared to integral windup, the phenomenon of directionality has
received less attention. In purely analytical control methods, a feasible
plant input, it, has been obtained by one of the following methods:
Clipping [8,28]:
A (u{n Wi <Uie
Ui = sa,ti[w] = < tu, uif <wt< Uht, ( = 1, , m
(uht, we > Uhe
where w is the controller output, and uie and Uht are respectively the
lower and upper limits on a plant input HI.
Direction preservation [4,6,12,17]:
. (sati(iy) satm(iy) 1
Ui = Wi mm < , , > , t = 1, , m
{ U!i Wm }

In [4], the direction preservation approach has been suggested for

directionality compensation in plants with ill-conditioned steady state
gain matrix.
Optimization formulation of the conditioning technique [7,27]. In this
approach, when a controller output is infeasible, a feasible controller
output is obtained by calculating (via optimization) a new setpoint
value that is closest (in the setpoint space) to the original setpoint
value and yields a feasible controller output.
Controller detuning [15]. When controller output is infeasible, an op-
timization problem is solved to obtain the values of controller tunable
parameters that result in a feasible controller output.

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 230 Soroush and Daoutidis

Optimal directionality compensation of Soroush and Valluri [21,22].

Unlike SISO plants, in MIMO plants clipping and direction preservation

may lead to completely different feasible plant inputs and may steer the
plant in wrong directions, leading to very poor closed-loop performance.
The directionality compensation via optimization formulation of the condi-
tioning technique may also lead to poor performance, because it is based on
the controller being used but not on the plant being controlled. For exam-
ple, when completely decentralized control is used, this method is identical
to clipping, irrespective of the nature of the plant under control.

9.2.2. Windup

Windup is another controller performance degradation phenomenon as-

sociated with actuator saturation, which is typically exhibited by dynamic
controllers with slow or unstable modes [6] (a special case being the PI/PID
controllers that can exhibit integrator windup). Although this phenomenon
has been studied extensively, only a few attempts have been made to define
it precisely. Furthermore, while closed-loop-response quality indices such
as response time and overshoot have been used to document the presence of
windup, at the present time there is no specific measure to quantify windup.
We note two criteria that have been proposed to check whether a dynamic
controller does/does not exhibit windup:

Criterion 1 [4]: A dynamic controller does not exhibit windup, if the

states of the controller are not driven by the error when the actuator
is in saturation.
Criterion 2 [9,10]: A dynamic controller does not exhibit windup, if
when the actuator is in saturation the closed-loop behavior under the
controller is identical to that under a 'reference' static state feedback
law (this characterization is based on the realization that windup is
not associated with static feedback controllers).

In linear analytical control, the issues of windup and constraint handling

as well as closed-loop stability in the presence of input constraints have been
studied extensively [1,2,5,7-9,13,15,23,27,28]. In nonlinear analytical
model-based control, these issues have also received considerable attention
in recent years. More specifically, there have been several approaches to
the problem of windup in input-output linearizing control methods. These
include:

Conditional integration (i.e. turning off integration when a constraint

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 Optimal Wind up and Directionality Compensation 231

is active). This approach was employed in real-time nonlinear control

of pilot-scale polymerization reactors (e.g. [20]).

Model predictive control (MFC) formulation of input-output lineariza-

tion [19,26].

Linear anti-windup (or MFC) schemes combined with control laws

that enforce input-output linearization, even in the presence of con-
straints, thus translating the constraints on the manipulated input
into state-dependent constraints on the input to the linearizing feed-
back loop [3,12,14,16].

An observer-based anti-windup approach with a nonlinear gain [10].

This approach allows attenuation of the effect of windup (due to the
controller dynamics), at a desired (arbitrarily fast) rate, with asymp-
totic closed-loop stability being ensured.

9.2.3. Organization of this Chapter

Section 9.3 presents the definition of directionality and the optimal di-
rectionality compensator. It also describes the scope of the work and the
application and performance of the directionality compensator via numeri-
cal simulations. In Section 9.4, a dynamic input-output linearizing control
law that can handle input constraints and constant disturbances and model
errors is derived. The performance of the nonlinear control law is shown
by a chemical reactor example.

9.3. Optimal Directionality Compensation

9.3.1. Scope
We consider the class of general but afnne-in-control, nonlinear multi-
variable plants described by a state-space model of the form

x = f(x)+g(x)u
y = h(x)
where x = [xi xn}T e X C 3? nxl , u = [HI um]T U C 5? mxl
and y = [yi ym]T 3 mxl are the vectors of state variables, plant
inputs (manipulated inputs), and controlled outputs respectively. Here
U = {u\uie <ue< Uhn =!,-,m}, where u\n u^n t = 1, ,m, are
scalar constant quantities, and X is an open connected set. A controller
output w is said to be feasible, if and only if w e U. It is assumed that:

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 232 Soroush and Daoutidis

gi(x), , gm(x), h(x) and f ( x ) are smooth vector functions, where gj(x]
represents the j th column of the matrix g(x); the plant is minimum phase
(has asymptotically stable zero dynamics); each controlled output yi has
a finite relative order (degree) r, which is the smallest integer for which
ri l ri l
locally \Lg
[ lL jf ~ hi(x}--LqrnL
r nf ~ hi(x]\
/ j
^ 0; the characteristic (decou-
pling) matrix of plant is locally nonsingular. The characteristic matrix is
an m x m matrix whose r/ th entry is Lg}Lr,'~1hi(x}] it will be denoted by
C(x). Here Lf and Lg/ are Lie derivative (in the directions of the vectors
/ and gj respectively) operators.

9.3.2. Directionality

Whether a plant is SISO or MIMO, clipping a controller output it;,

sat(u>), yields the feasible plant input which is closest (in the plant input
space) to the unconstrained controller output w. In mathematical terms,
u sat (it;) is the solution to the quadratic program
o
min u w\\
u
subject to
uif < ue < Uhf, 2 = 1, , m
where \\rj \ denotes the Euclidean norm of a vector ??. This feasible plant
input, which is closest (in the plant input space) to it;, may not lead to an
optimal response (i.e. one that is closest to the response of the same plant
to the unconstrained controller output it;). We will refer to this performance
degradation as the directionality, a precise definition of which is given here.

Definition 9.1. A plant in the form of (9.1) does not exhibit direc-
tionality, if and only if for every plant input it; G 5RTO x 1 the response of the
plant to sat (it;) is closest (in the output space) to the response of the same
plant to it;.

It is worth noting that in several aspects this notion of plant direction-

ality is different than the one known as the dependence of plant gain on
the direction of plant input vector. For example, the latter does not exist
when the condition number of plant gain matrix is one, while as we will
see, the former is not present when plant characteristic matrix is diagonal.

9.3.3. Optimal Directionality Compensation

For a plant in the form of (9.1), let y|(r) and yt(r}, r > t, represent
the predicted values of a controlled output yi when the plant is subjected

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 Optimal Windup and Directionality Compensation 233

to a given unconstrained controller output w and to a feasible plant input

u, respectively. The objective is to calculate a feasible plant input, w, that
renders the predicted value of every output ye, ye, as close as possible to
y\. In mathematical terms, we seek a feasible plant input u that is solution
to the constrained minimization problem:

(9.2)

subject to the input constraints

uie < ut(t) < uhf, = 1, m (9.3)

where t represents the present time, |(r)| Pe is the p^-function norm of

a scalar function (r) over a sufficiently short time interval of the form
[t, t + The} with 7)^ > 0 :

t+Th.
A
\pt.

and gi, , g m , are adjustable positive scalar parameters whose values are
set according to the relative importance of the controlled outputs: the
higher the value of a qg, the smaller the mismatch between the constrained
and unconstrained plant responses in ye (the lesser the effect of the con-
straints on the ye response).

Theorem 9.1. [18,21] For a plant of the form of (9.1), at each time
instant t given an unconstrained controller output w, the optimal feasible
plant input, denoted by w + , that minimizes the performance index in (9.2)
subject to the constraints of (9.3), is the solution to the m-dimensional
quadratic program:
mm \QC(x}u - QC(x)w (9.4)
u
subject to
m (9.5)
where Q is a constant m x m diagonal matrix given by

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 234 Soroush and Daoutidis

Controller
w Optimal
Directionality
Compensator
u
>
r
i(
X

Figure 2: Optimal directionality compensation in nonlinear systems.

The quadratic program of (9.4) and (9.5) is trivially solvable. For ex-
ample, one can use the computationally efficient, simple method described
in [21].
Theorem 9.1 indicates that at each time instant, the optimal feasible
plant input, u+, is calculated on the basis of C(x) and a given unconstrained
controller output w. Let J-[C(x),w] denote the solution to the quadratic
program of (9.4) and (9.5). Then, u+ ^"[C(x),^] represents the optimal
directionality compensator. Thus, the characteristic matrix plays a key role
in the optimal directionality compensation: to calculate an optimal feasi-
ble plant input in a nonlinear plant, given a controller output, one needs to
know the characteristic matrix and measurements of the state variables of
the plant (see also Figure 2). It is the nature of the characteristic matrix,
not that of steady state gain matrix, that determines when it is optimal to
use the clipping approach for directionality compensation. It is noteworthy
that the characteristic matrix and the steady state gain matrix characterize
two different aspects of plant behavior; the former characterizes the sensi-
tivity of plant to input changes over a very short horizon and the latter over
an infinite horizon. Structural properties such as singular values and the
condition number of the steady state gain matrix and relative gain array
also characterize the plant response over an infinite horizon. Using steady
state structural properties as a basis for selecting either of the approaches
may lead to a very poor closed-loop performance, unless one uses a steady
state controller.

Remark 9.1. For the class of plants with diagonal characteristic ma-
trix, the optimal directionality compensator is identical to m limiters (clip-
pers), i.e.

Thus, for this class of plants the feasible plant input that is closest (in
the plant input space) to the unconstrained controller output iy, yields
an optimal plant response (i.e. one that is closest to the response of the

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 Optimal Wj'ndup and Directionality Compensation 235

Optimal
Controller
w
Directionality u
Compensator
>
_T
Figure 3: Optimal directionality compensation in linear systems.

same plant to the unconstrained controller output w). In other words, this
class of plants do not exhibit the directionality, and thus in the presence
of input constraints their closed-loop performance is not degraded by the
directionality.

Remark 9.2. In the case that the weights qi, , qm are chosen such
that
qt\\(T-ty

that is, when the controlled outputs are of equal importance irrespective
of the values of their relative orders (ri, , r m ), the quadratic program of
(9.4) and (9.5) takes the simple form:

nun C(x)u-C(x)w\ (9.6)

subject to uif <Ui< Uhf, i 1, , m.

Remark 9.3. For multivariable time-invariant linear systems described

by a state-space model of the form

x = Ax + Bu
(9.7)
y =Cx
where A, B and C are nxn, n x m and mxn matrices respectively, the
characteristic matrix

C =

is independent of x. Thus, as shown in Figure 3, in time-invariant linear

plants the optimal directionality compensator does not require information
on the state of plant: u = F[C,w\. In linear plants, the characteristic

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 236 Soroush and Daoutidis

matrix will be diagonal, if and only if the diagonal element of every row of
the transfer function matrix has the absolutely lowest relative order in that
row, where relative order of a rational function is the difference between
the orders of the numerator and denominator polynomials of the function.

Remark 9.4. A feasible plant input value calculated by the optimal

directionality compensator of (9.4) and (9.5) is optimal over a very short
time horizon. Because the time horizon was chosen to be very short, the
resulting constrained optimization problem is a quadratic program that
is easy to solve. However, a calculated optimal plant input value that is
optimal over a very short time horizon may not be optimal over a very long
(infinite) time horizon, for example, when the plant under consideration is
non-minimum-phase.

Remark 9.5. The optimal directionality compensator allows one to

minimize the effect of input constraints on the controlled outputs that are
more important in a given plant, by choosing higher values for the weights,
< ? ! > ? < ? m ) corresponding to more important controlled outputs.

9.3.4. Application to Two Plants

Decentralized PI Control of a Linear Plant. Consider the linear
two-input two-output plant [18]:

I 40 -300
P(8) = (9.8)
100s + 1 -1 40

with Ui\ < 1. i 1, 2. For this example, r\ = 1, r<i 1, and

[ 0.4 -3.0
-0.01 0.4

Two completely decentralized PI controllers with kCl 2.1, kC2 = 0.36,

TJI 174.0 s and r/2 = 22.1 s, and with conditional integration (to prevent
integral windup) are used to track asymptotically the set point changes
yspi 8 and ysp2 = 3. The conditional integration involves turning off the
integrator of the ith loop when the input Ui saturates. Figure 4 depicts
the closed-loop output response under the same two PI controllers but
three different directionality compensators; it shows that while clipping
and direction preservation approaches lead to poor responses in 7/1, the
optimal direction compensator provides a significantly better closed-loop
performance.

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 Optimal Windup and Directionality Compensation 237

iu-
7-
4- (^-r----~'-~*""""~ '"-^
1- .-'' ..---"'

-2-
5_ 1 1 1 1
5_
-
CM 1- f.^-
>
-1-
-3-
5_ 1 1 1 1
1 0i

1.1- 1
1 "-l.^
'
0.9- \ /^

0.8-
0 7_
vyi i i i
1 R^

0.5-
1-

U-
v' __
O c_

(D
i
60
i
120
i
180
i
240 3C

Figure 4: Controlled outputs and plant inputs of the linear example of

(9.8): solid = no bounds on the inputs; dotted = clipping, dotted-dashed =
direction preservation, and dashed = optimal directionality compensation,
when \Uj\ < 1, j = 1,2.

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 238 Soroush and Daoutidis

I-O Linearizing Control of a Nonlinear Bioreactor. Consider a

continuous stirred-tank bioreactor described by a mathematical model [18]:

I
dt y*
dS2 (9.9)
dt YX / S 2
dX
= fjL(S1,S2)X-X(D1+D2]

where the specific growth rate

S2

Si and 52 denote the outlet concentrations of substrates 1 and 2 respec-

tively, S/l and 5/2 respectively represent the concentration of substrate 1
in feed 1 and the concentration of substrate 2 in feed 2, and DI and D2 are
respectively the dilution rates of feed streams 1 and 2. The values of the
model parameters are the same as in [24].
The controlled outputs and manipulated inputs are as follows: y\ =
5i, y2 = 52, MI = DI, and u2 = D2 with the bounds 0 < D\,D2 <
0.4 h~1. The control objective is to operate the reactor at the set points
yspl 2.0 kg.m~3 and ysp2 = 4.9 kg.m~3, by an input-output linearizing
controller with an optimal integral windup compensator.
For this plant. r\ 1, r2 1, and

n _
Sf1 Si Si
S2 5/2 52

Figure 5 depicts the startup profiles of the controlled outputs and plant
inputs of the bioreactor under the same nonlinear controller but three dif-
ferent directionality compensators. In the presence of the input constraints,
clipping (dashed line) cannot operate the plant at the steady state (leads a
very poor closed-loop response), and direction preservation (dotted-dashed
line) results in a relatively better performance compared to that of clipping.
However, the closed-loop performance under the optimal directionality com-
pensator is of higher quality; it is the closest response to that represented
by the solid line (obtained in the absence of the plant input bounds).

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 Optimal Wmdup and Directionality Compensation 239

4
<?- 3-
g> 2-

4-
CN
co 2

0-
0.8-
^ 0.6-

0
Tima h

Figure 5: Profiles of the controlled outputs and plant inputs of the biore-
actor example of (9.9): solid = no bounds on the inputs; dotted = clipping,
dotted-dashed = direction preservation, and dashed = optimal directional-
ity compensation, when 0 < Di,D2 < 0.4 h'1.

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 240 Soroush and Daoutidis

9.4. Windup Compensation

9.4.1. Scope

We consider the class of multi-input multi-output, continuous-time, non-

linear plants described by a state-space model of the form of (9.1). We
make the following additional assumptions: (al) every variable of the sys-
tem of (9.1) is in the form of deviation from its nominal steady-state value,
and thus the origin is an equilibrium point; (a2) the output set-point, de-
noted by ysp. is achievable at steady state in the sense that there exists
a u0 G interior (U), which satisfies /((") + g((,}uo = 0, where G X and
/i(C) = ysp\ (a3) det[C(x}} ^ 0. Va: G X; and (a4) the plant is minimum
phase (has asymptotically stable zero dynamics) on X.

9.5. Nonlinear Controller Design

For a nonlinear plant with a model in the form of (9.1), let us request
a linear input-output response of the form:

u~rrr 11ysp CQ 10")
i^y.iuj

where /^ G 5? mxl . i = l , - - - , m , j = l , - - - , r j , are adjustable constants

that are chosen such that all roots of

det< I -\

lie in the left half plane and the matrix

is nonsingular. It is straightforward to show through a direct calculation

of the output derivatives that this closed-loop response translates into the
following relation:
Q(X u\ ys (9-11)
where m, r,
3>(x,u] = h(x)

Let the solution for u of (9.11) be denoted by the following state feedback
law:
(9.12)

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 Optimal Windup and Directionality Compensation 241

where
A 1-1
ysp - h(x]

and the time-invariant linear system:

(9.13)

be a minimal-order state-space realization of

(9.14)

where 7^- e 3?mx , i = l , - - - , r a , j = l , - - - , r j are adjustable constants

that are chosen such that all roots of

det{I 5
^ 7m j =0
j 1

lie in the left half plane and the matrix

[Tin ' ' ' TmrJ

is nonsingular. The above system calculates 77 which is the vector of esti-
mates of the plant outputs. It essentially acts as an observer for the plant
outputs and their derivatives up to order (r^ 1), i = 1, , m.

Theorem 9.2. Consider a plant with a model of the form of (9.1) and
the dynamic controller:
,U), 77(0) =
w= (9.15)

A
where e ysp y. Then,

(a) the controller has integral action, i.e. in the presence of constant
disturbances and model errors, induces an offset-free closed-loop re-
sponse [24,25].
(b) when the constraints are not active, and (3u = 7^, ^ = 1, , TJ, i =
1, ,ra, and its states are initialized consistently, the controller in-
duces the linear input-output closed-loop response of (9.10) [24,25].

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 242 Soroush and Daoutidis

(c) when the constraints are active, fin 7^, ^ = l , - - - , r j , i =

1, , m, and its states are initialized consistently, the controller min-
imizes a constrained quadratic performance index [24,25].

(d) when the constraints are active, the origin of the closed-loop system is
locally asymptotically stable, provided that the state feedback law of
(9.12) is also locally asymptotically stabilizing and the poles of (9.14)
are placed sufficiently far left in the complex plane [10].

9.5.1. Application to a Nonlinear Chemical Reactor

We consider the same chemical reactor described in [20]:

CA = f i ( C A , T ) + T-CAi
l (9-16)
r=/ a ( c A ,r) + Q
where CA and CAi are respectively the outlet and inlet concentrations of
the reactant, T is the outlet stream temperature, and Q is the rate of heat
input to the reactor. The controlled outputs and manipulated inputs are:
y\ = CA, y-2 T, HI = CA,-, and u^ = Q with the bounds 5 < u\ <
15 kmol.m~3 and 10 < 112 < 10 k,J.s~l.
Here r\ = r% 1 and C = diag{l/r, l/(pcV)}. Application of the
control law of Theorem 9.2, 7^ = 7^ = /^ = /3|i = 100 and 7^ = 7^ =
j3^ = fyi 0, to this chemical reactor leads to the following mixed error-
and state- feedback controller:

& = 0.01 |-& + T + 100 ( / 2 (

<
u\ , <. r
= sati
100

pcF
100
Figure 6 depicts the startup profiles of the controlled outputs and manipu-
lated inputs under the nonlinear controller of (9.17). In the absence of the
input constraints, the closed-loop plant output responses (solid line) are
exactly two completely decoupled, first-order responses.

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 Optimal Wine/up and Directionality Compensation 243

co
3.0n

2
6 -H
1.0-
O
0.(

10-
8-
I 6-
4-

40-
30-
3 20-
$ 10
OH
0 200 400 600 800 1000
Time, s

Figure 6: Profiles of the controlled outputs and plant inputs of the reactor
example of 9.16: solid = no bounds on the inputs; dashed = when 5 < u\ <
15 kmol.m"3 and -10 < u2 < 10 kJ.s~1.

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 244 Soroush and Daoutidis

Acknowledgement
The authors gratefully acknowledge financial support from the National
Science Foundation through the grants CTS-9703278 and CTS-9624725.

References

[1] K. J. Astrom and L. Rundqwist. Integrator Windup and How to Avoid

It, in Proc. of American Control Conference, Pittsburgh, PA (1989)
1693.

[2] D. S. Bernstein and A. N. Michel. A Chronological Bibliography

on Saturating Actuators, International J. of Robust and Nonlinear
Contr., 5 (1995) 375.
[3] J. P. Calvet and Y. Arkun. Feedforward and Feedback Linearization
of Nonlinear Systems and its Implementation Using Internal Model
Control, Ind. Eng. Chem. Res., 27 (1988) 1822.
[4] P. J. Campo and M. Morari. Robust Control of Processes Subject to
Saturation Nonlinearities, Comput. & Chem. Eng., 14 (1990) 343.
[5] E. Coulibaly, S. Maiti, and C. Brosilow. Internal Model Predictive
Control (IMPC), Automatica, 31 (1995) 1471.
[6] J. C. Doyle, R. S. Smith, and D. F. Enns. Control of Plants with Input
Saturation Nonlinearities, in: Proc. of American Control Conference,
(1987) 1034.
[7] R. Hanus and M. Kinnaert. Control of Constrained Multivariable Sys-
tem Using Conditioning Technique, in: Proc. of American Control
Conference, (1989) 1712.
[8] N. Kapoor and P. Daoutidis. Stabilization of Systems with Input Con-
straints, International J. of Contr., 66(5) (1997) 653.
[9] N. Kapoor, A. R. Teel, and P. Daoutidis. An Anti-Windup Design for
Linear Systems with Input Saturation, Automatica, 34 (1998) 559.
[10] N. Kapoor and P. Daoutidis. An Observer-based Anti-Windup Scheme
for Nonlinear Systems with Input Constraints, International J. of
Contr., 72(1) (1999) 18.
[11] N. Kapoor and P. Daoutidis. Stabilization of Nonlinear Processes with
Input Constraints, Comp. Chem. Engng., 24 (2000) 9.

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 Optimal Wind up and Directionality Compensation 245

[12] T. A. Kendi and F. J. Doyle. An Anti-Windup Scheme for Multivari-

able Nonlinear Systems, J. of Process Contr., 7(5) (1997) 329.
[13] M. V. Kothare, P. J. Campo, M. Morari, and C. N. Nett. A Unified
Framework for the Study of Antiwindup Designs, Automatica, 30
(1994) 1869.
[14] J. M. Kurtz and M. A. Henson. Input-Output Linearizing Control of
Constrained Nonlinear Processes, J. of Process Contr., 7(1) (1997) 3.
[15] S. Mhatre and C. Brosilow. Multivariable Model State Feedback, in:
Proc. of IFAC World Congress, San Francisco, M (1996) 139.
[16] S. L. Oliveira, V. Nevistic, and M. Morari. Control of Nonlinear Sys-
tems Subject to Input Constraints, in: Preprints of Nonlinear Control
Systems Design Symposium (1995) 15.
[17] J.-K. Park and C.-H. Choi. Dynamic Compensation Method for Mul-
tivariable Control Systems with Saturating Actuators, IEEE Trans.
Auto. Contr., 40(9) (1995) 1635.
[18] M. Soroush. Directionality and Windup Compensation in Nonlinear
Model-Based Control, in: Nonlinear Model-Based Process Control, R.
Berber and C. Kravaris (eds.), NATO ASI Series, Kluwer Academic
Publishers, Dordrecht (1998) 173.
[19] M. Soroush and C. Kravaris. A Continuous-Time Formulation of Non-
linear Model Predictive Control, in: Proc. of American Control Con-
ference, Chicago (1992) 1561.
[20] M. Soroush and C. Kravaris. Nonlinear Control of a Batch Polymer-
ization Reactor: an Experimental Study, AIChE J., 38 (1992) 1429.
[21] M. Soroush and S. Valluri. Optimal Directionality Compensation in
Processes with Input Saturation Nonlinearities, International J. of
Contr., 72(17) (1999) 1555.
[22] M. Soroush and S. Valluri. Calculation of Optimal Feasible Controller
Output in Multivariable Processes with Input Constraints, in: Proc.
of American Control Conference (1997) 3475.
[23] A. R. Teel. A Nonlinear Small Gain Theorem for the Analysis of Con-
trol Systems with Saturation, IEEE Trans. Automat. Contr., 41(9)
(1996) 1256.
[24] S. Valluri. Nonlinear Control of Processes with Actuator Saturations,
Ph.D. Thesis, Drexel University, Philadelphia (1997).

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.

 246 Soroush and Daoutidis

[25] S. Valluri and M. Soroush. A Nonlinear Controller Design Method for

Processes with Actuator Saturation Nonlinearities, Submitted to Inter.
J. of Control (2000).
[26] S. Valluri, M. Soroush, and M. Nikravesh. Shortest-Prediction-Horizon
Nonlinear Model Predictive Control, Chem. Eng. Sci., 53(2) (1998)
273.
[27] K. S. Walgama and J. Sternby. Conditioning Technique for Multi-Input
Multi-Output Processes with Input Saturation, in: it Proc. IEE Part
r^ 140
D, ^1 OOQ\ 231.
1 /in (1993) OQI

A. Zheng, M. V. Kothare, and M. Morari. Anti-Windup Design for

Internal Model Control, International J. Contr., 60 (1994) 1015.

Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.