You are on page 1of 23

Linear Control of Nonlinear Processes:

Recent Developments and Future Directions

Michael Nikolaou and Pratik Misra


Chemical Engineering Dept.
University of Houston
Houston, TX 77204-4004

nikolaou@uh.edu
http://athens.chee.uh.edu

CEPAC 2001, Guarujá, Sao Paulo, Brazil, http://pqi.ep.usp.br/eventos/cepac2001.htm

Abstract
Virtually all chemical processes are nonlinear, but for several of them linear feedback
control is adequate. Therefore, before nonlinear controller design is attempted, it is
natural to ask “When is linear control adequate for a nonlinear process?” to ensure
favorable cost/benefit ratio. Consequently, methods are needed that quantify the
nonlinearity of a process within the context of assessing whether linear control is
adequate or nonlinear control is warranted. In this work we summarize efforts towards
this end and elaborate on our latest research in this area. Specifically, we present a
rigorous and general theoretical framework as well as an associated, heuristically
refined computational methodology that allow not only analysis but also synthesis of a
linear control system for a nonlinear process. Application to the multivariable case is
presented. Potential future developments within this framework are discussed.

1 Introduction

1.1 A basic question: Linear or nonlinear control?


Interest in nonlinear feedback control of chemical processes has been steadily increasing over the
last several years. This is due both to the pronounced nonlinear nature of several chemical
processes (whether in mature or emerging fields) and to the increased sensing and computational
capabilities afforded by modern sensors, computers, algorithms, and software. Such capabilities
have been claimed and at times proven to offer benefits in better operation and control of
chemical processes. However, nonlinear control systems usually pose substantially higher data,
design, implementation, and maintenance demands than linear control systems. Therefore,
before one develops and implements a nonlinear control system, one must carefully examine the
potential advantages of such a system in comparison to a linear one, by resolving the following
basic question, which will be the central theme of this proposal:
Q1: Is linear control adequate or is nonlinear control necessary for a given process?
Rather than a mere Yes/No answer, insight is sought into how the answer to Q1 depends on

1
various problem parameters, such as process model and control structure, experimental or routine
modeling information, process operating range, constraints, disturbances, etc., as discussed in the
sequel.

1.2 All chemical processes are nonlinear, but not all of them require nonlinear control
Linear feedback control of chemical processes has a long history of research and diverse
industrial applications (Qin and Badgwell, 2000; Nikolaou, 1997; Kayihan, 1997; Edgar et al.,
1999). From single-input-single-output proportional-integral-derivative (SISO PID) controllers
to plantwide model-predictive control (MPC) systems (Qin and Badgwell, 1997), there is an
abundance of feedback control systems which implicitly or explicitly assume that process
dynamics are either inherently linear or almost linear owing to process operation close to a
steady state. However, there are important instances for which the linearity assumption may be
violated. Such instances are not uncommon in chemicals, polymers, natural gas processing,
pharmaceuticals, microelectronics, and pulp and paper plants (Qin and Badgwell, 2000), thus at
times necessitating nonlinear control algorithms (Sidebar 1 ).
It can be argued that virtually all chemical processes are in principle nonlinear. However,
some are evidently “more nonlinear” than others. Moreover, feedback may drastically reduce
the nonlinearity of a system, a fact recognized as early as during the invention of the feedback
amplifier by Black in the 1920s (Brittain, 1997). Therefore, methods are needed that quantify the
nonlinearity of a process within the context of assessing whether (and what) linear control is
adequate or whether (and what) nonlinear control is warranted (Harris and Seppala, 2001).

1.3 The interplay between nonlinearity and feedback can be quantified


Several and diverse sporadic attempts to quantify the nonlinearity of a dynamic process have
appeared over the last decade, as reviewed in section 3.1.
The PI’s group (Eker and Nikolaou, 2000 and 2001; Nikolaou, 2001a; Misra and
Nikolaou, 2001) recently proposed a rigorous and general theoretical framework as well as an
associated, heuristically refined computational methodology that address the interplay between
nonlinearity and feedback. Within this framework, as explained more concretely in section 3.2,
the need for nonlinear control is assessed in terms of a closed-loop nonlinearity measure.
Bounds on the closed-loop nonlinearity measure can be relatively simply evaluated in terms of
(a) process nonlinearity and (b) a linear feedback controller. More importantly, this framework
allows not only analysis but also synthesis of a linear control system for a nonlinear process.
This framework will provide the starting point of the proposed research.

1.4 Objectives of proposed research


The objective of the research work proposed in the sequel is to apply and expand the framework
referred to in section 1.3, in order to develop rigorous results and computational methodologies
that answer the following nontrivial questions of both theoretical and practical importance:
Q2: Should constrained MPC employ a linear or nonlinear model of a controlled process?
Q3: Can linear control stabilize an unstable nonlinear process? If so, over what operating
range?
The above questions Q2 and Q3 entail a number of specific research directions that will be
elaborated on in section 5
Sidebar 1 – The need for control of nonlinear processes
Over the latest few decades, studies on nonlinear control have generated a voluminous

2
body of work, within both chemical engineering and other disciplines (Bequette, 1991;
Rawlings et al., 1994; Kumar and Daoutidis, 1999 and Christofides, 2001b – emphasis on
geometric control of distributed parameter systems; Henson and Seborg, 1997 – collection
of various contributions on nonlinear systems; Krstic et al., 1995 – emphasis on
backstepping; Khalil, 1992; Slotine and Li, 1991; Nijmeijer and van der Schaft, 1990 and
Isidori, 1999 – emphasis on geometric control; Vidyasagar, 1993; de Figueiredo and Chen,
1993; Brockett, 1996).
It is interesting to note that an early academic publication that introduced what
today would be called nonlinear MPC, explicitly recognized and dealt with the issue of
nonlinearity for model-based control of a distillation column through on-line optimization
(Rafal and Stevens, 1968).
Nonlinearity, by definition referring to absence of linearity, can manifest itself in
various ways, such as nonlinear dynamics, constraints (saturation nonlinearity) and
changing modes of operation (see also Pearson (1999) for an interesting classification of
nonlinear behavior). While such nonlinearities are widespread, experience indicates that
they do not always require nonlinear control. On the other hand, important cases are also
known (e.g., polymer production changeover policies (Pavilion, 2001; Young, 2001)) for
which nonlinear control provides tangible advantages.
Discussed below are industrially important cases for which nonlinearity is usually
present and, consequently, knowing whether linear or nonlinear control is adequate would
provide obvious benefits.
- Biochemical production of chemicals. Enzyme kinetics (e.g., Michaelis-Menten) depend
nonlinearly on susbstrate concentration. With the rapidly increasing importance of
biotechnology, control of processes that produce and separate chemicals with the aid of
reactions in living cells will have obvious implications.
- Systems operating near constraints (e.g., petroleum refining, natural gas processing).
The saturation nonlinearity renders constrained MPC nonlinear, even if the model
employed is linear. While constrained MPC with linear model has been amply validated
in industrial practice, transition to constrained MPC with nonlinear model is all but trivial
and requires careful justification and execution.
- Non-routine operation situations (e.g., start-ups, shut-downs, change-overs, flares,
relief valve emissions). These are extremely important as the main contributors to
safety, quality, resource use, and environmental problems (Allen and Shonnard, 2001).
Because a process moves far from a steady state during non-routine operation,
nonlinear behavior is usually pronounced.
- Nonlinear distributed process systems (Kumar and Daoutidis, 1999; Christofides,
2001b) such as control of spatial profiles (CVD, etching, crystal growth, packed-bed
reactors) control of size distributions (aerosol production and particulate processes (Chiu
and Christofides, 1999), crystallization (Braatz and Hasebe, 2001), emulsion
polymerization, cell cultures (Daoutidis and Henson, 2001)), control of fluid flows
(mixing, wave suppression, drag reduction, separation delay, control of material
microstructure (thin-film growth, nano-structured coatings processing).
- Batch processes, particularly important for the production of fine chemicals and
pharmaceuticals.

2 Motivating Examples
As already suggested, the main thesis of the proposed research is that lack of linearity in the
behavior of a process to be controlled does not necessarily make nonlinear control
indispensable. On the other hand, the nature of process nonlinearity may make linear control
completely inadequate and may require nonlinear control of a kind that may not even be
immediately obvious. The following two examples make the case (for more details and other

3
interesting examples see Eker and Nikolaou, 2001).

2.1 Example 1 – (The right) linear control may be adequate for a nonlinear process
Consider the exothermic reaction A → B in a system of two jacket-cooled continuous stirred-
tank reactors (CSTR) in series (Henson and Seborg, 1990). The concentration of the reactant at
the exit of the second CSTR, CA2 , is the controlled
variable, and the coolant flow rate qc (common for both 7 ×10
-3

reactors) is the manipulated variable.

CA2
6
Figure 1 – Open-loop responses of CA2 (in 5
deviation form) to qc step change of +9.3, for 4
(a) the nonlinear system of two CSTRs (dashed 3
line) and (b) its linearized model around the 2
nominal steady state 5.3 × 10-3 mol/L (solid 1
line). 00 2 4 6 8 10 12 14 16 18 20
time
The step-response simulation in Figure 1 demonstrates ×10
-3

that the open-loop system is “fairly nonlinear”. 5

CA2
4
Figure 2 – Closed-loop responses of CA2 to
3
pulse setpoint change +4.2 × 10−3 mol/L for (a)
2
the nonlinear process with linear IMC (filter
time constant λ = 1 ) (dashed line), and (b) the 1

ideal linear IMC loop (solid line). 0


-1
However, when linear internal model control (IMC, 0 100 200 300 400 500
time
Morari and Zafiriou, 1989, p. 65) is used to steer CA2 to ×10
-3

the same value as in Figure 1, the resulting closed loop 5


is much less nonlinear or virtually linear (Figure 3). 4
CA2

Figure 3 – Similar to Figure 2 except that the 3


IMC filter time constant λ = 1 . 2

But increasing the pulse setpoint change to +5.2 × 10−3 1

mol/L generates an unstable closed-loop response far 0

from the ideal linear response (Figure 4). Similar -1


0 100 200 300 400 500
behavior is eventually observed if λ is further reduced. time

Figure 4 – Similar to Figure 3, except that


×10-3
pulse setpoint change is +5.2 × 10−3 mol/L. 8
CA2

Note nonlinear loop oscillations that would 6


have been trivially missed if the pulse had
lasted less than ~70 time units. 4

Moreover, if measurement delay of 5 time units is 2


added, the ideal linear closed loop changes little in
0 100 200 300 400 500
comparison to absence of delay, while the nonlinear time
closed loop shows drastic deterioration (Figure 5).

4
This example suggests that
(a) Open- and closed-loop nonlinearities can be very different;
(b) Process nonlinearity and controller design interact tightly and not monotonically;
(c) Process operating range profoundly affects nonlinearity and control;
(d) Similarly to (if not more importantly than for) linear processes, modeling uncertainty must be
carefully taken into account when controlling
9×10
-3
nonlinear processes.
8

CA2
Figure 5 – Closed-loop responses to setpoint
7
step change +4.5 × 10−3 for (a) the nonlinear 6
closed loop with linear IMC, λ = 0.1 (dashed 5
line), and (b) the ideal linear closed loop with 4
the same IMC (solid line). 3
2
The many different facets of behavior of this 1
system can be very well predicted by the theory 00 1 2 3 4 5 6 7 8 9 10
mentioned in section 1.3 and briefly discussed in time
section 3.2.

2.2 Example 2 – There are nonlinear processes that cannot be globally stabilized by any
continuous feedback law, linear or nonlinear
Meadows et al. (1995) showed that it is impossible to find any continuous static state feedback
law uk = F ( xk , yk ) that can globally asymptotically stabilize the system
xk +1 = xk + uk , yk +1 = yk + u3k (1)
around (0,0), while the constrained nonlinear MPC
min ∑ ( xl + i + y l + i + u l+ i )
2
2 2 2
(2)
i= 0
subject to
( xl , y l ) given ; xl + i+1 = xl+ i + ul+ i , yl + i+1 = yl+ i + ul3+ i , i = 0,1,2 ; x l+ 3 = yl + 3 = 0 (3)
can globally asymptotically stabilize the above system around (0,0).
This example suggests that seemingly innocent (e.g., polynomial) nonlinearities may
create great control challenges.

3 Related Background and Prior Results


Sidebar 2 – Nonlinear operator analysis basics and notation
Basics of the input-output system framework can be found in Willems (1971), Desoer and
Vidyasagar (1975), and Zeidler (1990), among others. Within this framework the dynamic
behavior of a nonlinear system is described by an unbiased nonlinear operator (mapping)
N : U → Y : u a y = Nu : 0 a N (0) = 0 (4)
which maps input signals u in the space U to output signals y in the space Y. Note that
there is no unanimity in literature regarding uniqueness of y given u. We will not assume
uniqueness in this work. The operator N is commonly realized through a set of ordinary or
partial differential equations and algebraic equations, such as
dx ( t) dt = f ( x( t ), u( t)), y( t ) = h( x( t), u(t )) (5)
Note that partial differential equations may be approximated by ordinary differential
equations such as (5) after spatial discretization using finite differences, finite elements,

5
Karhunen-Loeve decomposition, spectral decomposition, etc. The norm (gain) and
incremental norm (incremental gain or local Lipschitz constant (Willems, 1971, p. 93)) of
N : U → Y over the set V ⊆ U are defined (Nikolaou and Manousiouthakis, 1989) as
N V
= sup Nu u (6)
u ∈V , u≠ 0

and
N ∆V
= sup Nu1 − Nu2 u1 − u 2 (7)
u1 , u2 ∈V , u1 ≠ u2

respectively, where the norm functions on the right-hand sides of eqns. (6) and (7) are
defined on the spaces U and Y. The set V identifies these input signals that are physically
important for the operator N, e.g., mole fractions in the interval [0,1] . An operator
N : U → Y is bounded (stable) over the set V when
N ∆V < ∞ . (8)
Note that the above definition of stability in eqn. (8) supercedes the standard stability
definition N V
< ∞ , because N V
≤ N ∆V
. Note also that even for very simple nonlinear

operators it is possible to have N ∆V1


= ∞ and N ∆V 2
< ∞ (or N V1
= ∞ and N V2
< ∞ ) for
two different sets V1 and V2 (Nikolaou and Manousiouthakis, 1989). The linearization of an
operator N : U → Y around the input trajectory u0 is defined as the linear operator
Lu0 : U → Y satisfying the equation lim N ( u0 + u) − Nu0 − Lu0 u u = 0 . For N defined via
u→ 0

eqn. (5), Lu0 is a linear time-varying operator, defined by the equations


d ∆x
(t ) = A(t )∆ x( t ) + B (t )u( t ), y ( t ) = C (t ) ∆x ( t ) + D (t )u( t) (9)
dt
where ( A( t ), B( t ), C ( t ), D (t ) ) =( ∂f ∂ f
)
ˆ ∂x , ∂u , ∂∂hx , ∂∂hu x0 ( t ) ; x0 (t ) , u0 (t ) satisfy eqn. (5); and ∆x (0) = 0 .
u0 ( t )

3.1 Efforts to quantify nonlinearity have spawned a plethora of creative approaches


As stated in section 1.2, nonlinearity quantification is needed in order to assess whether linear
control is adequate for a nonlinear process. Let us stress that nonlinearity quantification goes
beyond mere detection of lack of linearity, a topic that has been amply studied (e.g., see Haber
and Unbehauen (1990), Pearson (1999) and references therein).
Efforts have appeared in literature to quantify the nonlinearity of a nonlinear operator N
(eqn. (4)) by computing its distance from a suitably defined linear operator. The basis of this
idea has been that a “fairly” nonlinear operator would require nonlinear control.
Desoer and Wang (1980) defined the nonlinearity measure of an operator N as
v =ˆ inf N − L (10)
L∈Λ
where the above minimization in eqn. (10) must be performed over all linear operators L in the
set Λ, and the norm function can be any suitable norm. Computation of v in eqn. (10) can be
extremely complicated, e.g., if the norm in eqn. (10) is an induced norm (eqn. (6)).
To address the computational issue, Nikolaou (1993) constructed an inner-product and
corresponding 2-norm theory for a class of nonlinear operators. Based on that theory, the
quantity v in eqn. (10), corresponding to the average discrepancy between outputs of N and L for
inputs within an explicitly specified set, can be trivially computed via Monte Carlo simulations.
In addition, explicit formulas for the optimal L in eqn. (10) can be derived. Using this theory,

6
Nikolaou and Hanagandi (1998) quantified the nonlinearity of several open- and closed-loop
chemical process systems, and showed how different tunings of a linear IMC controller used to
control a nonlinear process may result in closed loops of significantly different nonlinearity
magnitudes. While making computations easy, the 2-norm is not an induced norm, therefore it
does not satisfy the submultiplicativity property ( N1 N2 ≤ N1 N 2 ), thus limiting its use in
direct feedback controller synthesis.
Allgöwer (1996) tackled the computational problems of using an induced norm in eqn.
(10), by parametrizing the input signal u in eqn. (6) and the linear operator L in eqn. (10) though
finite-dimensional approximations, and by directly performing the optimization in eqn. (10), i.e.
v =ˆ inf sup Nu − Lu u (11)
L∈Λ u ∈V , u≠ 0

corresponding to the worst possible discrepancy between outputs of N and L. For a number of
examples, he found that the value of v is insensitive to the particular parametrization of u.
Helbig et al. (2000) defined a nonlinearity measure as
φ =ˆ inf sup inf N [u, x N ,0 ] − L[ u, x L,0 ] N [u , x N ,0 ] ∈ [0,1] (12)
L∈Λ u∈U , x ∈ X x ∈ X L ,0
N ,0 N ,0 L ,0

corresponding to the worst discrepancy between outputs of N and L as a function of both initial
conditions and inputs. Exact computation of φ is practically infeasible. However, Helbig et al.
(2000) have shown how to efficiently calculate good approximations or bounds of φ by finite-
dimensional parametrization of u and convex optimization.
To avoid having to directly optimize with respect to L in nonlinearity measures such as in
eqns. (10) and (12), Sun and Kosanovich (1998) proposed to quantify nonlinearity as

{
max sup Nu − Lupperu ,sup Nu − Llower u
u∈U u∈U
} (13)
where Lupper and Llower are linear operators such that they provide the smallest bounding envelop
on the output of N as ( Lloweru ) ( t ) ≤ ( Nu ) ( t ) ≤ ( Lupper u ) ( t ) for u ∈ U . This approach has many
similarities to identification for robust control (Helmicki et al., 1991, 1992).
To better assess the need for nonlinear control, as opposed to just assessing the distance
of a nonlinear plant from a linear one, Stack and Doyle (1997) proposed to focus on the
nonlinearity magnitude of an optimal nonlinear controller designed for a nonlinear process.
Quantification of the nonlinearity of that controller, using any method, was proposed as a
measure of the need for nonlinear control, the assumption being that a highly nonlinear controller
would result in a highly nonlinear closed loop, hence rendering linear control inadequate and
necessitating nonlinear control. For static state feedback laws, these authors proposed to use
coherence analysis as the nonlinearity measure. (A related approach for static systems appeared
in Yana et al. (1994) and for time series in Nielsen and Madsen (2001)). Stack and Doyle (1999)
also applied coherence analysis to assess the closed-loop nonlinearity of a nonlinear system
controlled by linear IMC. Emphasis was placed on observing the effect of different IMC tunings
on closed-loop nonlinearity. An advantage of coherence analysis is that it entails trivial
computational load, and the analysis may be conducted using experimental data, without detailed
knowledge of a process model.
Departing from the notion of nonlinearity measures based on the distance (norm of
difference) between a nonlinear operator from a suitable linear operator, a number of
investigators took different pathways towards quantifying nonlinearity.
Guay et al. (1995) proposed to quantify the static nonlinearity of a system described by

7
eqn. (5) in terms of the local geometry of the steady-state locus, i.e., by considering the first and
second derivatives of the steady-state map 0 = f ( xs , us ) with respect to us. Guay (1996)
extended these results to quantification of dynamic nonlinearity.
Harris et al. (2000) proposed to quantify the nonlinearity of system as in eqn. (5) first by
approximating nonlinearities in the time domain by polynomials, and then by expressing the
solution as a functional infinite series in the Laplace-Borel domain. Their approach relies on
repeated application of the shuffle product (convolution). The nonlinearity of a system is large if
“many” and “large” higher-order terms are needed. Bounds of an accordingly defined
nonlinearity measure can be easily computed and used in controller design.
Hahn and Edgar (2001) compared the controllability and observability covariance
matrices of a system that is linearized at its steady state operating point to covariance matrices
that are computed from data collected within an operating region of the nonlinear process. This
comparison results in two measures for the nonlinearity of the input-output behavior of a
process: input-to-state (controllability) and state-to-output (observability). In addition to
identifying when a model is severely nonlinear, this approach can identify a model as being
Wiener-like, Hammerstein-like, or both.
Choudhury (2001) used higher-order statistics on feedback error to detect closed-loop
nonlinearities based on operating data. The approach is based on detecting nonlinearities in
time-series, a topic with rich past (Kantz and Schreiber, 1997; Tong, 1990).
Rajput (2001) used nonlinearity measures for process monitoring and fault diagnosis.
The premise of the majority of the above approaches is that if a nonlinear process (or the
optimal nonlinear controller) is “close” to a linear one, then a linear controller will be sufficient.
While that may frequently be true, proximity of a nonlinear process to a linear one is neither
necessary nor sufficient for good closed-loop performance. For example, Nikolaou and
Hanagandi (1998) have shown that a highly nonlinear process controlled by linear IMC may
result in an almost linear closed loop, if IMC is suitably designed. Conversely, Schrama (1992)
has shown that, even for a linear process, a controller design based on a linear model with close
proximity to a process may result even in closed-loop instability. The discussion in section 2
corroborates the preceding point. Nevertheless, one would intuitively expect that there must be
process- and controller-dependent connections between open- and closed-loop nonlinearity.
That intuition is indeed correct, as shown next.

3.2 Closed-loop nonlinearity depends on both the controlled process and the controller
disturbance, d
setpoint, r ε input, u output, y
- Q N
noise, w

L -
Controller

Figure 6 – Block diagram of IMC for a Nonlinear Process N.


Consider the IMC loop structure of Figure 6. The operator N corresponding to the stable-over-
set controlled plant is nonlinear, the (linear or nonlinear) operator L corresponds to the plant

8
model, and the (linear or nonlinear) operator Q is the Youla parameter of the controller. Note
that Q may have a closed form or may be defined implicitly, e.g., via on-line optimization, as is
the case for MPC (see section 5.1). Note also that this structure allows the development of
small-gain type of theorems (e.g., eqns. (17) and (18); Zheng and Zafiriou, 1999; Koung and
MacGregor, 1992; Kothare and Morari, 1999) that are much less conservative than similar
theorems for classical feedback structures (Willems, 1971, p. 108; deFigueiredo and Chen, 1993,
p. 96; van der Schaft, 2000, p. 11). Indeed, small-gain theorems on classical feedback structures
fail to capture important classes of stabilizing controllers, such as controllers with integral action.
If the model L and Youla parameter Q are linear, i.e. the classical feedback controller
C = Q( I − LQ ) −1 is linear, then one can compare the resulting closed loop to the ideal linear
closed loop that would result if the controlled plant were linear and equal to the model L. The
difference between the actual nonlinear and the ideal linear closed loop can be quantified by the
magnitude of the operator
∆N =ˆ NQ ( I + NQ − LQ ) − LQ .
−1
(14)
Eker and Nikolaou (2001) have proposed to use the incremental norm over a set (local Lipschitz
constant) to quantify W ∆N (where W is a weighting filter) and have shown that if
γ =ˆ ( N − L)Q ∆E < 1 (15)
then the closed-loop nonlinearity measure W ∆N ∆Z
is bounded as
W ( I − LQ)( N − L)Q ∆E
W ( I − LQ)( N − L)Q ∆E

1 + ( N − L)Q ∆E
I + ( N − L )Q ∆E

W ( I − LQ )( N − L)Q ( N − L)Q ∆E
≤ W ∆N ∆ Z ≤ ≤ W ( I − LQ )
∆E
(16)
1424 3 1 − ( N − L)Q ∆E 14 4244 3 1 − ( N − L)Q
Closed-loop α 1442443 ∆ E

γ
nonlinearity
1−γ
where the set E contains the signals ε (Figure 6) and the set Z = [ I + ( N − L)Q ]( E ) .
Note that the bounds of the closed-loop nonlinearity depend on both the process
nonlinearity, N − L , and the controller elements Q and L. Note also that if Q can be designed in
a way that makes γ<< 1 , then 1−γγ << 1 , and the closed loop is virtually linear (e.g., Figure 3).
The inequality of eqn. (16), predicts the results of Example 1 in a non-conservative way.
For a thorough justification of the use of incremental norms in the above results as well as for
discussion on the significance, graphical interpretation, and implication for controller design of
the above inequality (16) see Eker and Nikolaou (2001).

3.3 Two crucial elements for assessment of closed-loop nonlinearity


There are two crucial elements associated with the above results of section 3.2:
(a) Eqn. (16) is based on the following general result (Eker and Nikolaou, 2001): If
R ∆U < 1 (17)
−1
then the inverse ( I + R ) :Y →U of the nonlinear operator I + R : U → Y exists on
Y = [ I + R ](U ) , and is bounded as

9
1
( I + R ) −1 ≤ (18)
∆Y 1− R ∆U

(b) Incremental norms can be computed using the approximation (Eker and Nikolaou, 2000)
M ∆V (= sup L ) ≈
u0 ∈V
u0 sup
u0 ∈V
u0 constant
L'u0 (19)

where the operator L'u0 appearing in the right-hand side of eqn. (19) is the linearization of the
nonlinear operator M (eqn. (9)) around steady states (i.e. constant) u0 in the set V.
The importance of element (a) is that the existence and boundedness of the inverse of a
nonlinear operator are guaranteed over a set without explicit computation of that inverse. No
Banach space setting is needed either (Willems, 1971, p. 29). Note also the importance of the
sets U and V that are associated with ranges of input signals.
The importance of element (b) is that the associated computations, albeit approximate, do
not require a nonlinear dynamic process model but rather its linearization around a number of
steady states, a fact that significantly eases modeling efforts.
The work proposed in the following section 5 will rely heavily on the above two elements.

4 The Multivariable Case


An important feature of the framework captured by eqns. (17), (18), and (19) is that it can be
easily applied to multivariable systems. Shown next are three examples that demonstrate that the
predictions of the theoretical framework for the multivariable case can range from conservative
to tight.

4.1 Example 3 – Two nonisothermal CSTRs in series


Consider again the system of two -5
x 10
CSTRs studied in Example 1, with 0

coolant flow rate and feed flow rate


concentration

to reactor 1 as process inputs, and -2

concentration and temperature of


the second reactor as process -4

outputs.
-6
0 10 20 30 40 50 60 70 80 90 100
Figure 7 – Setpoint step-
-3 time
change responses of (a) x 10
20
nonlinear closed loop with
15
linear IMC controller (solid
temperature

lines) and (b) ideal linear 10

closed loop (dashed lines). 5


Solid and dashed lines are 0
virtually indistinguishable. -5
0 10 20 30 40 50 60 70 80 90 100
Linear IMC, for certain setpoint time
changes and for certain tuning, may
produce a closed loop that is virtually linear, as Figure 7 indicates. However, for different
setpoint changes, the same linear controller generates instabilities, as Figure 8 shows. Figure 7

10
-3
x 10
15
and Figure 8 correspond to setpoint
changes for which eqn. (15) is or is 10

concentration
not satisfied, respectively.
5
Figure 8 – Similar to
0
Figure 7 for different
setpoint changes. Notice -5
0 20 40 60 80 100 120 140 160 180 200
the discrepancies between time
solid and dashed lines. 0

Figure 9 shows the values of γ -5

temperature
computed for various setpoint
-10
changes in temperature and
concentration. The stability -15
domain indicated by Figure 9 is
-20
conservative; setpoint changes of 0 20 40 60 80 100 120 140 160 180 200
time
much higher magnitude than
indicated by Figure 9 had to be
effected for instabilities to be noticeable. Note that
deviation variables are used throughout.
Figure 9 – Contour plot of the value of γ,
eqn. (15), as a function of temperature and
concentration setpoint changes.
It is also possible to work in the frequency domain
in order to gain qualitative insight into the behavior
of the closed loop, although the results will be
quantitatively conservative for the multivariable
case. Nevertheless, the following analysis may be
helpful in understanding the effect of controller design and operating range on closed-loop
stability and performance. Similar analysis may be performed for the subsequent Example 4 and
Example 5, but is omitted for brevity.
Eqn. (15) is satisfied over an operating range E if
1
( Lε i − L) L−1 < for all ε i ∈ E (20)
F
where the IMC filter F is designed as
 1 
 (λ s + 1) 2 0 
F=   (21)
 1 
 0
 (λ s + 1) 2 
While eqn. (20) is conservative, it has a nice graphical interpretation: The magnitude Bode plot
of the the the reciprocal of the maximum singular value of F ( jω) should be above the
maximum singular value of ( Lεi ( jω ) − L( j ω) ) L−1 ( jω ) at all frequencies and for all ε i in the set
E. Figure 10 shows how inequality (20) may or may not be satisfied. Figure 10 is reminiscent of

11
10

similar plots that have been developed in linear control 10

theory (Morari and Zafiriou, 1989) and provides a link λ=10

for linear controller design for nonlinear systems. 5


10 λ=5

Figure 10 – ( Lε i − L) L−1 (dashed lines for λ=1

0
10
1
different operating points) and (solid lines
F -5

for different values of the filter time constant λ,


10

eqn. (21)) as a function of frequency.


-10
10
-4 -2 0 2 4
10 10 10 10 10

4.2 Example 4 – A fictitious problem


Consider the system described by the differential equations
x&1 = −2 x1 − 4 x1 x22 + u1
(22)
x& 2 = −4 x12 x2 − 2 x2 + u2
where u1 , u2 are the inputs and y1 , y2 are
the outputs. Figure 11 shows closed- 0.4

loop responses for setpoint step 0.3


changes around the equilibrium point
x1

0.2
(0,0). Note that the closed loop is only
mildly nonlinear, if at all. 0.1

Figure 11 – Setpoint step- 0


0 100 200 300 400 500 600 700 800
change responses of (a) time

nonlinear closed loop with 0.4

linear IMC controller (solid 0.3


lines) and (b) ideal linear
x2

0.2
closed loop (dashed lines).
Solid and dashed lines are fairly 0.1
close. 0
0 100 200 300 400 500 600 700 800
However, for different setpoint time
changes, the instabilities of Figure 12
emerge.
1.5

Figure 12 – Similar to Figure


1
11 for different setpoint
x1

changes. Notice the 0.5


discrepancies between solid and
dashed lines. 0
0 100 200 300 400 500 600 700 800
time
Such instabilities can be fairly well
1.5
predicted by the contour plot of Figure
13, which depicts the value of log γ as
1
a function of setpoint changes on x 1 and
x2

x2 . Note the difference between Figure 0.5

0
0 100 200 300 400 500 600 700 12
800
time
13 and Figure 11 regarding the regions of setpoint
changes for which the closed-loop stability
condition, eqn. (15), is satisfied: In Figure 13 that
region appears to be open, while it is closed in
Figure 11. Note also that the stability region
predicted by Figure 13 is much less conservative
than that predicted by Figure 11.
Figure 13 – Contour plot of the value of
logγ, eqn. (15), as a function of x1 and x2
setpoint changes.

4.3 Example 5 – Nonisothermal CSTR


Consider the CSTR studied in Nikolaou and Hanagandi (1988), with feed flow rate and heat
removal rate as inputs and concentration and
temperature as process outputs. Figure 14 and 0

Figure 15 are the counterparts of Figure 11


concentration
-50
and Figure 12.
-100
Figure 14 – Setpoint step-change
responses of (a) nonlinear closed loop -150
0 50 100 150
with linear IMC controller (solid lines) time

and (b) ideal linear closed loop 0

-1
(dashed lines). Solid and dashed lines
temperature

-2
are virtually indistinguishable.
-3

-4

-5
0 50 100 150
time

6000
Figure 15 – Similar to Figure 14 for 4000
concentration

different setpoint changes (-200


mol/m 3 , -30° K).
2000
Notice the
0
discrepancies between solid and
-2000
dashed lines. 0 10 20 30 40 50 60 70 80 90
time

The behavior of observed in Figure 14 and 600

Figure 15 can be explained by the contour plot 400


temperature

of Figure 16. Note that predictions of the 200

stability region are less conservative in this 0

Example than in Example 3. -200


0 10 20 30 40 50 60 70 80 90
time

13
Figure 16 – Contour plot of the value
of logγ, eqn. (15), as a function of x1
and x2 setpoint changes.

5 Future Directions
The value of the approach discussed in the
preceding sections is that it is valid for general
classes of nonlinear operators, and relies on
only two crucial elements, presented in section
3.3. The results discussed above can be
further refined along a number of promising
directions. We discuss below two such directions addressing two interrelated classes of systems,
namely constrained MPC and control of unstable nonlinear processes. Issues particular to each
of these classes as well as issues common to both classes will be indicated.

5.1 Constrained MPC: Linear or nonlinear model?

5.1.1 Problem formulation


MPC is currently enjoying enormous popularity in both academia and industry, as the workhorse
of advanced process control (Nikolaou, 2001b). While static nonlinear elements (e.g., nonlinear
valves, inferential sensors, etc.) are included in several industrial MPC implementations, the vast
majority of them employ a linear dynamic model of the controlled process. Academic research
has produced a substantial amount of work on MPC with nonlinear model (Allgöwer and Zheng,
2000). Despite their obvious advantages in accuracy, nonlinear models have the disadvantages
of (a) necessitating on-line nonlinear optimization that is frequently non-convex, and (b)
seriously complicating model development. Would there be any advantages in using a (partly or
totally) nonlinear model in MPC for a given process or would a linear model be sufficient?
We propose to use the framework discussed in section 3 to address the above question.
We demonstrate next that this framework can serve as basis for constrained nonlinear MPC
analysis and synthesis.

5.1.2 Preliminary analysis


The starting point of the research proposed in this section is the realization that an MPC system
can be cast in the form of Figure 6. To see that, consider the prototypical MPC on-line
optimization
 p m−1

∑ ( ) ∑
2
min  r − y( k + i | k ) + R ∆u ( k + i | k )2  (23)
u (k | k ),..., u ( k+ m−1|k )
 i =1 i =0 
subject to
y min ≤ y (k +i | k ) ≤ y max (24)
umin ≤ u( k +i | k ) ≤ umax , ∆umin ≤ ∆u ( k + i | k ) ≤ ∆umax (25)
where a linear finite-impulse-response (FIR) model is used to predict future outputs as

14
64
4744
ymodel
8
n n
y ( k + i k ) = ∑ h j u( k + i − j k ) + y ( k ) − ∑ h j u( k − j ) (26)
j =1
14442444 j =1
3
d (k k )

After substitution of eqn. (26) into eqns. (23) and (24), the above constrained MPC algorithm
corresponds to the IMC structure of Figure 6, with the operator Q : ε a u realized by
 p  
2
m −1 
2
n
min
u (k | k ),..., u ( k+ m −1|k )
 ∑  ε( k ) + ∑ h j u( k + i − j k )  + R ∑ ∆u ( k + i | k )  (27)
 i =1  j =1  i =0 
subject to eqn. (25) and
n
r − ymin ≥ ε( k ) + ∑ h j u( k + i − j k ) ≥ r − ymax (28)
j =1

where (Figure 17) ε =ˆ r − [ y ( k ) − ymodel ( k )] , and the operator L : u a y model realized by


n
y model ( k ) = ∑ h j u( k − j ) (29)
j =1

d
r ε QL Eqns. (27), (28), (25) u
N
y
-

ymodel -
Controller L Eqn. (29)

Figure 17 – Closed-loop MPC block diagram in IMC form, for a nonlinear process N.
The corresponding closed-loop process output is
y LinearModel = d + NQ L ( I + NQ L − LQ L ) −1 ( r − d ) (30)
Now, if MPC employs a nonlinear model to predict future process outputs as
6444 474444
ymodel
8
y ( k + i k ) = f ( u( k + i − 1 k ),..., u ( k + i − n k ) ) + 144444244444
y ( k ) − f (u ( k − 1),..., u( k − n3
)) (31)
d (k k )

then a structure analogous to that of Figure 17 is trivially obtained, and the closed-loop process
output, assuming no process/model mismatch, is
y NonlinearModel = d + NQN (r − d ) (32)
Therefore, eqns. (32) and (30) yield
y LinearModel − y NonlinearModel = [ NQL − NQN [ I + ( N − L)QL ]] [I + ( N − L)QL ]−1 ( r − d ) =
(33)
=ˆ ∆N MPC ( r − d )
To compare MPC with linear model and MPC with nonlinear model, we can use an
approach similar to that in section 3.2 and the two crucial elements of section 3.3 to show and
perform computations on the counterpart of eqn. (16), i.e.

15
NQL − NQN [ I + ( N − L)QL ] ∆E NQL − NQN [ I + ( N − L)QL ] ∆E
≤ ∆N MPC ≤ (34)
I + ( N − L)QL ∆E
∆Z
1 − ( N − L)QL ∆E

provided that ( N − L )QL ∆E < 1 . If a satisfactory controller QL can be designed that makes the
bounds in eqn. (34) “small enough”, then MPC with nonlinear model is not necessary.

5.1.3 MPC stability revisited


For the analysis presented in the preceding section 5.1.2 to be valid, closed-loop stability must be
ensured. For that, it is necessary and sufficient that both QL : ε a u and QN : ε a u be input-
output stable over corresponding sets, i.e. there exists a constant K > 0 such that
u ≤K ε . (35)
Constrained MPC stability has been amply studied in literature (Nikolaou, 2001b) after
the seminal paper of Rawlings and Muske (1993). Virtually all analyses take a Lyapunov
stability approach. The main result of this approach can be summarized as “feasibility of the on-
line optimization guarantees closed-loop stability” (Mayne et al., 2000). While the Lyapunov
stability approach is an extremely important advancement in our understanding of MPC
behavior, an input-output stability approach (Willems, 1971, p. 101) could also provide
fundamental insight, by determining sets of signals (e.g., disturbances, setpoint changes, etc.)
over which stability is maintained.

5.1.4 Issues to address and potential paths of development


- Establishing input/output stability over set for QL and QN, is not trivial, because no closed-
form expressions exist for either QL or QN. Any computational approach attempting explicit
enumeration of all controllers resulting from corresponding sets of constraints being active
would be hopeless for problems of industrial size. A more promising approach could be to
use the Karush-Kuhn-Tucker optimality conditions and attempt to find (preferably the
smallest) K that satisfies eqn. (35) using parametric programming ideas (a concept
successfully used by Bemporad et al. (2000) for small problems). The input-to-state stability
(ISS) concept (Sontag, 2000) may also be used to blend input/output and Lyapunov stability
over sets of inputs ε.
- Computation of incremental norms through eqn. (19) must account for the fact that QL and
QN, realized through on-line optimization, are almost everywhere linearizable, i.e. points
corresponding to changes in the set of active constraints have one-sided derivatives.
- Given that MPC is a multivariable control scheme, it is natural to ask what parts of the model
employed by MPC need to be nonlinear and what parts can be handled as linear. Blending of
physical process understanding (e.g., sparsity, monotonicity) with algorithms might prove
promising in addressing the computational complexity of the problem.
- Simulation tests on multivariable test problems should reveal the limits of the approach.

5.2 Unstable nonlinear processes: Over what range can they be stabilized by linear control?

5.2.1 Problem formulation


Quantification of nonlinearity and determination of whether linear control is adequate for an
unstable system is complicated by the fact that unstable systems produce unbounded signals and,
consequently, distances between unstable operators cannot be measured in terms of standard

16
norms, but require different metrics. For example, eqn. (10) yields an unbounded quantity. The
gap metric, introduced by Zames and El-Sakkary (1980) is a suitable metric that relies on
fractional representation of operators (Vidyasager, 1985) and quantifies distances of operators in
terms of distances of fraction components. We demonstrate next that the fractional
representation approach fits naturally within the framework of section 3, and can be used to
assess when linear control is adequate for an unstable nonlinear process.

5.2.2 Preliminary analysis


To stabilize a nonlinear process by a linear controller C, one can start with the standard Youla
parametrization of all linear stabilizing controllers for a linear plant L and examine whether this
controller stabilizes the actual nonlinear plant. In addition, the operating domain over which
stabilization is possible should also be examined.
To see how the basic results of section 3 can be used in this analysis, parametrize all
linear controllers C that stabilize L as C = (Y r − QAl ) −1 ( X r + QBl ) = ( X l + B rQ )(Yl − Ar Q) −1 . The
Youla parameter Q must be a proper and stable operator and the proper stable linear operators Ar,
Al, Br, Bl, Xr, Al, Yr, Yl, for which construction formulas are well known (Vidyasagar, 1985) ,
constitute a doubly coprime factorization of L as L = Ar Br−1 = Bl− 1 Al .

Figure 18 – Block diagram of linear v


control for a nonlinear process N. r e q u y
Because N is unstable, input disturbances - C N
v must be explicitly considered to ensure
closed-loop stability.
When the above controller C is used with the plant N as in Figure 18, the closed-loop operators
from the external signals r and v to the signals u and y (needed to assess closed-loop stability and
performance) can be explicitly expressed as
u = ( I + CN ) −1 ( v + Cr ) , y = N ( I + CN )− 1( v + Cr ) (36)
Note that because N is nonlinear, the effects of r and v on u and y in the right-hand side of eqn.
(36) cannot be separated.
To proceed with the analysis, an obvious first step is to cons ider the cases v = 0 and
r = 0 separately, and establish closed-loop stability and performance on the basis of non-singular
perturbation arguments. Thus, on the basis of linearity of the controller C, it can be shown after
a series of manipulations that,
v = 0 ⇒ y = ArN D −1( X r + QBl ) r and u = BrN D −1 ( X r + QBl )r (37)
r = 0 ⇒ y = ArN D −1 ( Yr − QAl ) v and u = BrN D −1 ( Yr − QAl )v (38)
where N = ArN ( BrN ) = ( BlN ) AlN is a doubly coprime factorization of the nonlinear operator N
−1 −1

(van der Schaft, 2000, note 9, p. 161; Ball and Verma, 1994; Baramov and Kimura, 1997) and
D =ˆ (Yr − QAl ) BrN + ( Xr + QBl ) ArN (39)
Eqns. (37) and (38) can be used in closed-loop stability and performance analysis, as follows.
Stability. The stability of the closed loop depends on the existence and stability of D −1
over a corresponding set. According to element (a) in section 3.3, if
γ =ˆ D − I ∆E < 1 (40)

17
1
then D −1 exists, is stable, and D −1 ≤ . Element (b) in section 3.3 can be used in all
∆Z 1− γ
computations.
Performance. If the controlled plant were equal to L, then a series of manipulations can
show that it would be
−1
DLinearPlant = DLinearPlant =I (41)
and that the resulting ideal linear closed loop would be
yLinearPlant = Ar (Yr −QAl )v + Ar ( X r + QBl )r , uLinearPlant = Br (Yr −QAl )v + B r ( X r + QBl )r (42)
Comparison of eqns. (42) with eqns. (37) and (38) provides a measure of closed-loop
nonlinearity for unstable nonlinear plants. For example, when v = 0 , we have
y − y LinearPlant = ( ArN − Ar D) D −1 ( X r + RBl ) r =ˆ ∆Mr (43)
and if eqn. (40) holds, then we can show that
1
W ∆M ∆Z ≤ W ( ArN − Ar D) X r + QBl (44)
∆R 1 − γ

which is the counterpart of eqn. (16) for unstable nonlinear systems. Again, as in section 5.1.2,
if a linear controller Q can be designed that makes the bound in eqn. (44) “small enough”, then
nonlinear control is not necessary.

5.2.3 Issues to address and potential paths of development


- Eqns. (40) and (44) indicate that linear control will have to render W ( ArN − Ar D) and γ
∆R

“small”. Design limitations stemming from this fact should be investigated both in terms of
general rigorous results development and simulation tests.
Computing a nonlinear coprime factorization N = ArN ( BrN ) = ( BlN ) AlN may be feasible
−1 −1
-
but cumbersome (van der Schaft, 2000). However, use of element (b) (p. 10) can greatly
simplify computations by requiring only a series of linear coprime factorization of N (cf.
element (b), p. 10).
- The relationship between the sets R and Z in eqn. (44) will have to be better quantified, in
order to determine sets over which stability and performance are ensured (cf. Kapoor and
Daoutidis, 1997; Scibile and Kouvaritakis, 2000; Cantoni, 1999; Zheng and Morari, 1995).
- Simulations should test and compare the approach on unstable systems that have been
studied in literature (e.g., Downs and Vogel, 1993).

5.3 Issues common to both directions of research


- How conservative is the approximation in eqn. (19) for the computation of the incremental
norm of a nonlinear operator?
- Nikolaou and Manousiouthakis (1989) have shown how the incremental norm can be
computed through the exact equality in eqn. (19), using nonsmooth optimal control. Their
approach is rigorous but cumbersome for large problems. Could it be made more efficient?
- How conservative are the inequalities (17) and (18) in the various instances of element (a) (p.
10) such as eqns. (15) and (16), (34), (40) and (44)?
- How can the proposed approach be generalized when disturbances are not additive?
- Is it more efficient (in terms of experimental data needed) to linearize a nonlinear model or to
develop a series of empirical linear models in eqn. (19)?

18
- Could steady-state information obtainable from commercial process simulators be combined
with monotonicity arguments to make the proposed approach easier to apply (see Vinson and
Georgakis (2000) for related work)?

6 Conclusions
In this paper we addressed the issue of linear controller design for nonlinear plants. We gave a
brief exposition of a general theory that quantifies the interplay between nonlinearity and
feedback control. The main ingredients of this theory, summarized in section 3.3, are the indices
γ
γ, eqn. (15), and β =ˆ , eqn. (16). The theory presented is important for controller analysis
1− γ
and, more importantly, efficient controller synthesis. A number of potential extensions and
refinements were suggested. We believe that the proposed framework can be widely applicable,
especially if it is improved according to theoretical possibilities and industrial needs. As
discussed in Sidebar 1 (p. 2), the need for control of nonlinear systems exists both in mature
high-volume industries (for which the tendency towards wider integration of operations
inevitably leads to encompassing nonlinearities) as well as in emerging industries (for which
control may be essential for feasible rather than efficient process operation). However, control
systems for nonlinear processes do not have to be overly complex. Moreover, if such systems
have to be internally complex, the complexity of the design (e.g. translation of qualitative
engineering requirements to design parameter specifications), operation, and maintenance of
such systems by process engineers and operators should be low, to ensure successful
implementation (Birchfield, 1997; Downs, 2001). Our vision is to develop related tools that are
based on rigorous concepts but are not more complicated than necessary, and delegate
cumbersome computations to the computer while allowing the designer to concentrate on
important design concepts.

Acknowledgement – Partial financial support from the National Science Foundation (Grant CTS
– 9896231) is gratefully acknowledged.

7 References Cited
[1] Allen, D. T., and D. R. Shonnard, Green Engineering: Environmentally Conscious
Design of Chemical Processes, Prentice-Hall (2001).
[2] Allgöwer ,F., “Definition and computation of a nonlinearity measure”. Proceedings of
Nonlinear Control System Design Symposium. NOLCOS95, Pergamon, Oxford, UK,
257-262 (1996).
[3] Allgöwer, F., and A. Zheng (Eds.), Nonlinear Model Predictive Conrol, Birkhäuser
(2000).
[4] Ball J.A., and M. Verma, “Factorization and feedback stabilization for nonlinear
systems”, Systems & Control Letters, 23, 187-196 (1994).
[5] Baramov, L. and H. Kimura “Nonlinear coprime factorizations and parameterization of a
class of stabilizing controllers”, International Journal of Control, 66, 413-434 (1997).
[6] Bemporad, A., M. Morari, V. Dua, and E. N. Pistikopoulos, “The Explicit Linear
Quadratic Regulator for Constrained Systems”, American Control Conference, Chicago,
pp. 872–876 (2000).
[7] Bequette, B. W., “Nonlinear Control of Chemical Processes: A Review”, Ind. Eng.
Chem. Res., 30, 1391-1413 (1991).

19
[8] Birchfield, G. S., “Trends in Optimization and Advanced Process Control in the Refinery
Industry”, Chemical Process Control V, Tahoe City, CA (1997).
[9] Braatz, R., and S. Hasebe, “Particle size and shape control in crystallization processes”,
preprints of CPC VI (2001).
[10] Brittain, J. E., “Scanning the Past: Harold S. Black and the Negative Feedback
Amplifier,” Proceedings of the IEEE, Vol.85, No.8, 1335-1336 (1997).
[11] Brockett, R., “Characteristic Phenomena and Model Problems in Nonlinear Control”,
IFAC World Congress, 135-140 (1996).
[12] Cantoni, M., “Gap Metric Performance Bounds for Linear Feedback Systems”,
Proceedings of the 38th Conference on Decision & Control, Phoenix, Arizona (1999).
[13] Chiu, T., and P. Christofides, “Nonlinear Control of Particulate Processes”, AIChE J., 45,
6, 1279-1297 (1999).
[14] Choudhury, M. D. S., “Detection and Diagnosis of Valve Problems using Higher Order
Statistics”, paper 282f, AIChE Annual Meeting, Reno, NV (2001).
[15] Christofides, P. D., “Control of Nonlinear Distributed Process Systems: Recent
Developments and Challenges”, AIChE J., 47, 3, 514-518 (2001).
[16] Christofides, P. D., Nonlinear and Robust Control of PDE Systems: Methods and
Applications to Transport-Reaction Processes, Birkhäuser (2001).
[17] Daoutidis, P., and M. Henson, “Dynamics and Control of Cell Populations”, preprints of
CPC VI (2001).
[18] de Figueiredo, R. J. P., and G. R. Chen, Nonlinear feedback control systems : an operator
theory approach, Academic Press (1993).
[19] Desoer, C. A., and M. Vidyasagar, Feedback Systems: Input-Output Properties,
Academic Press (1975).
[20] Desoer, C. A., and Y.-T. Wang, “Foundations of Feedback Theory for Nonlinear
Dynamical Systems”, IEEE Trans. Circ. Syst., vol. CAS-27, 2, 104-123 (1980).
[21] Downs, J. J. and E. F. Vogel, “A plant-wide industrial control problem”, Comput. Chem.
Engng., 17, 245–255 (1993).
[22] Downs, J. J., “Linking Control Strategy Design and Model Predictive Control”, preprints
of CPC VI (2001).
[23] Edgar,T. F., W.J. Campbell, and C. Bode, “Model-Based Control in Microelectronics
Manufacturing,” Proceedings of the 38th Conference on Decision & Control Phoenix,
Arizona USA, 4185-4191 (1999).
[24] Eker, S. A., and M. Nikolaou, "Ensuring Co-primeness in Least-Squares Identification of
ARX Models: The SICLS algorithm", Automatica (submitted) (2000).
(http://athens.chee.uh.edu).
[25] Eker, S. A., and M. Nikolaou, "Linear Control of Nonlinear Systems - The Interplay
between Nonlinearity and Feedback", AIChE J., submitted (2001).
(http://athens.chee.uh.edu).
[26] Eker, S. A., and M. Nikolaou, "Simultaneous Model Predictive Control and
Identification: Closed-Loop Properties", Automatica (submitted) (2000).
(http://athens.chee.uh.edu).
[27] Eker, S. A., and M. Nikolaou, “Adaptive Control through On-line Optimization: The
MPCI Paradigm and Variants,” Appl. Math. And Comp. Sci., 9, 1, 101-128 (1999).
[28] Genceli, H., and M. Nikolaou, “New Approach to Constrained Predictive Control with
Simultaneous Model Identification,” AIChE J., 42, 10, 2857-2869, 1996.

20
[29] Guay, M., McLellan,J., Bacon, D.W. “Measurement of nonlinearity in chemical process
control systems: the steady state map”, Canadian Journal Chemical Engineering. 73, pp
868-882 (1995).
[30] Guay, M., Measurement of Nonlinearity in Chemical Process Control, Ph.D. Thesis,
Queen’s University, Kingston, ON (1996).
[31] Haber, R. and H. Unbehauen, “Structure Identification of Nonlinear Dynamic Systems -
A Survey on Input/Output Approaches”, Automatica, 26, 4, 651-677 (1990).
[32] Hahn, J., and T. F. Edgar, “Nonlinearity Quantification and Model Classification using
Gramians and other Covariance Matrices”, paper 286d, AIChE Annual Meeting, Reno,
NV (2001).
[33] Harris K. R., Colantonio M. C., Palazoglu A., “On the computation of a nonlinearity
measure using functional expansions”, Chem. Eng. Sci, 55 (13) 2393-2400 (2000).
[34] Harris, T. J., and C. T. Seppala, “Diagnosis and Analysis of Control Systems”, preprints
of CPC VI (2001).
[35] Helbig, A., W. Marquardt, and F. Allgöwer, “Nonlinearity measures: definition,
computation and applications”, Journal of Process Control.10, 113-123 (2000).
[36] Helmicki, A., C. Jacobson, and C. Nett, “Identification in H8 ”, IEEE Trans. Auto. Contr,
1163 (1991).
[37] Helmicki, A., C. Jacobson, and C. Nett, “Identification in H8 ”, IEEE Trans. Auto. Contr,
604 (1992).
[38] Henson, M.A, Seborg, D.E, “Input-Output Linearization of General Nonlinear Process”,
AIChE Journal. 36,11, pp 1753-1757 (1990).
[39] Henson, M.A., and D.E. Seborg (eds.), Nonlinear Process Control, Prentice Hall, Upper
Saddle River, NJ (1997).
[40] Isidori, A., Nonlinear Control Systems, Springer, NY (1999).
[41] Kantz, H., and T. Schreiber, Nonlinear time series analysis, Cambridge University Press,
Cambridge (1997).
[42] Kapoor N, and P. Daoutidis, “Stabilization of nonlinear processes with input constraints”,
Comp. Chem. Eng., 24 (1): 9-21 (2000).
[43] Kayihan, F., “A Review of Modeling and Control in the Pulp and Paper Industries”, in
Fifth International Conference on Chemical Process Control, (J. C. Kantor, C. E. García,
B. Carnahan, Eds.) 117-132 (1997).
[44] Khalil, H. K., Nonlinear Systems, MacMillan (1992).
[45] Kothare M. V., and M. Morari, “Multiplier theory for stability analysis of anti-windup
control systems”, Automatica, 35: (5) 917-928 (1999).
[46] Koung, C.-W., and J. MacGregor, “Robustness of Multivariable Linear Controllers to
Process Nonlinearities”, Ind. Eng. Chem. Res., 31, 1085-1096 (1992).
[47] Krstic, M., I. Kanellakopoulos, and P. Kokotovic, Nonlinear and Adaptive Control
Design; John Wiley & Sons, Inc.: New York (1995).
[48] Kumar, A., and P. Daoutidis, Control of nonlinear differential-algebraic-equation
systems with applications to chemical processes, Chapman & Hall/CRC Research Notes
in Mathematics Series (1999).
[49] Mayne D.Q., J. B. Rawlings, C. V. Rao, P. O. M. Scokaert, “Constrained model
predictive control: Stability and optimality”, Automatica, 36 (6): 789-814 (2000).
[50] Mead, N., MPC vs. PID Control Strategy Comparison for a Process with Inherent
Feedback Characteristics, Honors Thesis, Chemical Engineering Department, University

21
of Houston (2000).
[51] Meadows, E. S., M.A. Henson, J.W. Eaton, and J. B. Rawlings, “Receding Horizon
Control and Discontinuous State Feedback Stabilization”, Int. J. Control, 62 (5), 1217-
1229 (1995).
[52] Misra, P., and M. Nikolaou, “Input Design for Model-Order Determination in Subspace
Identification”, AIChE J., submitted (2001). (http://athens.chee.uh.edu).
[53] Misra, P., and M. Nikolaou, “Linear Control of Multivariable Nonlinear Processes”,
manuscript in preparation based on paper 252g, AIChE Annual Meeting (2001).
[54] Morari, M., and E. Zafiriou, Robust Process Control, Prentice-Hall (1989).
[55] Nielsen, H. Aa., and H. Madsen, “A generalization of some classical time series tools”,
Computational Statistics and Data Analysis, 37, 13-31 (2001).
[56] Nijmeijer, H., and A. J. van der Schaft, Nonlinear dynamical control systems, Springer-
Verlag (1990).
[57] Nikolaou, M. and V. Manousiouthakis, “A Hybrid Approach to Nonlinear System
Stability and Performance”, AIChE Journal, 35, 4, 559-572 (1989).
[58] Nikolaou, M., “Computer-Aided Process Engineering in the Snack Food Industry”, , in
Fifth International Conference on Chemical Process Control, (J. C. Kantor, C. E. García,
B. Carnahan, Eds.) 61-70 (1997).
[59] Nikolaou, M., “Identification and Adaptive Control,” Comp. Chem. Eng., special issue on
NSF/NIST Vision 2020 Workshop, 23/2, 215 - 225 (1998).
[60] Nikolaou, M., “Linear Control of Nonlinear Processes: Recent Developments and Future
Directions”, CEPAC 2001, Guarujá, São Paulo, Brazil,
http://pqi.ep.usp.br/eventos/cepac2001.htm (2001).
[61] Nikolaou, M., “When is Nonlinear Dynamic modeling necessary?”, Proceedings of the
American Control Conference, June 2-4, San Francisco, CA, 1067-1071 (1993).
[62] Nikolaou, M., and V. Hanagandi, “Nonlinearity Quantification and its Application to
Nonlinear System Identification”, Chem. Eng. Comm., 166, 1-33 (1998).
[63] Nikolaou, M., Model Predictive Controllers: A Critical Synthesis of Theory and
Industrial Needs, Advances in Chemical Engineering Series, Academic Press (2001).
[64] Pavilion Technologies, http://www.pavtech.com/web/content.nsf/ISChemicals (2001).
[65] Pearson, R., Discrete-Time Dynamic Models, Oxford (1999).
[66] Qin, S. J., and T. A. Badgewell, “An Overview of Industrial Model Predictive Control
Technology”, Chemical Processes Control V, AIChE Symposium Series, No. 93, 232-
256 (1997).
[67] Qin, S.J. and T.A. Badgwell, “An overview of nonlinear model predictive control
applications”, in Nonlinear Model Predictive Control, edited by F. Allgöwer and A.
Zheng, Birkhauser, SWITZERLAND (2000).
[68] Rafal, M. D., and W. F. Stevens, “Discrete Dynamic Optimization Applied to On-Line
Optimal Control”, AIChE J., 14, 1, 85-91 (1968).
[69] Rajput, S., S. Daw, and D. D. Bruns, “Non-linear Measure Based Process Monitoring and
Fault Diagnosis”, paper 275i, AIChE Annual Meeting, Reno, NV (2001).
[70] Rawlings, J. B., and K. R. Muske, “The stability of constrained receding horizon
control”, IEEE Trans. AC, AC-38, 1512-1516 (1993).
[71] Rawlings, J. B., E. S. Meadows, and K. R. Muske, “Nonlinear Model Predictive Control:
A Tutorial and Survey”, Proceedings of ADCHEM ’94, 203-214, Kyoto, Japan (1994).
[72] Schrama, R., Approximate Identification and Control Design, Ph.D. Thesis, Delft

22
University of Technology, Faculty of Mechanical Engineering (1992).
[73] Schwarm, A. T., S. A. Eker, and M. Nikolaou, “Model Predictive Control and
Identification: A New Approach to Closed-Loop Identification and Adaptive Control,” in
Pekny, J. F., and G. E. Blau, editors, Third International Conference on Foundations of
Computer Aided Process Operations, volume 94 in AIChE Symposium Series 320, 303
(1997).
[74] Schwarm, A., and M. Nikolaou, “Chance Constrained Model Predictive Control,” AIChE
Journal, 45, 8, 1605-1844 (1998).
[75] Scibile, L., Kouvaritakis, B., “Stability region for a class of open-loop unstable linear
systems: Theory and application”, Automatica, Vol. 36, 1, pp. 37-44 (2000).
[76] Slotine, J.-J. E., and W. Li, Applied nonlinear control, Prentice Hall (1991).
[77] Sontag, E. D., “The ISS philosophy as a unifying frame- for stability-like behavior”, in
Nonlinear Control in the Year 2000 (Volume 2) (Lecture Notes in Control and
Information Sciences, A. Isidori, F. Lamnabhi-Lagarrigue, and W. Respondek, eds.),
Springer-Verlag, Berlin, 443-468 (2000).
[78] Stack, A.J., Doyle, F.J., “Local nonlinear performance assessment for single-controller
design”. IFAC World Congress (1999).
[79] Stack, A.J., Doyle, F.J., The optimal control structure: an approach to measuring control-
law nonlinearity. Computers Chem. Engng. 21, 9, pp1009-1019 (1997).
[80] Sun, D., and K. Kosanovich, “Nonlinearity measures for a class of SISO nonlinear
systems”, Proceedings of American Control Conference, 2544-2549 (1998).
[81] Tong, H., Nonlinear Time Series: A Dynamical Systems Approach, Oxford University
Press, Oxford (1990).
[82] University of Houston - Fact Book 1999-2000,
http://www.uh.edu/oppa/factbook99/sashtm1.5A.htm#engin.
[83] Van der Schaft, A. J., L2-gain and passivity techniques in nonlinear control, Springer
(2000).
[84] Vidyasagar, M., Control system synthesis: a factorization approach, MIT Press,
Cambridge, MA (1985).
[85] Vidyasagar, M., Nonlinear systems analysis, Prentice Hall (1993).
[86] Vinson D. R., and C. Georgakis, “A new measure of process output controllability”, J. of
Process Control, 10 (2-3): 185-194 (2000).
[87] Willems, J., The Analysis of Feedback Systems, MIT Press (1971).
[88] Yana, K., H. Yoshida, and M. Komai, “Measurement and Estimation of System
Nonlinearity via a Neural Network”, Electronics and Communications in Japan, part 3,
vol. 77, no. 2, 35-43 (1994).
[89] Young, R. E., “Evolution of an Industrial Nonlinear Model Predictive Controller”,
preprints of CPC VI (2001).
[90] Zames, G., and A.K. El-Sakkary, “Unstable systems and feedback: The gap metric,”
Proc. Allerton Conference, 53-59 (1980). (Introduction of the gap metric to control.)
[91] Zeidler, E., Nonlinear Functional Analysis, Springer (1990).
[92] Zheng Q. S., and E. Zafiriou, “A local form of small gain theorem and analysis of
feedback Volterra systems”, IEEE Trans. AC, 44: (3) 635-640 (1999).
[93] Zheng, A., and M. Morari, “Control of Linear Unstable Systems with Constraints”,
Proceedings of the American Control Conference, Seattle, WA, 3704-3708 (1995).

23

You might also like