FeedforwardIFeedback Control of

Multivariable Nonlinear Processes

This paper concerns general MlMO nonlinear processes, whose
dynamic behavior is described by a standard state-space model of
arbitrary order, including measurable disturbances. The concept of
relative order of an output with respect to an input, extended to include
disturbance as well as manipulated inputs, is generalized in a MlMO
context and it is used to obtain a characterization of the dynamic
interactions among the input and the output variables. A synthesis Prodromos Daoutidis
formula is calculated for a feedforwardlstate feedback control law that Masoud Soroush
completely eliminates the effect of the measurable disturbances on the
Costas Kravaris
Department of Chemical Engineering
process outputs and induces a linear behavior in the closed-loop University of Michigan
system between the outputs and a set of reference inputs. The Ann Arbor, MI 48 109
input/output stability and the degree of coupling in the closed-loop
system are determined by appropriate choice of adjustable para-
meters. A MIMO linear controller with integral action completes the
feedforward /feedback control structure. The developed control method-
ology is applied to a continuous polymerization reactor and its perfor-
mance is evaluated through simulations.

Introduction (Hunt et al., 1983); input/output linearization (Kravaris and

Chemical processes, multivariable in nature, exhibit as a rule Soroush, 1990). It was the differential geometric framework
a highly nonlinear and highly interacting behavior. Efficient that allowed an elegant formulation and geometric interpreta-
control of such processes is necessary, especially under the tion of these problems and their solutions. Research in this area
recently increasing environmental, operational and energy restric- has reached such a point that the synthesis of nonlinear compensa-
tors for stabilization, set point tracking, and disturbance rejection
tions. In fact, rnultivariable control has emerged as a major
can and must be addressed in a unified framework. At the same time,
research area during the last 20 years. The huge existing
recent advances in model development from first principles and
literature (for a survey see Ray, 1983) offers a variety of
increasing computational capabilities encourage the development of
methodologies, both from a state-space and from an input/
such a framework in this direction.
output perspective. The majority of these methodologies, how- In this work, we present a unified methodological framework
ever, is based on a linear approximation of the nonlinear process, for the feedforward/state feedback control of a large class of
which is valid only in a small neighborhood around the operating multivariable nonlinear processes, with the main emphasis on
steady state. For this reason, the presence of severe nonlineari- the regulatory aspect of the problem. While the main result is a
ties translates into large robustness margins in the linear generalization of a previous result for SISO systems (Daoutidis
controller design, leading therefore to conservatism and, in some and Kravaris, 1989a), the proposed methodology naturally
cases, to extremely poor performance. On the other hand, the identifies the fundamental nature of the multivariable control
inherent complexity of M I M O nonlinear systems makes it problem and provides a synthesis framework to address it.
almost impossible to use empirical methods to synthesize nonlin- In particular, we consider general MIMO nonlinear processes
ear controllers of acceptable generality for set point tracking with a state-space description of the form:
and disturbance rejection.
On the nonlinear systems theory front, transparent solutions
for typical multivariable control problems were obtained only
recently which include: invertibility (Hirschorn, 1979); decou- 1 , . . . , rn
yi = hi(x), i =
pling via state feedback (Freund, 1975; Ha and Gilbert, 1986);
noninteracting control and disturbance decoupling (Hirschorn, wheref, gj, w, are smooth vector fields on [w", hi are smooth
I98 1 ; Isidori et al., 198 1); exact state equation linearization scalar fields on [w", and

AIChE Journal October 1990 Vol. 36,No. 10 1471

 Yl
-
Y2
x= EW,d= E R",Y = ERm

Ym,

are the vectors of the states, the disturbance inputs, the feedforward/state feedback control structure. An alternative
manipulated inputs and the outputs (to be controlled), respec- feedforward/output feedback control structure will be proposed
tively, with m 5 n. For simplicity, we assume equal number of for the case of unavailable state measurements. A discussion will
manipulated inputs and controlled outputs. follow, in which specific design objectives in the closed-loop
The general servo and regulatory control problem for such system will be associated with the choice of some adjustable pa-
processes will be addressed and solved in two steps, in analogy rameters. Finally, the developed feedforward/feedback method-
with the SISO case (Daoutidis and Kravaris, 1989a) and the ology will be applied to a continuous polymerization reactor and
disturbance-free MIMO case (Kravaris and Soroush, 1990). At its performance will be evaluated through simulations.
the first step, which is the main synthesis problem, all the
available process information (i.e., model structure and parame-
ters, measurements of the states and the disturbances) will be
used to synthesize a feedforward/state feedback control law Relative Order and the Fundamental Nature of the
which: Control Problem
Eliminates the effect of measurable disturbances on the An extension of the concept of relative order to include
process outputs y , k disturbance inputs as well as manipulated inputs was recently
Induces a prespecified linear behavior between the outputs introduced in the context of SISO systems (Daoutidis and
and a set of reference inputs u;s. Kravaris, 1989a); it was proven to be a very useful tool in
Once this synthesis problem is solved, the second step involves analyzing the dynamic structure of a SISO nonlinear system
using linear control theory to design a linear M I M O controller and developing a feedforward/feedback control methodology. A
with integral action around the linear v - y system. This will similar extension will now be introduced in the context of
ensure satisfactory servo and regulatory behavior, despite the M I M O systems and will be used to characterize the nature of
presence of unmeasured disturbances and/or model uncer- the regulatory control problem, based on a structural analysis of
tainty. The overall control configuration is shown in Figure 1 the system. The definitions that follow refer to MIMO systems
and it clearly depicts the resulting two-step control methodol- in the form of Eq. 1.
ogy. We will start by reviewing the definition of the relative order
We will begin by introducing a generalization of the concept of an output variable with respect to the manipulated input
of relative order in a M I M O setting to incorporate disturbance vector.
as well as manipulated inputs. The fundamental nature of the Definition 1. The relative order of the output y i with respect to
regulatory control problem will then be discussed, using the the manipulated input vector u, ri, is defined as the smallest
relative order as a structural analysis tool. In the following integer for which there exists a j E [ 1, m]such that:
section, the main theoretical result will be developed, i.e., a
feedforward/state feedback control law that solves the posed
synthesis problem. This will lead to the development of a

1472 October 1990 Vol. 36,No. 10 AIChE Journal

Equivalently, ri is the smallest integer for which: lated input vector and the individual disturbance inputs for a
particular output. More specifically,
Disturbances that belong to class Ai have a less direct effect
on the output yithan the manipulated input vector.
The above concept has been extensively used in multivariable Disturbances that belong to class 48, have the same effect on
the output y i as the manipulated input vector.
systems theory, with a variety of names depending on the
*Disturbances that belong to class ei have a more direct
context (e.g., Hirschorn, 1979; Ha and Gilbert, 1986). It can be
effect on the output y , than the manipulated input vector.
easily shown (Kravaris and Soroush, 1990) that ri is the smallest
Based on the intuition that has been obtained from the SISO
order of derivative of the output y i that explicitly depends on the
treatment of the subject (Daoutidis and Kravaris, 1989a), the
vector u . In this sense, the above concept of relative order
following properties are expected to hold concerning the nature
characterizes the dynamic effect of the manipulated input vector
of the regulatory control problem:
on each process output. It is therefore extremely meaningful in a
Bi = ei = 0 : pure feedback compensation will suffice to
multivariable control context, where we are interested in the
eliminate the effect of the disturbances on the output yi.
effect of the manipulated input vector, rather than the effect of
Bi # 0 , ei = 0 : static feedforward/state feedback
each individual manipulated input, on the process outputs.
compensation will be necessary to eliminate the effect of the
Clearly, in any well-formulated control problem, a finite relative
disturbances on the output yi.
order r, must exist for every output yi. ei # 0 : dynamic feedforward/state feedback compensa-
We now generalize the concept of relative order of an output
tion will be necessary to eliminate the effect of the disturbances
variable with respect to a disturbance input variable, originally on the output yi.
introduced for SISO systems by Daoutidis and Kravaris (1989a), The overall control action must compensate for each class of
in a multivariable setting: disturbances and for each output in an appropriate way.
DeJinition2. The relative order of the output y i with respect to Remark 1 . In general, definitions 1 and 2 are valid locally, in a
the disturbance input d,, piK,is defined as the smallest integer for region around a point of interest in state space. Singular points
which: may exist throughout the state space.
Remark 2. The concept of relative order concerns the input/
(3) output behavior of the system; therefore, it is invariant under
coordinate change and in a linear setting it is associated with the
Note that each disturbance input is treated individually in transfer function description of the system and not with the
definition 2. The reason for this is that in a multivariable control specific state-space realization.
context we are concerned with the effect of each individual
disturbance on each individual output. It can be easily shown
Synthesis of the FeedforwardIState Feedback
that a finite relative order pi, represents the smallest order of Control Law and Development of the Control
derivative of the output yi that explicitly depends on the Structure
disturbance d,. In this sense, the above concept of relative order
The main result of this section, which generalizes a previous
characterizes the dynamic effect between disturbance and
result for SISO systems (Daoutidis and Kravaris, 1989a), is
output variables; small relative orders imply a direct effect of the
summarized in the following theorem.
disturbances on the outputs, rendering thus the regulation of the
Theorem 1. Consider the MIMO nonlinear process described
process an important control issue.
by Eq. 1. Let r;s be the relative orders of each output yi with
The above assertion can be made more precise through the
respect to the manipulated input vector and p,'s be the relative
following structural characterization of disturbances:
orders of each output with respect to each disturbance input.
Referring to the nonlinear process described by Eq. 1, we Also, consider the partitions of the set of disturbances according
introduce the following partition of the set of disturbance inputs to Eq. 4 and assume that the characteristic matrix
into the classes .Ai, B,and ei,associated with the output yi:

(4)

Note that for each output, a different partition of the set of the
disturbances will be obtained in general. The above partition
captures the relative dynamic interactions between the manipu- is nonsingular. Then, the control law:

I I I. m

AIChE Journal October 1990 Vol. 36, No. 10 1473

 Completely eliminates the effect of the disturbances on the input/output linearity and eliminates, for each i , the effect of
process outputs. the disturbances in .Ai on the output yi:
Induces the linear v - y closed-loop response:

(7)

where Pik = /$k

constant parameters with
- . @$IT E R" are vectors of adjustable

A static feedforward/state feedback part which eliminates,

for each i, the effect of the disturbances in Bi on the output y,:

and v = [vI v2 -- + '

v,] E R"' is a vector of reference inputs. m

We are going to omit the detailed proof of the above theorem,

because it involves a notationally complicated, but conceptually
straightforward, procedure. The procedure of the proof goes
through the following steps:
a. Derive explicit expressions for the derivatives of each
output yi up to rith order, in terms of x, u and d,'s.
b. Using these expressions, calculate the sum A dynamic feedforward/state feedback part which elimi-
nates, for each i, the effect of the disturbances in eion the output
Yi:

in terms of x, u and d,'s.

c. Calculate the expression for the manipulated input vector
u that makes the above sum equal to v.
The interested reader may refer to the SISO case (Daoutidis
and Kravaris, 1989a) for the expressions of the derivatives of the
outputs that are used in the calculations. It is worth noting that The dynamic element in the above part can be implemented
the whole development of the control law is carried out in an using a lead-lag type of approximation of appropriate order.
algebraic context, with the differential geometric character In each one of the above parts of the control law, the control
being implicitly present through the concept of relative order action results by superimposing explicit compensation terms for
and the Lie derivative notation. each output and for each disturbance, depending on the corre-
Remark 3. The nonsingularity assumption for the characteris- sponding partition of the set of disturbances. It is exactly this
tic matrix C ( x ) and the condition given by Eq. 8 guarantee that explicit character of the control law that results in its rather
the inverse complicated form. More compact expressions can be written, by
adopting a more compact vector notation.
F m m 1-1 At this point, let us summarize the basic characteristics of the
proposed control law. By
Calculating the relative orders r, and p,, for every output y,
*Implementing the control law given by Eq. 6 for an
appropriate choice of the adjustable parameters fi{k's
we use all the available process information so that:
The input/output behavior of the closed-loop system for
changes in the reference inputs is linear
The regulatory behavior of the closed-loop system is perfect
with respect to the measurable disturbances
under the assumption, of course, of a perfect model and perfect
and consequently the control law are well-defined. The nonsingu- implementation of the controller. In the next section, we will
larity assumption for the matrix C ( x ) is somewhat restrictive, in associate the choice of the adjustable parameters p:k's with the
the sense that it imposes a structural restriction on the nonlinear stability characteristics and the degree of coupling in the
system. On the other hand, the majority of practical applications closed-loop system. It should be mentioned that an implicit
conforms with the above assumption. More sophisticated tech- assumption in the previous development is that measurements of
niques than the ones presented here have to be employed for the system's states are available. In fact, this assumption is a key
those that don't. one in obtaining the perfect disturbance rejection property on
Despite the apparent complexity of the control law, a simple the output y, for the disturbances that belong to the class A,,
structure is present. In particular, it is composed of three without using measurements of these disturbances. The measure-
distinct parts: ments of the states completely capture the effect of these
* A pure static state feedback part, which accounts for disturbances and since the manipulated input vector has a more

1474 October 1990 Vol. 36, No. 10 AIChE Journal

direct effect on the output yi than these disturbances, it 3,= ei = 0 for every i:
completely compensates for their effect.

Simplifications and generalizations

The control law given by Eq. 6 simplifies greatly in the
following two cases:
@, = 0 for every i:
which is a pure state feedback control law, exactly the same as if
no disturbances were present (Kravaris and Soroush, 1990).
The above results conform with the intuitive arguments of the
previous section concerning the nature of the control law,
depending on the nonvoid classes of disturbances present in each
partition.
In the case of m = 1 (SISO system), the control law reduces
which is a static feedforward/state feedback control law. to:

which, as expected, is exactly the same with our previous result applies to a much more general class of systems than distur-
for SISO systems (Daoutidis and Kravaris, 1989a). bance decoupling and gives as a byproduct a much simpler and
The nonsingularity of the characteristic matrix C ( x ) (Eiq. 5 ) more transparent necessary and sufficient condition for solvabil-
is a sufficient condition for a static state feedback input/output ity of the disturbance decoupling problem.
linearizing control law to exist. It has been shown (Kravaris and In the context of MIMO systems of the form given by Eq. 1
Soroush, 1990). however, that input/output linearization can be and assuming that the characteristic matrix C ( x )is nonsingular,
achieved for a larger class of disturbance-free systems than the necessary and sufficient condition found in the literature
those satisfying this assumption. Generalization of theorem 1 (e.g., Isidori et al., 1981) for solvability of the disturbance
for this class of systems is possible, but it would require decoupling problem is:
complicated notation and several technicalities that would dilute
the focus of the present paper.
The proposed methodology can also be trivially generalized to WAX)E ,b
I-
[ker {dhi(x)j n ker {dL/hj(x)I
achieve any nonlinear input/output closed-loop response of the
form: n . . . (7 ker{dLj'-'h,(x)}] (9)

for every x E R" and every K .

Within our proposed feedforward/feedback formulation, the
above condition can be easily shown to be equivalent to:

Finally, consider the more general class of nonlinear processes

with a state-space description of the form:
for every i and K , which is much simpler and easier to verify than
Eq. 9.
Clearly, the above condition is extremely restrictive and is
rarely met in real systems. On the other hand, our proposed
methodology provides a systematic way of treating measurable
disturbances, allowing for them to be incorporated in the control
where 4,(x, u,, d:) is a scalar function solvable for u,, and d: is a law in an appropriate way.
vector of additional measurable disturbances. The above class Remark 5. The disturbance rejection capability of the pro-
incorporates cases where some manipulated inputs appear in the posed methodology with respect to the disturbances in class 34,
state equations coupled with some measurable disturbances. In can find an interesting robustness interpretation. In particular,
this case, the proposed methodology can be applied by simply consider a localized perturbation (model uncertainty and/or
letting U, = +,(x, u,, d: ), calculating the control law (Eq. 6) for unmeasured disturbance) of arbitrary magnitude which enters
the new manipulated inputs U, and then solving for the original the system dynamic structure in an additive way at a certain
manipulated inputs u,. location (i.e., a certain state equation). Such a perturbation can
Remark 4 . The relation of the feedforward/feedback problem be viewed as an unmeasurable disturbance and can be assigned
to the classical disturbance decoupling problem was discussed an "equivalent relative order." The particular perturbation
for SISO systems in Daoutidis and Kravaris (1989b). It was can then be included in one of the classes of disturbances defined
shown that our proposed feedforward/feedback formulation by Eq. 4.In case it belongs to the class .Ai, it will not have any

AIChE Journal October 1990 Vol. 36,No. 10 1475

 Figure 2. Feedforward/output feedback control structure.

effect on the output y i in the closed-loop system. The above will converge to the correct estimates of the states for large t ;
inherent robustness feature of our methodology is extremely any initialization error will eventually die out. The resulting
meaningful in chemical systems, where a model uncertainty can control structure after incorporating the open-loop state ob-
often be identified with errors in certain system parameters, e.g., server becomes a feedforward/output feedback control struc-
kinetic rate constants and heat transfer coefficients. ture, as it is shown in Figure 2. Obviously, in such a control
structure one has to use measurements of all the disturbances,
Feedforwardlfeedback control structures even the ones that belong to the classes .Ai, since there is some
lost information from the states that are not measured.
Theorem 1 provides an explicit solution to the synthesis
problem posed in the introduction. In general, as was discussed, Closed-Loop Design Considerations
the control structure has to be completed by a linear MIMO As indicated by the control structure, the proposed control
controller with integral action, designed appropriately to ac- methodology is a two-step procedure that involves: a) the design
count for unmeasured disturbances and/or model uncertainty. of the inner feedforward/feedback loop (v - y system); and
The resulting overall feedforward/state feedback control struc- b) the design of a multivariable external linear controller to
ture is depicted in Figure 1. In the case where the disturbances ensure satisfactory regulatory behavior of the closed-loop sys-
are not modeled explicitly and therefore are not used in the tem in the presence of unmeasurable disturbances and model
control law, the above control structure reduces to the MIMO uncertainty.
GLC control structure (Kravaris and Soroush, 1990).
In many practical situations, on-line measurements of the Design of the feedforward/feedback inner loop
states are not available or economicallyjustified. In this case and Under the control law given by Eq. 6, the dynamics of the u -
for an open-loop asymptotically stable system, a simple method y system is governed by:
of reconstructing the states is the use of an open-loop state
observer. In particular, given estimates of the initial states
+ ' * * + P l r ,d'lY1) ...
k(0) = ko,and driven by on-line measurements of the process
inputs (manipulated and disturbance), a replica of the process
model
(BlOY, dt" +

or expanding the @-columnnotation:

1476 October 1990 Vol. 36, No. 10 AIChE Journal

or in the Laplace domain and using a matrix fraction descrip- [(rl + -- + rm + m),in the case of full decoupling] and the
tion: use of SISO controllers in the external loop, in which case their
tuning is straightforward. Physical constraints on the manipu-
lated inputs and/or physical importance of the controlled
outputs may often dictate whether decoupling is realistic and/or
desirable. In general, despite the extensive research effort in this
The order of the closed-loop system is rl + r2 + ---
+ rm.In area, there is a lack of systematic methods of fundamental rigor
for assessing when decoupling is favorable, even in the case of
analogy with the SISO case, the closed-loop system does not linear systems. It should be noted that, within our proposed
possess any finite zeros. On the other hand, the poles of the synthesis framework, any degree of coupling can be. achieved by
closed-loop system are the roots of the characteristic equation: simply an appropriate choice of the adjustable parameters
(without any modification in the actual synthesis procedure).
det [(2 (2
k=O -.
B I k s k )k - 0 BZks*) (5
k=O Bmksk)] = 0.
This fact allows a significant degree of flexibility to the designer,
who can incorporate his/her own intuition and experience in the
tuning procedure and test the resulting performance characteris-
Consequently, the B I B 0 stability characteristics of the closed- tics.
loop system depend on the values of the m(r, + ---
+ rm) +
m2adjustable parameters fijk's.
The issue of closed-loop asymptotic stability of the states Design of the external linear controller
under no external inputs is not going to be addressed in detail in The design of a multivariable linear controller for the linear
this work, for brevity. The procedure to be followed in order to v - y system can be performed using techniques from linear
obtain a precise characterization for this issue is conceptually control theory. Of course, if the v - y system is decoupled, the
similar to the one followed in the SISO case (Daoutidis and designer has a much simpler task of synthesizing and tuning of
Kravaris, 1989a), but the treatment must be done in a M I M O the corresponding SISO linear controllers. In any case, the
setting using a multivariable notion of nonlinear zeros. In external linear controller must be designed to ensure
particular, one can generalize the disturbance-free concept of Stability of the overall closed-loop systemyp - y
M I M O zero dynamics and obtain a concept of zero dynamics for *Satisfactory tracking of set points and rejection of the
MIMO nonlinear systems with disturbances; then, appropriate unmeasurable disturbances
stability conditions on the zero dynamics will guarantee the That the magnitude of the manipulated inputs uis does not
asymptotic stability of the unforced closed-loop system. exceed the bounds imposed by practical constraints.
In some cases, it may be desirable to achieve input/output
decoupling in the closed-loop system, i.e., to have each reference Application of the Feedforward/Feedback Control
input u, affect only the output yi. In this case, the postulated Methodology to a Continuous Polymerization
closed-loop response is: Reactor
Consider the CSTR shown in Figure 3, where free-radical
polymerization of methyl methacrylate (MMA) takes place,
dt with azo-bis-isobutyronitrile(AIBN) as initiator and toluene as
solvent. The reaction is exothermic and a cooling jacket allows
the heat removal. The standard mechanism of free-radical
polymerization is assumed, together with the resulting rate laws
(Ray, 1972; Ray et al., 1971; Congalidis et al., 1989; Schmidt
and Ray, 1981; Tsoukas et al., 1982). We also make the
following assumptions:
Perfect mixing in the reactor

and one simply sets

Monorner+Solvent
n Initiator

r
a;, = 0,i # j F Gin T i n

in the control law.

Any kind of partially decoupled closed-loop response can also
be achieved by appropriate choice of the adjustable parameters
P j k ' S , as long as the nonsingularity condition of Eq. 8 is satis-
fied. By requesting any kind of input/output decoupling, we
may impose an additional structural constraint on the closed-
loop system which may cause deterioration in its perfor-
mance characteristics. On the other hand, several advan-
tages are present, such as fewer adjustable parameters Figure 3. Simulated continuous polymerlzatlon reactor.

AIChE Journal October 1990 Vol. 36,No. 10 1477

 Constant density of the reacting mixture (no volume shrink- 0 Constant reactor volume (constant volumetric flow rate of

age) the monomer stream)

Constant heat capacity of the reacting mixture Negligible flow rate of the initiator solution compared to
Uniform coolant temperature in the jacket that of the monomer stream
Insulated reactor and cooling system Negligible inhibition and chain transfer to solvent reactions
Constant density and heat capacity of the coolant Quasisteady state and long-chain hypothesis.
No polymer in the inlet streams The dynamic behavior of the process is then described by the
No gel effect (because of low monomer conversion) following mass and energy balances:

where following the standard procedure and letting

and

Control of the temperature Tand the number average molecular

weight Dl/Doof the polymer product is considered, by manipu-
lating the volumetric flow rate of the initiator F, and the
volumetric flow rate of the cooling water F,. The concentration
of monomer in the inlet stream Cminand the temperature of the the dynamic equations of the process are put in the form of Eq.
inlet stream Tin are the major measurable disturbances. Thus, 1, where n = 6, m = 2 , p = 2 and

/(x) =

1478
 -
0
0
0

0 , w2(x) =
0

Tw* - .I
- VO

Based on definitions 1 and 2, the relative orders are easily where

found to take the following values:
.Outputy,:r, = 2 , p , , = 2,p,, = 2
~ O u t p u t y , r2
: = 2,pz,= 2,p2, = 1
Consequently, the set of disturbances is partitioned as follows:
A , = a,%,= {d,,dz},el = 0
= {d,],eZ= {d,}
34, = 0,3,
Clearly, as a result of the dynamic structure of the particular
process, we will need static feedforward/feedback compensation
to eliminate the effect of d, and d, on y , and the effect of d, on y,,
while dynamic feedforward/feedback compensation will be
necessary to eliminate the effect of dzon y,.
The characteristic matrix C(x) of this system defined by Eq. 5
becomes:
It can be easily checked that C(x) is generically nonsingular; the
feedforward/state feedback control law given by Eq. 6 can
therefore be applied in a straightforward fashion, yielding:

where

AIChE Journal October 1990 Vol. 36,No. 10 1479

The control law given by Eq. 16 has a static feedforward/state we simply let
feedback component and a dynamic feedforward/state feedback
component; this agrees with the intuitive arguments made
earlier. In order to obtain full input/output decoupling of the
form

in the control law (Eq. 16), which takes the form

I
Y2 + p:, + p:, -
dt2
d2Y2
= 212

The kinetic and physical parameters and the operating steady tude of the manipulated input variables (F, 2 0 and F,, z 0) be
state conditions for the particular process are given in Tables 1 satisfied. The external linear controllers in the FF/FB control
and 2. Integration of the system dynamic equations was per- structure were chosen as two PI controllers with settings K , =
formed (after appropriate dedimensionalization) by using the 15 and 7, = 0.4 h. A number of simulation runs verified the
subroutine LSODA from the ODEBACK Library, on the stability of the open-loop system around the operating steady
Apollo network of The University of Michigan. The values for state and the internal stability of the closed-loop system.
the adjustable parameters in the FF/FB control law (Eq. 17) The performance of the proposed feedforward/feedback con-
were chosen as trol methodology was tested in terms of rejection of step changes
at the two measurable disturbances. The process was initially
assumed to be at steady state. At time t = 1 h a step change a t
the inlet monomer concentration Cm,,was applied, from 6 to 5
to place the closed-loop poles at -2.5 and -25.0 for the two kmol/m3. The process was allowed to reach a new steady state,
decoupled v I - y I and v2 - y 2 systems. The location of the and at time t = 6 h a step change at the inlet temperature Tin
closed-loop poles was chosen so that the constraints on the magni- was applied, from 350 to 345 K.

Table 1. Kinetic Parameters

i, =I,
Reference
T, 3.8223 x 10" kmol/ml.h 2.9442 x 101kJ/kmol Schmidt and Ray, 1981
Td 3.1457 x 10" kmol/m'.h 2.9442 x IO'kJ/kmol Schmidt and Ray, 1981
I 3.7920 x lo'* h-.' 1.2877 x lo5 kJ/kmol Tobolsky and Baysal, 1953
P 1.7700 x lo9 kmol/m3.h 1.8283 x lo4 kJ/kmol Mahabadi and O'Driscoll, 1977
f", 1.0067 x lOI5kmol/m'. h 7.4478 x lo4 kJ/kmol Stickler and Meyhoff, 1978
f* = 0.58 Tobolsky and Baysal, 1953

1480 October 1990 Vol. 36, No. 10 AIChE Journal

 Table 2. System Parameters and Steady-State Values

F = 1.00 m3 F, 0.01679
= m 3 .h-'
v = 0.1 m3 R= 8.314 .
kJ kmol-' K - '
p = 866 kg.m-3 M, = 100.12 kg. kmol-I
C,,"= 6.0 krnol.m--l C,,"
= 8.0 kmol. m-3
cp= 2.0 .
kJ kg - I K-I. -AHp = 57,800 kJ. kmol-'
A = 2.0 m2 U= 720 .
kJ. h-' K - ' . m-
p , = 1.000 kg.m-' C, = 4.2 kJ*kg-' * K - '
v, = 0.02 m3 T,, = 350 K
y:" = 25,000 kg * h o l - I y q = 335 K
Fcy = 3.26363 m'.h-' Two= 293.2 K

Figures 4-7 illustrate the profiles of the two controlled outputs c) Two linear SISO PI loops (coolant flow rate/temperature,
and the two manipulated inputs, under the assumption of perfect initiator flow rate/number average molecular weight).
model and perfect measurements. The figures provide a compar- The derivative term in the FF/FB control law (Eq. 17) was
ison of the output and input responses under approximated by a lead-lag element with transfer function
a) The MIMO FF/FB control structure s/(O.OOls + 1). In the implementation of the GLC structure,
b) The MIMO GLC structure the same values of &'s were used as in the FF/FB structure,

'1
028
In
Y

026

0 2 2 4 . .
0
I
3
. . ,
6
. . ,
9
. . O.OM I J
Time, hr Time, tr

Figure 4. Number-average molecular weight profiles un- Figure 6. Initiator flow rate profiles corresponding to
der different control methods. Figures 4 and 5.

i:1
c

i 333

332 o ! . . , . . , . . I . . 1
0 3 6 9 0 3 6 9 t7
Tm,h- Time. hr

Flgure 5. Reactor temperature profiles under different Figure 7. Coolant flow rate profiles corresponding to
control methods. Figures 4 and 5.

AIChE Journal October 1990 Vol. 36, No. 10 1481

 In

i
I

Time, hr

Figure 8. Number-average molecular weight profiles un- Figure 10. Number-average molecular weight profiles un-
der FF/FB control: effect of model lmperfec- der FF/FB control: effect of disturbance mea-
tions. surement noise.

while the external linear controllers were chosen as PI control- noise. In particular we compared the closed-loop behavior of the
lers with the same settings as in the FF/FB structure. Finally, process under
in the linear control approach, the two SISO PI controllers a) Perfect model and perfect disturbance measurements
were tuned through a trial-and-error procedure which resulted b) 20% error in the frequency factor Z, and the heat of
in the values K, = -1 x lo-’ m3/h, r, = 0.075 h and K, = reaction AHp
.
-0.1 m’/h K, T , = 0.075 h, respectively for “best” closed-loop c) Sinusoidal noise in the measurements of the disturbances
performance. Due to the severe nonlinearity of the process, the -
d, and d, of amplitudes 0.05 kmol m-3 and 0.5 K, respectively,
response characteristics were found to be very sensitive to the and period of oscillation of 10 minutes.
values of the PI controllers’ settings. Figures 8 and 9 depict the excellent performance of the
Clearly, as the theory predicts, the FF/FB control law results FF/FB structure in rejecting the applied step changes in the
in perfect regulation of the outputs, i.e., an obvious improvement disturbances for the case when the above model uncertainties
of the closed-loop behavior compared with the one under the exist. As shown in the two figures, even in the presence of the
GLC structure (where no measurements of the disturbances are modeling errors, the output profiles are very close to the ones
used in the control law), or the linear PI controllers. In another obtained when a perfect model is available. Figures 10 and 11
set of simulation runs, assuming the same disturbance changes depict the performance of the FF/FB structure in rejecting the
as previously, we tested the robustness characteristics of the applied step changes when the disturbance measurements are
FF/FB method in the face of modeling error and measurement corrupted with the above noise. Clearly, although the output

Y 35554 I

334

Time. hr

Figure 11. Reactor temperature profiles under FF/FB

Figure 9. Reactor temperature profiles under FF/FB con- control: effect of disturbance measurement
trol: effect of model imperfections. noise.

1482 October 1990 Vol. 36,No. 10 AICbE Journal

regulation is not perfect in the presence of measurement noise, nation by coupling and termination by dispropor-
the proposed method performs very satisfactorily. .
tionation reactions, respectively, kmol/m’ h
Z,= frequency factor in Arrhenius equation, for the
initiation reaction, h-’
d = vector of disturbance inputs
Conclusions f * = initiator efficiency
A general synthesis methodology for the feedforward/ J g,, w, = vector fields
feedback control of a large class of multivariable nonlinear h, = output scalar field
processes with measurable disturbances was presented. An r, = relative order of the output y, with respect to the
manipulated input vector
appropriate formulation of the concept of relative order, to s = the Laplace domain variable
include disturbance inputs as well as the manipulated input t = time
vector, led to a transparent characterization of the regulatory u = vector of manipulated inputs
aspect of the problem. A feedforward/state feedback control v = external input vector
x = vector of state variables
law that results in complete elimination of the effect of measur-
y = vector of outputs
able disturbances on the outputs and a linear input/output y ’ p = vector of set points
behavior in the closed-loop system was calculated. T h e closed- -AHp = heat of propagation reaction, kJ/kmol
loop stability and design objectives were associated with the A,,B,, @, = partition of disturbances for the output yi
choice of the adjustable parameters of the control law. The
proposed methodology was successfully applied to a continuous Greek letters
polymerization reactor, for the regulation of the number aver- ,9ik = parameters of the feedforward/state feedback
age molecular weight of the product and the reactor tempera- law
T , = reset time
ture, in the presence of measurable disturbances at the composi-
p = density of the reacting mixture, kg/m’
tion and temperature of the inlet stream, modeling error and p , = density of water, kg/m’
measurement noise. pi, = relative order of the output y, with respect to the
disturbance dK

Acknowledgment Math symbols

Financial support from the National Science Foundation through the =. = implies
grant CTS-8912836 is gratefully acknowledged. 0 = is equivalent to
E = belongs to
n = intersection
Notation 0 = void set
A = heat transfer area, m* det = determinant of a matrix
C,,, = molar concentration of the monomer, kmol/m3 ker = kernel of an operator
C,,,,”= molar concentration of monomer in the mono- dh, = gradient of a scalar field hi
mer inlet stream, kmol/m3 W = real line
C, = molar concentration of the initiator, kmoI/m’ R“ = n-dimensional Euclidean space
C,,”= molar concentration of the initiator in the initia- L/hi =‘Lie derivative of the scalar field hi with respect
tor inlet stream, kmol/m3 to the vector fieldf
c,, = heat capacity of the reacting mixture, kJ/kg K Ljh, = k-th order Lie derivative of the scalar field hi
c, = heat capacity of water, kJ/kg K- with respect to the vector fieldf
Do= molar concentration of the dead polymer chains,
kmol/m3 Acronyms
D ,= mass concentration of the dead polymer chains,
kgp’
E/.,.E, ETc,E,, E , = activation energies for chain transfer to mono-
-
BIB0 = bounded input bounded output
SISO single input single output
MIMO = multiple input multiple output
mer, propagation, termination by coupling, ter- GLC = globally linearizing control
mination by disproportionation and initiation FF/FB = feedforward/feedback
reactions, respectively, kJ/kmol CSTR = continuous stirred tank reactor
F = volumetric flow rate of the monomer stream into
the reactor, m’/h
F, = volumetric flow rate of the inlet initiator stream, Literature Cited
m3/h Congalidis, J. P., J. R. Richards, and W.H. Ray, “Feedforward and
F,, = volumetric flow rate of the cooling water, m3/h Feedback Control of a Copolymerization Reactor,” AIChE J., 35,
K, = proportional gain 891 (1989).
M, = molecular weight of the monomer, kg/kmol Daoutidis, P., and C. Kravaris, “Synthesis of Feedforward/State-
Po = molar concentration of the live polymer chains, Feedback Controllers for Nonlinear Processes,” AIChE J., 35, 1602
kmol/m’ (1989a).
R = Ideal Gas constant, kJ/kmol K . -, “Feedforward/State-FeedbackControl of Nonlinear Processes,”
T = reactor temperature, K Proc. A.C.C., 1082, Pittsburgh (1989b).
7‘, = jacket temperature, K Freund. E., “The Structure of Decoupled Nonlinear Systems,” Int. J. of
T,, = temperature of the inlet streams in the reactor, Contr.. 21,443 (1975).
K Ha, I. J., and E. G. Gilbert, “A Complete Characterization of
Two= temperature of the inlet coolant stream, K Decoupling Control Laws for a General Class of Nonlinear Systems,”
V = overall heat transfer coefficient, kJ/m2 h K IEEE Tram. Autom. Contr.. AC-31,823 (1986).
V = reactor volume, m3 Hirschorn, R. M., “Invertibility of Multivariable Nonlinear Control
V, = overall effective volume of the cooling sub- Systems,” IEEE Tram. Autom. Contr., AC-24,855 (1979).
system, m3 -, “(A, B)-Invariant Distributions and Disturbance Decouplingof
Zffl,,2,. Z,, 2, = frequency factors in Arrhenius equation, for Nonlinear Systems,” SIAM J. Contr. Optimiz.. 19, 1 (1981).
chain transfer to monomer, propagation, termi- Hunt, L. R., R. Su,and G. Meyer, “Design for Multi-Input Nonlinear

AIChE Journal October 1990 Vol. 36,No. 10 1483

 Systems,” Differential Geometric Control Theory, R. W. Brockett, -, “On the Mathematical Modeling of Polymerization Reactors,”
R. S. Millman, and H. J. Sussman, eds., Birkhauser, Boston, 268 J. Macromol. Sci.-Revs. Macromol. Chem.. C8,l (1972).
(1983). Ray, W. H., T. L. Douglas, and E. W. Godsalve, “Molecular Weight
Isidori, A,, A. Krener, C. Gori-Giorgi, and S. Monaco, “Nonlinear Distributions in Copolymer Systems. 11. Free Radical Copoly-
Decoupling via Feedback a Differential Geometric Approach,” merization,” Macromol.. 4,166 (1971).
IEEE Trans. Autom. Contr., AC-26,331 (1981). Schmidt, A. D., and W. H. Ray, “The Dynamic Behavior of Continuous
Kravaris, C., and M. Soroush, ‘Synthesis of Multivariable Nonlinear Polymerization Reactors-I,” Chem. Eng. Sci., 36,1401 (1981).
Controllers by Input/Output Linearization,” AIChE J., 36, 294 Tobolsky, A. V., and B. Baysal, “A Review of Initiation in Vinyl
(1990). Polymerization: Styrene and MMA,” J. of Polym. Sci., 11, 471
Mahabadi, H. K., and K. F. O’Driscoll, “Absolute Rate Constants in ( 1953).
Free-Radical Polymerization: 111. Determination of Propagation and Tsoukas, A., M. Tirrell, and G. Stephanopoulos, “Multiobjective
Termination Rate Constants for Styrene and Methylmethacylate,” J. Dynamic Optimization of Semibatch Copolymerization Reactors,”
Macromol. Sci.-Chem.. A l l , 967 (1977). Chem. Eng. Sci., 37,1785 (1982).
Ray, W. H., “Multivariable Process Control-A Survey,” Comp. and
Chem. Eng., 7,367 (1983). Manuscript received Dec. 18. 1989. and revision received July 23, 1990.

1484 October 1990 Vol. 36,No. 10 AIChE Journal