Chemical Engineering Science, Vol. 47, No. 4, pp. 755-762, 1992. OmP-2So9/92 ss.

00 + mo
hinted in Great Britain. &31992pergmonReMpk

AN ADAPTIVE NONLINEAR PREDICTIVE CONTROLLER*

J. DUANE MORNINGRED,’ BRADLEY E. PADEN,: DALE E. SEBORG’

and DUNCAN A. MELLICHAMPt6
‘Department of Chemical and Nuclear Engineering;*Department of Mechanical and Environmental
Engineering,University of California, Santa Barbara, CA 93106. U.S.A.

(Receivedfir publication 6 August 1991)

AbstrPct--The design and implementation of a new adaptive nonlinear predictive controller is presented
using a general nonlinear model and variabletransformations.The resultingcontrolleris similarin form to
standard linear model predictive controllersand can be tuned analogously. Alternatively,the controller
can be tuned using a single parameter.The design is computationallyefficient.The controller is updated
on-line without recalculatingthe controller gain matrix, which involves a matrix inversion. The new
controller is compared to a PI controller and to an adaptive linear predictive controller through
simulationsof a continuous stirred-tankreactor. The effectsof modeling errors on the new controller are
also shown with simulations.

1. MTRODUaION controller is formulated in a discrete-time nonlinear

Computer process control, beginning in the 196Os, predictive control framework.
initially used traditional linear control algorithms.
However, the highly nonlinear characteristics of many
processes caused problems for these controllers. Con- 2. MODELS AND PARAMEI’ER ESTIMATION

sequently, in the 197Os, self-regulating controllers Many SISO processes can be accurately repres-
Cl, 23 were developed to enhance explicit and implicit ented by the discrete, time-invariant model of eq. (I):
model-based controllers and controller tuning. The
y(t) = F [y(t - l), u(t - k - I), Z(t - l), M(t - l)]
process industries, however, found few successes with
the early, hard-to-tune adaptive techniques [3]. (1)
Meanwhile, moving-horizon and iinear-pro- where
gramming methods were being revived in the
nonadaptive, model-based predictive controllers y(t - 1) = Cy(t - l), y(t - 2), . . . , Y(t - n,V (2)
(MPC), dynamic matrix control (DMC) [4], and u(t - k - 1) = [u(t - k - l), u(t - k - 2), . . .,
model-predictive heuristic control [SJ. Industrial suc-
cesses with MPCs renewed academic interest in these u(t - k - %.)I= (3)
methods and new formulations of MPCs emerged [6].
Z(t - 1) = [z(t - l),z(t - 2), . . . , z(t - n.)] (4)
These multivariable controllers were based on easily
understood process models, could incorporate con- M(t - 1) = [m(t - l), m(t - 2), . . . , m(t -.n,) J (5)
straints, and were relatively easy to tune. In addition,
their performance seemed less sensitive to varying t is the discrete time, k is the time delay, y(t) is the
time delays, one of the major limitations of the early controlled output, u(t) is the manipulated input, z(t) is
adaptive controllers. To improve the robustness of the a vector of other process outputs, m(t) is a vector of
adaptive controllers, some researchers began to em- other process inputs, and F C-1 is the general, nonlin-
ploy extended-horizon strategies [7]. Likewise, pre- ear output function with stable zero dynamics. This
dictive controllers were improved by incorporating model is called the NARMAX model and can require
adaptive techniques [87. significantly fewer terms than an equivalent Volterra
During the past several years, some researchers series model. Chen and Billings [19] discuss this and
have started to focus on strategies that directly com- other properties of the NARMAX model.
pensate for process nonlinearities in controller design The NARMAX model form represents process
[9, lo]. Naturally, these nonlinear control concepts nonlinearities in ways that allow computationally efii-
have been introduced into adaptive-type controllers cient parameter estimation methods, such as recursive
[ 11, 127 and into predictive-type controllers [ 13-161. least squares (RLS), to be used if the model is linear in
The nonlinear control problem has also been ap- the parameters. Otherwise, ‘quasi-Newton or other
proached numerically using nonlinear optimization gradient-based methods must be used to solve the
software packages [ 17, 181. In this paper, an adaptive resulting nonlinear optimization problem. In this
paper, the assumption is made that the nonlinear
model, chosen with a priori process knowledge, rep-
*An earlier version of this paper was presented at the 1990
resents the characteristics of the process over a wide
American Control Conference. range of operating conditions. If the model is linear
‘Author to whom correspondence should be addressed. with respect to the parameters, the process model in

755
756 I. DUANE MORNIN~RED et al.
eq. (1) then can be rearranged to give u(t - k - 1) = G[g(t - l),
y(t) = e=(t - l)#(t - 1) + E(t) (6) u(t - k - l), 9;(t - l), M(t - I)] (12)
where f?(t - 1) is the parameter vector estimated at so that eq. (10) can be expressed as a linear model
time t - 1, #(t - 1) is the information vector at time
A(q- ‘)9(t) = u(t - k - 1) (13)
t - 1, and E(t) is the estimation error. With a
NARMAX model, 4 contains functions of past values and eq. (11) can be expressed as
of y, U, z, and m. RLS estimation of B is used to adjust
the model to local process conditions. This paper Q(t + i) = q’[l - A(q-‘)lt+‘9(t)
employs RLS with directional forgetting, as described
by Kulh&$ and Karnjr [20]. + qi i [l - A(q- ‘)]‘u(t - k - 1). (14)
j=o
The use of eq. (12) provides an exact transformation
3. CONTROLLER DESIGN METHODS
of the nonlinear model to a linear model. Thus,
eq. (14) can be rewritten to represent the nonlinear
Two types of adaptive nonlinear predictive con-
prediction in the matrix form:
trollers (ANPC) will be designed. The first type as-
sumes a restriction on the NARMAX model and is 9(t) = pp(t) + &Av(t) (15)
tuned in a manner similar to that of linear MPCs. The
where y(t) is an R-vector of predicted outputs; and
second type removes this restriction and is tuned by
yP(t) is an R-vector of the effect of past input changes
selecting a parameter to place closed-loop poles.
on the R future outputs:

3.1. Desigb based on a NARMAX model with separ- 9(t) =C9(t + I), 9(t + 3, . . . ,9(t + WI= W)
able linear dynamics 9P(t) = [gP(t + l), gqt + 2), . . . , gP(t + RW. (17)
Predictive controller designs that use a “dynamic
In eq. (15) Av(t) is an L-vector of the form
matrix” [4] require a linear-dynamics model. Without
linear dynamics, the subsequent controller design Av(t) = [Av(t), Au(t + l), Av(t + 2), . . . ,
could not be tuned using the optimization and control
horizon lengths. Consequently, the first ANPC design Au(t + L - 1)-J= (18)
assumes that linear dynamics can be separated from and _a? is an R x L “dynamic matrix”:
the NARMAX model in eq. (l), as follows:
al 0 ... 0
F [y(t - l), u(t - k - l), Z(t - l), M(t - l)] = ...
a2 a1 0
= qC1 - A(q-‘)ly(t - 1) + GMt - 11,
Jaf= * * *** * (1%
aL aLdl e-e aI
u(t - k - l), Z(t - l), M(t - l)] (7)
. . .-. .
where [ aR aR_-i -.. aR-L+l 1.
A(q-‘) = 1 + aIq-’ + a2qe2 +. . - + a,,q-“A (8) In eq. (15) A is the backward difference operator. The
optimization horizon R is the number of future process
fi$I la.lfQ nA21. (9) moves over which L future control moves are
optimized. L is called the control horizon. The follow-
G[*] is a general, nonlinear function, and q- 1 is the ing relationships for ai and jP(t + i) can be derived:
backwards-shift operator. The poles of A(q-I) must
gp(t + i) = q’[l - A(q-‘)]‘+‘p(t)
also be stable. The NARMAX model in eq. (7) is then
expressed in a prediction form as iN,+k+l
+ f: h,u(t - 1) + c hju(t + i - j)
-Q-‘)9(t) = GC9(t - I), j=o j=i+l

u(t - k -
(20)
l), &(t - l), M(t - l)] (10)
a, = c h. (21)
where the caret denotes a prediction. Equation (10) n=o
can be rearranged to give an i-step ahead prediction:
where h, = 0 for 0 > n 2 k and h, = the coefficient of
j(t + i) = q’[l - A(q-‘)]‘+lg(t) the q-” term of [q-kA(q-l)]-l for n > k. Conse-
quently, h, is merely the nth impulse-response coeffi-
+ d i Cl - A(q-')]'G[9(t - l), u(t - k - l), cient of eq. (13) for a unit impulse in u(t), and a, is the
j=o jth step-response coefficient for a unit step in u(t).
A predictive controller can be designed by minimiz-
at - l), M(t - l)]. (11)
ing the objective function
Next, following the linearization strategy of Hunt et
al. [21], a new variable, u(t - k - I), is defined to J = [r(t) - S(tWQ=QCW - 9Wl
represent the nonlinearity of eq. (10): +j-Av(t)=Av(t) (22)
An adaptive nonlinear predictive controller 757

where r(t) is the R-vector of future set-points, Q is eqs (7)--(25), a predictive controller can be designed
a weighting matrix, and f is the move suppression using this process model. The roots of x(q-‘) can
factor. Substituting eq. (15) into eq. (22) and minimiz- be specified to shape the closed-loop response.
ing with respect to Av(t) yields For example, choosing a(q-‘) = 1 - pq-I, where
0 > p > - I, and combining with the variable trans-
Av(t) = K,E(t) (23) formation of eq. (12), gives the following prediction
K, = (.#Q=Q& +SI)-‘&=Q=Q (24) model:

where K, is the constant controller gain matrix and $(t + i) = q’[pq - 1-j+ 1$(t)
E(r) is the “no-future-change-in-u” set-point error vec-
tor, + qi 2 [pq-‘]‘o(t - k - 1) (30)
j=o
E(t) = r(t) - fP(t). (25) from which the controller is designed. The values of h.
In analogy with linear MPCs, only the first element of for this controller design are simply
Av(t) needs to be calculated. Using Au(t), u(t - I), and
h, = 0 for 0 b n 2 k
eq. (12), the controller output u(t) can be calculated.
(31)
In addition, the tuning parameters for the “control h, = pnmkV1 for n > k.
law” in eq. (23), R, L,J and Q are analogous to those
Thus, little computational effort is needed to calculate
of linear MPCs.
hn*
However, the major computational advantage is
3.2. Design based on a general NARMAX nwdel not in the calculation of h,. Note that the controller
For the second type of controller, the restriction of gain matrix is now a function of the chosen A(q-l),
eq. (7), that linear dynamic terms can be separated and is no longer a function of the output model F [*I.
from F [a], is removed. Removing this restriction Thus, the controller gain matrix does not need to be
corresponds to setting A(q-‘) of eq. (13) to 1 when recalculated when parameters of the output model are
such a polynomial cannot be separated. In such updated on-line. This approach avoids having to do
a case, the prediction model of eq. (14) reduces to a matrix inversion each time the controller is re-
designed.
Y(t + i) = o(t - k - 1 + i). (26)
For this model, each aj = 1, for j > k. In effect, the 3.3. Controller tuning
“dynamic” matrix & of eq. (18) reduces to a “static” For the first ANPC controller design, the tuning
matrix. In the linear analogy, all of the model poles methods are analogous to those of linear MPCs
and zeros are cancelled. This analogy provides insight [4,22]. The second controller design seemingly adds
as to why R and L are not useful tuning parameters in another tuning parameter that must be selected. In
this case. practice, R is often chosen to correspond to the open-
In the design of a linear pole-placement controller, loop settling time of a process response to a step
all poles are cancelled except for the ones that are change in the manipulated input (41. For a linear
placed to specify the speed and shape of the closed- first-order process, the 99% response is about 5 times
loop response. A pole-placement ANPC will be de- the dominant time constant plus the time delay of the
signed in a similar manner. Regardless of whether process. Because the tuning pole, p, ideally determines
or not linear dynamics can be separated from the the settling time of the process, R can be chosen as the
general NARMAX model eq. (I), a polynomial, following function of p:
[I - A(q-‘)]y(t), can be added to and subtracted
from the right-hand side to yield -5
R= ~ + km,
log, (PI
y(t) = q[l - Al(q_‘)y(t - 1) - q[l - A”(q_‘)]y(t - 1)
where k,,, is the maximum expected process time
+ F[y(t - I), u(t - k - I), Z(t - I), M(t - I)].
delay. Thus, the choice for p specifies R. As the effects
(27) of p on the closed-loop response are easily under-
Combining eq. (27) with the following definition of stood, the more difficult to selectfand Q can be set to
their limits of zero and the identity matrix, respect-
G[y(t - l), u(t - k - I), Z(t - l), M(t - I)] = ively. The result is that only p and L need to be
selected. Consequently, the objective function in
= F[y(t - l), u(t - k - l), Z(t - I), M(t - I)]
eq. (22) can be restated as
- 4c1 - &?-‘)Mt - 1) (28)
J’ = [r(t) - 9(OlTC~(~)
- f(t)1 (33)
yields the process model,
and the minimization solution for Av(t) as
Y(l) = 4Cl - z(q-‘)]Y(t - 1) + G[Y(~ - l), ‘_&E(t).
Av(t) = (s&=s~)- (34)
u(t-k-l),Z(t-l),M(t-I)] (29)
Several researchers have reported [IS, 221 that the
which is identical in form to eq. (7). As shown in performance of a predictive controller designed with
758 J. DUAL MORNINGRED~~ al.

L > 1 is often not significantly better than one de- ( - AEf)koC~(t + 4

signed with L = 1. Choosing L = 1, eq. (34) can be m =F [T,(t) - z-(t)] -
PC,
rewritten in terms of p as

.=$+, f-k-1
Au(t) = ; R-k-_? E(t + i) (35)
i=k+l
c P2j
j=o

where E(t + i) is the ith element of E(t).

The measured concentration has a time delay
4. CSTR SIMULATION EXAMPLE d = 0.5 min. The objective is to control the measured
concentration of A, C,, by manipulating the coolant
The simulated continuous stirred-tank reactor
flow rate qc. This model is a modified version of the
(CSTR) process consists of an irreversible, exothermic
first tank of a two-tank CSTR example by Henson
reaction, A + B, in a constant volume reactor cooled
and Seborg [23]. In the original model, the time delay
by a single coolant stream which can be modeled by
was zero. The nominal parameter values appear in
the following equations:
Table 1. The highly nonlinear characteristics of the
CSTR can be seen in Fig. 1, which shows the open-
C,(t + 4 = q CC,,(t) - C,(t + 41 loop concentration responses to step changes in the

-
coolant flow rate. The coolant flow rate was changed

I 1
-E from an initial value of 100 1 min- i to 110, to 100, to
koCA(t + 4 exp RT(t)

90, and back to 100, at 8min intervals.

Table 1. Nominal CSTR parameter values

Measured product concentration CA 0.1 mol/l

Reactor temperature T 438.54 K
Coolant flow rate 4c 103.411 min-’
Process flow rate 4 1001 min-’
Feed concentration C A0 1 mol/l
Feed temperature To 350K
Inlet coolant temperature Tco 350 K
CSTR volume V 1001
Heat transfer term hA 7 x 10s calmin- K-’
Reaction rate constant ko 7 2 x 101Omin-l
Activation energy term E/R 1.x 1W K
Heat of reaction - 2 x 10s cal/mol
Liquid densities E$%, 1 x 103 g/l
Specific heats lcalg-‘K-’

I I 1 I I I
0.04!
5 10 15 20 25 30

time (minutes)
Fig. 1. Open-loop composition response.
An adaptive nonlinear predictive controller 759

4.1. The controller designs Parameter estimation

The sampling period of all process measurements Because the output model was an approximation to
was assumed to be 0.1 min. This sampling period the process, an assumption was made that its mis-
allowed the faster dynamics of the higher purity re- match with the process over different operating condi-
gion to be sampled about 4 times per dominant time tions could be accounted for with condition-depend-
constant. The time delay of 0.5 min (k = 5) was as- ent coefficients fli and Bz on the derivatives and a bias
sumed known. The feed flow rate 4, the feed temper- term 6. Thus, the output model was reexpressed as
ature T,, the coolant inlet temperature T,,, and the
reactor temperature T, were assumed to be accurately c_&) = e,,(t - 1) + BlAtfi(t) - /?2$&t)+&
measured. The inlet feed concentration CA0 was un-
measured and assumed to be constant at its nominal (42)
value. All controllers were designed with a rate-of-
The three coefficients, pII f12, and 6, were then estim-
change constraint of 2 1 min- ’ per sampling period
ated by RLS with directional forgetting. The initial
on the controller output.
values of the three parameter estimates were set to 1,
For the adaptive controllers, the first 6 min of each
1, and 0, respectively. The covariance matrix was reset
run were used to commission the estimator. During
whenever the estimation error exceeded 0.0004 mol/l.
this period, the process was run open loop with an
input PRBS of f 1 1min- I. The diagonal elements of
the covariance matrix were set to lo6 initially and Controller design
reset to this value whenever the estimation error The pole-placement method of design was used.
exceed a specified error tolerance. A PRBS of For this controller, eq. (43) was used as the output
f 0.5 lmin-l was also added to the controller out- function F [-I, and A(q-‘), the tuning polynomial,
put from the time the covariance matrix was reset was chosen to be (1 - 0.8q-I)‘. This choice attempts
until the parameters converged sufficiently to update to give the closed-loop system a 99.9% response time
the controller model. Following a covariance reset, of about 4 min. Based on the chosen A(q- I), R was set
another reset was not permitted for 2 min. The to 40. The remaining tuning parameters were L = 1,
forgetting factor for the estimator was chosen to f = 0, and Q = I. The prediction model was updated
be 0.8. with the estimates of Bi, p2, and 6 whenever the
estimation error remained below 0.0002 moljl for five
consecutive samples. The following equations de-
4.1.1. The adaptive nonlinear predictive controller scribe the prediction model for which the controller
Output model was designed.
Because the input q=(t) does not appear in eq. (36),
CA(t) is differentiated again. The resulting equation is d,(t) = 1.6c,(t - 1) - 0_64c,(r - 2) + v(t - k- 1)

%,(t+d)= - $+k,exp
[_ I>
-E
RT(t)
CAt + d)
where
(43)

F(t)C,(t + d). v(t - k - 1) = c,(t - 1) + BlAtfi(t) - p2 y,, + 6

(38) - 1.6&,(t - 1) + 0.64c,(t - 2). (44)

q=(t) appears in eq. (38) after the substitution of 4.1.2. The adaptive linear predictive controller
eq. (37) for p(t). A discrete output model in which qc (A LPC)
appears can be written using an implicit second-order
interpolation equation [24]: Prediction model
(At)2 . . A loo-term step response model was used for the
C,(t) = C,(t - 1) + A&,(t) - 2 C,(t). (39) predictions and the controller design. One hundred
terms were needed to provide a 93% complete re-
Equation (39) can be rewritten in a predictive form as sponse (180 for a 99% response). This prediction
the following implicit output model: model was obtained by long division of a second-
order difference equation model that was updated
(A02 on-line.
e”(t) = eA(t - 1) + Atfi(t) - Ffi(t) (40)

where Parameter estimation

^ .C Six parameters of a linear difference-equation
fi(t) = C,(t), fi(t) = C,(t) (41)
model were estimated: three numerator coefficients,
and fi(t) and f2(t) are functions of past outputs and two denominator coefficients, and a bias term. The
inputs. Equation (40) and the future values offi (t) and initial parameter estimates were zero. The covariance
f2(t) are evaluated using eqs (36)-(38) with the im- matrix was reset whenever the estimation error ex-
proved Euler method [24]. ceeded 0.00025 mol/l.
760 J. DUANE MORNINGRED et ui.

0.16

........ PI
- - set point

0.08

rime (minutes)

Fig. 2. Concentration set-point tracking.

0.16 -

--
.. . . . ALpc
. ... .. .. p1

, -

0 1

time (minutes)

Fig. 3. Measured and unmeasureddisturbance rejection.

Controller design parison with the ANPC:

The ALPC was designed with the following tuning
(45)
parameters: L = 1, R = 100, Q = I, and f = 9 times
the sum of the squares of the step response coefficients Auf f is the compensator output and Kff =
(fw 0.01). f was chosen so large because the control - 1.0 1min- ’ K-l is the compensator gain.
action was too vigorous for the set-point change to
increase CA. The prediction model parameters were 4.1.3. The PI controller. The initial PI controller
updated whenever the estimation error remained be- tuning parameters were selected using the ITAE cri-
low 0.0001 mol/l for five consecutive samples. In addi- terion for set-point changes. The controller gain was
tion, a static reactor-temperature compensator was adjusted to give a good response for the set-point
added to the controller output to provide a fair com- change to increase CA from 0.1 to 0.15 mol/l. The
An adaptive nonlinear predictive controller 761

0.16

0.06

0.04 I t I
10 20 30 443

time (minutes)

Fig. 4. Set-point tracking with modeling errors.

0.16
I I I

0.14

% 0.12
8
u’

0.10

0.08

I I I
10 20 30 44)

time (minutes)

Fig. 5. Measured and unmeasured disturbance rejection with modeling errors.

controller gain was 52 l* mol-l min-’ and the inte- the prediction model is as accurate as the second-
gral time was 0~46 min. In addition, the static reactor- order discretization in eq. (41) permits. As can be seen
temperature compensator in eq. (45) was again used. in Fig. 2, the corresponding response of the ANPC is
excellent: the responses to all of the set-point changes
4.2. Simulation results are symmetric and the response times are as designed
4.2.1. Small modeling errors in the ANPC. The first (about 4 min).
comparison is between the ANPC, the ALPC, and the Figure 3 compares the disturbance rejection per-
PI controller for set-point tracking performance. The formance of the same three controllers. The un-
set point was changed at 8 min intervals from measured feed concentration disturbance changes
0.1 mol/l to 0.15, to 0.1, to 0.05, and back to 0.1. The from 1.0 mol/l to 0.95 at 8 min, and back to 1.0 at
ANPC prediction model was designed using the cor- 24 min. The measured coolant temperature disturb-
rect values of all process parameters. ConsequentIy, ance decreases by 10°C at 16 min and returns to its
762 J. DUANE MORNINGRED etal.

nominal value at 32 min. Again, the nonlinear con- c41 Cutler, C. R. and Ramaker, B. L., 1980, Dynamic
troller easily outperforms the other controllers. matrix control-a computer control algorithm. JACC
Preprints, San Francisco, Paper WP5-B.
c51 Richalet, J. A., Rault, A., Testud, J. D. and Papon, J.,
4.2.2. Large modeling errors in the ANPC. The 1978, Model predictive heuristic control. Automatica
next simulation compares the set-point tracking per- 14, 81.
formance of the ANPC and its corresponding PI Clarke, D. W. and Mohtadi, C., 1989, Properties of
(nonadaptive) nonlinear predictive controller (NPC). generalizedpredictivecontrol. Automatica 25, 859.
c71 Ortega, R. and Tang, Y., 1989, Robustness of adaptive
Both controllers were designed with significantly er- controllers-a survey. Automatica 25, 651.
roneous model parameters: the activation energy was Bl Garcia, C. E., Prett, D. M. and Morari, M., 1989,
assumed to be 5% higher than the actual value, the Model predictive control: theory and practice--a sur-
rate constant was 15% lower, the heat of reaction was vey. Automatica 25, 335.
20% higher, and the heat transfer coefficient was 20% c91 Hoo, K. A. and Kantor, J. C., 1985, An exothermic
continuous stirred-tank reactor is feedback equivalent
higher. As can be seen in Fig. 4, the NPC controller to a linear system. Chem. Engng Commun. 37, 1.
becomes unstable without adaptation. The ANPC, Cl01 Wong, S. K. P. and Seborg, D. E., 1988, A control
although not yielding symmetric, critically damped strategy of SISO systems with time delays. Znt. J.
responses, still settles fairly quickly with reasonable Control 4S, 2303.
Cl11 Saatry, S. S. and Isidori, A., 1989, Adaptive control of
amounts of overshoot. linearizable systems. ZEEE Trans. autom. Control 34,
Figure 5 compares the disturbance rejection per- 1123.
formance of the ANPC and the NPC. Both con- WI Agarwal, M. and Seborg, D. E., 1987, A multivariable
trollers perform well under these circumstances. non-linear self-tuningcontroller. A.Z.Ch.E. J. 33, 1379.
Neither controller performs significantly better than Cl31 Economou, C. G., Morari, M. M. and Palsson, B. O.,
1986, Internalmodel control 5. Extension to nonlinear
the other for three of the four disturbances. The systems. Znd. Engng Chem. Proc. Des. Dev. 25, 403.
ANPC, however, provides a much better response to Cl41 Yeo, Y. K. and Williams, D. C., 1987, Bilinear model
the - 10°C changes in coolant temperature. predictivecontrol. Znd. Engng Ckem. Res. 26. 2267.
Cl51 Brengel, D. D. and Seider, W. D., 1989, Multistep
nonlinearpredictivecontrol. Znd. Engng Chem. Res. Zs,
5. CONCLUSIONS
1812.
A new adaptive nonlinear predictive controller has WI Georgiou, A., Georgakis, C. and Luyben, W. L.. 1988,
been proposed_ The controller is tuned in a manner Nonlinear dynamic-matrixcontrol for high-purity dis-
similar to linear model-predictive controllers or by tillation columns. A.Z.Ch.E. J. 34, 1287.
Cl71 Jang, S., Joseph, B. and Mukai, H., 1987, On-line
placing poles of the closed-loop response. The new optimization of constrained multivariable chemical
controller is computationally efficient and can per- processes.A.Z.Ch.E. J. 33, 26.
form well, even when initially designed with modeling WI Beouette, B. W. and Sistu, P. B. 1989, Disturbanceand
inaccuracies. Simulations show that the new control parameterestimationfor honlineardynamic systemsin
a predictive control framework. A.1.Ch.E. Annual
strategy can outperform PI control, adaptive, linear
Meeting. San Francisco.
predictive control, and nonadaptive, nonlinear pre- Cl91 Chen, S. and Billings, S. A., 1989, Representationsof
dictive control strategies. non-linear systems: the NARMAX model. Znt. J.
Control 49, 1013.
Acknowledgements-The authors gratefullyacknowledgethe
financialsupport of the National Science Foundation (NSF PO1 Kulhavy, R. and Kamy. M., 1984, Tracking of slowly
CBT-8605233). Appreciation is expressed to Mike Henson varying-parameters by- directional forgetting, in Proc.
for helpful discussions. ZFAC Congress, Budapest, p. 687.
WI Hunt, L. R., Su, R. and Meyer, G., 1983, Global trans-
formations of nonlinear systems. IEEE Trans. nutom.
REFERENCES Control 28, 24.
[l] &&rem, K. J. and Wittenmark, B., 1973, On self-tuning c221 Maurath, P. R., Mellichamp, D. A. and Seborg, D. E.,
regulators.Automatica 9, 185. 1988, Predictive controller design for SISO systems.
[2] Clarke, D. W. and Gawthrop, P. J., 1975, Self-tuning Znd. Engng Chem. Res. 27, 956.
controller. Proc. IEE 122, 929. ~231 Henson, M. A. and Seborg, D. E., 1990, Input-output
[33 Seborg, D. E., Edgar, T. F. and Shah, S. L., 1986, linearizationof generalprocesses.A.Z.Ch.E. J. 36, 1753.
Adaptive control strategies for process control: a ~241 Finlayson, B. A., 1980, Nonlinear Analysis in Chemical
sut~ey. A.Z.Ch.E. J. 32, 881. Engineering, p. 27. McGraw-Hill, New York.