Sie sind auf Seite 1von 9

J. Proc. Cont. Vol. 5, No 1, pp.

19-27, 1995

Copyright 1995 Elsevier Science Ltd


Printed in Great Britain. All rights reserved
0959-1524/95 $10.00 + 0.00

E M A N N

Model predictive control of an industrial packed


bed reactor using neural networks
Kwaku O. Temengfl* Phillip D. Schnelle* and Thomas J. McAvoy*
*E. L du Pont de Nemours & Co. Inc, Wilmington, Delaware, USA
~Department of Chemical Engineering, University of Maryland, College Park,
Maryland, USA
This paper discusses an industrial application of a multivariable nonlinear feedforward/feedback model
predictive control where the model is given by a dynamic neural network. A multi-pass packed bed
reactor temperature prone is modelled via recurrent neural networks using the backpropagation through
time training algorithm. This model is then used in conjunction with an optimizer to build a nonlinear
model predictive controller. Results show that, compared with conventional control schemes, the neural
network model based controller can achieve tighter temperature control for disturbance rejection

Keywords: model predictive control; packed bed reactor; neural networks


In the chemical process industry, the drive to reduce
operating costs and develop new markets has frequently
emphasized improvements in product quality, better use
of energy resources and reduced environmental emissions. These objectives, in turn, have placed stringent
requirements on the available process control systems.
These control systems must usually cope with multivariable process interactions, constraints on manipulated and controlled variables, as well as time delays
and other problematic dynamic characteristics.
Model predictive control (MPC) algorithms have
been recognized as effective tools for handling some of
the difficult control problems in industry 1,2. MPC
schemes derive some of their industrial appeal from
their ability to handle input and output constraints,
time delays, non-minimum phase behaviour and multivariable systems. Two popular variations of the model
predictive control algorithm are dynamic matrix control
(DMC) 3, and model algorithmic control (MAC) 4. The
underlying strategy of MPC algorithms is to use a
model to predict the future output trajectory of the
process and compute a controller action to minimize the
difference between the predicted trajectory and a userspecified trajectory.
Despite the success enjoyed by MPCs in industry,
there are some processes which pose a challenge for the
standard, linear model-based algorithms. For example,
batch and semi-batch processes are carried out over a
wide dynamic range; hence, the concept of operation
around a steady state becomes invalid. Also, there are
some continuous processes which undergo frequent
transitions to permit the manufacture of several grades

of a basic product. Such processes operate at several


steady state levels and may experience start-ups and
shutdowns on a daily basis, making the use of linear
models impractical. Lastly, there are some chemical
processes (e.g. some polymer reactors) which are so
severely nonlinear that small to moderate perturbations
around the steady state can render a linear model-based
controller inadequate or even unstable 5. Thus, there is
an incentive to develop extensions of MPC to tackle
nonlinear systems.
The goal of this paper is to demonstrate the nonlinear
model predictive control of a spent acid recovery
converter (a packed bed reactor), where the model of
the process is given by a dynamic neural network. We
call this scheme neural model predictive control. We use
nonlinear MPC because of the nonlinear and highly
interactive nature of the process. Furthermore, the lack
of some key measurements rendered the construction of
a reliable first principles model virtually impossible.
The paper is organized as follows: first we offer
reviews of MPC algorithms and nonlinear system identification methods, including neural networks and their
application to MPCs. We then describe the process and
the control problem, as well as the system identification
of the process. We devote the remainder of the paper to
the control structure, implementation details and
discussion of results.

M o d e l predictive control

Linear Jbrmulation
The essential elements of the various predictive control
schemes are: (1) at sampling time k, use an appropriate

t A u t h o r to w h o m correspondence should be addressed

19

Model predictive control using neural networks: K.O. Temeng et al.

20

model to predict future plant output over a predetermined horizon, assuming no further control action, (2)
compute the present and future control actions such
that some performance index is minimized over the
horizon, and (3) implement the first control action and
repeat the procedure at time k + 1.
Consider a single-input, single-output case without
loss of generality. Then the control calculation consists
of the following optimization problem:
mJn

~72(i)(y~P(k+i)-2(k+i)) 2

Au(k), ...,Au(k-M-1) i=l

(1)

+ 2~ A2 ( j ) ( A u ( k + M -

j))2

j=l

where ySp is the setpoint, )9 is the model prediction, y


and h are weighting factors, P is the prediction horizon
and M is the number of future moves to be optimized,
Au is the change in manipulated variables, defined as
au(k) = u(k) - u(k - 1). In general, M < P and therefore
u(k + M - 1) = u(k + M ) = ... = u(k + P - 1). Also, we
A
can write the model prediction, y, as:
J

.p(k + j ) = y * (k + j ) + SmiAU(k + j - i)
i=1

+ w ( k + j)

(2)

(j = 1,2,...P)

where y * ( k + .1) is the contribution to the future values


due to past input moves (up to time k - 1), Sm~ are the
step response coefficients for the manipulated variables
and w(k + j) captures all unmodelled effects, y * ( k + j)
can be written as:
N

y * (k + j ) = Y. SmiAu(k + j - i) + SdjAd(k )
i=j+l

(3)

( j = 1,2...P)
The second term on the RHS of Equation (3) is the
contribution of the most recent measured disturbance
change. To predict the output values using Equation
(2), w(k + j) must be estimated. This is carried out as
follows:
w(k+j)=w(k)=y(k)-y*(k)

( j = 1,2.... P)

(4)

where y(k) is the current measurement. The solution to


Equation (1) subject to Equation (2), reduces to a least
squares solution. When linear constraints are imposed,
the problem can be formulated as a quadratic program,
as was proposed by Garcia and Morshedi 6 in their
extension of DMC to obtain the popular variation
known as quadratic dynamic matrix control.
Nonlinear formulation

Nonlinear effects are introduced into the problem


formulation when the model prediction (~) becomes

nonlinearly dependent on any of the three terms on the


right-hand side of Equation (2) or when other nonlinear
constraints are imposed. Several attempts have been
made to extend MPC techniques to nonlinear systems.
There are essentially three major approaches: i.e. scheduled linearizati0n, extended linear MPC and explicit
nonlinear MPC.
In scheduled linearization7, the nonlinear model is
linearized around a steady state to compute the effect of
future manipulated variables on the predicted output.
However, the full nonlinear model is integrated at every
sampling step to compute the effect of past inputs and
future disturbances on the predicted output (i.e. y*(k +
j) in Equation (2) is obtained from a linear model). A
quadratic program (QP) is solved as in the linear case
to obtain the future manipulated variables. The algorithm has been successfully applied to a semi-batch
reactor.
In their extended DMC approach, Peterson et al. 8 update a linear model by computing a time-varying disturbance vector (w(k + j) in Equation (2)) which
accounts for the nonlinearities. The algorithm involves
the iterative solution of a quadratic program at each
sampling point until the linear model approximates the
output of the nonlinear model. The recommended
controller actions become perturbations around the
linear solution.
Finally, a number of control algorithms have been
proposed in which the model predictive control
problem is posed as a nonlinear program (NLP), incorporating an explicit nonlinear model (which can take a
variety of forms) and possibly nonlinear constraintsT M .
A ,
In this case y in Equation (2) becomes nonlinearly
dependent on the optimization decision variables
(Au(k + i), i = 0,1 . . . . M - 1). Posed in this fashion, the
problem becomes an NLP that can be solved via any of
several efficient nonlinear programming algorithms. In
this work, we represent the prediction model by a
dynamic neural network.

Dynamic nonfinear system identification


Review

A prerequisite of any explicit nonlinear MPC algorithm


is the availability of a model that captures the salient
nonlinearities of the plant under control. When the
dynamic nonlinear model structure is known (e.g. from
first principles) the problem of system identification
reduces to a parameter estimation problem which can
be solved with the extended Kalman filter, for example.
When the model structure is unknown or uncertain, a
relatively general model structure needs to be assumed.
Some nonlinear MPC schemes rely on detailed
fundamental models which are usually unavailable (due
to lack of fundamental knowledge of the process) or
very time consuming to build. For this work, the lack
of some key process measurements made it difficult to
construct a first principles model. Indeed, the cost of
model development and validation usually accounts for

M o d e l p r e d i c t i v e c o n t r o l u s i n g n e u r a l networks: K.O. T e m e n g

a substantial portion of the expenditure in industrial


advanced control projects. Hence the possibility of identifying nonlinear empirical models can become an
appealing alternative. Another motivation for using
empirical models is that, once constructed, these models
are much less computationally intensive than fundamental models which can comprise dozens of differential equations.
To identify systems with unknown or uncertain structures, it is common to choose a model set which can be
used to approximate a large class of nonlinear systems.
For example, the Hammerstein model structure, first
investigated by Narendra and Gallman 12, comprises a
static nonlinear function of the inputs in series with a
dynamic linear model. Indeed, Hammerstein models
have been successfully applied to the identification of
simulated distillation columns and an experimental heat
exchanger process 13. However, the model structure does
not accommodate systems which exhibit output nonlinearity.
Volterra series have been used as model sets for
nonlinear identification 14, but the dimensionality of the
parameter vector can be quite large for high order
series. Nonlinear auto regressive moving average
models with exogenous inputs (NARMAX) have also
been used to describe a large class of nonlinear systems.
These N A R M A X models may be obtained via polynomial expansions 15 or radial basis functions.
Recently, there has been interest in the use of neural
networks to learn the underlying phenomena of nonlinear systems u',lT. Since neural networks can learn by
example, and since they can approximate any nonlinear
continuous function 'arbitrarily' welP 8,19, they can offer
a cost effective way of developing useful process
models.
In fact the idea of incorporating neural network
models in the MPC algorithm has been proposed by a
number of researchers. Bhat and McAvoy 17 presented a
scheme whereby the linear convolution model is
replaced with a dynamic neural network. They offered
some hints for carrying out the required nonlinear optimization. Willis et al. 2 performed simulations showing
the superiority of a neural network based model predictive controller over a linear controller in distillation
column control. Also Hernandez and Arkun 21 used a
neural network in their extended D M C approach to
stabilize a system around an open-loop unstable point.
The corresponding linear version of the controller was
not successful in the same task. Lee and Park 22
employed a neural network to learn disturbance dynamics in an MPC algorithm. The feedforward control
scheme proved effective in simulated distillation column
and reactor control. Their results also show that the
neural network based controller exhibits desirable fault
tolerance and generalization capabilities.

etal.

21

computing units called neurons. A typical feedforward


network architecture is depicted in Figure 1. The
network generally consists of an input layer and an
output layer with one or more hidden layers between
them. Each layer contains an appropriate number of
neurons and all neurons in a layer are usually fully
connected to neurons in adjacent layers. The inputs to
the network are passed to neurons in the first layer,
whose outputs become inputs to the second layer and so
on.

The input-output relationship for a neuron can be


written as:
N

y = a ( E w i u i +0)

(5)

i=1

where y is the output, u are the inputs, w are the connection weights, 0 is the basis term, N is the number of
inputs and a is the activation function. A commonly
used activation or squashing function is the sigmoid:
1
a(x) - - l+e x

(6)

In the context of process control, a neural network


will be considered a means of approximating unknown
nonlinear, multivariable functions by a collection of
simple known functions. Consider a system whose
input-output relationship can be written as:
y =f(u)

(7)

Here, y is the output, u is a vector of inputs and f is an


unknown, possibly nonlinear function. We desire to
design a network to generate a function h which closely
approximates the unknown function f in Equation (7)
by using input-output data [u(i), y(i)]. To carry out
the identification, one can write:
= h (u, w)

(8)

where ;~ is the network prediction of y (Equation (7))


and h is a vector of known nonlinear functions obtained

Outputlayer
~

cWenlngh~
~idn "-I~
neuron

(~

idden layer

(~ ,nputlayer

Nonlinear identification using neural networks


Artificial neural networks are computing paradigms
which consist of adjustable weights linking simple

I~1
Figure 1

1~2

I ~3

Architecture of feedforward neural network

22

Mode/predictive control using neural networks: K.O. Temeng et al.

by the superposition of special functions and w is the


weight vector to be determined. These special functions
are usually sigrnoidal (see Equation (6)) but can be
polynomials or linear functions.
The estimation problem then involves the determination of the best (in a least squares sense) weight vector
w, such that h(u,w) closely matchesf(u). This training of
the network, using measurement data is achieved via a
solution of an optimization problem:
Q

J = min Y~(Yi- h(ui,w)) 2


W i=1

(9)

where Q is the number of input-output pairs in the


training set. Thus the weights are chosen to minimize
the sum of squares of the difference between the
observed output and the network prediction. In general,
a gradient descent type algorithm is used to solve
Equation (9). An efficient method for updating the
weights (i.e. training the network) is the backpropagation algorithm which offers a convenient way of
computing derivatives for evaluating the gradient
descent direction. A derivation of the algorithm is available in Rumelhart eta/. 23.
The feedforward architecture shown in Figure 1 is
typically used for steady state functional approximation
or one-step-ahead dynamic prediction. Recall that the
MPC methodology requires a dynamic model which
can predict with reasonable accuracy over a horizon.
However, feedforward networks using the one-stepahead structure generally perform poorly over a trajectory because errors are amplified when inaccurate
network outputs are recycled to the input layer24.
Schemes which use a long vector of past inputs and
outputs (i.e. where measured data from k - P to k are
used for prediction up to k + P) tend to suffer from
over-parameterization, even though methods such as
partial least squares can be used to reduce the dimensionality of the input vector to improve performance.
To improve prediction over a horizon, we use timelag recurrent networks25. The concept is illustrated in
Figure 2. Unlike a conventional feedforward network
where information flows only from the input layer to
the outputs, recurrent networks incorporate delayed
information flow back to preceding layers. The functionality is similar to that of a NARMA model whose
output depends solely on past predictions and external
stimuli. A network that is trained in this mode retains
the ability to predict process behaviour with a consistent degree of accuracy over the entire horizon 26.
Various algorithms for training such recurrent networks
a r e a v a i l a b l e 25,27,28.

An effective training algorithm for recurrent


networks is the backpropagation-through-time algorithm which has been applied successfully to the long
term prediction of the dynamic behaviour of a wastewater treatment facility24. In this work, we model the
dynamic behaviour of an industrial packed bed reactor
using a recurrent neural network. The training of the

network was carried out via the backpropagationthrough-time algorithm, a derivation of which can be
found in Werbos 25 and Suet a[. 26.

Process description
The spent acid recovery (SAR) converter is shown in
Figure 3. Cold feed containing sulfur dioxide (SO2),
oxygen (O2) and inerts is sent to a series of four vanadium pentoxide catalyst beds where reaction to sulfur
trioxide (SO3) takes place. Conversion is carried out in
four stages or passes to maximize reaction efficiency,
and in each stage, the inlet temperature is controlled.
The reaction is exothermic, necessitating heat exchangers to cool the hot gases leaving each stage. The transferred heat is then used to heat the cold feed to the
proper temperature to initiate reaction in the first pass.
The gases leaving the third pass are cooled by transferring heat to steam, thereby generating superheated
steam.
The reactor system also provides a means to divert
portions of the cold feed through a network of valves,
dampers and piping to control the temperature inlet of
each pass. For example, the temperature of the feed to
the first pass can be lowered by opening valve C. This
permits a fraction of the cold feed to bypass the heat
exchangers completely, and depending on the split, it
can have an appreciable effect on the temperature of the
gas entering the first pass of the converter.
To achieve optimum conversion, the inlet temperature for each pass must be controlled. In general, the
optimum inlet temperature for each pass is that which
gives the greatest temperature rise across the bed. The
optimum temperature depends on the conversion
achieved in the preceding passes, the gas strength (SO 2
concentration) and production rate. Furthermore, to
prevent quenching of the reaction, the inlet temperatures are maintained above some threshold values at all
times. Also, to extend the active life of the catalyst, the
bed temperatures are constrained below some
maximum levels.

Feedback. , ~

Outputlayer

~
(~
Figure 2

idden

(~

Architecture of recurrent neural network

layer

Inputlayer

Model predictive control using neural networks: K.O. Temeng et al.

valve C

Xchg(I
~a%~ Interm.(~
.~i~ xchg
valve B

S02 + 02~.

Cold(,
Xchg"

rf

"o

SO3

23

The control problem is to regulate the inlet temperatures of the first, second and third passes of the spent
acid recovery converter by manipulating the flow rate of
cold feed through valves A, B or C. Note that since
there are no flow rate measurements, the air loadings to
the valves are used as the manipulated variables. The
complication that arises is that there is no unique relationship between the loading to a valve and the flow
rate through that valve. However, a unique set of loadings to the valves gives rise to a reproducible flow rate
through each valve, assuming a constant total flow rate.

@,,@,@

Figure 3 The process schematic

To gain a sense of the interaction that can be induced


by manipulation of a single valve or damper, consider
the following scenario. Suppose the temperature inlet of
the first pass is higher than the target and must be
lowered. An operator would typically open valve C to
bypass more of the cold feed directly to the inlet of the
first pass. Initially, the second and third passes should
heat up due to increased heat exchange between the hot
reaction products and the cold feed. Howeve;', when the
inlet temperature of the first pass has been sufficiently
lowered, the exit temperature of the first bed will be
slightly lower, causing the inlet and outlet temperatures
of the second, third, and fourth passes to be lowered as
well. But, since all passes are cooler, the cold feed is
unable to pick up enough heat from the heat exchangers, further cooling the first pass inlet temperature. This
effect cascades down the column once again and may
not be fully evident for about two hours. Under some
operating conditions, the interactions give rise to open
loop instability where the progressive cooling results in
extinction of the reaction.
The configuration of valves and piping also creates
interactions among the manipulated variables. Since the
valves merely redirect portions of the cold feed through
different heat exchangers, a change in the air loading to
a single valve affects the flow rate through all other
valves (even when the loadings to the remaining valves
are unchanged). In other words, the manipulated variables are not independent.
These interactions pose a challenge for decentralized
controllers, which partly accounts for the unsatisfactory
control the plant experienced with schemes based on
single loop controllers. Other process problems include
slow dynamics and frequent disturbances. These disturbances include changes in gas strength, blower speed,
and oxygen concentration. Also, environmental and
economic constraints place high demands on the controller performance. These environmental constraints
limit the amount of unreacted sulfur dioxide that can be
discharged into the atmosphere. An inefficient control
system can result in low conversion and a high emissions rate, leading to cutbacks in the production rate.
Also, all unreacted SO 2 represents a yield loss that must
be made up by either burning sulfur or importing
oleum.

Process identification
Various dynamic response tests were carried out in open
loop to generate data for constructing the dynamic
model of the spent acid recovery converter, The identification experiments entailed the use of pseudo random
sequences (PRS) to drive the three valves (A, B and C)
that comprised the manipulated variables. In general,
PRS signals are preferred over step testing because of
the latter's susceptibility to disturbances, even though
the duration of a PRS test is generally much longer. The
wide dynamic range of data obtained from a PRS test
can also lead to more accurate models. Finally, plant
personnel are more tolerant of PRS tests, since the
process is typically kept near the nominal operating
point.
The pulses had a duration of 5 min and each experiment lasted at least 12 h. The duration of the PRS test
had to be at least five times the dominant time constant
of the process. In this application, the longest time
constant was 40 min, which ensured that 12 h would be
sufficient to excite the process. The proper amplitudes
of the PRS were estimated from historical process data,
with a general objective of achieving at least 5-10C
change in the first pass inlet temperature. Preliminary
experiments were then used to refine the initial estimates. The experiments generally consisted of manipulating a single valve at a time. The other valves were set
at levels dictated by process conditions. A small number
of multivariable tests, where all three valves were
manipulated simultaneously, was also conducted to gain
a sense of interactions among the inputs. The tests were
run over a wide range of operating conditions to
capture process nonlinearities. However, to minimize
excursions in the output variables that could violate
standard operating conditions, operators were empowered to manually manipulate the remaining valves,
but only as a last resort. The set of variables shown in
Table 1 were used in modelling the process.
To gain insight into the extent of process excitation
during the experiments, refer to Figures 4 and 5. Figure
4 shows the test signals applied to valve A for a portion
of an experiment, whereas Figure 5 illustrates the
responses of the first, second and third pass inlet
temperatures, respectively.
The backpropagation-through-time algorithm was
used to generate the seven-input, three-output recurrent

Mode/predictive control using neural networks: K.O. Temeng et al.

24

neural model. The architecture comprised three layers,


with four nodes in the hidden layer and unit time
delayed output feedback connections to the input layer.
To permit cross-validation, the entire set of data was
partitioned into training and testing sets. Generally, the
training data were used to adjust network parameters,
but the convergence of the network was determined by
the output error of the test data. The training and test
data sets generally consisted of at least 150 exemplars.
For a representative test set, the network weights that
produced the global minimum of the output error were
considered to be the optimal set of weights. This crossvalidation was used to minimize over-parameterization
and over-training of the network.
Model validation was performed by prediction on
novel data. The quality of the model can be gauged

from its prediction for the first pass inlet temperature as


shown in Figure 6. The solid line tracks the model
output while the dashed line indicates actual process
reading. (Note: the model in this case is initialized with
actual process data at time -- 0 and receives no state
information thereafter.)

The controller
The general philosophy of neural model predictive
control is the same as that of any available MPC
package. The control law is essentially the optimization
of an objective function, such as is given in Equation
(1). In this work, however, ~ the prediction model, is
given by:

(k+j)=f(~,u,d)+w(k)

Table 1 Modelling variables


Input variables

Output variables

Feed forward:
SO2 concentration
0 2 concentration
Blower speed
Loading of valve D

First pass inlet temperature


Second pass inlet temperature
Third pass inlet temperature

Manipulated:
Loading of valve A
Loading of valve B
Loading of valve C

100

90

!o
80

g 7o
50

10] J

. . . . . . . . . . .

200

,
400

600

'

'

800

'

'

'

1000

1200

Time (minutes)

Figure 4 Loading to valve A during identification experiment


510

,
1

500-~

( j = 1,2.... P)

(10)

where f represents a dynamic neural network function,


d is a vector of measured disturbances and u is a vector
of manipulated variables.
Figure 7 shows a general schematic of the controller.
At every sampling step, the past and current measurements of the controlled and manipulated variables are
input into the neural network model. Using the last
vector of recommended manipulated variables, the
model calculates the trajectory of the process outputs
over the horizon. The predictions are fed into the optimizer where the objective function is evaluated. The
optimizer computes a new set of manipulated variables
and passes them back to the neural network model.
This iteration continues until the calculation converges.
The first set of recommended control moves are then
implemented and the procedure is repeated at the next
sampling step.
The nonlinear optimizer used for selection of the
future control moves was Automated Design Synthesis
(ADS), a public domain software package obtained
from the Naval Postgraduate School in Monterey,
California. The optimization strategy used was sequential unconstrained minimization using the quadratic
extended interior penalty function method. The specific
optimization method was the Broydon-FletcherGoldfarb-Shanno (BFGS) variable metric method. The

3rdpass temp
470

490~
480-t

~"

460

2nd pass temp

460

",'

F~

..... Model O u t p u t

450

450.

",, [

ProcessMeasurement

"L/ < .....


440

430 :
0

'

'

'

i
200

'

'

'

i
400

'

'

'

'

'

'

600

i
500

'

/
1000

'

, ,
1200

Time (minutes)

430
.

JO0
j

20~0

Time Steps

Figure 5 Temperature responses during manipulation of valve A (see


Figure 4)

Figure 6 Model prediction for first pass inlet temperature

3O0

Model predictive control using neural networks: K.O. Temeng et al.


Controlled, Feedforw~rd
and
Manipulated Variables

Model
Prediction

_l

- Dynamic Neural
~_ Network Model

-t
Future
Manipulated
Variables

Nonlinear
Optimizer

To Plant

composition disturbance. This is due to the slow


dynamics of the third pass temperature. However, the
controller is able to reject the effects of the disturbances
and maintain satisfactory control for the remainder of
the period.
The spike in the SO 2 concentration profile at t = 1300
coincided with the daily calibration of the analyser. A
rule had been incorporated in the expert system to
prevent the controller from reacting to such rapid loss
of the signal. In this case, the signal from the analyser
was not updated in the control calculation until normal
conditions were re-established.
The profiles of the manipulated variables are shown
in Figure 10. An important observation is that between
t = 600 and t = 1000 (during the period of both the SO 2
concentration and blower speed upsets) valves B and C
are constrained at 30% and 100%, respectively. This
means that only valve A is actively being manipulated
to reject the severe disturbance. The feedforward controller's success in regulating three outputs with a single
manipulated variable lies partly in the fact that the
manipulated variables are not independent. Variation of
the loading for valve A also affects the flow of cold feed
through valves B and C and permits a measure of feedback control for all three outputs.

t Disturbances
fl~,+ Setpoint
MJ'~

Present
Manipulated
Variables

Figure 7

25

Controller structure

one-dimensional search method was minimization by


polynomial interpolation/extrapolation.
For the control calculation, P (the prediction
horizon) and M (the control horizon) were 20 and 5,
respectively. The optimizer operated on 15 independent
optimization parameters (three variables at five points
in time). Input constraints (usually with bounds of 30%
and 100% open) were enforced for all valves. We did
not implement any output constraints. Using the neural
model, this optimization consumed about 15 s of real
time on a VAXstation 3100. This was an adequate
performance since the controller was only required to
run once every 10 min. An on-line expert system was
configured to handle timing, check for availability and
validity of process measurements and coordinate other
controller functions.
The optimizer was initialized with the most recent
optimum values for the decision variables. If the optimizer failed to converge in some specified number of
control and model iterations (say, 2000), the controller
action defaulted to the 'old' or last optimization results.
This happened in practice only when the process was
run outside of where the model was defined.

Comparison o] conventional control approaches


The neural MPC performance was compared, via simulation, to conventional process control methods. Two
480
460-

440-

set point

420(a) 1st Pass Inlet Temp


400
200

400

600

800

10'00

12'00

14'00

'1600

500-

Discussion

-,,..7,'=-

480-

The neural MPC application was placed in closed-loop


service in August 1991 and enjoyed a period of high
utility rate. Figure 8 depicts the profiles of the process
outputs during the first 24 h of operation. Closed-loop
service was initiated at t -- 170 min. An inspection of
the profiles reveals that the controller performance is
quite satisfactory for the regulatory problem. In
general, the response of the first pass temperature is
fastest whereas the third pass temperature exhibits the
slowest dynamics. The ability of the feedforward
controller to reject disturbances is borne out by Figure
9. Between t = 500 and t = 900, the SO 2 concentration
drops by roughly 15% (a major feed upset by plant
standards) while the blower speed (an indication of
throughput) shows variation throughout the period.
Figure 8 shows that the greatest offset is suffered by the
third pass inlet temperature in response to the feed

set point

~460440-

(b) 2nd Pass Inlet Temp

420

200

400

600

800

10'00

,2'00

,4'00

1600

14'00

'1600

480 460-

o'440420(c) 3rd Passlnlet Temp


4OO
200

400

600
800
10'00
Time(minutes)

Figure 8 Temperature responses in closed loop

12'00

Mode/predictive control using neural networks: K.O. Temeng et al.

26
8

100

7=

80-

6~

: 60-

5.

~ 40-

4~

20-

3.

(a) S02 Concentration

(a) ValveA
2()0

400
'

600
'

800
'

10100

12100

14100

1600

10100

12~00

1'lO0

1600

1 O0

1200 1400 1600

lOO

1.
0

200

400

600'

800

IOIO0 12100 14'00 1600

60o
-#

6000 ,

40-

200

58001

(b) ValveB

2(~0

400

600

800

100

80-

5400

ol

_ 60o

5200

- -

(b) Blower Speed

5000

2;0

400' 600' 800' 10'00 12'00

40(c) ValveC

201400 1600

Time (minutes)

200

400

600

800

Time (minutes)

Fi~aJre 9 Disturbance profiles in closed loop


Figure 10 Manipulated variables

conventional control strategies were tested, using the


neural network model of the packed bed reactor as the
process to be controlled.
The first approach was developed based on an understanding of how the process was manually operated. This
will be referred to as the 'seat of the pants' (SOP) design.
This control scheme was optimally tuned by nonlinear
optimization methods using the same objective function
and disturbance as the neural MPC test method. This
control design is shown schematically in Figure 11.
Another conventional control design is shown in
Figure 12. This design is referred to as the best conventional control (BCC) and is based on an in-house
control design method. The input-output pairings, as
well as the feedforward element and decouplers (collectively labelled impulse relays in Figure 12) were determined by optimization. Here also, the controllers were
optimally tuned to reject the effects of disturbances
using nonlinear optimization.
The performance of the various control schemes is
summarized in Table 2 using the integral squared error
(ISE) as the key measure.
For the open loop case, where no control action was
taken, the ISE was 0.805. The best performance was
obtained with the neural network MPC control
(NNMPC) system, with an ISE of 0.223. The SOP
scheme naturally outperformed the open loop case but
due to controller interactions and lack of disturbance
measurement, the ISE of 0.594 is the worst among the

closed loop simulations. The ISE for the BCC design is


0.277, which compares favourably with the neural MPC
controller.
The very good performance of the BCC system, relative to the neural MPC controller, is partly due to the
fact that the process model was essentially linear in the
range within which the simulations were carried out.
Despite its performance, the BCC scheme will be more
difficult to implement in the actual plant because tuning
and maintenance will be more time-consuming.

Xchg(l~- =

2nd Pass.

v,~,* interm.(d}=
Xchg
valve

,~
IIIIIIIIIIIlll

Ste

Cold( b,=

'

SO2 + 02~
(~

Figure 11 SOP control design

- Temp.Controller

SOs

Model predictive control using neural networks: K.O. Temeng et

al.

27

was carried out, for developing the Fortran-based software package used to train the recurrent network. We
also thank Eric Bauer and Wayne Schafllein, both at
DuPont, for their valuable contributions.
Q

Iv.2!

i:r .... o

References
~alve~

interrn.(t~)Xchg

'

S03

- impulSeRelay

Figure 12 Best conventional control design


Table 2

Results and comparisons of control schemes

Open loop (no control)


SOP design 'optimal' tuning
Best conventional control design
NNMPC

0.805
0.594
0.277
0.223

Summary
We have demonstrated a practical, industrial application of a multivariable nonlinear feedforward/feedback
model predictive control where the model is given by a
dynamic neural network. The plant is highly interactive
and has manipulated variables that are correlated. The
process outputs were modelled via recurrent neural
networks using the backpropagation through time
training algorithm. It was shown that identification
using recurrent structure permits prediction over a
horizon without degradation and is well suited to the
model predictive control scheme.
Closed loop results indicate that the neural-based
controller can achieve tighter regulatory control than is
possible with decentralized single loop controllers.

Acknowledgements
The authors are indebted to Hong-Te Su, a graduate
student at the University of Maryland when this work

1 Garcia, C. E,, Prett, D. M. and Morari, M. Automatica 1989, 25,


355
2 McAvoy, T. J., Arkun, Y. and Zafiriou, E. (Eds) 'IFAC
Workshop Proceedings on Model Based Process Control',
Pergamon Press, Oxford, 1989
3 Cutler, C. R. and Ramaker, B. L. AIChE National Meeting,
Houston, TX, 1979
4 Richalet, J. A., Rault, A., Testud, J. D. and Papon, J. Automatica
1978, 14, 413
5 Economou, C. G., Morari, M and Palsson, B. Ind. Eng. Chem
Process Des. Dev 1986, 25, 403
6 Garcia, C. E. and Morshedi, A. M. Chem. Eng. Commun. 1986,
46, 73
7 Garcia, C. E. AIChE Annual Mtg., San Francisco, CA, 1984
8 Peterson, T., Hernandez, E., Arkun, Y. and Schork, F. J. Proc.
Amer. Control Conf., 1989
9 Brengel, D. D. and Seider, W. D. Ind. Eng. Chem. Res. 1989, 28,
1812
10 Li, W. C. and Biegler, L. T. Chem. Eng. Dev. 1989, 67, 562
11 Sistu, P. B. and Bequette, B. W. AIChE Annual Mtg., Chicago,
IL, 1990
12 Narendra, K. S. and Gallman, P. G. IEEE Autom. Cont. 1966,
AC-11, 546
13 Eskinat, E., Johnson, S. H. and Luyben, W. L. AIChE J. 1991,
37 (2), 255
14 Fu, F. C. and Farison, J. B. Int. J. Cont. 1973, 18 (6), 1281
15 Leontaritis, I. J. and Billings, S. A. Int. J. Syst, Sci. 1988, 19, 519
16 Chen, S., Billings, S. A. and Grant, P. M. Int, J. Cont. 1990, 51
(6), 1191
17 Bhat, N. and McAvoy, T. J. Comput. Chem. Eng. 1990, 14 (5),
573
18 Cybenko, G, Math. Cont. Sig. Syst. 1989, 2, 303
19 Hornik K., Stinchcombe, M. and White, H. Neur. Netw. 1989, 2,
359
20 Willis, M. J,, Di Massimo, C. Montague, G. A., Tham, M. T.
and Morris, A. J. AIChE Annual Mtg., Chicago, IL, 1990
21 Hernandez, E. and Arkun, Y. Proc. Amer. Control. Conf., 1990,
2454
22 Lee, M. and Park, S. AIChE J. 1992, 38 (2), 193
23 Rumelhart, D. E. and McClelland, J. L. 'Parallel Distributed
Processing: Explorations in the Microstructure of Cognition',
MIT Press, Cambridge, MA, 1986
24 Su, H. T. and McAvoy, T. J. Proc. Amer. Control Conf. 1991
25 Werbos, P. J. Proc. IEEE. 1990, 78 (10), 1550
26 Su, H. T., McAvoy, T. J. and Werbos, P. J. I&ECRes. 1992, 31,
1338
27 Pineda, F. J, Neur. Computat. 1989, 1, 167
28 Williams, R. J. and Zipser, D. Neur. Computat. 1989, 1,270
29 Werbos, P. J. Neur. Netw. 1988, 1,339

Das könnte Ihnen auch gefallen