Predictive Control of Quality in Batch

Polymerization Using Hybrid ANN Models

Andy Yen-Di Tsen, Shi Shang Jang, and David Shan Hill Wong
Chemical Engineering Dept., National Tsing Hua University, Hsin-Chu, Taiwan, 30043 R.O.C.
Babu Joseph
Chemical Engineering Dept., Washington University, St. Louis, MO 63130

Two issues involving the methodology used for on-line control of product quality in
batch manufacturing processes are addressed: the generation of fast, data-driven process
models and the use of such process models for on-line feedback control of product
quality. The methodology is investigated using the example of the control of dispersity
and molecular weight distribution in a batch reactor for emulsion polymerization of
vinyl acetate. A n artificial neural network (ANN) is used as a model to predict the
quality as a function of the manipulated variables and on-line measurements. This
model is constructed using an augmented dataset that integrates experimental informa-
tion and knowledge from a mathematical model. The proposed model is compared with
other types such as a theoretical model whose key parameters are fitted to experimental
data. The hybrid A N N is superior to the parameter-fittingapproach for this case. Exper-
imental and simulation studies confirm the advantage of using the proposed model and
the predictive control algorithm.

Introduction
Objective
Batch processes are especially suited for manufacturing of Artificial neural networks have been suggested as an alter-
small amounts of high value added products. Quality control native for generation of process models. These models are
in batch processes presents a uniquely challenging problem data-driven and have the advantages of computational effi-
because the product quality is not known until the batch ciency and ease of construction. They also provide conve-
processing is completed. One is naturally forced to use pre- nient means for on-line adaptation as shown by Joseph and
dictive models to compensate for undesirable disturbances Wang Hanratty (1993). Since batch processes are often poorly
detected through intermediate measurements while the batch understood physically, the artificial neural network (ANN)
is in progress. Due to the complex nonlinear nature of seems to be a promising modeling tool for modeling such sys-
batch-process dynamics, a model capable of accurately pre- tems.
dicting the final product quality is usually difficult to gener- Joseph and Wang Hanratty (1993) developed a “shrinking
ate. Feedback control therefore requires significant operator horizon” predictive control scheme applied to an autoclave
experience and input. Statistical process control (SPC) and curing process for manufacturing composites using an ANN
statistical quality control (SQC) provide systematic ways of model. In this approach, ANN predictive models were devel-
monitoring process status (Klein, 1991; Levinson, 1992) but oped from simulation runs. Corrective control actions were
does not address the issue of on-line feedback corrections based on predictions of the ANN model and intermediate
when the quality measurement is not directly available. Due measurements. They showed that tight quality control can be
to the strong nonlinearity exhibited by most batch processes, achieved using model-based on-line feedback corrections to
there is still much room for development of automated deci- the batch recipe. Their studies were limited to simulated
sion-making tools for batch quality control. processes. They also did not address an important issue of
generating the data required for training the ANN model.
Correspondence concerning this article should be addressed to S . S . Jang. Normal operating data cannot be used for fitting models be-

AIChE Journal February 1996 Vol. 42, No. 2 455

cause of the lack of variability and correlation among the Chan, 1989; Kramer and Leonard, 1990), data reconciliation
variables. On the other hand, the design and execution of a (Whitely and Davis, 1992), and process control (Pao, 1989;
large number of experiments to generate data for modeling is Bhat and McAvoy, 1990; Hernandez and Arkun, 1992; Psi-
expensive and time-consuming and may not be acceptable in chogios and Ungar, 1991, 1992). Sarle (1994) and Ripley
a production environment. (1993) interpret that ANNs are nothing more than nonlinear
The objective of this work is to evaluate the ANN model- regression and discriminate models.
based predictive control approach for the product quality A number of workers have addressed the issue of combin-
control in a batch reactor for the emulsion polymerization of ing fundamental mathematical models with data-driven re-
vinyl acetate. A new approach for creating a hybrid ANN gression models such as ANN. Thompson and Kramer (1994)
model that combines scarce experimental data with prior classified the approaches into two categories: the process
knowledge encapsulated in approximate fundamental, first- model “in series”to ANN and the model “in para1lel”to ANN.
principles models is developed and tested. The model is ob- Psichogios and Ungar (1992) implemented an “in series”
tained by “training” an ANN model based on an augmented hybrid ANN to perform tracking of a biochemical reactor. In
data set. The augmented data set combines the information their approach, a neural network is used to estimate parame-
from the experimental data set and the “trends” predicted by ters of a first-principles model. It was found that this hybrid
a rigorous model built from fundamental physical and chemi- ANN requires less training data, and extrapolates and inter-
cal phenomena taking place in the system. This approach is polates better than standard “black-box” models. However,
applied to a bench-scale unit for the batch polymerization of in this approach, it is implicitly assumed that all essential
vinyl acetate. characteristics of the process had been captured in the first-
principles model.
Polymerization model Kramer et al. (1992) classified prior knowledge into two
forms: “hard constraints” (such as mass balances) that the
The vinyl acetate polymerization reactor is an interesting network response must adhere to, and “default models,”
paradigm for studying batch process control. Considerable which are idealized theoretical models of system behavior.
past work has been done on trying to model this process. A Hard constraints are incorporated by formulating the train-
complex first-principle model for the batch polymerization of ing problem into a semiinfinite programming problem. A
vinyl acetate was derived by Penlidis et al. (1985). The model connectionist model is trained to learn the residual between
was implemented to compute the optimal control profile by training targets and default model predictions. In this case,
Jang and Yang (1989). The experimental results show that the ANN model is “in parallel” to the process model. Radial
the model is acceptable in predicting monomer conversion basis function networks are used to ensure that the combined
but unreliable in predicting quality of the product, that is, model extrapolates according to the default model for novel
number average molecular weight and dispersity. Penlidis et cases outside the training data set. This hybrid ANN model
al. (1988) found that impurities affect the performance of the was used to predict the behavior of a simulated vinyl acetate
reactor in two ways. Impurity dissolved in water will deter-
polymerization reactor.
mine the induction period before the polymerization starts to
take off. Impurities dissolved in the monomer may affect such
final conditions of the batch as end time and molecular
weights. In general it is difficult to develop mathematical The Batch Process Quality Control Problem
models to quantify the precise effects of these impurities on We start with a general statement of the batch product
product quality. Hence it is desirable to seek empirical or quality control problem, and then consider a simplified ver-
semiempirical models that can capture the actual behavior of sion that was used in the application study on hand. These
a given reactor system. Even if precise mathematical models simplifications are not necessary but used for convenience
are available on-line, estimation and adaptation may be only.
needed since critical measurements are often not available. Consider a batch process with an operating period 0 I t 5
In such cases, tools such as the extended Kalman filter must tr. Its operating objective Y’[ff,xf; ~ ( t )u(t),
, d ( t ) ]would be
be used in conjunction with the nonlinear model of the a function of the end time tf, state vector of the system at the
process (Kozub and MacGregor, 1992). end of the batch xf, the state vector trajectory x ( t ) , the ma-
In this work, the model developed by Penlidis et al. (1985) nipulated variable vector, u(t), and some unknown disturb-
is used as the basis for developing a hybrid ANN model. The ance vector d(t). Assume that the batch operation only con-
purpose is to demonstrate how experimental data can be used siders the quality constraints q [ y ( t f ) ] 20, where y are the
to augment model accuracy when prior knowledge is incom- measurements and functions of state vector x only. A general
plete. The hybrid ANN model is, in turn, used to perform optimal control problem for the batch process is to compute
feedback predictive control based on the strategy proposed the control policy u ( t ) such that:
by Joseph and Wang Hanratty (1993).

Hybrid Ah”
Recently, the use of an ANN to capture nonlinear dynam-
ics and serve as a “black-box” model has been extensively
subject to system dynamics:
researched. ANNs have been applied to diverse areas of
chemical process engineering such as fault diagnosis (e.g.,
Hoskins and Himmelblau, 1988; Venkatasubramanian and (OP1.C1)

456 February 1996 Vol. 42, No. 2 AIChE Journal

with the end product quality cal model cannot be evaluated accurately. Hence predictions
based on first principles only are generally not quantitatively
accurate enough for use in predictive control. It is very useful
if the product quality model can be augmented by a data-
and quality constraints driven model such as ANN. The most extreme case is to de-
velop the following ANN models:
(OP1 .C3)

In addition some intermediate measurements of the process

are also available:

based entirely on experimental data. Other approaches are

Model predictive control could be performed only if accurate also feasible. For example, one could consider building a sin-
models for the system dynamics, f, the relation between gle ANN model to predict the quality and the objective func-
measured variables and the quality of the final product q, as tion at the same time. Using the preceding approach, the op-
well as the relation between measured variables and the state timal control problem (OP2) can be rewritten as
of the system, g, were known. However, such models are
generally unavailable.
To evaluate the model predictive control strategy we con-
sider a simplified version of the preceding problem adapted
to the vinyl polymerization process. In batch polymerization, subject to the identification of
usually only discrete manipulative actions are taken. We con-
sider a one-shot control policy whereby the feedback control
is based on a single intermediate conversion measurement, z.
Further, only one primary disturbance in the feed is consid- and the quality constraints
ered in this study. The method can be extended to the case
of multiple control actions, multiple measurements, and mul-
tiple disturbances. The manipulative action is classified into q = ANN&, xf ,xo,uo,d , u1) . (OP3 .C2)
two classes, those taken before the intermediate measure-
ment {uo} and those taken after {uJ. Thus, the on-line pre- The model predictive control strategy is implemented as fol-
dictive control problem can be rewritten in the following form: lows. We start with a nominal control policy. After some time
has elapsed, a sample is taken from the polymerization reac-
tor and the conversion is measured. This measurement is used
in conjunction with the models just given to decide on the
corrective action necessary. The accuracy of the corrective
subject to the identification of action can only be evaluated after the batch is completed.
The main drawback of the preceding approach is the large
d = d(z), (OP2.Cl) number of experimental runs required to train such A ” s .
Hundreds of data points are needed to train a network with
the system model: five or six input variables. This is generally not feasible in
industrial processes. If a first-principles model that can pre-
dict the “trend” of the objective function and the constraints
is available, then it can be used to generate “augmented data
points” for training. The next section describes how this aug-
and the quality constraints: mented data set is created and used to train what we call a
“augmented data A“”model.

Augmented Data ANN Model

A further assumption made here is that the disturbance is
observable, that is, we can estimate it using z. Thus any dis- A rigorous fundamental model for the vinyl polymerization
turbances entering the system that do not affect z cannot be process is available in Penlidis et al. (1985). Details of the
corrected by this control strategy. If there are other disturb- model are available in their original work, and hence are not
ances on the process, then additional measurements will be reproduced here. This is a fairly complex model and its use
necessary so that the effect of these disturbances can also be for on-line control would require substantial computational
included in the prediction of product quality. resources.
Although a qualitatively sound first-principles model may
not produce predictions that are accurate in absolute terms,
Neural network model predictive control it is usually able to predict the relative trends in process be-
In many cases, the effect of disturbances, d , cannot be fully havior. Mathematically this can be interpreted to mean that
understood. Even worse, many key parameters in the physi- while the model predictions may not be quantitatively pre-

AIChE Journal February 1996 Vol. 42, No. 2 457

cise, the model captures the gradients with respect to I I I I I
I 1
processing variables at least qualitatively. This leads to the 0augmented data
concept for building the hybrid ANN built using data extrap-
olated and interpolated from experimental points using the
first-principles model. Hundreds of new points can be gener-
ated this way from a small experimental data set and the ANN
can then be trained using this augmented data set.
The basic idea is illustrated in Figure 1. Let m , .. . mM rep-
resent M independent variables in the process model, f rep-
resents the output variable, (mil,m51,...) (my2,m52,...), and
(m;,, m&, ...) represent three experimental data points (dark
circles in the figure). The extrapolation equation from an ex-
perimental data (rn?,, me,,,...) to an augmented data point
(my,m ; , . . .) (open circles in the figure) can be written as fol-
lows:

I I I I
f ” ( m ; , m ; , . ..) = f e ( m ; l , r n & , ..) + - (my - m f , ) I I I I
m1
model
... .
( m ; - me,,>+ Figure 1. Extrapolation from different experimental
”f/
+ dm2 m h , ... points.

Since the mathematical model is often encapsulated in a sim-

ulation, extrapolation is done using numerically generated judgment in deciding what is best for a given application. In
this study the data points were all spaced sufficiently close,
gradients of the output function. and hence all of the experimental data points were employed
Due to measurement errors and other noise, extrapolation in the generation of each augmented data point.
from different experimental points will give different results.
The average of these values, weighted by the proximity of
experimental conditions to the input grid position (see Figure
l), is used as the training data at the particular grid point.
Development of ANN Models for Batch Emulsion
Polymerization of Vinyl Acetate
The augmented data set can be given as:
The mechanism of batch emulsion polymerization of vinyl
acetate had been extensively investigated by Lindemann
(1967), Hamielec and coworkers (Friis et al., 1974; Penlidis et
al., 1985, 1988), among others. As mentioned earlier, the
mathematical model developed by Penlidis et al. is used in
this work. This model is able to track conversion/time history
of the batch process fairly accurately. However, it predicts a
sharp rise in molecular weight of product at high conversion
(see, e.g., Jang and Yang, 1989). Therefore, prediction of the
with wik being a weighting factor that is inversely propor- quality of the product is not reliable.
tional to the distance of the augmented data point from the For this application we have taken the control objective as:
experimental data: achieve a desired monomer conuersion (92% in our case) in the
least possible time. The manipulative actions are the reaction
temperature, T , and amount of initiator added, I. The prod-
uct quality constraints are number average molecular weight,
M,,, and dispersity, D , the ratio of weight average to number
average molecular weights ( D = M,,,/M,). The initial condi-
tions-monomer charge Mo = 400 mL, Zo = 40 mL, and To =
50°C-are kept constant in all experiments of this study for
simplicity, though this assumption is not necessary.
For a given set of Zo, M,, and To,the monomer conversion
Equations 4 and 5 represent a simple intuitive way of inte- depends primarily on the unknown disturbance, the amount
grating experimental data and first-principle model predic- of inhibitor present in the initial monomer charge. The
tions into a neural network model. There is a question of monomer conversion is a monotonically decreasing function
which experimental data points to use in the extrapolation. If of the inhibitor at fixed time, plus it is a function of monomer
the data points are widely scattered, then it is probably more soluble impurity (Penlidis et al., 1988). Therefore, an inter-
appropriate to use only the closest experimental points for mediate measurement of monomer conversion can be used to
the interpolation (nearest neighbor policy). There are many identify the unknown disturbance, which includes the lumped
possibilities here and the engineers will have to use their effect of the inhibitors and impurities. Ideally, the intermedi-

458 February 1996 Vol. 42, No. 2 AIChE Journal

ate measurement should be taken as early as possible. Since
conversion can only be accurately measured when it exceeds
about 20% using an on-line gas chromatograph (GC), the
earliest time intermediate measurement for reliable estimate

+
of the disturbance variable is about 45 min ( Y ~ ~ ) .
The completion of intermediate measurement using a GC I l
requires another 20 minutes. Therefore, manipulative action
is taken at the 65th minute, with a one-shot addition of initia-
tor 165 and the temperature of the reactor controlled at T65
Initial condition
4 8% A I I
throughout the later stage of the process. Since Z,, M a , and
To are kept constant in this study, A “ s for predicting the
end time, the number average molecular weight (Mn), and Model predictive
dispersity ( D ) have the following simplified forms: controller

Begin batch
Initial condition

An on-line density meter can be used to measure the conver- measurement, y

45
sion continuously and faster than an on-line GC. The avail-
ability of such an instrument will definitely improve the con-
trollability of the process. However, in many batch processes,
time =65 mins _ _ _ _ _ ._ MPC Correction,
only a limited set of intermediate measurements is available
and the situation we have described is typical and represents ‘65 and 65
a more challenging control problem.
The parameters chosen for the ANN structure (number of
hidden layers, number of hidden nodes, activation function,
etc.) were chosen after some initial experimentation. There
are no strict guidelines provided in the literature for choosing
Conversion=O.92 u
End batch
Final products

these parameters. It is possible to improve the efficiency and ~

accuracy of the networks by fine-tuning these parameters. An Figure 2. Control policy.

ANN model with two hidden layers of 10 nodes was chosen
to remesent all three network models.
The upper and lower bounds of the molecular weight and
dispersity are chosen based on typical industrial practice. The
dD upper bound of the initiator added is due to the limitation of
- were computed numerically from a computer simula- the electrolyte composition in a latex system. The lower bound
I65
tion of the mathematical model. Finite difference approxima- of TS5 is implemented because, below that temperature, the
tion was used at each experimental data point. polymerization reaction is too slow. The upper bound is fixed
by the limitations of the apparatus on hand.
A plot of the control policy is shown in Figure 2. As the
Model Predictive Control of the Batch Latex temperature increases, the exothermic reaction rate in-
Reactor creases. It becomes increasingly difficult to control the tem-
Given the preceding hybrid ANN model, the model predic- perature of the reaction due to the limitation of heat removal
tive control problem can be formulated as follows: equipment. Although experiments can be conducted up to
6YC, a more conservative limit of 62°C is used in optimal
control.
(OP4) This control problem can be solved using any one of nu-
merous optimization codes. For this work we used a general-
subject to ized reduced gradient (GRG) method. For a more detailed
discussion of the nonlinear model predictive control algo-
rithm and the modifications necessary for batch process con-
trol problems, see Joseph and Wang Hanratty (1993).

Experiments
The experimental setup is illustrated in Figure 3. The reac-
tor is a l-liter glass vessel placed in a constant temperature

AIChE Journal February 1996 Vol. 42, No. 2 459

 Motor the training set. The training set was designed to sample the
expected range of the operating variables, T,, and I,,, rea-
sonably well. In the training data set, y , ranges from 0.24 to
a high of 0.49. In addition, four sets of control experiments
were used as the test set. The conditions and results of these
experiments are tabulated in Table 1.
Based on the number of parameters that should be fitted
in the ANN model, 1,000 augmented data sets were gener-
ated at evenly spaced input grid points to train the ANNs in
Eqs. 6, 7, and 8. Back-propagation ANN (Rumelhart et al.,
1986) with two hidden layers of 10 nodes were used. The
ANN was trained to a precision level such that the total sum
of the squared differences between ANN output and training
data is less than As stated earlier, the hybrid ANN uti-
lized the gradients computed from the theoretical model for
the effects of T65 and Z65. A smaller ANN that uses a single
3-node hidden layer network, trained directly by the experi-
Figure 3. Experimental setup. mental data, was implemented to predict the "gradient" of
the intermediate measurement y45.This approach was found
bath. The content is kept well mixed using a motor-driven to be superior to using the theoretical model for this particu-
stirrer (speed 120 rpm). Nitrogen is used to maintain the lar variable.
pressure inside the reactor. Initiator is added using a meter-
ing pump with a maximum capacity of 150 mL/min. Comparison of the augmented data neural network with
The monomer vinyl acetate used is of industrial grade. It other models
was purified by extraction using sodium hydroxide, and wash-
ing with deionized water several times. The product was then The conventional approach to empirical modeling is to ob-
distilled under subatmospheric pressure at 70°C. Reagent- tain a much smaller ANN, termed ANN-EXP, by fitting only
grade sodium dodecylsulfate and potassium peroxodisulfate experimental data. The number of nodes must be kept small
were used as emulsifier and initiator, respectively. to avoid overfitting the data. An ANN with one hidden layer
At the beginning of an experiment, fixed quantities of containing three nodes was used in this study.
On the other hand, the conventional approach to modeling
monomer, deionized water, and emulsifier were added to the
using prior knowledge is to adapt the theoretical model by
reactor and stirred under nitrogen for about one hour until
the mixture was completely emulsified. The reactor tempera- fitting some of the key model parameters to the experimental
ture was controlled initially at 50°C. Initiator was then added. data. In this article we refer to this third model as the fitted
Monomer concentration was measured 45 minutes after the theoretical model. It is natural to assume that the polymeriza-
addition of the initiator. About 1 g of the reactor content was
sampled, to which 5 g of sodium chloride solution was added Table 1. Conditions and Results of all Experimental Runs
to destroy the emulsion. The solution was then extracted with
12 g of dimethyl ether. Six g of the organic phase was taken
Training Set
out, to which 0.8 g methanol was added as the internal stan- 1 0.46 0 60 89 2.01 2.8
dard. The sample was then analyzed by a gas chromatograph 2 0.33 0 67 94 2.13 2.9
using a Porapak Q column, a FID detector, and nitrogen as 3 0.30 40 67 82 1.49 2.7
the carrier gas. The entire procedure takes about 20 min. 4 0.28 20 50 117 2.06 2.3
After analyzing monomer conversion at 45 min and identi- 5 0.33 0 50 126 [125] 1.64 L1.671 2.7 [2.7]
fying the amount of inhibitor in the charge, initiator was 6 0.34 40 58 96 1.98 3.7
added again and the temperature of the reactor was raised 7 0.31 40 55 105 1.60 3.3
8 0.49 0 50 118 1.74 2.6
according to the predictive control strategy. Samples were
again taken at 5-min intervals, starting about 20 min before 9 0.28 0 55 116 1.83 2.4
10 0.31 40 55 102 1.69 2.6
the predicted end time. Monomer concentration was mea- 11 0.28 40 61 93 1.48 2.9
sured and the actual end time (92% monomer conversion) 12 0.24 40 50 125 1.73 2.6
was interpolated. The molecular weight and dispersity of the
13 0.30 30 50 123 1.87 2.5
samples taken at three intervals closest to the end time were 14 0.44 20 55 114 1.89 3.0
measured using gel permeation chromatography. 15 0.28 0 60 102 2.09 3.1
Analytical experimental errors in measurement of the end 16 0.41 30 66 88 2.10 2.9
time, number average molecular weight, and dispersity are Test Set
estimated at lo%, 5%, and lo%, respectively. 17" 0.32 39 58 106[991 1.92[1.81 3.2L3.01
18* 0.28 37 57 101 11021 2.00[I.81 3.213.01
19* 0.31 39 57 103 I1001 1.82 uI.81 2.9l3.01
20r 0.26 40 62 93[961 1.98[1.71 3.8l3.71
Results
In order to identify the ANN models to be used in predic- * ANN model predictive control.
Aggressive control.
tive control, sixteen experiments were performed to establish [ I Predictive values.

460 February 1996 Vol. 42, No. 2 AIChE Journal

 Table 2. Updated Kinetics Parameters
~ ~~~

40
Training set A d
Original Value Updated Value
AP 0.89 x 107 2.21 x 107 I I t
- 5,650 - 5,563 W Augmented Doto ANN
EP E Z 2 Z ANN-EXP
Aim 0.43 X lo6 0.41 X lo6 30 0Fitted Theoretical Model
Eh - 9,020 - 11,146

l l
fP 3.55 x 106 4.04 X lo6
EfP - 9,950 -9,168

55
i

tion reaction and chain transfer reaction constants are incor-

10
rect and should be updated. There are six such constants:
1. A , and E fo; the polymer propagation rate constant:
k , = A , X e-EJT.
0
‘ 2. A f mand Efm for the rate constant of chain transfer to 1 2 3 4 5 6 7 8 9 10 1 1 12 13 14 15 16 17 18 19 20

4obT[
monomer: kfm= A , X e-€fflT. RUN NUMBER
3. A f p and Efp for the rate constant of chain transfer to (a)
polymer: k f p= A f pX e-Ef@T.
The following optimization problem was solved to “update”
these parameters: Training set Test s e t >

m Augmented Data ANN

azZa ANN-EXP
0Fitted Theoretical Model

,I’)?(
30

-
. . I t I
+ (OP.5)

subject to

RUN NUMBER

(b)

Test set
Q------ Traininqset _____() k
Y
40
1 I
m Augmented Data ANN
C&2a ANN-EXP
30 .
The “updated”parameters are shown in Table 2.
Figure 4 shows the error in the predictions of the three
models. Table 3 gives the mean and standard deviation of the
--
A-
u

correlation and prediction error of the three models. It was g 2o

0
found that the A”-EXP model fits the training set very well,
but error for the test set is substantially higher. This is natu- 10
ral because there are only 16 data points in the training set,
and no guarantee of successful extrapolation and interpola-
tion is incorporated in the model development. On the other
0
hand, the fitted theoretical model performs poorly for the 1 3 5 7 9 11 13 15 17 19
training set, since only partial knowledge has been captured RUN NUMBER
in the model. No matter how we adjust the key parameters in (C)
the model, agreement with data can only be limited. How-
ever, error for the test set is of the same order as that for the Figure 4. (a) Absolute values of percent error in pre-
training set, indicating that the fitted theoretical model ex- dicted end time; (b) absolute values of per-
trapolates and interpolates better than the A”-EXP, as ex- cent error in predicted number average
pected. The augmented data ANN achieve a correlation ac- molecular weight; (c) absolute values of per-
curacy in the training data set that is only slightly inferior to cent error in predicted dispersity.

AIChE Journal February 1996 Vol. 42, No. 2 461

 Table 3. Mean and Standard Deviations of the Absolute
Values of Percent Errors in Model Prediction
Augmented Fitted
Data Theoretical
% ANN A”-EXP Model
Runl-16 tf p 1.16 1.20 9.11
[training set] u 0.78 1.15 5.53
M“ P 4.62 1.30 10.03
U 3.65 1.20 6.10
D P 5.58 2.92 10.88
U 4.97 3.88 6.59
Run17-20 tf p 4.04 5.25 4.92
[test set] U 1.95 2.70 2.54
Figure 6. Parallel hybrid ANN.
Mn P 7.64 16.64 6.27
U 4.80 7.58 3.51
D P 5.15 17.74 8.97 neural network are obtained by fitting the training data set.
U 1.25 9.57 6.47 This neural network is then used in series with the default
model to give prediction of end-product quality as is illus-
Runl-20 tf I.L 1.74 2.01 8.27 trated in Figure 5. In the “parallel hybrid A”,” the residual
U 1.60 2.27 5.35
of experimental data and default-model predictions is used
Mn P 5.23 4.37 9.28 to train an ANN. Radial basis transfer functions (RBFNs) are
U 4.09 7.09 5.87
used for the ANN model to ensure that the extrapolation is
D P 5.49 5.89 10.50 dominated by the default model. The plot of the implementa-
u 4.48 8.09 6.61 tion of “the parallel hybrid ANN model”is given in Figure 6.
I*: mean value; u :standard deviation. The results are shown in Table 4 and Figure 7. Although
both the series and parallel hybrid A ” s improve over the
fitted theoretical model, the performance is not as good as
A”-EXP. However, by using the knowledge contained in the hybrid approach using augmented data points. Both the
the theoretical model, the augmented data ANN is able to series and parallel hybrid ANNs assume that all essential
extrapolate and interpolate better for the test data set. characteristics of the system have been captured by the de-
fault model, which is not correct in this example. The aug-
Comparison with other types of hybrid model mented data hybrid ANN seems to be a better reconciliation

Two other types of hybrid ANN model proposed separately

by Psichogios and Ungar (19921, that is, “the series hybrid Table 4. Mean and Standard Deviations of the Absolute
ANN model,” and Kramer et al. (1992), “the parallel hybrid Values of Percent Errors in Model Prediction Using
ANN mode1,”were also adapted to this system and compared Different Types of Hybrid ANN Models
with the augmented data ANN. Augmented Series Parallel
In the “series hybrid A ” , ” w e implement the fitted theo- Data Hybrid Hybrid
retical model as the default model. We further assume that % ANN ANN ANN
the frequency factors are functions of d, or alternatively yd5. Runl-16 tf p 1.16 8.39 3.20
A back propagation neural network (BPN) with one input [training set] U 0.78 7.48 4.26
node and a hidden layer of three nodes and three output MI I* 4.62 7.78 3.58
nodes is used to represent this relation. The weights of this U 3.65 6.51 3.78
D P 5.58 11.15 4.46
u 4.97 5.55 4.82
I 1
Run 17-20 tf fi 4.04 4.58 3.01
[test set] L7 1.95 1.60 2.66
Mn P 7.64 5.03 16.26
u 4.80 2.80 6.85
D P 5.15 13.01 14.70
u 1.25 9.67 8.20

Mo
‘ 6 5 3 T65
I
Fitted Theoretical

Model -1 xn
t

D
Runl-20 tf

Mn

D P
U

u
P
1.74
1.60
5.23
4.09
5.49
4.48
7.63
6.90
7.23
6.06
11.52
6.63
3.16
3.99
6.12
6.82
6.51
6.99
Figure 5. Series hybrid ANN. +: mean value; o:standard deviation.

462 Februarv 1996. Vol. 42, No. 2

~ AIChE Journal
 - Augmented Data ANN
ezzZa Series Hybrid ANN
0Parallel Hybrid ANN

1 3 5 7 9 11 13 15 17 19
Run number

(4
Y45

Figure 8. Control actions vs. intermediate measure-

ment.
EZZd Series Hybrid ANN
D Parallel Hybrid ANN
30
between empirical data and theoretical knowledge when both
A
h:
v
abundance of data and adequate prior knowledge are lack-
e
T
20 ing.
I

10
Model predictive control of latex reactor
In this work, a one-shot policy, as shown in Figure 2, is
0
implemented to perform the model predictive control. It is
1 3 5 7 9 11 13 15 17 ID possible to obtain the control profiles of T6, and Z65 as func-
Run number
tions of y,, based on the hybrid A" model described ear-
(b) lier by solving the model predictive problem OP4. The result
is shown in Figure 8, which shows that the control actions
est se
can be more aggressive to shorten the batch time if the inter-

-
mediate measurement shows that the amount of feed impu-
rity is small (high conversion), and the product qualities can
still be guaranteed as shown in Figure 9. In polymerization
Augmented Data ANN reaction the dispersity of the polymer product will be larger
ezzzd Series Hybrid ANN
0Parallel Hybrid ANN with higher reactor temperature and higher concentration of
free radicals (initiator). Figure 8 shows that if y, is low, that
h
x
v is, inhibitor is high, then the control action i65 and T65 can-
g0 2o not be aggressive to shorten the batch time since the con-
straint on dispersity will be violated (see also Figure 9c).
Based on the results shown in Figure 8, three test batch
10 runs (runs 17, 18 and 19 in Table 1) were carried out to eval-
uate the predictions of the hybrid ANN model. The results
are given in Figure 9. It shows that the control actions taken
0
1 3 5 7 9 11 13 15 17 19
do shorten the batch time while maintaining the product
Run number quality within its tolerances. If one simply wants to shorten
the batch time and ignore the existence of the impurity by
(c) setting the control actions into their upper bounds (62°C for
Figure 7. (a) Absolute values of percent error in pre- T6, and 40 mL for &5r called aggressiue control), the upper
dicted end time with different hybrid ANN bound on the dispersity of the polymer product is violated as
models; (b) absolute values of percent error predicted by the model. This is confirmed by run 20. How-
in predicted number average molecular weight ever, if no predictive control action is taken, and a fixed recipe
with different hybrid ANN models; (c) abso- is followed, the batch time can be very long and the number
lute values of percent error in predicted dis- average molecular weight of the polymer product is out of
persity with different hybrid ANN models. the lower bound as shown in Figure 9 and Table 1 (run 5).

AIChE Journal February 1996 Vol. 42, No. 2 463

 150 Conclusion
140 A model predictive control scheme for batch processes us-
ing an augmented data hybrid ANN has been developed and
130
applied to the latex polymerization process. The control
120 strategy uses this augmented data hybrid ANN to identify the
unknown and unmeasured disturbances in the initial charge
W 110 of the batch. In the application to the batch polymerization
5
F 100 of vinyl acetate, the amount of inhibitor present in the
0
Z
W
monomer charge was estimated by measuring monomer con-
90 version. Based gn the information of the intermediate mea-
1 surement, the augmented data hybrid ANN is used to predict
I
~ Model prediction and hence control the batch: end time and the molecular
-0- Model predictive control, Run
t Model predictive control, Run
t Model predictive control. Run
+ Aggressive control, Run #20
X No control, Run #5
3 17
18
19
weight and dispersity of the final polymer product.
The advantages of using the ANN-based predictive control
strategy are:
50 I I 1. The effect of unmeasured disturbances can be incorpo-
0.1 0.3 0.5
rated easily using intermediate measurements.
(a) Y45
2. The ANN structure allows us to capture nonlinear re-
lationships among the variables.
3. The predictive capability of the model allows us to cor-
I I 1 I rect for changing raw material properties and output specifi-

I
Model prediction
cations.
The augmented data hybrid ANN is trained using a combi-
nation of actual experimental data and augmented data points
2.3
generated through extrapolation and interpolation using gra-
ro dient information from the theoretical model. This procedure
I
0 of augmenting the data allows development of ANN models
X when experimental data are scarce and expensive to obtain.
z
3 The augmented data hybrid ANN was also shown to be
superior in terms of its correlative capability to other hybrid
ANN approaches. It also performed better than a theoretical
model whose parameters were fitted to the experimental data.
Although we restricted ourselves to specific types of mea-
surements, models, and control actions, the methodology
1.5 I I presented is quite general and can be applied to the model-
0.1 0.3 0.5 ing and control of other batch processes where product qual-
Y45 ity control is difficult to achieve using conventional tech-
niques and algorithms.

Acknowledgment
Model prediction
5- I

Financial support provided by the National Science Council, Tai-

-0-
t
-+-
Modal predictive control, Run 1 7
Model prcdictivccontrol. Run 18
* Modalpredictivtcontrol.620
Run
Aggressive control. Run
--+c No control, Run #5
I 19 wan, through grant NSC82-0402-E007-249, and by the National Sci-
ence Foundation, USA-Taiwan Cooperative Research, through grant
INT-93-11390, are gratefully acknowledged.

4 -
11 Upper bound Notation

3 T 4 .__..
A f m=frequency factor of the rate constant of chain transfer to
monomer
A f p=frequency factor of the rate constant of chain transfer to
polymer
A , =frequency factor of the polymer propagation rate constant
Derr0,=relative error of the predictive dispersity
Dj=dispersity of the final product at run j
DT =predictive dispersity of the final product at run j
2 I Efm =activation energy of the rate constant of chain transfer to
0.1 0.3 0.5 monomer
Efp =activation energy of the rate constant of chain transfer to
Y45
polymer
Figure 9. (a) End time: Experimental and simulation re- Ep =activation energy of the polymer propagation rate con-
sults; (b) number average molecular weight: stant
f" =augmented data function
experimental and simulation results; (c) dls- f'=experimental result
persity: experimental and simulation results. h =intermediate states of the system

464 February 1996 Vol. 42, No. 2 AIChE Journal

 kfm =rate constant of chain transfer to monomer Kramer, M. A., and J. A. Leonard, “Diagnosis Using Back Propaga-
k -rate constant of chain transfer to polymer
4I polymer propagation rate constant
M,, er,Or =relative error of the predictive number average molecular
tion Neural Networks-Analysis and Criticism,” Comput. Chem.
Eng., 14, 1323 (1990).
Kramer, M. A., M. L. Thompson, and P. M. Bhagat, “Embedding
weight Theoretical Model in Neural Networks,” ACC/WAZI, 475 (1992).
Mn,, =number average molecular weight of the final product at Levinson, W. A., “Make the Most of Control Charts,” Chem. Eng.
run f Process, 88, 86 (1992).
M:,, =predictive number average molecular weight of the final Lindemann, M. A. “The Mechanism of Vinyl Acetate Polymeriza-
product at run j tion,” in Kinetics and Mechanisms of Polyme~atwn,Vol. 1, Knyl
N =number of experimental data points Polymerization, Part I, Marcel Dekker, New York, p. 207 (1967).
t =time Pao, Y., Adaptiw Pattern Recognition and Neural Networks, Addison-
t,, =relative error of the predictive end time Wesley, Reading, MA (1989).
t,,, =end time of run j Penlidis, A., J. F. MacGregor, and A. E. Hamielec, “A Theoretical
t z , =prelictive end time of run j and Experimental Investigation of the Batch Emulsion Polymeriza-
xo =initial state variables tion of Vinyl Acetate,” Polym. Process Eng., 13, 185 (1985).
yd5 =monomer conversion at 45th minute Penlidis, A., J. F. MacGregor, and A. E. Hamielec, “Effect of Impu-
y45,,=monomer conversion at 45th minute at run j rities on Emulsion Polymerization: Case I Kinetics,” J. Appl. Pofy.,
y&,, =predictive monomer conversion at 45th minute at run j 35, 2023 (1988).
Psichogios, D. C., and L. H. Ungar, “A Hybrid Neural Net-
work-First Principles Approach to Process Modeling,” AIChE J.,
Literature Cited 38, 1499 (1992).
Bhat, N., and T. J. McAvoy, “Use of Neural Nets for Dynamic Mod- Psichogios, D. C., and L. H. Ungar, “Direct and Indirect Model Based
eling and Control of Chemical Process System,” Comput. Chem. Control Using Artificial Neural Networks,” Znd. Eng. Chem. Res.,
Eng., 14, 573 (1990). 30, 2564 (1991).
Friis, N., D. Goosney, J. D. Wright, and A. E. Hamielec, “Molecular Ripley, B. D., “Statistical Aspects of Neural Networks,” Networks and
Weight and Branching Development in Vinyl Acetate Emulsion Chaos: Statistical and Probabilistic Aspects, Chapman & Hall, Lon-
Polymerization,” J. Appl. Poly. Sci., 18, 1247 (1974). don, p. 40 (1993).
Hernandez, E., and Y. Arkun, “Study of the Control Relevant Prop- Rumelhart, D. E., G. E. Hinton, and R. J. Willams, “Learning Inter-
erties of Back Propagation Neural Network Models of Nonlinear nal Representation by Error Propagation,” Parallel Distributed Pro-
Dynamical Systems,” Comput. Chem. Eng., 16, 227 (1992). cessing, Vol. 1, MIT Press, Cambridge, MA, p. 318 (1986).
Hoskins, J. C., and D. M. Himmelblau, “Artificial Neural Network Sarle, W. S., “Neural Networks and Statistical Model,” Proc. of the
Models of Knowledge Representation in Chemical Engineering,” SAS Users Group Znt. Conf, Dallas, TX (Apr., 1994).
Comput. Chem. Eng., 12, 881 (1988). Thompson, M. L., and M. A. Kramer, “Modeling Chemical
Jang, S. S., and W. L. Yang, “Dynamic Optimization of Batch Emul- Processes Using Prior Knowledge and Neural Networks,” AIChE
sion Polymerization of Vinyl Acetate-An Orthogonal Polynomial J., 40, 1328 (1994).
Initiator Policy,” Chem. Eng. Sci., 44,515 (1989). Venkatasubramanian, V., and K. A. Chan, “A Neural Network
Joseph, B., and F. Wang Hanratty, “Predictive Control of Quality in Methodology for Process Fault Diagnosis,” AIChE J., 35, 1993
a Batch Manufacturing Process Using Artificial Neural Network (1989).
Models,” Ind. Eng. Chem. Res., 23, 1951 (1993). Whitely, J. R., and J. Davis, “Knowledge-Based Interpretation of
Klein, R. A., “Achieve Total Quality Management,” Chem. Eng. Sensor Patterns,” Comput. Chem. Eng., 16, 329 (1992).
Process, 87, 83 (1991).
Kozub, D., and J. F. MacGregor, “State Estimation for Semi-Batch
Polymerization Reactors,” Chem. Eng. Sci., 47(5), 1047 (1992). Manuscript receiued Aug. 22, 1994, and revision receiwd Apr. 7, 1995.

AIChE Journal February 1996 Vol. 42, No. 2