You are on page 1of 6

Recent Researches in Circuits and Systems

Prediction of Semi-Batch Reactor Using Multilayered Feed-Forward


Neural Networks
LUBOMIR MACKU1, DAVID SAMEK2
1
Department of Electronics and Measurements, Faculty of Applied Informatics
Tomas Bata University in Zlin
nam. T. G. Masaryka 5555, 76005 Zlin
Czech Republic
macku@fai.utb.cz http://www.fai.utb.cz
2
Department of Production Engineering, Faculty of Technology
Tomas Bata University in Zlin
nam. T. G. Masaryka 5555, 76005 Zlin
Czech Republic
samek@ft.utb.cz http://www.ft.utb.cz

Abstract: - The contribution focuses on artificial neural networks application for prediction of chemical semi-
batch reactor output. There are studied two approaches to prediction. The first methodology uses classical time
series prediction apparatus, when the prediction is based on the one signal only. The second approach uses
two signals and corresponds to model predictive control methodology. Both predictors are based on artificial
multilayered feed-forward neural networks.

Key-Words: - artificial neural network, prediction, chemical reactor

1 Introduction networks [22, 23]. Interesting generalization of


Prediction of nonlinear and complex systems can be neural networks, which is achieved by using multi-
performed by various methods. One of them is argument and learnable functions, bring functional
simplification and linearization that leads to linear networks [24].
models [1, 2]. Some authors use wavelet filtering in Generally, it can be distinguished between two
order to divide stochastic and deterministic parts main approaches in the prediction. The first
that are modeled separately [4, 5]. Javadi et al. and approach deals with the predicted signal only. It
Phaiboon present successful application of fuzzy means that the predicted value must be predicted
logic to prediction [6, 7]. However, probably the from known older values of the same signal. This
most popular approach in these cases is based on approach is very common in financial engineering
artificial neural networks. and it is usually denoted as a signal prediction or
Artificial neural networks are commonly used in time series prediction (in case of time series).
various fields, for example weather forecasting [8], The second approach uses other signals (usually
time series prediction of financial data [9, 10], system inputs) for the prediction and it is commonly
biology and medicine [11, 12], power engineering called system or process prediction. This approach
[13] and process control [14, 15]. There is lot of is widely used in the process industry, especially in
types of artificial neural networks, but not all of the control (predictive control).
them are usable for prediction. The most common In this contribution both approaches to prediction
are multilayered feed-forward neural networks [8, are studied and the multilayered feed-forward neural
11, 16, 17]. Fairly wide group of artificial neural networks are tested on the prediction of temperature
networks belongs to recurrent neural networks [18, in the chemical semi-batch reactor. The motivation
19]. Very popular due to their fast training are radial of this work comes from our long time research
basis function neural networks [18, 19]. Al-Shayea focused on predictive control of chemical reactors.
and El-Refae use generalized regression neural Usually, the predictor must be adapted during the
network for prediction of Spanish banks data [16]. function, because of its inaccuracy or changing
There are also other less frequent methods such as parameters of the system to be controlled. The goal
fuzzy-neural networks [20], adaptive linear of the presented work is to design and test the
networks [21], unsupervised Kohonen neural proper predictor based on multilayered feed-forward

ISBN: 978-1-61804-108-1 453


Recent Researches in Circuits and Systems

neural network that would not require adaptation 3 Multilayered feed-forward neural
during control of the given reactor.
networks
Multilayer feed-forward neural networks (MFFNNs)
are very often called backpropagation networks
2 Chemical Semi-Batch Reactor because of the typical training algorithm. Some
In this paper, a semi-batch reactor model is used to authors prefer name multilayer perceptrons (MLP)
study the prediction abilities of multilayered feed- [25, 26], because MFFNNs have been developed by
forward neural networks and their applicability in generalization from Rosenblatt perceptron with
the model predictive control. The model input data binary transfer function.
comes from a real process - the chromium waste The signals flow only in the one direction (from
recycling process [15]. The task of the predictor is the input to the output) in these neural networks. All
to predict the in-reactor temperature. neurons are structured into layers and, typically, all
Let us consider a single input – single output neurons in the specific layer use same transfer
(SISO) system of the chemical exothermic semi- function. In contrast to Rosenblatt perceptron
batch reactor, which is described by the following multilayer feed-forward neural networks use various
mathematical model: transfer functions, usually continuous (e.g. linear,
hyperbolic tangent, sigmoidal functions, etc.).
d m(t ) By applying of Kolmogorov theorem it has been
= FI (1)
dt proved [27] that two-layered MFFNN (with one
hidden layer) can approximate any function with
E certain accuracy while non-polynomial transfer
d a (t ) FI −
= − A ⋅ e R⋅T (t ) ⋅ a(t ) (2) function in the hidden layer is used. This
dt m(t ) methodology is adopted in the paper.
E

R ⋅T (t )
d T (t ) FI ⋅ c I ⋅ TI A ⋅ e ⋅ ∆H r ⋅ a(t )
= + − 4 Simulations and results
dt m(t ) ⋅ c c MATLAB with Neural Network Toolbox and
K ⋅ S ⋅ T (t ) K ⋅ S ⋅ TC (t ) Simulink were used for all simulations. There were
− + (3)
m(t ) ⋅ c m(t ) ⋅ c studied two approaches: signal prediction and
system prediction.
d TC (t ) FC ⋅ TCI K ⋅ S ⋅ T (t ) In the first case the artificial neural network used
= + − five past values of the predicted signal as the input
dt mC mC ⋅ cC
vector and predicted only one step ahead. In other
K ⋅ S ⋅ TC (t ) FC ⋅ TC (t ) words, when it was needed the MFFNN repeatedly
− − (4)
mC ⋅ cC mC used its own predictions as inputs. Thus, in the input
(zero) layer of the artificial network were five
where m is the total weight of the reaction neurons and the output layer consisted one neuron.
components in the reactor, a is the mass
concentration of the reaction component in the
reactor, c = 4500 J.kg.K-1 is the specific heat
capacity of the reactor content, T is the temperature
of the reactor content. FI, TI = 293.15 K and cI =
4400 J.kg.K-1 is the reaction component input mass
flow rate, the temperature and the specific heat Fig. 1 – Scheme of MFFNN with 5 input neurons
capacity. FC = 1 kg.s-1, TCI = 288.15 K, TC, cC = (net1)
4118 J.kg.K-1 and mC = 220 kg is the cooling water
mass flow rate, the input temperature, the output As is depicted in the Fig. 1, the network had five
temperature, the specific heat capacity and the neurons in the hidden layer. This number was
weight of the cooling water in the cooling system of obtained by many experiments as sufficient for this
the reactor, respectively. case. The hidden layer used hyperbolic tangent
Other constants: A = 219.588 s-1, E = sigmoid as the transfer function, while the output
29967.5087 J.mol-1, R = 8.314 J.mol-1.K-1, ∆Hr = layer employed linear transfer function.
1392350 J.kg-1, K = 200 kg.s-3.K-1 and S = 7.36 m2 is Three simulation results from our previous work
the effective heat-transfer area [15]. [15] were chosen as the training and testing data.

ISBN: 978-1-61804-108-1 454


Recent Researches in Circuits and Systems

Only one simulation data were used for training 0.3


(furthermore they will be symbolized as data1),
whilst the rest two simulation results served as 0.2

testing group (data2 and data3). The training was 0.1


performed using built-in Matlab/Neural Network

prediction error [K]


0
Toolbox Levenberg-Marquardt backpropagation
algorithm. This artificial neural network will be in -0.1

the following text denoted as net1. -0.2


After the network training, the obtained predictor
was tested to all three signals (Fig. 3, 5, 7). For -0.3

better comparability the prediction error for all -0.4


testing signals was computed (Fig 4, 6, 8).
-0.5
In order to compare influence of the input signal 0 0.5 1 1.5 2 2.5 3
time [s] 4
knowledge to the prediction quality, the second x 10

predictor was also tested. This predictor used not Fig. 4 – Prediction error of net1 for the training data
only the predicted signal old values but also old (data1)
values of the system input. In this case it was the
chromium sludge dosing speed. Thus the
multilayered feed-forward neural network involved 380
ten neurons in the input layer (see Fig. 2). As well predicted signal
reference signal
as in the previous approach, the hidden layer used 370

hyperbolic tangent sigmoid as the transfer function,


while the output layer employed linear transfer 360

function. This artificial neural network will be in the


T [K]

following text denoted as net2. 350

With the intention of obtaining comparable


340
results the same methodology as in case of net1 was
used. It means that same training data, the same
330
training algorithm and same testing data were
applied. Results are depicted in Fig. 9-14. 320
0 500 1000 1500 2000 2500 3000
time [s]

Fig. 5 – Net1 test on the testing data (data2)

0.4

Fig. 2 – Scheme of MFFNN with 10 input neurons 0.3


(net2)
0.2

0.1
prediction error [K]

380
predicted signal 0
370 reference signal
-0.1
360
-0.2

350
-0.3
T [K]

340 -0.4

330 -0.5
0 500 1000 1500 2000 2500 3000
time [s]
320
Fig. 6 – Prediction error of net1 for the testing data
310
(data2)
300
0 0.5 1 1.5 2 2.5 3
time [s] 4
x 10

Fig. 3 – Net1 test on the training data (data1)

ISBN: 978-1-61804-108-1 455


Recent Researches in Circuits and Systems

390 0.3
predicted signal
380 reference signal
0.25

370
0.2
360

prediction error [K]


350 0.15
T [K]

340 0.1

330
0.05
320

0
310

300 -0.05
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
time [s] 4 time [s] 4
x 10 x 10

Fig. 7 – Net1 test on the testing data (data3) Fig. 10 – Prediction error of net2 for the training
data (data1)

0.5

380
predicted signal
0
reference signal
370

-0.5
prediction error [K]

360

-1
T [K]

350

-1.5

340

-2

330

-2.5
0 0.5 1 1.5 2 2.5 3
time [s] 4 320
x 10 0 500 1000 1500 2000 2500 3000
Fig. 8 – Prediction error of net1 for the testing data time [s]

(data3) Fig. 11 – Net2 test on the testing data (data2)

380 2
predicted signal
370 reference signal
1.5

360

1
prediction error [K]

350
T [K]

340 0.5

330
0

320

-0.5
310

300 -1
0 0.5 1 1.5 2 2.5 3 0 500 1000 1500 2000 2500 3000
time [s] 4 time [s]
x 10

Fig. 9 – Net2 test on the training data (data1) Fig. 12 – Prediction error of net2 for the testing data
(data2)

ISBN: 978-1-61804-108-1 456


Recent Researches in Circuits and Systems

390
predicted signal
380 reference signal Table 1 – Prediction quality criterions
370
net1 net2
360
C1 C2 C1 C2
350
data1 0.0024 4.529e-05 9.828e-05 2.218e-06
T [K]

340
data2 0.0133 0.0013 0.0115 0.0107
330
data3 0.0565 0.0806 0.0155 0.0080
320

310
As can be seen from results, the prediction is in both
300
0 0.5 1 1.5 2 2.5 3 cases very good. However, if the best method
time [s] 4

Fig. 13 – Net2 test on the testing data (data3)


x 10
should be judged, the second approach (net2) must
be chosen, because it provided much better results
2 with the one and only exception (criterion C2 for
1.5 data2).
1

0.5
5 Conclusion
prediction error [K]

0
Motivation of this paper was to develop effective
-0.5
predictor for the chemical semi-batch reactor that is
-1 used in chromium recycling process. The accuracy
-1.5
and settings of the predictor should fulfill the
requirements for the applicability in the model
-2
predictive control of the reactor.
-2.5
0 0.5 1 1.5 2 2.5 3 The proposed predictor (net2) proved the
time [s] 4
x 10 abilities in the all selected tests and validation. Thus
Fig. 14 – Prediction error of net2 for the testing data
it can be concluded that it can be supposed for the
(data3)
model predictive controllers of the given chemical
reactor.

4 Comparison and discussion


For better comparison two prediction quality References:
criterions were defined. The first criterion C1 [1] E. F. Camacho, C. Bordons, Model Predictive
describes total sum of absolute values of prediction Control in the Process Industry. Springer -
errors relative to number predictions, whilst the Verlag, 2004.
second criterion function C2 characterizes total sum [2] P. Chalupa, V. Bobal, Modelling and Predictive
of squares of prediction errors relative to number Control of Inverted Pendulum, In 22nd
predictions. The C1 criterion gives same importance European Conference on Modelling and
to all errors. On the other hand C2 emphasizes Simulation, European Council for Modelling
higher errors and lower prediction errors are and Simulation, 2008, pp. 531-537.
suppressed. [3] T. Minerva, Wavelet Filtering for Prediction in
Time Series Analysis, In: Non-Linear Systems
N
and Wavelet Analysis, WSEAS Press, 2010, pp.
∑ t (i ) − p(i )
i =1
89-94.
C1 = (5) [4] M. Dong, D. Yang, Y. Kuang, D. He, S. Erdal,
N
N
D. Kenski, PM2.5 concentration prediction
using hidden semi-Markov model-based times
∑ (t (i ) − p(i ))
i =1
2

series data mining, Expert Systems with Apps,


C2 = (6)
N Vol. 36, No. 5, 2009, pp. 9046-9055.
[5] J. Yan, C. Guo, X. Wang, A dynamic multi-
N stands for number of predictions (length of scale Markov model based methodology for
predicted signal), t is target (original) signal, p remaining life prediction, Mechanical Systems
and Signal Processing, Vol. 25, No. 4, 2011,
denotes predicted signal and i is number of the
pp. 1364-1376.
prediction.

ISBN: 978-1-61804-108-1 457


Recent Researches in Circuits and Systems

[6] S. Javadi, Z. Hojjatinia, Wind Speed Modeling Systems, In RECENT ADVANCES in


and Prediction in Wind Farms Using Fuzzy AUTOMATION & INFORMATION, WSEAS,
Logic, In Applied Mathematics in Electrical 2010, pp. 110-115.
and Computer Engineering, WSEAS, 2012, pp. [18] R. Zemouri, P. C. Patic, Prediction Error
98-103. Feedback for Time Series Prediction: a way to
[7] S. Phaiboon, RF Macro-cell Prediction Using improve the accuracy of predictions, In
Fuzzy Logic: Case Study in Bangkok City – Proceedings of the 4th EUROPEAN
Thailand, In Latest Trends on Communications COMPUTING CONFERENCE, WSEAS, 2010,
and Inf. Tech., WSEAS, 2010, pp. 105-109. pp. 58-62.
[8] I. Maqsood, M. R. Khan, A. Abraham, An [19] M. H. B. Karami, Application of Neural
ensemble of neural networks for weather Networks In Short-Term Load Forecasting, In
forecasting, Neural Computing & Applications, 7th WSEAS Int. Conf. on Mathematical
Vol.13, No.2, 2004, pp. 112-122. Methods and Computational Techniques In
[9] E. M. Azoff, Neural network Time Series Electrical Eng., WSEAS, 2005, pp. 37-41.
Forecasting of Financial Markets, John Wiley [20] P. M. Patil, S. N. Kulkarni, D. D. Doye, U. V.
& Sons, 1994. Kulkarni, Modular General Fuzzy Hypersphere
[10] E. Diaconescu, The use of NARX Neural Neural Network, In Proc. of the 4th WSEAS/
Networks to Predict Chaotic Time Series, IASME Int. Conf. on Syst. Sc. and Simulation in
WSEAS Transactions on Computer Research, Engineering, WSEAS, 2005, pp. 211-216.
Vol.3, No.3, 2008, pp. 182-191. [21] D. Samek, D. Manas, Artificial neural networks
[11] G. Brion, Ch. Viswanathan, T. R. Neelakantan, in artificial time series prediction benchmark,
S. Ligireddy, R. Girones, D. Lees, A. Allard, Int. J. of Math. Models and Methods in Applied
A. Vantarakis, Artificial Neural Network Sciences, Vol. 5, No. 6, 2011, pp. 1085-1093.
Prediction of Viruses in Shellfish, Applied and [22] F. Ortega, et al. An Hybrid Approach to
Enviromental Microbiology, Vol.71, No.9, Prediction of Electric Load with MARS and
2005, pp. 5244-5253. Kohonen Maps, [online] Available:
[12] E. Pasomsub, C. Sukasem, S. Sungkanuparph, http://neuron.tuke.sk/competition/reports/
B. Kijsirikul, W. Chantratita, The application FranciscoOrtega.pdf
of artificial neural networks for phenotypic [23] F. J. Marin, F. Garcia-Lagos, G. Joya, F.
drug resistance prediction: evaluation and Sandoval, Peak Load Forecasting using
comparison with other interpretation systems, Kohonen Classification and Intervention
Japanese journal of infectious diseases, Vol.63, Analysis, [online] Available:
No.2, 2010, pp. 87-94. http://neuron.tuke.sk/competition/reports/Javier
[13] A. D. Papalexopoulos, H. Shangyou, P. Tie- Marin.pdf
Mao, An implementation of a neural network [24] E. Castillo, B. Guijarro, A. Alonso, Electricity
based load forecasting model for the EMS, load forecast using functional networks,
IEEE Transactions on Power Systems, Vol.9, [online], Available: http://neuron.tuke.sk/
No.4, 1994, pp. 1956-1962. competition/reports/ BerthaGuijarro.pdf
[14] D. Samek, P. Dostal, Artificial Neural [25] S. F. Crone, N. Kourentzes, Feature selection
Networks in Prediction and Predictive Control. for time series prediction – A combined filter
In Proc. of the 22nd European Conf. on Model. and wrapper approach for neural networks,
and Sim. ECMS 2008. European Council for Neurocomputing, Vol. 73, 2010, pp. 1923-
Modelling and Simulation, 2008. pp. 525-530. 1936.
[15] L. Macků, D. Sámek, Two step, PID and model [26] E. Guresen, G. Kayakutlu, T. U. Daim, Using
predictive control using artificial neural artificial neural network models in stock
network applied on semi-batch reactor. WSEAS market index prediction, Expert Systems with
TRANSACTIONS on SYSTEMS. Vol.9, No.10, Apps, Vol. 38, No. 8, 2011, pp. 10389-10397
2010, pp. 1039-1049. [27] M. Leshno, V. Y. Lin, A. Pinkus, S. Schocken,
[16] Q. K. Al-Shayea, G. A. El-Refae, Evaluation of Multilayer feedforward networks with a non-
Banks Insolvency Using Artificial Neural polynomial activation can approximate any
Networks, In Evaluation of Banks Insolvency function. Neural networks. 1993, Vol. 6, No. 6,
Using Artificial Neural Networks, WSEAS, pp. 861-867.
2012, pp. 113-118.
[17] M. Oprea, A. Matei, Applying Artificial Neural
Networks in Environmental Prediction

ISBN: 978-1-61804-108-1 458

You might also like