Neural Network based Optimal Control of a Fed-batch Reactor

glibdoadingΤεχνίτη Νοημοσύνη και Ρομποτική

20 Οκτ 2013 (πριν από 3 χρόνια και 11 μήνες)

98 εμφανίσεις

CPACT Confidential Proprietary. Photocopying

not permitted. Copy requests through CPACT (Newcastle)

Neural Network
based Optimal Control of a Fed
-
batch Reactor
§


Background

The o
ptimal control of
a fed
-
batch
bio
-
reactor
based on
an
augmented recurrent neural network model
has been studied.


Methodology

An augmented

recurrent
neural network was used to m
odel a fed
-
batch bio
-
reactor from process
operational data. During a batch,
reaction

volume was obtained through integration of the feeding
policy and was used as an important network input variable to enhance long range predictions. Based
on the augmented

neural network model, constrained sequential quadratic programming was used to
calculate the optimal
reactor
feeding policy.

Results

Figure 1 shows the
long
-
range

predictions of protein concentration (
P
M
) from the augmented recurrent
neural network on an
unseen batch
.
It can be seen that the predictions are quite accurate. Figure 2
compares the optimal control policies calculated from the augmented neural network model and a
mechanistic model. It can be seen that two control policies are very close indicat
ing the applicability
of the neural network model.



Figure
1
.
Long range

predictions
of
P
M

Figure
2
. Comparison
of

control policies


Conclusions

This study
demonstrated
that

by

incorporating
relevant

available

process knowledge into a neural
network m
odel can
enhance

the neural network model representation

and its predictive capabilities
.
In
particular, t
he augmented recurrent neural network modelling approach significantly improves the
model long range prediction
accuracy

leading to reliable
reactor
o
ptimal control policies.




§

Tian, Y., J. Zhang, and A. J. Morris, “Neural Network
Based Optimal Control of A Fed
-
batch Reactor
”,

CPACT Confidential Proprietary. Photocopying

not permitted. Copy requests through CPACT (Newcastle)








Proceedings of EUFIT

98
,
September 7


10, Aachen, Germany, Vol.1, pp308
-
312
.