Development of adaptive neural networks for flexible control of batch processes

This paper deals with the application of a neural controller for temperature control of a batch reactor. The term ''neural controller'' is used to refer to a multilayer neural network which computes the control values to be applied to the process. We present the design and the development of the neural network: architecture, learning database and learning procedure. In a first step, the learning phase consists in teaching the neural network to map the dynamics of a classical adaptive controller (generalized predictive control with double model reference) implemented on the process. Although the neural controller performance is good for operating conditions included in the learning set (interpolation), it exhibits limitations on extrapolation. In this work, two methods for the on-line adaptation of the network's weights are developed: one of them is the ''specialized'' learning technique, whereas the other uses another neural network in order to model the reactor dynamics. Several results are shown and prove the good capacities of neural networks for controlling batch processes.