A HYBRID NEURAL-NETWORK FIRST PRINCIPLES APPROACH TO BATCH UNIT OPTIMIZATION

Achieving an optimal operation of batch processes is a difficult probl em due to issues such as imperfect modelling, scarcity of information and delayed access to key measurements of process states. In this pape r, a hybrid approach for batch process optimisation that integrates to gether inductive learning (neural networks) and first principles knowl edge is proposed. The latter is expressed as derivative constraints th at impose learning bias for inductive modelling and generalisation. Th e result of such integration is a value function that is incrementally learned, and later used to implement a near-optimal control policy th at is geared to guarantee product quality and to improve process perfo rmance. Artificial batch runs are simulated by means of a predictive p rocess model to help constrain and speed up the learning procedure for the value function. The efficacy of the proposed methodology is demon strated using a semi-batch autocalytic reactor. (C) 1998 Elsevier Scie nce Ltd: All rights reserved.