Pii: S0925-2312(01)00680-4

The paper presents a structured approach to building neural network models for a fed batch bioreactor to allow the development of reactor optimal control policy. Since the ultimate interest in batch bioreactor control is on the end-of-batch product quality, accurate long range predictions are essential in developing optimal control policy. To address the long range prediction issue, an augmented recurrent neural network is used to build long range prediction models which can predict the product quality over the batch trajectory. The augmented recurrent neural network contains two recurrent neural networks in series and utilises available process knowledge to improve long range prediction accuracy. The 3rst network predicts an important process state variable while the second network predicts the product quality variable. Based on the augmented neural network model, constrained optimisation techniques are used to 3nd the optimal control policy. The proposed technique is applied to a simulated fed-batch bioreactor. The results obtained are shown to be comparable to those computed using a full phenomenological model which is usually di5cult to obtain, demonstrating that the proposed approach can contribute to the optimal control of some batch processes where detailed mechanistic models are di5cult or infeasible to develop. c © 2002 Elsevier Science B.V. All rights reserved.