Online-optimized feed switching in semi-batch reactors using semi-empirical dynamic models

Short prediction horizons, as required by continuous processes, or accurate process models allow adequate on-line model-predictive input correction us ing conventional filter-based state estimation and model integration. For i nadequately modeled time-varying or high-order systems requiring long predi ction horizons, e.g. batch processes, a reliable, neural-network-based, sem i-empirical predictor previously proposed by the authors (Schenker & Agarwa l, 1995b, International Journal of Control, 62(1), 227-238) is better suite d for model-predictive control. This work achieves satisfactory on-line opt imization of the feed switching in a simulated semi-batch chemical reactor using the proposed scheme and compares it with other neural-network-based s chemes and with conventional state filtering followed by model integration. (C) 2000 Elsevier Science Ltd. All rights reserved.