Optimal control of batch processes incorporating model prediction .confidence bounds based on multiple neural networks

Abstract

The performance of batch process optimal control is strongly effected by the model prediction reliability at the end-point of the batch. A new method to solve the optimisation problem of batch processes incorporating model prediction confidence bounds is proposed in this paper. The empirical model of a batch process is built based on multiple neural networks and model prediction confidence bounds can be calculated based on the individual network predictions. Using the penalty function technique, the optimisation objective function is augmented with the model prediction confidence bounds at the end-point of a batch. The non-linear programming problem with augmented objective function can be solved by using iterative dynamic programming for discrete-time systems. The proposed optimisation scheme is illustrated on a simulated ethanol fermentation process. It is shown that the calculated optimal control profile using the proposed technique is much reliable in the sense that it can give consistently good performance when applied to the actual simulated process.