Optimal Control of Fed-Batch Processes Based on Multiple Neural Networks

The performance of empirical model based fed-batch process optimal control is strongly affected by the model prediction reliability at the end-point of a batch. An optimal control profile calculated from an empirical model may not give the best performance when applied to the actual process due to model-plant mismatches. To tackle this issue, a new method for improving the reliability of fed-batch process optimal control by incorporating model prediction confidence bounds is proposed. Multiple neural networks (MNN) are used to build an empirical model of fed-batch process based on process operation data. Model prediction confidence bounds are calculated based on predictions of all component networks in an MNN model and the model prediction confidence bound at the end-point of a batch is incorporated into the optimization objective function. The modified objective function penalizes wide prediction confidence bounds in order to obtain a reliable optimal control profile. The non-linear optimization problem based on MNN with augmented objective function is solved by iterative dynamic programming. The proposed control strategy is illustrated on a simulated fed-batch ethanol fermentation process. The results demonstrate that the optimal control profile calculated from the proposed approach is reliable in the sense that its performance degradation is limited when applied to the actual process.