Reliable optimal control of a fed-batch fermentation process using ant colony optimization and bootstrap aggregated neural network models

Abstract

This chapter presents a study on using bootstrap aggregated neural networks and ant colony optimization for the reliable optimization control of a fed-batch fermentation process. In order to overcome the difficulty of developing detailed mechanistic models, neural network models are developed from process operation data. Instead of developing a single neural network, multiple neural networks are developed on bootstrap re-sampling replications of the original training data. These networks are then combined to give the overall model. A further benefit of using bootstrap aggregated neural networks is that model prediction confidence bound can be calculated from individual network predictions. Model accuracy and reliability have significant impact on the optimization results. Due to the inevitable model plant mismatches, optimization results are only optimal on the model and may not be optimal on the plant. To address this issue, model prediction confidence bound is incorporated in the optimization objective function. In addition to optimizing process operation objectives, the proposed strategy also searches for solutions that lead to narrow prediction confidence bounds (i.e., reliable model predictions) under the calculated control policy. In order to overcome the local minima problems, ant colony optimization is used to solve the optimization problem. Application results on a simulated industrial scale fed-batch fermentation process demonstrate that the proposed strategy is very effective.