Reliable on-line re-optimization control of a fed-batch fermentation process using bootstrap aggregated extreme learning machine

Abstract

This paper presents a reliable on-line re-optimization control of a fed-batch fermentation process using bootstrap aggregated extreme learning machine. In order to overcome the difficulty in developing detailed mechanistic models, extreme learning machine (ELM) based data driven models are developed. In building an ELM model, the hidden layer weights are randomly assigned and the output layer weights are obtained in a one step regression type of learning. This feature makes the development of ELM very fast. A single ELM model can lack of robustness due the randomly assigned hidden layer weights. To overcome this problem, multiple ELM models are developed from bootstrap re-sampling replications of the original training data and are then combined. In addition to enhanced model accuracy, bootstrap aggregated ELM can also give model prediction confidence bounds. A reliable optimal control policy is achieved by means of the inclusion of model prediction confidence bounds within the optimization objective function to penalize wide model prediction confidence bounds which are associated with uncertain predictions as a consequence of plant model-mismatch. Finally, in order to deal with unknown process disturbances, an on-line re-optimization control strategy is developed in that on-line optimization is carried out while the batch process is progression. The proposed technique is successfully implemented on a simulated fed-batch fermentation process.