NEURAL NETWORK MODEL PREDICTIVE CONTROL APPLIED TO FERMENTATION AND DISTILLATION PROCESSES

Abstract

In this work, Artificial Neural Networks (ANN) were used in combination with model-based predictive control (Model Predictive Control, MPC) for the fermentation and distillation processes. The relationship between saturation and the value of the ANN objective function used during the control simulation was evaluated. The ANNs showed to be able to represent the process well, with the control being robust and without the presence of offset. The RNA saturation did not influence the controller result, its value being typically less than 0.4 for structures with linear activation functions and greater than 0.6 for the sigmoid. The saturation value is more related to the structure and weight adjustment during ANN training, without a direct relationship with controller performance.