Multistep Model Predictive Control Based on Artificial Neural Networks

Abstract

This work studies the effect of different models on the performance of multistep model predictive control (MMPC) via simulation examples and bench- and pilot-scale experiments. The models used in the study are two common types of artificial neural networks (ANNs), namely, feedforward networks (FFNs) and external recurrent networks (ERNs). The steady-state offset of MMPC using FFN models is observed throughout simulation cases and experiments in case that prediction horizon is longer than the control horizon. This study further explains the FFN-induced offset phenomena mathematically. In the experimental part of this work, we compare the performances of MMPC using these two ANN models, conventional proportional−integral controllers and linear model predictive control in the dual-temperature control problems, which include a bench-scale ethanol and water distillation column and a pilot-scale i-butane and n-butane distillation column.