Input convex neural networks in nonlinear predictive control: A multi-model approach
Maciej Ławryńczuk
Abstract
The presented input convex neural multi-modelling approach to Model Predictive Control (MPC) has two essential advantages. Firstly, the MPC algorithm solves only convex optimisation tasks; complex not convex and multi-modal problems are not considered. Secondly, the model is trained in a simple one-step-ahead configuration and it is not used recurrently for prediction. Two neural multi-model structures are considered: fully input convex and partially input convex ones. The fully convex model structure turns out to be too restrictive for the considered polymerisation reactor benchmark process. The partially convex model leads to very good prediction accuracy and excellent control quality. Development of fully and partially input convex multi-models is thoroughly described, particularly selection of the number of the hidden nodes. Additionally, low effectiveness of a linear multi-model is discussed for comparison.