Efficient Model Predictive Control Implementation via Machine Learning: An Algorithm Selection and Configuration Approach

Abstract

Model Predictive Control (MPC) is a widely used optimization-based control strategy for constrained systems. MPC relies on the repeated online solution of an optimal control problem, which determines the operation of the underlying system. However, the online solution of the optimal control problem can be computationally expensive. This necessitates a compromise between solution quality and solution time. In this paper, we propose a machine learning-based automated framework for algorithm selection and configuration for MPC applications. This framework aids the online implementation of MPC by selecting the best solution strategy and its tuning while accounting for solution quality and time. The proposed approach is applied to a mixed-integer economic MPC problem that arises in the operation of multiproduct process systems. The proposed approach allows us to (1) decide whether to use a heuristic or exact solution approach and (2) tune the exact algorithm if needed. The results show that machine learning can be used to guide the implementation of MPC and ultimately lead to lower average solution time while maintaining solution quality.