Machine Learning-Based Initialization of Generalized Benders Decomposition for Mixed Integer Model Predictive Control

Model predictive control (MPC) has been widely used to control and operate complex systems. However, the efficient implementation of MPC depends on the efficient solution of the underlying optimization problem. In this work, we propose a machine learning (ML) based warm start initial-ization strategy for Generalized Benders Decomposition (GBD) to be used for the solution of mixed integer MPC problems. Specifically, the mixed integer MPC problem is first solved using an ML-based branch and check GBD algorithm, which provides a set of high-quality integer feasible solutions. These solutions are subsequently used to compute the exact Benders cuts, which are used to warm start the GBD algorithm. We apply the proposed approach to a case study on the operation of chemical processes where a mixed integer MPC problem is solved to compensate for disturbances. The results show that the proposed approach leads to 39% reduction in solution time compared to the standard application of GBD.