Weighting Factors' Real-time Updating for Finite Control Set Model Predictive Control of Power Converters via Reinforcement Learning

The selection of weighting factors (WFs) is a common obstacle for the finite control set model predictive control (FCS MPC) of power converters. This paper proposes a generic approach to update the WFs via reinforcement learning (RL). The WFs' selection is self-taught online with full consideration of user-defined requirements. The trained policy is deployed to update the WFs in real-time. The self-taught process can be reactivated anytime in case of parametric variations or load change. This idea is verified on FCS MPC-regulated stand-alone inverters cascaded with LC input filters. Simulation results demonstrate that RL significantly improves the load-voltage tracking accuracy without sacrificing dc-link voltage stabilization.