Weighting Factor Design of FCS-MPC in Power Electronics Using DDPG and TD3: A Reinforcement Learning Approach

This study addresses the challenge of tuning a Finite Control Set Model Predictive Controller (FCS-MPC) in power electronics using Reinforcement Learning (RL). The focus is on a two-level voltage source converter (2L-VSC) equipped with an LC filter and resistive load, regulated by FCS-MPC. The primary strategy is to minimize a cost function that includes terms for reference tracking and current constraints, while an RL agent is used to adapt the cost function by also considering the Total Harmonic Distortion (THD). Two types of actor-critic RL agents, the Deep Deterministic Policy Gradient (DDPG) and Twin Delayed DDPG (TD3), are examined and compared. The study highlights the advantages of TD3, which exhibits more stable and rapid convergence in the training process.