Generalized Model Predictive Control for Multilevel Inverter Fed PMSM Drives Using DRL Agent

ABSTRACT Multilevel inverters (MLIs) are increasingly used in electric drives to improve power quality, reduce harmonic distortion, and enhance system efficiency. However, conventional model predictive control (MPC) schemes face challenges such as increased computational burden and limited scalability with higher inverter levels. This study aims to develop a generalized and real‐time feasible MPC framework for PMSM drives across any multilevel inverter configuration. The proposed method restricts the control action to a maximum of six nearest switching states, significantly reducing the combinatorial complexity, while a TD3‐based deep reinforcement learning (DRL) agent is employed in the outer speed loop to replace conventional PI control. Real‐time experiments using a 4.5 kW PMSM with three‐ and five‐level inverters and hardware‐in‐the‐loop validation up to nine levels demonstrate the method's effectiveness. The results show consistent execution time (~43 μs), over 10% reduction in torque and current ripple, minimal overshoot, and improved speed tracking, confirming the proposed strategy's suitability for scalable, high‐performance PMSM drive applications.