Robust Model-Free Reinforcement Learning Based Current Control of PMSM Drives

Data-driven control approaches of permanent magnet synchronous machines (PMSMs) have gained significant attention due to their ability to eliminate reliance on analytical machine models and parameters. Reinforcement learning (RL) has emerged as a viable method for achieving a data-driven current control of the PMSM drive without the necessity of machine parameter information. In this approach, RL is trained offline to learn an optimal control policy, resulting in a computationally efficient controller compared to other data-driven control methods. However, standard RL methods struggle to adapt to new operating conditions and different parameter sets, reducing system performance and robustness. This research proposes a multi-set robust reinforcement learning (MSR-RL) based current control method for PMSM drives. MSR-RL leverages multi-task RL to optimize a single policy that can generalize and provide robust performance across multiple parameter sets. The parameter sets, referred to as contexts, are represented as Contextual Markov decision processes (CMDPs), capturing the dynamics associated with each parameter set. During the training phase, CMDPs with shared information are clustered into models. These models are then utilized to generate a unified policy that remains robust to all clustered and unseen models. The effectiveness of MSR-RL is validated through comparison with standard RL based on numerical simulations, experimental tests, and robustness evaluation. The findings highlight the advantages of MSR-RL in terms of adaptability, robustness, and performance of PMSM current control.