Structural Optimization of Meta-Reinforcement Learning-based Finite-Control-Set Direct Torque Control of Permanent Magnet Synchronous Motors

Efficient and adaptable control strategies are important for optimal performance in modern electric drive systems. Unlike conventional methods, meta-reinforcement learning (MRL) offers a data-driven, adaptable approach that generalizes control across various setups with minimal comissioning time. Previous investigations have presented an MRL approach for direct torque control of permanent magnet synchronous motors (PMSMs) through online system identification, generating a context representation of the identified drive to enable targeted control. Although proven effective for real-world control, the presented MRL-based control approach has not yet been optimized systematically. This paper presents a hyperparameter optimization (HPO) and a comparative analysis of context designs. HPO results show performance improvements and a significant reduction in network size, enhancing transferability to cost-effective real-time-capable controller hardware, and bringing MRL-based PMSM control closer to industrial application. Additionally, the comparison of dynamic and quasi-static context designs reveals that both designs are valid for the presented problem.