A Deep Reinforcement Learning-Based Direct Switching Controller Design for Permanent Magnet Synchronous Motors

Direct switching control methods for permanent magnet synchronous motor (PMSM) drives offer advantages stemming from the elimination of an intermediate modulator. However, they often feature a complex control structure, increased online computations and memory burden, and require high sampling frequencies for satisfactory steady-state performance. This paper introduces a novel direct switching controller design for torque control of PMSM drives based on model-free reinforcement learning. A real-time-capable explicit control policy is trained to directly manipulate the switching times of the inverter terminals, thereby eliminating the necessity for cascaded control architectures, as well as resource-intensive computations and storage. Three trained agents are evaluated and compared to a state-of-the-art direct switching model predictive control approach, demonstrating fast and precise torque tracking with reduced current harmonics at a lower sampling frequency.