Reinforcement Learning Control of Three-Level Converter Permanent Magnet Synchronous Machine Drives

Data-driven strategies present a suitable alternative for controlling electric drives. By learning from interaction with the plant, these strategies can theoretically find optimal solutions for the given objective and inherently incorporate the intrinsic nonlinearity and complexity of the controlled system within their policies. This investigation demonstrates the successful transfer of a reinforcement learning strategy for torque control of permanent magnet synchronous machine drives, originally proposed for three-phase two-level inverter applications, to a three-level topology. With minor a priori knowledge of the system parameters, the proposed holistic approach simultaneously achieves torque tracking and voltage balancing of the DC-link capacitors, targets to minimize copper and switching losses, and maintains the operating point within safeguarded margins at all times. The feasibility of the proposed method is tested through real-time hardware-in-the-loop validation and compared with a model predictive control approach.