Reinforcement Learning-Based Current Controller for Switched Reluctance Motor Drives

This paper presents a machine learning-based current controller for a switched reluctance motor (SRM) using a reinforcement learning strategy. The proposed approach eliminates the need for a pre-existing training dataset, leveraging a self-learning process based on the motor’s nonlinear simulation, derived from its static characteristics. The control policy (control action) is determined using the Proximal Policy Optimization (PPO) method, where the value function (reward) is based on the current tracking error with a reference obtained from the sinusoidal Torque Sharing Function (TSF). Comparative analyses between the hysteresis controller and the proposed controller, operating at half the sampling frequency, are conducted under both steady-state and dynamic conditions. The results demonstrate near-optimal current tracking and smooth torque performance, even under control saturation conditions and operating scenarios distinct from the training phase.