Evaluation of Classical and Reinforcement Learning Controllers for Motor Drives

This paper presents an analysis of reinforcement learning (RL) controllers for power converters applied in motor drives. A comparison between a classical proportional-integrative controller (PI), a finite control-set model predictive controller (FCS-MPC) and two reinforcement learning controllers is performed. These reinforcement learning techniques are: the Deep Q-network (DQN) and the Proximal Policy Optimization (PPO). An internal permanent magnet synchronous machine (IPMSM) is considered for evaluation of the four controllers. The mathematical theory of the neural networks training is presented, along with an analysis of the reward function applied to find the optimal control action. Simulation results are presented to show a detailed steady-state and transient analysis and comparison of the four control techniques evaluated.