Design of Controller for three-phase Grid-Connected Inverter Using Reinforcement Learning

This paper introduces a method for designing a controller for a voltage source inverter (VSI) using Reinforcement Learning (RL). The proposed controller is based on RL Deep Deterministic Policy Gradient (RL-DDPG) algorithm, which independently updates its control gain online by the actor-critic structure without the knowledge of exact system model. The traditional controllers have a problems of model dependence, inadequate compromise between steady-state (SS) precision and dynamic responsiveness, resulting in complexity in controller gain tuning. The RL-DDPG agent is trained, tested, and validated. Following its implementation in the control system, it can furnish correction signals for robust control, leading to superior performance in minimizing SS error, ripple, and Total Harmonic Distortion (THD). Through simulation, the dynamic and SS performance of the RL-DDPG controller is compared with conventional proportional-Integral (PI) control. The simulation results show a reduction in THD and an improvement in current tracking performance.