Behavior Cloning based Deep Reinforcement Learning for Control of Grid-Connected Photovoltaic Inverter

This study introduces an intelligent control approach that integrates Behavior Cloning (BC) with Deep Reinforcement Learning (DRL)for the control of grid-connected photovoltaic (PV) inverters. This strategy is intended to overcome difficulties like slow system response and significant voltage fluctuations under nonlinear and dynamic load conditions. In the current loop, the BC is employed during the initial phase to pre-train the Actor network, providing an initial control policy that accelerates the convergence of the Deep Deterministic Policy Gradient (DDPG) algorithm. Through the on-line learning of DDPG, the control strategy is further optimized to ensure fast and accurate tracking of the reference current, while maintaining stability under load variations and disturbances. This strategy effectively regulates active and reactive power while suppressing voltage fluctuations. Simulation results demonstrate that the proposed BC-DDPG strategy significantly outperforms conventional PI control and standard DDPG in terms of convergence speed, control accuracy, and harmonic suppression.