An Improved Distributed Secondary Control Scheme in Islanded DC Microgrid based on Deep Reinforcement Learning

This paper proposes an enhanced secondary control method based on deep reinforcement learning (DRL), building upon the conventional secondary control framework. A data-driven approach, specifically the Deep Deterministic Policy Gradient (DDPG) algorithm, is introduced. Except the voltage reference compensation provided by traditional secondary control for primary control, an additional voltage compensation value is generated by the DDPG agent. The proposed method operates in a distributed manner, requiring only information from neighboring units, which constitutes a subset of the data needed by the Dynamic Consensus Algorithm (DCA), thus imposing no additional communication overhead. Simulation results demonstrate that the proposed reinforcement learningbased secondary control significantly accelerates the restoration of bus voltage and the equalization of current compared to conventional secondary control.