Decentralized Reinforcement Learning for Adaptive Power Sharing in Hybrid DC Microgrids

Abstract This paper proposes a decentralized voltage control strategy for islanded DC microgrids that replaces conventional droop control with a reinforcement learning (RL)-based approach. Using a Deep Deterministic Policy Gradient (DDPG) agent, the controller learns to generate real-time voltage references based solely on local measurements, eliminating the need for inter-unit communication. Compared to droop control, the proposed method reduces power sharing error from +30% to +8% and halves bus voltage deviation under high line impedance scenarios. The framework adapts to dynamic load and network conditions, offering a scalable and resilient control solution for next-generation microgrids.