Intelligent Voltage Regulation and Fault Recovery in Active Distribution Networks Using Physics-Guided Graph Reinforcement Learning

The increasing penetration of photovoltaic generation and battery energy storage is making distribution networks more dynamic and increasingly sensitive to renewable variability and feeder disturbances. This study proposes a Physics-Guided Graph Reinforcement Learning (P-RL) framework for coordinated voltage regulation and adaptive fault recovery in renewable-rich active distribution networks. The proposed method represents the distribution feeder as a dynamic electrical graph, enabling the learning agent to capture bus connectivity, voltage conditions, distributed-generation output, feeder status, and BESS state of charge. A physics-guided evaluation mechanism further filters control decisions according to voltage, feeder-loading, battery, power-loss, energy-not-supplied, and switching constraints. The resulting policy coordinates DG reactive-power control, BESS charging/discharging, and feeder reconfiguration within a unified control strategy. The framework is evaluated on the IEEE 33-bus distribution network under photovoltaic variability, load changes, and feeder contingencies. The results achieve a minimum voltage of 0.980 p.u. and reduce network power loss to 115.7 kW, corresponding to a 46.6% reduction. For the 9–10 feeder outage, the proposed method restores service within 11 min, limits ENS to 65 kWh, and attains a resilience index of 0.97. Stable policy convergence is reached at approximately episode 540. These results demonstrate that integrating network topology with physical operating knowledge can strengthen both voltage regulation and adaptive self-healing of renewable-rich distribution networks.