Topology-Aware Physics-Informed Reinforcement Learning for Resilient Operation and Adaptive Restoration of Renewable Integrated Active Distribution Networks

Abstract

The increasing penetration of photovoltaic generation in active distribution networks (ADNs) introduces significant operational variability and challenges for resilient network restoration. This paper proposes a Topology-Aware Physics-Informed Reinforcement Learning (PIRL) framework for adaptive operation and self-healing of renewable-integrated ADNs. The proposed approach integrates topology-aware network representation, physics-informed reward guidance, coordinated distributed generation (DG) and battery energy storage system (BESS) control, and feeder reconfiguration within a unified learning framework. The controller is evaluated on the IEEE 33-bus ADN under different combinations of PV generation, load demand, and feeder contingencies. Simulation results show that PIRL achieves a minimum voltage of 0.976 p.u. and reduces network power loss to 121.3 kW. Under a feeder fault, the proposed framework restores the interrupted load within 13 min with an energy not supplied of 72 kWh and achieves a resilience index of 0.96. In addition, PIRL reaches stable policy convergence at approximately 560 training episodes with a final normalized cumulative reward of 0.91. These results demonstrate that combining physical operating information with adaptive reinforcement learning and coordinated DG-BESS control can improve voltage security, network efficiency, restoration capability, and learning efficiency in renewable-integrated ADNs.