Fault Recovery Strategy for Urban Distribution Networks with Distributed Photovoltaic Generation Based on Dual-Dimensional Graph Reinforcement Learning

With the rapid development of renewable energy, the large-scale integration of distributed photovoltaic (PV) generation into urban distribution networks has resulted in increasingly complex topologies and frequent PV power fluctuations, posing challenges for fault-state assessment and time-constrained restoration decision-making. To address these issues, this paper proposes a dual-dimensional graph reinforcement learning framework, termed NWD, which integrates a dual-dimensional state assessment mechanism and a reward-based decision-making mechanism. Specifically, topological features and node features are jointly utilized to improve fault-state representation, while a dual-dimensional reward function guides decision-making from topology and load recovery perspectives. Based on this framework, an enhanced graph reinforcement learning algorithm, termed E-GRL, is developed. First, a topology-driven priority evaluation algorithm, NEM, is introduced to assess node topology priorities based on connectivity and topological relevance. Then, NEM-based topology information is combined with node features to construct the dual-dimensional fault-state assessment algorithm WSA, enhancing state perception. Subsequently, E-GRL designs the state space, action space, and reward function, and employs DDQN to achieve interactive training between the agent and the PV-integrated distribution network environment. Simulation results demonstrate that the proposed strategy achieves high load recovery rates with fewer switching operations under various fault and operating conditions.