Learning To Resilience: A Gnn-Enabled Optimal Dynamic Network Reconfiguration of Active Distribution Networks Under Typhoons

Typhoon-induced failures pose significant threats to active distribution networks (ADNs), necessitating rapid dynamic network reconfiguration (NR) for operational resilience. Traditional mixed-integer programming (MIP) approaches for the NR suffers from prohibitive computational burdens that impede realtime emergency response. This paper proposes a novel graph neural network (GNN)-enabled optimization framework (GraNR) for fast and near-optimal dynamic reconfiguration under typhoons. The framework incorporates a switch-as-gate message passing mechanism to model physical switching operations, topologyagnostic local predictors to ensure scalability across network configurations, pre-ML filtering to identify non-critical switches, and post-ML confidence-based selection to strategically fix highconfidence decisions while retaining uncertain variables for subsequent MIP optimization. By integrating neural predictions with mathematical programming, GraNR leverages the speed of machine learning and the optimality guarantees of optimization solvers. Numerical experiments on the IEEE 33-bus system demonstrate that GraNR achieves 1.64× computational speedup with only 0.67% average optimality gap, validating the practical viability for enhancing distribution system resilience against extreme weather events.