Dynamic reconfiguration of large-scale distribution networks using GNN-guided B&B

Introduction The high penetration of renewable energy sources has posed significant operational challenges to modern distribution networks. Dynamic network reconfiguration, as a critical optimization technique, plays a vital role in reducing power losses and enhancing system flexibility. However, the dynamic distribution network reconfiguration (DNR) problem is often formulated as a large-scale Mixed-Integer Second-Order Cone Programming (MISOCP) model, where conventional branch-and-bound (B&B) algorithms suffer from the curse of dimensionality and typically fail to converge within the limited time window required for real-time decision-making. To address this issue, this paper proposes a graph neural network (GNN)-guided branch-and-bound approach. Methods First, a high-quality training dataset is constructed by collecting optimal branching decisions made by expert-level solvers across a large number of reconfiguration instances. Then, a GNN model is designed and trained to accurately predict the most promising branching variable under arbitrary network conditions. Finally, the trained GNN is integrated into the B&B framework to guide the branching process with learned decision policies, effectively pruning the search tree and accelerating convergence. Results and Discussion Comprehensive case studies on IEEE standard test systems demonstrate the superior performance of our approach, achieving on the 1354-bus system a 60.1% reduction in solving time, a 59.5% reduction in optimality gap, and a 0.39% reduction in operational cost. On the larger 2383-bus system, the improvements are even more substantial, with a 65.0% reduction in solving time, a 47.6% reduction in optimality gap, and a 0.65% reduction in operational cost compared to the standard SCIP solver.