Scenario-Aware Feeder Grouping with Transfer-Path Preservation for Post-Fault Load Restoration Using Graph Neural Networks

Abstract

Post-fault load restoration in active distribution networks requires fast switching decisions that remain consistent with topology and operating constraints. Feeder grouping can accelerate restoration by simplifying weak inter-feeder connections, but a fixed grouping rule may either remove valuable transfer paths or leave the mixed-integer linear programming (MILP) search space too large. This paper proposes a scenario-aware feeder-grouping framework with transfer-path preservation for MILP-based distribution-network restoration. The post-fault network is encoded as an attributed graph, and the learning module selects the feeder-grouping granularity and boundary-switch fallback configuration before optimization. The final restoration actions are generated by a constrained MILP, so voltage, connectivity, source, switch, and load-pickup requirements remain explicitly enforced. The studied 70-bus Hard N-3 results show that the proposed method preserves restoration quality close to a relaxed-budget global MILP reference while reducing online solution time. Additional robustness and validation checks cover random distributed-generation output, decomposition-parameter sensitivity, AC power-flow reconstruction, and equivalent flexible-source support. Mechanism and selector analyses further show that graph-guided grouping supports scenario-aware transfer-path preservation while screening low-value boundary interactions.