Resource Allocation for Resilient Distribution Systems Based on Multisource Data-Driven Failure Scenario Generation

Extreme weather increasingly triggers widespread distribution outages. Resilience planning requires realistic multiple line outage scenarios and restoration aware investments that remain valid under frequent topology changes. This paper couples inductive failure probability learning with a two stage resource allocation model. Each distribution line is modeled as a node in an edge graph, and its feature vector fuses meteorological intensity, geographic exposure, asset conditions, and connectivity cues. An inductive GraphSAGE predictor estimates line fault probabilities and transfers to unseen switching configurations. Independent Monte Carlo sampling converts these probabilities into a finite set of outage scenarios. The scenarios drive a two stage mixed integer linear program that sites dispatchable DG, PV, and energy storage, and schedules post event restoration actions, including optional network reconfiguration, under linearized power flow and radiality constraints. Tests on a modified IEEE 123 bus feeder show improved prediction accuracy and a 61.9% reduction in expected energy not served when reconfiguration is enabled.