AC-Screened Robust Restoration of Weather-Stressed PV–Storage–EV Distribution Networks via Graph Learning and Multi-Agent Control

Extreme weather couples spatially correlated component damage with photovoltaic (PV) derating, changing electric-vehicle (EV) demand, repair delay, and time-varying network topology. This paper develops a coordinated restoration architecture for multi-area feeders containing PV, battery energy storage, and charging stations. Its weather-facing layer constructs joint outage-risk, renewable-error, charging-demand, and voltage-vulnerability descriptors. Those descriptors parameterize a two-stage mixed-integer second-order-cone program with a finite-support optimal-transport ambiguity set that remains well defined for discontinuous mixed-integer recourse. Regional actor–critic agents propose five-minute corrections around the hourly robust schedule; constrained projection, non-linear AC power-flow screening, emergency fallback, and margin-tightened re-optimization retain the authority to accept or reject each proposal. The evaluation uses public 33-node and 123-node feeders together with synthetic 240-node and 850-node stress networks. A pre-fit manifest allocates 240 records to training, 80 to validation, and 320 to final testing, while aggregate operational outcomes cover 50 random streams. Within this controlled benchmark, accepted schedules restore 93.6% of critical-load energy (SD 2.1 percentage points), serve 96.7% of total demand (SD 1.8 percentage points), retain 82–86% of EV service across hazard classes, and reduce the modeled 24 h objective by 25.8% relative to deterministic dispatch. The full pipeline records two to four candidate-stage voltage-limit events by hazard, and 4.9% of candidates undergo tightened re-optimization before accepted schedules reach zero reported AC voltage-limit violations. Between-method comparisons are descriptive and unpaired; the larger synthetic cases are structural stress tests rather than feeder-transfer tests.