Coordinating Time-Coupled Industrial Microgrid Flexibility for Resilient Distribution Network Restoration Under Extreme Events

Extreme-event restoration requires damaged distribution feeders to coordinate sequential line repair with time-dependent support from industrial microgrids. Temporary industrial load reduction releases capacity during outages but creates recovery obligations that compete with later load restoration and storage charging. We study a topology-aware restoration controller with staged correction of discrete network decisions and continuous industrial and storage actions, followed by bounded candidate screening. The modified IEEE 33-bus and 69-bus feeders are each evaluated over 440 matched physical scenarios and three independently trained policies, giving 3960 complete 96-step cases per feeder. Mean weighted grid energy not supplied is 7.268 versus 7.863 MWh for GS-HF and GS on IEEE33 and 5.640 versus 5.800 MWh on IEEE69. Their mean online decision times are 566.2 versus 18.8 ms and 3503.0 versus 24.2 ms, respectively. A separately certified 48-scenario IEEE33 encoder comparison with training seed 2027 yields 2.79% lower mean weighted EENS after removing inter-node message passing, so the complete-controller gains are not attributed to graph message passing. A matched 48-scenario industrial-storage sensitivity study and a completed CEPP action-chain audit provide additional implementation evidence. The demonstrated guarantee remains limited to the stated active-power, topology and industrial-resource model; online voltage, reactive-power and comprehensive AC feasibility are outside its scope.