Abstract
Unplanned power outages create major economic costs. To better predict and manage these events, we present a multilayer graph neural network (GNN) framework that captures spatial, short-term co-occurrence, and statistically enriched co-failure patterns using 7 years of Oklahoma Gas & Electric data from 347 substations. These relations are encoded with weighted graph convolutions and fused by attention into a unified representation. For predictive maintenance, the model flags substations requiring near-term intervention across 30-, 60-, and 180-day horizons, achieving a peak 30-day F1 score of 0.8935. The same representation supports resilience planning by clustering substations into eight risk groups, each with distinct incident rates and recovery times. The highest-risk group has over six times higher incident rates than the low-risk groups. Combining prediction and clustering in a single framework provides utilities with an integrated basis for scheduling maintenance, prioritizing inspections, and targeting grid-hardening investments to minimize outage impacts and enhance reliability.