AI framework speeds disaster power restoration by 3.6%

Researchers from Lin Jiang, Dahai Yu, Rongchao Xu, Tian Tang, and Guang Wang have developed EPOPR (Equity-aware Predict-Then-Optimize Framework for Post-Disaster Power Restoration), a machine learning framework designed to address systemic inequities in power restoration after disasters like hurricanes. The work, published as arXiv:2508.04780, highlights a critical flaw in current restoration strategies: disadvantaged communities often submit fewer restoration requests, leading to prolonged outages due to underrepresentation in decision-making processes.

EPOPR tackles this by introducing two key innovations. First, it uses Equity-Conformalized Quantile Regression to predict repair durations with uncertainty awareness, addressing data heteroscedasticity (where uncertainty varies across data points). Second, it employs Spatial-Temporal Attentional Reinforcement Learning (RL) to make equitable decisions that adapt to uncertainty levels across regions. In experiments, EPOPR reduced average power outage duration by 3.60% and decreased inequity between communities by 14.19% compared to state-of-the-art baselines. The framework is positioned for real-world deployment in collaboration with utility providers and disaster response agencies.