Enhancing Power System Resilience through Explainable Hybrid Ensemble Modeling of Power Outage Severity

This record accompanies the manuscript “Enhancing Power System Resilience through Explainable Hybrid Ensemble Modeling of Power Outage Severity”. It contains (1) a merged dataset of 4,375 state-level U.S. power outage events (DOE OE-417, January 2000 – March 2023) linked to 32 onset-available climatic, socioeconomic, demographic, temporal and electricity-sector features and the outage duration in minutes; (2) the exact train/test split indices (row_id lists) – a chronological split (training 2000–2018, 2,663 events; test 2019–2023, 1,712 events) and a class-stratified random 80/20 split (3,500/875 events, seed 42) provided for comparison – with SHA-256 checksums; and (3) a documented Python pipeline that reproduces the full study: three-class duration severity construction (quartile thresholds Q1 = 80 min, Q3 = 2,268 min), preprocessing, 14 benchmark models, a CatBoost–LightGBM–LSTM stacking ensemble with out-of-fold meta-learning, isotonic probability calibration, evaluation metrics, ablation controls, leave-one-NERC-region-out validation, SHAP and partial-dependence analyses, and a dispatch-triage simulation. No post-event variable is included. Underlying sources are public: U.S. DOE OE-417, EIA, NOAA, BLS, U.S. Census Bureau and BEA. See README.md for the data dictionary and run instructions.