A Probabilistic Framework for Reconstructing Weather-Related Power Outage Frequency

Abstract Weather-related power outage models are usually evaluated on how well they predict a select number of individual events, whereas long-term resilience planning also requires understanding of how outage frequencies are affected by climate variability and change. We develop a probabilistic framework that reconstructs weather-related outage frequency by summing calibrated daily class probabilities. It combines EAGLE-I outage observations with hydrometeorological data across all 67 Florida counties from January 2015 to September 2024. A multiclass gradient-boosted tree model separates days with no weather-related outage from those with minor (1–2 h), moderate (2–8 h), and major (at least 8 h) outages, and is evaluated by leave-one-year-out cross-validation. The predicted probabilities are well calibrated, with expected calibration errors of 0.005 or less, and summing them recovers every class total within 3%. Including static attributes in the model raises the spatial Kling–Gupta efficiency from 0.69 to 0.97 while leaving temporal skill unchanged, showing that weather controls when outages occur and local attributes control where. For nine water years, predicted and observed annual totals correlate at 0.82 across counties. Across 26 tropical cyclones, totals correlate at 0.95 for all outages and 0.85 for major outages, far exceeding a county-by-month baseline. Skill is lowest for major and summer outages, reflecting convective and restoration processes that daily county-scale predictors do not resolve. The framework offers an open and transferable route to reconstructing outage climatology and to isolating climate-driven changes in outage frequency under a stationary-grid assumption.