Interpretable Severity Stratification of Weather-Associated County Power-Outage Events

Authors

  • Huachen Zhang School of Energy, Belarusian National Technical University

Keywords:

power system resilience, extreme weather, power outages, event severity, EAGLE-I; VTEC, machine learning

Abstract

This study examines whether information available at the first observed outage hour can help distinguish severe weather-associated county power-outage events from less severe events. The analysis uses the public EAGLE-I and National Weather Service Valid Time Event Code (VTEC) harmonized dataset. The downloaded primary CSV contains 3,572,394 county-hour records from 2 January 2015 to 3 January 2023. Because the final year is incomplete, the main analysis is restricted to 1,373,715 reconstructed events in complete calendar years 2015–2022 across 3,010 counties and 49 state-level jurisdictions. Event severity is defined as the event peak customer-outage count divided by a county-specific 75th-percentile outage baseline. A severe event is one at or above the 90th percentile of this ratio in the 2015–2020 training period. Regularized logistic regression and random forest models use only first-record information, including the matched VTEC phenomena, county location, calendar timing, number of simultaneous phenomena, and source-record availability. Models are trained on a reproducible 200,000-event sample and evaluated on a 100,000-event sample drawn from the 2021–2022 holdout period. The random forest achieves a ROC-AUC of 0.693 and average precision of 0.186, compared with a held-out severe-event prevalence of 0.100. When the highest-risk 10% of events are selected, precision is 0.221 and recall is 0.220. The results indicate modest ranking value, but also show that first-record warning and contextual features alone miss most severe events. The analysis is conditional on a positive outage already being recorded and does not estimate outage incidence or establish weather causation.

Downloads

Published

2026-11-17

How to Cite

Zhang, H. (2026). Interpretable Severity Stratification of Weather-Associated County Power-Outage Events. International Journal of Advanced AI Applications, 3(1), 16-28. https://www.dawnclarity.press/index.php/ijaaa/article/view/212

Similar Articles

1-10 of 74

You may also start an advanced similarity search for this article.