Texas Power Outage Prediction Dataset with NWS/EAGLE-I Geographical and Lag Features (Lag 1, 12, 24)

This dataset supports the development and evaluation of machine-learning models for predicting county-level power outages caused by multiple types of extreme weather events in Texas. It contains hourly, county-level data covering the period from 2014 to 2023 and integrates power outage records with meteorological, geographic, vegetation, transportation infrastructure, and extreme weather event information. The dataset includes historical variables at three temporal lag intervals: 1 hour (lag 1), 12 hours (lag 12), and 24 hours (lag 24). Separate data configurations are provided for each lag interval, enabling the evaluation of short-term and longer-term power outage prediction horizons. The lagged variables include historical power outage counts, weather conditions, and vegetation information, where applicable. The primary target variable is the number of customers experiencing power outages in each county. Power outage data were obtained from the Environment for the Analysis of Geo-Located Energy Information (EAGLE-I) dataset. Meteorological variables were derived from the National Aeronautics and Space Administration (NASA) Prediction of Worldwide Energy Resources (POWER) application programming interface (API) and include hourly weather indicators such as temperature, precipitation, wind speed, humidity, and atmospheric conditions. Additional variables include county-level population and geographic characteristics, the Leaf Area Index (LAI) derived from Moderate Resolution Imaging Spectroradiometer (MODIS) data, and roadway information used as a proxy for electrical infrastructure density. Extreme weather event information was obtained from the National Weather Service (NWS) and includes event types, affected counties, and event timing. The dataset focuses on major weather hazards associated with power outages in Texas, including thunderstorm winds, winter storms, hail, flash floods, drought, winter weather, ice storms, cold/wind chill, and frost/freeze events. All source datasets were spatially harmonized at the county level and temporally aggregated into hourly intervals. The data can be used to reproduce power outage prediction experiments, compare model performance across different forecast horizons, investigate temporal dependencies in outage occurrence, and examine the relationships among extreme weather, infrastructure, environmental conditions, and electric power disruptions. The corresponding study applies Light Gradient Boosting Machine (LightGBM) and Extreme Gradient Boosting (XGBoost) models using training data from 2014 to 2022 and held-out testing data from 2023.