Weather-electricity demand relationships: A comparative analysis of linear regression and machine-learning approaches in Texas

Weather is a major driver of electricity demand, and modelling this relationship is increasingly important as weather variability intensifies. This study examines zone-level weather–electricity demand patterns across Texas and evaluates whether machine-learning models improve performance relative to ordinary least squares (OLS) regressions. The analysis used hourly electricity demand from ERCOT and meteorological data from NASA POWER covering 2015-2024. Daily summaries were derived and split into training (2015-2019) and testing (2020-2024) periods. Two OLS models were developed using daily mean temperature and temperature-derived Cooling and Heating Degree Days (CDD/HDD) as predictors. The OLS models were compared with XGBoost and Random Forest models trained first on degree-day variables, and on extended meteorological feature set. Results showed that mean temperature is a weak predictor of electricity demand, while the degree-day OLS model outperformed all models in temperature-dominated zones. The zone-level analyses revealed heterogeneity in demand and model performance that aggregate analyses may obscure. XGBoost and Random Forest models provided benefits in zone with structurally complex demand characteristics; however, within the predictor set evaluated, they did not outperform the degree-day OLS models in temperature-dominated zones. Among the machine-learning models, Random Forest outperformed XGBoost in most zones. Incorporating additional weather variables did not yield significant improvements. Electricity demand exhibits an upward trend statewide, with the fastest growth rate (13.04%/year) in Far West Texas. The findings validate the utility of degree-days as suitable metric for capturing weather-driven component of electricity demand and underscore the need for geographically explicit, physically informed approaches to demand modelling.