Enhancing residential load forecasting accuracy through dynamic feature selection and ensemble machine learning models: A real-world scenario in Southern Finland

• Real-world data from 200 Finnish buildings used in seasonal evaluation • Dynamic feature selection applied to short-term residential load forecasting • Stacking ensemble model with Ridge meta-learner achieves the highest accuracy • Dynamic method outperforms static and full-feature models in 96–98.5% of cases • Up to 7% improvement in prediction error with adaptive feature selection This study aims to improve the accuracy of residential electricity consumption forecasts, a key component of modern energy management systems, especially in settings with high demand variability. The novelty of this work lies in combining dynamic feature selection–adapting to recent data patterns–with a stacking-based ensemble model that leverages multiple predictors. A comprehensive hourly dataset from 200 residential buildings in southwestern Finland was collected throughout 2023, including meteorological variables and real-time electricity prices. The forecasting task was defined as a day-ahead (24-hour ahead) prediction. The year was divided into cold (October–March) and warm (April–September) periods. Four machine learning models were employed per season, including XGBoost, Random Forest, Voting, and a Stacking ensemble with Ridge regression as the meta-learner. The stacking model achieved the best performance in 163 buildings in the cold period and 178 buildings in the warm period. Feature importance was assessed using SHAP values, comparing static and dynamic feature selection strategies. The dynamic approach reduced average prediction error from 11.85% to 9.31% in cold months and from 11.67% to 9.14% in warm months, outperforming the static method in over 96% of buildings. These findings underscore the effectiveness of adaptive feature selection and ensemble learning in capturing seasonal and behavioral dynamics in residential electricity usage.