Adaptive Ensemble Learning for Short-Term Energy Consumption Prediction Using Smart Building Data

Abstract

Accurate short-term energy forecasts can help schedule, monitor and make data-driven energy management decisions in smart environments. This paper proposes a leakage-aware adaptive ensemble framework that combines a persistence baseline, linear regression, histogram-based gradient boosting and XGBoost for short-horizon appliance-energy prediction. The study uses the publicly available UCI Appliances Energy Prediction dataset containing 19,735 ten-minute observations from a low-energy house. This approach constructs timestamp variables and lagged consumption features from only information available before the forecast instant. A chronological 70:15:15 train-validation-test split was used to prevent future information from entering model selection. Base-model contributions were determined from inverse validation MAE and normalized prior to generating the final forecast. On the held-out test period, the adaptive ensemble achieved MAE 28.99 Wh, RMSE 59.83 Wh, MAPE 27.12% and R² 0.567. It outperformed the individual machine-learning models evaluated when considering RMSE and R², although the persistence baseline achieved lower MAE and MAPE. This result highlights that adaptive combination can improve overall variance-sensitive forecasting while also underlining the need for multiple metrics and leakage-safe evaluation.