Residual-Guided Ensemble Learning for Short-Term Building Energy Demand Prediction

Abstract

Accurate short-term building energy demand prediction is crucial to energy management, load scheduling, and efficient smart building operation. Individual machine learning models can exhibit different error patterns since energy demand is determined by the nonlinear interactions of historical consumption, indoor environment, outdoor weather, and temporal operating patterns. This paper proposes a Residual-Guided Ensemble Learning approach, which combines Random Forest, Extra Trees, and Gradient Boosting together and leverages the behavior of validation residuals to guide ensemble weighting that is followed by residual correction. The next 10-minute appliance energy demand is predicted with a chronological train-validation-test protocol. Experiments are carried out on the UCI Appliances Energy Prediction dataset with 19735 observations and 28 recorded variables. The proposed method achieves a test MAE of 32.96 Wh, RMSE of 63.10 Wh, MAPE of 31.28%, and R² of 0.519. Compared to the best individual baseline, Extra Trees, the proposed approach reduces the MAE by 21.27%, the RMSE by 10.91%, the MAPE by 25.48%, and improves R² from 0.394 to 0.519. The result suggests that validation-aware combination with residual correction can enhance the short-horizon building energy prediction in a manner that does not require a complicated deep-learning architecture.