An Explainable Machine Learning Framework for Healthcare Building Energy Demand Forecasting

Abstract:

Accurate building energy forecasting is essential for efficient energy management and operational optimization in intelligent building systems. This study presents an explainable machine learning framework for high-resolution healthcare building energy demand forecasting. Random Forest, XGBoost, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) models were evaluated alongside a naïve persistence baseline using engineered temporal and meteorological features. Random Forest achieved the best predictive performance, with MAE = 12.99, RMSE = 29.32, R² = 0.9930, and SMAPE = 8.17%, followed closely by XGBoost. In contrast, the recurrent neural models produced comparatively higher prediction errors. The naïve persistence model achieved a competitive SMAPE of 12.48%, highlighting strong short-term temporal dependence in healthcare energy demand. SHapley Additive exPlanations (SHAP) analysis identified recent energy demand and building size as the dominant predictors, while meteorological, seasonal, and calendar variables had smaller contributions. Ablation studies, cross-building validation, statistical significance testing, and residual diagnostics further supported model robustness. Overall, the findings demonstrate that explainable ensemble learning provides an accurate, interpretable, and scalable approach to healthcare building energy forecasting.