An Interpretable Stacking Ensemble Learning Framework for HVAC Energy Consumption Prediction

Abstract

To improve prediction accuracy and interpretability in heating, ventilation, and air conditioning (HVAC) energy forecasting, this study develops an interpretable stacking ensemble framework for fan power consumption prediction in smart building energy systems. The model integrates LightGBM, XGBoost, random forest, and CatBoost as base learners and uses ridge regression as the meta-learner under fivefold cross-validation. Operational data from an air-handling unit are used, covering environmental conditions, thermal states, humidity variables, and control settings. Results show that the stacking model outperforms individual models, achieving an R2 of 0.983 with mean squared error, root mean square error, and mean absolute error of 0.101, 0.317, and 0.106, respectively. Shapley Additive Explanations analysis identifies outdoor temperature, outdoor humidity, and return air temperature setpoint as the dominant factors. Temperature variables show nonlinear and piecewise effects, whereas humidity and control parameters exhibit more stable regulatory influences. The proposed framework provides an accurate and interpretable data-driven tool for HVAC operation optimization and energy-efficient building management.