Application of Machine Learning Algorithms Based on Active Learning Strategies and Interpretable Models for <scp>HVAC</scp> System Energy Consumption Prediction

ABSTRACT Effective energy management in buildings is crucial for reducing emissions and achieving sustainability goals. However, traditional consumption prediction models often exhibit decreased accuracy with limited data, heightened sensitivity to outliers, and inadequate interpretability. To address these shortcomings, this study proposes a novel approach that integrates active learning strategies with commonly used machine learning algorithms—CatBoost and Random Forest—to predict and optimize HVAC energy consumption in office buildings. The method systematically evaluates predictive performance under various iteration counts, contrasting active learning with random sampling, and identifies the optimal model using metrics such as R 2 , RMSE, MAE, and MSE. Additionally, SHAP analysis is adopted to elucidate key input features and their interactions, enabling targeted optimization measures for HVAC operations. Experimental findings demonstrate that active learning improves accuracy under scarce data conditions, while the Random Forest model reduces outlier impacts, enhancing robustness and generalization. Furthermore, the SHAP‐based interpretability framework highlights major drivers of energy consumption and guides dynamic fan power control strategies. Compared with traditional methods, the proposed approach delivers high accuracy and stability despite small sample sizes or outliers, and offers more reliable, actionable energy‐saving recommendations for building operations. Hence, it presents a robust solution for advancing sustainable building management.