SMART BUILDING CLIMATE CONTROL: MACHINE LEARNING APPROACH FOR INDIVIDUAL THERMAL PREFERENCE PREDICTION

Abstract

Modern building management systems rely on uniform climate settings that fail to accommodate individual occupant preferences, resulting in energy waste and reduced comfort satisfaction. This study presents a data-driven approach for personalized thermal comfort prediction using machine learning algorithms integrated with multimodal sensor networks. We developed and evaluated three classification models (Random Forest, XGBoost, and Artificial Neural Network) using environmental parameters (air temperature, humidity, CO₂ concentration) and physiological measurements (heart rate variability, blood pressure, oxygen saturation) collected from controlled experiments with eight participants under various thermal conditions. The optimized Random Forest model achieved 95% accuracy in predicting seven-level thermal sensation votes using only ten key features identified through SHAP analysis. Indoor air temperature emerged as the dominant predictor, while physiological parameters provided complementary information for personalized comfort assessment. The proposed system demonstrates significant potential for integration into smart building automation, enabling dynamic climate control that adapts to individual preferences while optimizing energy consumption. Implementation of such personalized HVAC systems could reduce energy usage by up to 20% compared to conventional static temperature control, while simultaneously improving occupant satisfaction and productivity in commercial buildings.