Understanding thermal comfort using self-reporting and interpretable machine learning

Standard thermal comfort models often fail to capture individual thermal sensations and offer limited interpretability for practical use. This study presents a building-specific, occupant-centric approach that combines self-reported comfort data with interpretable machine learning. The methodology is demonstrated through a case study involving self-reporting campaigns conducted during summer and winter seasons, accompanied by the development of a random forest regression (RFR) model. We employ three IML techniques namely Partial Dependence Plots (PDPs), SHAP values, and surrogate models to enhance the understanding of this RFR model. These interpretative tools facilitate a deeper understanding of the factors influencing thermal comfort, enabling targeted interventions for energy savings and improved occupant satisfaction. While the methodology offers a replicable framework for occupant-centric building control systems, it acknowledges limitations such as reliance on subjective self-reporting and the exclusion of architectural features. This research emphasizes the importance of integrating interpretable machine learning techniques to balance accuracy and usability, laying the groundwork for energy-efficient and occupant-focused indoor environmental management.