Interpretable AI model for predicting indoor thermal comfort: a SHAP-based analysis using a benchmark dataset

Purpose Ensuring indoor thermal comfort is critical for occupant well-being, productivity and energy efficiency in buildings. Traditional models, such as the Predicted Mean Vote (PMV), often fall short due to their reliance on static assumptions that fail to reflect the dynamic and individualized nature of human thermal perception. Design/methodology/approach This study proposes an interpretable artificial intelligence (AI) framework for predicting thermal comfort using machine learning techniques applied to the ASHRAE II benchmark dataset. The model is built using eXtreme Gradient Boosting (XGBoost), which demonstrates strong performance in predicting both thermal sensation and comfort levels. To address the critical challenge of model interpretability, the SHapley Additive exPlanations (SHAP) technique is employed, offering both global and instance-level insights into feature importance. Findings Key features influencing predictions include operative temperature, air velocity, relative humidity, clothing insulation and metabolic rate. The SHAP analysis not only validates the model's decision-making process but also identifies nonlinear and interactive effects among varied environmental conditions, which are typically overlooked in conventional models. Originality/value The research highlights how interpretable AI can bridge the gap between data-driven precision and practical usability, paving the way for occupant-centric, adaptive climate control strategies that align thermal comfort with energy conservation goals. This work contributes to the emerging field of human-in-the-loop intelligent building systems, where transparency, personalization and performance are equally prioritized.