Advancing personal thermal comfort prediction: A data-driven framework integrating environmental and occupant dynamics using machine learning

Various individual thermal preferences do not always lead to optimal thermal comfort, despite the high energy consumption of Heating, Ventilation, and Air Conditioning (HVAC) systems. Thermal comfort models often struggle to accurately predict occupants’ thermal sensations due to their inability to adapt to diverse scenarios. This research employs machine learning techniques to develop personalized data-driven thermal comfort models using a global database of 17,814 samples from the ASHRAE Global Thermal Comfort Database II. Specifically, the study implements random forest and k-nearest neighbor algorithms to analyze the influence of environmental factors (such as air temperature, outdoor temperature, relative humidity, and air velocity) and personal factors (including clothing insulation, activity level, and individual differences) on thermal comfort perception. The study systematically compares the predictive performance of different output types (thermal sensation vs. preference vote), input parameters (6 vs. 12), and machine learning methods. The optimized random forest model with 12 inputs demonstrated an accuracy of over 70% in predicting thermal preference votes, which was significantly higher than the 34% accuracy of the PMV model. The results present new optimized data-driven models and offer insights into the relative influence of parameters such as climate, building cooling strategies, and occupant age and gender. This research aims to enable personalized thermal comfort predictions to improve satisfaction and energy efficiency in various buildings. •Developing extensive and accurate models to predict personalized thermal comfort.•Systematic evaluation of algorithms, inputs, and outputs to optimize predictors.•Employing hierarchical feature selection to find the optimal set of mixed features.•Providing a framework for developing accurate, practical, and adaptable models.