Progress in data-driven thermal comfort analysis and modeling

Abstract

Thermal comfort modeling has become increasingly reliant on data-driven approaches, utilizing the potential of machine learning algorithms to predict and improve indoor environmental quality. This paper provides a comprehensive review of methodologies and approaches employed in thermal comfort research, focusing on input variables, output predictions, applied algorithms, and performance evaluation metrics. The study systematically analyzes key environmental inputs, i.e. air temperature, mean radiant temperature, air velocity, and relative air humidity, along with physiological inputs including heart rate and skin temperature which are used to predict thermal comfort indicators, i.e. Thermal Comfort (TC), Thermal Preference (TP), Thermal Sensation (TS) and Predicted Mean Vote (PMV). A variety of machine learning algorithms, such as Random Forest, Support Vector Machine, Artificial Neural Networks, K-Nearest Neighbors have been applied in both regression and classification tasks. The assessment of thermal comfort models developed through the literature was carried out using common performance metrics; Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Accuracy, Precision, Recall and F1-score. The most common algorithms and performance metrics were considered from the reviewed papers and the best results are presented within the herein presented work reflecting the complexity of predicting personal thermal comfort. The paper also highlights the importance of dataset diversity, analyzing data sources from multiple geographical regions, including Europe, North America, and Asia. Finally, the future direction for thermal comfort modeling is discussed, emphasizing the need for developing a framework that can support the widespread application of thermal comfort models in energy-efficient building management systems.