Identification and application of the best-suited machine learning algorithm based on thermal comfort data characteristic: A data-driven approach

Abstract

Integrating machine learning (ML) techniques in thermal comfort analysis is pivotal in contemporary research. This study systematically evaluates and applies ML algorithms for predicting thermal comfort parameters, emphasizing data imputation, parameter-specific behaviour analysis, and the development of an Independent Parameterised Modelling Ensemble (IPME) framework. A data-driven approach addresses missing values in the ASHRAE Global Thermal Comfort Database II using ML algorithms tailored to each thermal comfort (TC) parameter. Performance metrics, such as coefficient of determination (R2) for numerical data and accuracy for categorical data, guide the selection process, resulting in optimised ML-TC pairs for imputation. The analysis reveals parameter-specific behaviours, with numerical parameters exhibiting varying correlations with environmental conditions and parameters related to individual preferences posing challenges. ML models show improved performance with categorical data, potentially because of their ability to capture subjective aspects effectively. The IPME framework was proposed to address the conventional ML limitations by employing a hybrid ensemble strategy that selects tailored TC parameter-specific ML algorithms. This approach aims to enhance prediction accuracy by acknowledging the uniqueness of parameters and harnessing diverse ML capabilities. Future research directions include further exploration of the IPME framework and post-imputation retraining of ML models to capture actual behaviour. This study underscores the potential of ML in enhancing thermal comfort analysis, offering insights into imputation strategies, parameter-specific behaviours, and ensemble modelling approaches for more reliable predictions in building environments.