A comprehensive method for exploratory data analysis and preprocessing the ASHRAE database for machine learning

Thermal comfort prediction is crucial for building energy efficiency and occupant comfort. ML methods are commonly used to predict thermal comfort. This research presents a comprehensive process for exploring and preprocessing the ASHRAE Database, providing a substantial dataset comprising 107,583 records of thermal comfort observations to create ML algorithms that can estimate Fanger’s PMV. With the most detailed cleaning and preprocessing stages in the literature, which included the imputation of missing values and the management of outliers, the final dataset is reduced to 55,443 records for the analyses. For practical applications and indoor comfort assessments, its estimation offers significant advantages due to its speed, ease of use, and cost-effectiveness. This study aimed to investigate which parameters are important in Fanger’s PMV model and which subset of variables is best for variable selection using different feature selection and analysis methods. The T_a and T_r had a high correlation value of 0.92, indicating a robust link between these two variables. The study employed Feature importance, the SelectKBest, SHAP, P-box, and PDP analyses, which showed consistency and suggested condensing the first six elements into three, and also was validated with the Chinese Database with 41,977 entries. The study targeted three parameters: T_a, clo, and M, using less expensive and simple measurement devices. To evaluate the accuracy of the research performance, RF and SVM models were created based on these three parameters. The results indicated that they have the accuracies of 85% and 70%, respectively, which are far better than the conventional models.