Parameter reduction for PMV prediction via data driven approaches using the ASHRAE global thermal comfort database II and Chinese dataset

Abstract

Thermal comfort significantly impacts building occupants' well-being, efficiency, and energy consumption. In this research, ASHRAE Thermal Comfort Database II, a comprehensive dataset consisting of 85,583 thermal comfort observation records, was used to develop ML models that can predict Fanger's PMV. While implementing the proposed model, feature selection analyses were used to indicate the importance of the parameters. The objective of the study is to reduce the factors to predict PMV from six (Tr, Va, Ta, clo, RH, and M) to three (Ta, clo, and M), keeping the predictions accurate and considering new, practical, and cost-effective aspects as the second work in the literature on the reduction of PMV factors having the higher accuracy. In the work, 10 ML techniques were evaluated, and the XGBoost model showed superior performance. The model attains remarkable outcomes for accuracy and interpretability by emphasizing critical characteristics: Ta, clo, and M. The research reveals that optimized version of XGBoost can improve the accuracy rate of PMV prediction to 96.29 % with 0.93 R2,0.124 MAE and 0.93 CV-R2 Mean values, while Lasso, the least accurate model among the 10 tested models, achieves an accuracy rate of 59.63 % with 0.22 R2, 0.478 MAE and 0.22 CV-R2 Mean values. The validation has been performed with various methods including cross-validation with ASHRAE Thermal Comfort Database II and fine-tuning using the Chinese Thermal Comfort Dataset containing 41,977 records. This work aims to increase the popularity of PMV usage, providing a practical and generalized solution suitable for a broad audience.