Optimal local skin temperatures for mean skin temperature estimation and thermal comfort prediction of seated person in thermally stratified environments.

Thermally stratified environments are universal in "real world" buildings. However, the studies on the machine learning model and mean skin temperature (MST), which was based on the analysis of Local Skin Temperatures (LSTs), were insufficient in thermally stratified environments. To create thermally stratified environments in this study, the air temperatures at the lower body parts in a climatic box were controlled independently from the upper body parts exposed in climate chamber, with 12 air temperature combinations of 22, 25, 28, and 31°C. Sixteen human subjects were recruited to collect thermal perceptions and measure their LSTs. The variations of LSTs and the optimal LSTs to estimate MST and predict thermal state were analyzed. Based on the classifications of LSTs and area of local skin, a new method using chest (0.42), forearm (0.21), thigh (0.30), and foot (0.07) was proposed to estimate MST. Its errors decreased by at least 22.8% as compared to the existing methods. Then, the model based on Random Forest was used to filter the optimal LSTs for the predictions of Thermal Sensation Vote (TSV) and Local Thermal Comfort (LTC). Results showed at least three LSTs were needed to reach a robust model prediction accuracy and generalization ability. The optimal LSTs for the predictions of TSV and LTC were (Forearm, upper arm, foot) and (Forearm, chest, thigh), respectively. This study contributes to provide the basic information of optimal LSTs to improve the accuracies of the thermal comfort predictions and MST estimation in the thermally stratified environments.