Machine learning algorithms applied to a prediction of personal overall thermal comfort using skin temperatures and occupants' heating behavior.

Thermal comfort modeling has been of interest in built environment research for decades. Mostly the modeling approaches focused on an average response of a large group of building occupants. Recently, the focus has been shifted towards personal comfort models that predict individuals' thermal comfort responses. Currently, thermal comfort responses are collected from the occupants via survey. This study explored if the thermal comfort of individuals could be predicted using machine learning algorithms while relaying on the set of collected inputs from an experiment. The model was developed using experimental data including collected from a previously performed experiment in the climate chamber. Two different approaches based on the output data (thermal sensation and thermal comfort votes) and five different sets of input variables were explored. The algorithms tested were Support Vector Machine with four different Kernel functions (Linear, Quadratic, Cubic and Gaussian) and Ensemble Algorithms (Boosted trees, Bagged trees and RUSBoosted trees). The combination of occupants' heating behavior with a personal comfort system (PCS), skin temperatures, time and environmental data were used for the development of personal comfort models to predict individuals' thermal preference. The study investigated the novel combination of inputs such as the use of skin temperature and settings of the personalized heating system as parameters in predicting personal thermal comfort. The results showed that personal comfort models among all tested approaches and subjects showed the best median accuracy of 0.84 using RUSBoosted trees. Individually looking, the approach using thermal sensation output produced better prediction accuracy. On the other hand, the models based on inputs that consisted of PCS control behavior and mean and hand skin temperatures produced the best prediction accuracy when assessing all tested algorithms. The main limitation of the study is the number of test subjects, and further recommendation is to perform more experiments.