Rapid Establishment Method of a Personalized Thermal Comfort Prediction Model.

Abstract

In this paper, we address the challenge of predicting occupants' different thermal states (cold-uncomfortable, comfortable and hot-uncomfortable) with high accuracy and high flexibility. At present, most solutions are based on traditional average models or traditional personalized models, which generally fail to guarantee accuracy and flexibility at the same time. To address this issue, we introduce a rapid establishment method of a personalized thermal comfort model by using environmental and physiological parameters as inputs. When a model is being built for a new occupant based on pre-collected data of other occupants, the weights of the training data will be changed personally by quantifying the thermal sensation similarities between the target occupant and the occupants in the data set. In order to validate the method, 14 healthy subjects were recruited for experiments, during which four environmental and physiological parameters (air temperature, skin temperature, skin humidity, skin conductance) and their gradients were recorded. The model is based on LightGBM classifier and achieves an average weighted F1 score of 0.893 with a small amount of personal data. The results clarify the effectiveness of this method and also shows the possibility of applying this method to thermal environment control with wearable sensing technology.