Thermal comfort indices analysis using multiple linear regression and neural network

Abstract

Compared to methodology provided by standards concerning thermal comfort, by using models based on various approximation methods or artificial intelligence, it may be possible to ensure more time efficient and accurate calculation of thermal comfort indices. The aim of this study is to compare Predicted Mean Vote (PMV) computation model established by using multiple linear regression and trained artificial neural network, from the standpoint of accuracy. Both models are established on the basis of the same dataset which consists of 400 combinations of 4 thermal comfort parameters. These parameters are the air temperature, mean radiant temperature, relative humidity and clothing resistance, while activity level and air velocity are adopted as 1.1 met (office typing activity) and 0.05 m/s, respectively, and are considered constant values for selected type of indoor environment. Clothing resistance is adopted as 0.5 clo for summer period and 1.0 clo for winter period, while the air temperature, mean radiant temperature and relative humidity are values which are randomly generated within appropriately selected ranges. Taking into account that coefficients of determination which correspond to it are over 95%, resulting first degree polynomial relation obtained by using multiple linear regression can be considered a satisfactory approximation of PMV model as it is given in ASHRAE Standard 55-2020. Furthermore, there are certain input value combinations for which PMV values obtained by using this model coincide with the ones calculated by using algorithm which is provided by standard. However, results obtained by using trained neural network with one hidden layer coincide with PMV values calculated on the basis of ASHRAE Standard 55-2020 for each input value combination. Therefore, from the standpoint of accuracy, it is concluded that neural network provides significantly better approximation of PMV model.