A Multi-Domain Approach to Explanatory and Predictive Thermal Comfort Modelling in Offices

It is well known that physical variables, such as temperature, exert a significant
influence on occupants’ thermal comfort in office buildings. Despite this knowledge, models that are currently used to predict thermal comfort fail to do so accurately, resulting in a mismatch between design conditions and actual thermal comfort conditions. The assumption is that exclusive attention to physical variables is insufficient for understanding or predicting thermal comfort. The question arises as to how a multi-domain approach can aid in
explaining and predicting thermal comfort in offices. In this study, a unique dataset containing indoor environment, demographic, occupancy and personality related variables is used to construct two types of thermal comfort models. The dataset contains 524 observations, collected during summertime in two office buildings in the Netherlands. Firstly, structural equation modelling (SEM) is used to construct an explanatory model, with the aim to identify significant variables affecting thermal comfort, as well as the interactions between them. Secondly, machine learning is used to train four binary classification models to predict
thermal discomfort. For the investigated cases, SEM suggests that thermal discomfort is significantly affected by (i) temperature, (ii) sound pressure level, (iii) the interaction between temperature, sound pressure level and illuminance, and (iv) the interaction between gregariousness and occupancy count. The four predictive models are subsequently trained using only the significant variables. Nevertheless, the weighted F₁-score for all four models ranges between 0.55 and 0.59, indicating weak predictive performance. The results show that significant influencers are not necessarily good predictors of thermal discomfort. Future researchers are encouraged to combine explanatory and predictive modelling techniques, in order to test whether variables that are relevant to the domain are useful for prediction.