Prediction of Indoor Clothing Insulation Levels: A Comparison between Different Machine Learning Approaches

Accurate prediction of clothing insulation levels is imperative for reducing building energy consumption. Clothing insulation is a critical parameter in the prediction of occupant thermal comfort. Lack of this information may result in miscalculations in the comfort conditions required, which may result in poorly sized heating, ventilation, and air conditioning (HVAC) systems. Predicting thermal comfort via clothing insulation levels of occupants in indoor settings using machine learning (ML) is a hot research topic. The advances in ML opens new opportunities for occupant thermal comfort prediction to mitigate the challenges encountered by existing models. Diverse algorithms and data preprocessing methods get applied to predict thermal comfort indices in heterogeneous contexts. But limited studies have systematically analyzed how different algorithms and data processing methods can have repercussions on the prediction accuracy. We experimentally study the perspectives of predicted comfort indices, algorithms implemented, different input features, data sources, sample-size, training and test set proportion, and predicting accuracy. For the data collection, a Microsoft Kinect camera is deployed and created a database with different clothing patterns, see Figure 1 (a). Ground-truth labels were collected with a second camera to validate the data annotations on clothing patterns for the classification task. We have applied four ML algorithms (K-Nearest Neighbor, Catboost, Gradient Boosting, XGBoost) for the Clovalue estimation. We also investigated the clothing patterns in natural and dark light settings. The relationship between clothing and gender was also meticulously analyzed and came up with interesting conclusions. The results in Figure 1 (b) highlight that the KNN has the best performance among the tested algorithms with an accuracy of 84.50% in dark light setting and 91.68% or the natural light setting.