TA-23-C079 – An Evaluation of Ensemble Machine Learning Algorithms for Predicting Thermal Sensation using the ASHRAE Comfort Database II

Thermal comfort models, capable of predicting comfort sensation, can be integrated into HVAC systems to achieve both energy savings and enhanced comfort. Data-driven models serve as excellent tools for predicting thermal comfort in buildings, but they necessitate a substantial volume of training data. The ASHRAE Global Thermal Comfort Database II, comprising over 109,000 observations from field studies, facilitates the comparison of the performance of various data-driven algorithms for comfort prediction. Prior research indicates that the Random Forest (RF) ensemble algorithm surpasses other machine learning algorithms.

This study evaluates seven ensemble models, including Bagging, Random Forest, AdaBoost, Gradient Boosting, Xgradient Boosting, and LightGBM, to predict the three-point thermal sensation vote (TSV) for the ASHRAE Global Comfort Database II. Ten features, such as indoor and outdoor environmental conditions and age, were selected based on their importance in the prediction model as determined by Recursive Feature Elimination (RFE). Accuracy, precision, recall, and F1 scores were employed as performance metrics to evaluate the various algorithms. Unlike previous attempts, this study utilized the entire dataset rather than selecting subsets, to include the entire information of the set and better understand the extent to which this can be leveraged for better-generalized predictions that are not climate- or building-specific.

Additionally, the effect of similarity in observations was considered by employing a clustering function to identify similar observations in terms of indoor conditions and clothing levels. The probability of observations belonging to the same clusters was incorporated into the classification model, and the final TSV prediction was determined by considering similarity. The results demonstrated that LightGBM provided the best performance, with 57% accuracy, 55% recall, 57% precision, and 56% F1 on the test dataset. Incorporating clusters into the classification model improved performance by up to 3% across various ensemble models. Furthermore, as the entire dataset was utilized, the model is applicable across all climates, seasons, and building types.