Development of data-driven thermal sensation prediction model using quality-controlled databases

Predicting the thermal sensations of building occupants is challenging, but useful for indoor environment conditioning. In this study, a data-driven thermal sensation prediction model was developed using three quality-controlled thermal comfort databases. Different machine-learning algorithms were compared in terms of prediction accuracy and rationality. The model was further improved by adding categorical inputs, and building submodels and general models for different contexts. A comprehensive data-driven thermal sensation prediction model was established. The results indicate that the multilayer perceptron (MLP) algorithm achieves higher prediction accuracy and more rational results than the other four algorithms in this specific case. Labeling AC and NV scenarios, climate zones, and cooling and heating seasons can improve model performance. Establishing submodels for specific scenarios can result in better thermal sensation vote (TSV) predictions than using general models with or without labels. With 11 submodels corresponding to 11 scenarios, and three general models without labels, the final TSV prediction model achieved higher prediction accuracy, with 64.7%–90.7% fewer prediction errors (reducing SSE by 3.2–4.9) than the predicted mean vote (PMV). Possible applications of the new model are discussed. The findings of this study can help in development of simple, accurate, and rational thermal sensation prediction tools.