Prediction of individual thermal comfort based on ensemble transfer learning method using wearable and environmental sensors

Thermal comfort is a critical issue in achieving an acceptable indoor environment and managing building energy use. However, it is difficult to precisely recognize thermal comfort because its determination varies depending on the characteristics of humans and indoor spaces. Moreover, accumulating datasets of indoor environmental and individual features is challenging in terms of both collection time and cost, and is sometimes unrealistic. This study established a prediction model for individual thermal comfort to mitigate this challenge. This model is based on ensemble transfer learning (TL) to transfer knowledge from datasets of someone in different indoor spaces and thermal environments, even if the physiological and environmental data of the target subject are insufficient. First, the physiological data of each subject and the indoor environmental data were collected from wearable wristbands and sensors. Then, a pre-trained model was developed with the datasets by combining deep learning and machine learning algorithms. Based on the pre-trained model, the ensemble TL method was applied to overcome the weak generalization performance that occurred when the dataset of each target subject was insufficient. The results revealed that the ensemble TL more accurately predicted the thermal comfort of two target subjects using the pre-trained model from a source. The accuracy and F1-score were both 95% for the first subject. For the second subject, they were calculated as 85% and 83%, respectively. It was also found that the ensemble TL was suitable for application when using fewer and imbalanced datasets in the target domains. • Individual thermal comfort was predicted using ensemble transfer learning method. • A hybrid model (CNN-SVM) was utilized as a pre-trained model to be transferred. • The proposed model increased accuracy by up to 7% and the F1-score by up to 8%. • Proposed model improved generalized performance on high variance of target dataset. • Effect of data availability and fine-tuning on model performance was explored.