Deep transfer learning-based hybrid modelling method for individual thermal comfort prediction

Abstract

Data-driven thermal comfort models offer higher accuracy and applicability compared with traditional Predicted Mean Vote (PMV) model. However, they still face two significant challenges: insufficient modelling data and feature selection. To address these challenges, we propose a framework for thermal comfort modelling based on transfer learning, named deep transfer learning (DTL) based convolutional neural networks-long short-term memory neural networks (DTL-CNN-LSTM). To mitigate the impact of insufficient data, we used deep transfer learning for thermal comfort modelling. We retained the knowledge (parameter weights) from the high-level mappings of the pre-trained model and transferred it to the target domain, while retraining the lower levels of the model architecture with the target dataset to enhance the model's prediction performance. For feature selection, we utilized the Pearson correlation coefficient and deep reinforcement learning (DRL) to refine feature selection process, thus identifying the most relevant features subset. Furthermore, we employed the oversampling method (K-Means SMOTE) to synthesize minority classes data for handling class imbalance problem. Extensive experimental results on Medium US Office dataset demonstrate that DTL-CNN-LSTM can achieve over 55% accuracy with limited data in the target buildings. Additionally, we verified the effects of newly added relevant features and knowledge transfer on model performance.