Deep transfer learning strategy based on TimesBlock-CDAN for predicting thermal environment and air conditioner energy consumption in residential buildings

The deployment of data-driven deep learning black-box models for thermal environment and air conditioner energy consumption modeling has gained popularity due to their high accuracy in the residential building sector. However, extensive data collection in real-world residential settings is both time-consuming and costly, complicating the development of these models. Although traditional domain adaptation methods enable transfer learning with minimal data from target buildings, significant discrepancies in insulation performance, thermal capacity, and air conditioner types between different buildings result in substantial label shifts between source and target domains. In this study, a domain adaptation strategy based on the TimesBlock-CDAN approach is proposed. This method integrates conditional features, including air conditioner heat input and the resulting variations in temperature, humidity, and energy consumption, into the adversarial domain adaptation process, effectively minimizing the domain shift between the source and target domains and thereby enhancing the overall efficiency and accuracy of transfer learning. Specifically, two real-world tasks were designed: transfer learning between different setpoint temperatures within the same building, and transfer learning across different buildings. A comparative analysis of performance in single-step and iterative multi-step predictions within these two tasks demonstrates that the proposed approach improves predictive accuracy by 20% compared to models trained solely on target building data and shows superior performance compared to other transfer learning approaches.