Energy consumption prediction of residential area in cold regions: a deep transfer learning approach with limited data

Urban residential areas account for 20% of the total energy consumption in buildings. Accurate prediction of residential area energy consumption offers significant values for energy management and energy-saving design. Existing energy simulation engines are not well-suited for residential area scale, while black-box models typically face challenges such as data dependency, the complexity of buildings, and inter-building environmental impacts, leading to poor generalization ability. This study proposes a TabNet and transfer learning-based method for predicting energy consumption at residential area scale. The interpretability of the model assists energy-saving potential analysis and energy-saving design decisions. A cross-region case study is conducted to validate the generalization and to comparatively analyze different prediction models. The results indicate that the proposed method for predicting the energy consumption of residential areas has improved the R2 by 13.81% compared to the XGBoost model and 2.38% compared to the TabNet model in external validation set. The difference between the prediction accuracy of the TabNet-TL prediction model in the internal and external validation sets is only 3.37%. Through interpretable machine learning, 11 features including building density, floor area ratio and building type are identified as significant features in residential area energy consumption. This study enhances the accuracy and generalization ability of energy consumption forecasting for cross-region residential areas. The influence mechanisms of significant features directly help guide residential area design decisions.