Abstract. Accurate and scalable building energy prediction in early design remains challenging due to heterogeneous spatial configurations, climate variability, and limited data. This study proposes a hybrid framework integrating transfer learning with physics-embedded graph neural networks (GNNs) to improve generalization and interpretability. Buildings are modeled as heterogeneous graphs, where spatial units are nodes with geometric and material attributes, and thermal interactions are edges. A transfer learning strategy adapts knowledge from data-rich source domains to data-scarce target cases, enhancing cross-building and cross-climate applicability. Fundamental heat transfer mechanisms—conduction, convection, and radiation— are embedded into the GNN message-passing process to enforce thermodynamic consistency and mitigate overfitting. Experiments on multi-climate datasets with diverse typologies show that the proposed method outperforms purely data-driven GNNs and conventional machine learning models, while requiring fewer targetdomain samples. The framework also improves interpretability by linking learned representations with physical principles, supporting generalizable and data-efficient energy prediction for green building design.
Li et al. (Tue,) studied this question.