Answers are grounded in this paper and cite their sources.
Data-driven building energy forecasting typically requires large amounts of historical data. However, real-world deployment of forecasting models often faces cold-start problems where local data is scarce or entirely unavailable. Cross-climate transfer learning offers a promising solution to address this issue, yet it introduces distribution shifts that challenge model robustness. We conduct a controlled study across multiple climate regions using a zero-shot protocol and a factorial design. Specifically, we compare two widely used model components: a convolutional neural network (CNN) for local pattern extraction and a temporal attention mechanism for time-step weighting, both integrated into a shared Long Short-Term Memory (LSTM) backbone. We further include a Transformer encoder as a reference model and evaluate eight independent seeds across Beijing, Hefei, and Huizhou. The results show no uniformly dominant architecture. Among the controlled LSTM-backbone variants, the attention-based LSTM achieves the lowest average NMSE and shows comparatively stable improvements, whereas the CNN-based and combined variants are more target- or seed-dependent. Together with supplementary MAE/RMSE, attention-weight, and source-target distribution diagnostics, these findings support context-dependent evaluation under the intended transfer setting using multiple diagnostic criteria.
No takes yet. Share an insight, caveat, or question.
Bian et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: