Answers are grounded in this paper and cite their sources.
Machine learning (ML) surrogate models accelerate early-stage building energy prediction by replacing slow physics-based simulation, but a surrogate is often tied to the climate of its training data, and reusing it elsewhere normally requires regenerating datasets and retraining. Transfer learning (TL) offers a potential solution by reusing a surrogate trained in one city for another, yet its reliability under cross-climate weather domain shifts remains poorly understood, and practitioners lack guidance on when it is worthwhile. This study investigates how far an ML building energy surrogate can be transferred across climates under limited target domain data. An artificial neural network trained on Gothenburg data was transferred to five target cities spanning a range of weather domain distances, quantified from hourly EPW data using Maximum Mean Discrepancy. Three TL methods were evaluated through transfer gain, negative transfer rate, and equivalent target simulations saved, across multiple target-data regimes and transfer configurations and compared to multiple linear regression and XGBoost as baselines. The results indicate that TL offers a clear advantage when the source and target climates are similar, whereas for distant climates, a well-tuned XGBoost model trained on the target data alone performs better. The results are consolidated into practical, exploratory maps. By locating a target city on these maps, practitioners can judge when cross-climate transfer is reliable, when it requires validation, and when local retraining is preferable, providing evidence-based guidance for cross-climate surrogate reuse under early-stage design constraints.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: