Transformer trained on Nepal data forecasts energy use across South Asia

Summary

A transformer trained on 1,100 EnergyPlus simulations across Nepal's 72–2,300 m elevation range predicts weekly building energy consumption at test stations (SMAPE 3.1–8.7%). With a single architectural change—scaling weather inputs to fixed physical bounds instead of location-based statistics—the model generalizes zero-shot to ten cities across seven other South Asian countries, achieving 3.1–5.9% SMAPE.

Why it matters

Building performance simulation is nearly absent in energy code enforcement across South Asia. Data scarcity and computational costs pose significant barriers in this field. A lightweight surrogate model that transfers across climates without local calibration could enable retrofit screening and code compliance studies in data-poor regions. The finding that generalization capacity is determined by climate regime coverage rather than data volume offers a practical roadmap for deploying such tools in other geographies with weak service infrastructure.

Highlights

• Scaling weather features to fixed physical bounds instead of location-based min/max values is the single change that enables zero-shot transfer and improves cross-climate accuracy.
• Three training stations covering hot, temperate, and cold regimes keep all unseen Nepal stations within a 5.7% error margin. Removing a single cold mountain station from training negatively impacts predictions for all cold cities.
• In Ulaanbaatar, which lies outside the climate envelope, consumption is consistently underpredicted. Additional training data or data augmentation methods do not resolve this issue. The binding constraint is coverage, not data volume.

Critical take

The achieved 3–6% SMAPE values were measured against EnergyPlus simulations of a single DOE mid-sized office prototype, not against metered consumption in real buildings. Therefore, the difference in practical accuracy is unknown. Describing this approach as 'zero-shot' is also a generous characterization. Nepal's 2,300 m elevation spread happens to cover many South Asian climates. Consequently, the model benefits from a favorable training distribution rather than proving a generalizable architecture.