Weather-guided surrogate modeling of building energy use in Nepal and zero-shot generalization across South Asia

Abstract

Building performance simulation is a standard tool for energy codes, retrofit screening, and design, but it is seldom used in Nepal or elsewhere in South Asia. This paper applies a recently published weekly weather-guided surrogate framework to Nepal, where ten typical-meteorological-year stations cover 72–2,300 m of elevation within roughly 200 km. A total of 1,100 EnergyPlus simulations of the U.S. Department of Energy (DOE) medium-office prototype were run at 22 locations. A transformer trained on the Nepali stations predicted weekly energy use at fully held-out Nepali stations with a symmetric mean absolute percentage error (SMAPE) of 3.1–8.7%. One change to the reference method was required: weather inputs were scaled to fixed physical bounds rather than to per-location statistics. This is a precondition for zero-shot use and also improved transfer between climates. With no local simulation, the Nepal-trained model predicted ten cities in the other seven South Asian countries at 3.1–5.9% weekly SMAPE, and sliding-window training with a three-seed ensemble reduced the regional mean to 4.0%. When the training network was cut to three stations, every unseen Nepali station stayed within 5.7% as long as the three spanned the hot, intermediate, and cold regimes, whereas removing the single cold-mountain station degraded accuracy at every cold city. Ulaanbaatar, designated in advance as outside this coverage, was consistently underpredicted; additional Nepali stations or training windows did not correct this, and augmentation made it worse. Coverage of climate regimes, rather than the volume of training data, determines where the model can be trusted.