Non-autoregressive transformer for high-resolution time series downscaling of solar irradiance
Year: 2026
Type: article
Abstract: The temporal resolution of solar irradiance data significantly impacts the accuracy of photovoltaic (PV) system simulations, particularly in estimating inverter clipping losses and power output variability. While high-resolution minutely data is essential for accurate modeling, most widely available satellite-derived datasets are limited to hourly or coarser temporal intervals. This paper presents a novel non-autoregressive transformer-based approach for downscaling hourly Global Horizontal Irradiance (GHI) data to minutely resolution. We treat the downscaling problem as a sequence-to-sequence task, tokenizing the Clear Sky Index (kt) into 1024 discrete values and generating entire minutely sequences (1440 values per day) in a single forward pass, while a physics-informed loss function ensures energy conservation at the hourly scale. A two-stage postprocessing algorithm addresses high-frequency noise while preserving realistic variability. Comprehensive benchmarking against five state-of-the-art downscaling methods across 12 geographically distributed BSRN test stations demonstrates our model’s superior performance in distribution-based metrics (KLD, JSD, OVC, and KSI). Our transformer achieves the best average KLD (0.0020 vs. next best 0.0034) and JSD (0.0005 vs. next best 0.0010), while requiring only hourly GHI and geographic coordinates as input, unlike competing methods that need extensive historical data or additional meteorological variables.
Source: Solar Energy
Institution: Universidade Estadual de Campinas (UNICAMP)
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,900
OpenAlex ID: W7204254534