Answers are grounded in this paper and cite their sources.
Machine-learning weather prediction (MLWP) supports global forecasting, but solar radiation is often provided on coarse grids as multihour accumulations, creating a scale mismatch with local applications. We present SolarFINE, a physically constrained hierarchical framework that converts 0.25° six-hour accumulated radiation and concurrent meteorological states to surface solar irradiance at 0.05° and 10-min resolution. It reconstructs hourly coarse-grid irradiance from solar geometry and meteorological states under a six-hour integral-consistency constraint, then constructs an hourly high-resolution spatial background and generates subhourly variability using conditional diffusion. The temporal stage was trained on ERA5, the high-resolution stage was supervised with a satellite radiation product, and SolarFINE-T was adapted to Artificial Intelligence Forecasting System (AIFS) inputs. Evaluation covered hourly reconstruction, high-resolution spatial structure, and AIFS-driven end-to-end station performance. In controlled ERA5 experiments, SolarFINE-T achieved a daytime hourly root-mean-square error (RMSE) of 40.37 W m−2, 29.9 % lower than with solar-geometry allocation. At 0.05° and 10-min resolution, SolarFINE-J achieved an RMSE of 132.42 W m−2 and more closely reproduced the spatial power spectrum, spatial variability, and high-irradiance tail of the high-resolution reference product. In the independent station evaluation, RMSE skill scores relative to solar-geometry allocation were positive on forecast days 1–15, rising from 0.084 to 0.174. Improvements were largest and most consistent at intermediate clear-sky index. Adapting to AIFS inputs improved skill, whereas forcing hourly fields to match the original six-hour accumulation reduced it. The results indicate that MLWP-driven solar-irradiance downscaling requires exploiting atmospheric information beyond solar geometry while accounting for upstream forecast errors.
No takes yet. Share an insight, caveat, or question.
Dai et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: