Reconstruction of the surface solar radiation based on a physics-constrained spatiotemporal framework

Accurate high-resolution surface solar radiation (SSR) estimates are vital for regional photovoltaic planning and resource assessment, but coarse-resolution radiation products often provide insufficient spatial detail in complex mountainous terrain. This study introduces a Physics-Constrained Spatiotemporal Neural Operator (PC-STNO) for high-resolution SSR reconstruction over a mountainous region of F Province, China. The model integrates coarse-resolution reanalysis inputs at 0.25∘ with high-resolution meteorological and terrain information to reconstruct SSR on a 0.10∘ grid. By leveraging 0.10∘ dynamic meteorological predictors, this principal configuration performs high-resolution SSR reconstruction from multi-resolution inputs rather than strict coarse-input-only downscaling. It combines temporal attention with global spectral and local convolutional operators to capture evolving atmospheric conditions and spatial heterogeneity, while incorporating top-of-atmosphere constraints into irradiance reconstruction. Evaluation against the held-out 2023 gridded reference data, averaged over five random seeds under matched training budgets, shows that PC-STNO reduces mean absolute error and root mean square error by 7.71% and 7.76%, respectively, relative to the spatiotemporal Transformer baseline. Seasonal, cloud-regime, and spatial analyses further indicate consistent reconstruction performance and improved preservation of irradiance spatial patterns. These results demonstrate the potential of physically constrained spatiotemporal learning with multi-resolution inputs for high-resolution SSR reconstruction and regional solar resource assessment in complex terrain.