With the increasing integration of large-scale distributed photovoltaic systems into power grids, probabilistic forecasting of regional-scale solar irradiance presents significant challenges, primarily driven by the inherent stochasticity and complexity of cloud dynamics. Existing probabilistic forecasting models face two key limitations: long-horizon performance degradation caused by cumulative residuals, and underutilization of spatio-temporal features and prior knowledge-guided sampling, resulting in physically inconsistent predictions. To address these challenges, an innovative prior knowledge-guided residual diffusion model, termed DiffSolar, is proposed. The model decomposes the regional-scale solar irradiance prediction task into two components: deterministic prediction based on the smart persistence forecasting model and stochastic residual probabilistic forecasting based on prior knowledge-guided residual diffusion model. In DiffSolar, a denoising UNet based on the conditional guidance of multi-scale spatio-temporal features is proposed for efficient denoising of residual diffusion model. Furthermore, to improve the accuracy of residual prediction, a novel training-free sampling strategy is developed, which is guided by prior statistical knowledge derived from global residuals. Experimental validation on a satellite-derived solar irradiance dataset demonstrates DiffSolar's capability to generate physically consistent probabilistic predictions while outperforming existing methods across multiple evaluation metrics. Notably, DiffSolar achieves a root square mean error of 105.9 W/m$^2$ and 0.367 of CRPSS in 4-hour-ahead predictions.
Once the paper will be accepted, I immediately upload the pre-training weights as well as release the training and inference scripts.
You guys can download regional solar irradiance data throught CM SAF.
