Cloud-ground mapping matching for enhanced regional distributed PV power forecasting

Accurate forecasting of distributed photovoltaic (PV) power remains a significant challenge due to the lack of on-site meteorological measurements. Satellite cloud images (SCIs) offer a promising alternative. However, existing methods that directly utilize SCIs as input features often yield suboptimal performance in distributed PV scenarios. This is primarily attributed to the broad spatial coverage of SCIs, which introduces substantial irrelevant information unrelated to PV sites, thereby impairing model learning efficiency and increasing computational complexity. To address these challenges, this paper proposes a cloud–ground mapping framework to improve the forecasting accuracy of distributed PV cluster power. First, a cloud–ground mapping model is developed to identify SCI feature regions that directly obstruct the distributed PV sites. Second, a feature region attention enhanced 3D-CNN is constructed to forecast future SCIs accurately. Third, a similar cloud condition matching method is developed to extract similar historical power as meteorological features based on SCIs. Finally, both the similar historical power and temporally adjacent PV outputs are integrated as input features for the forecasting model. Extensive case studies on a real-world dataset comprising 638 distributed PV sites demonstrate that, compared with existing PV power forecasting methods, the proposed framework reduces nMAE by approximately 14.6% and nRMSE by up to 16.5%, confirming its effectiveness and superiority in regional distributed PV power forecasting.