Spatiotemporal solar radiation forecasting driven by satellite-based and reanalysis data for distributed PV integration using fully-convolutional neural network

Accurately forecasting solar radiation makes great difference to reducing solar photovoltaic curtailment rate and improving energy management and scheduling performance on microgrids and virtual power plants. Inspired by the extraordinary capability of deep neural network in dealing with non-linear features, in this work, a fully convolutional model named Res-MRE-UNet is proposed for spatiotemporal solar radiation forecasting where classical U-Net is employed as backbone. To validate the forecasting performance of Res-MRE-UNet, the spatiotemporal solar radiation forecasting issues in northwest of China are taken as example, and the results show that Res-MRE-UNet is of high stability and accuracy in solar radiation forecasting with RMSE reaching 23.96 W/m2, 32.18 W/m2 and 43.46 W/m2 for 1 h, 3 h, and 6 h ahead forecasting cases, respectively. The difference caused by data source type to the solar radiation forecasting performance is also investigated, on which basis a method for improving the forecasting performance when using satellite-based remote sensing data is proposed. Besides, the solar photovoltaic solar-to-electricity conversion is also considered along with solar radiation forecasting, with which PV generation output forecasting is achieved to provide strong support on improving energy management and control performance of microgrids and virtual power plants.