Accurately forecasting solar radiation distribution at both spatial and temporal dimensions simultaneously with fully-convolutional deep neural network model

Abstract

Accurately forecasting solar radiation is of great significance to solar energy utilization. To forecast the spatial and temporal distributions of solar radiation simultaneously, a deep neural network model named MRE-UNet is proposed, the solar radiation data in Heilongjiang province is taken as an example to test the forecasting performance in different periods ahead forecasting cases. According to the evaluation results, an 16 h historical solar radiation data was determined to be the best choice for input, and the minimum of MSE can reach 6.47 × 10−4, 1.38 × 10−3 and 2.69 × 10−3 for 1 h, 3 h and 6 h ahead forecasting cases, respectively. The transferability of the MRE-UNet is tested by performing the solar radiation nowcasting in Hubei province, China using the pre-trained MRE-UNet trained by the solar radiation data in Heilongjiang province. The robustness of MRE-UNet is tested by monitoring the effects of adding different level of noise, and MSE keeps to be 6.27 × 10−4 even though the measuring noise increase to be 50%. For further demonstration on the effectiveness of MRE-UNet in spatiotemporal forecasting, the performance in total cloud cover forecasting is also tested, and satisfactory forecasting results are obtained. Finally, spatiotemporal correlation analysis on solar radiation and total cloud cover data is carried out, a potential reason for satisfying forecasting performance of MRE-UNet is given. From this work, MRE-UNet proposed can be provided as an efficient tool for dealing with further solar radiation spatiotemporal forecasting problem, and the spatiotemporal correlation characteristics can be employed as the basis for further developing effective solar radiation forecasting approach to a degree.