Assessment of a New Solar Radiation Nowcasting Method Based on FY-4A Satellite Imagery, the McClear Model and SHapley Additive exPlanations (SHAP)

Abstract

The global warming effect has been accelerating rapidly and poses a threat to human survival and health. The top priority to solve this problem is to provide reliable renewable energy. To achieve this goal, it is important to provide fast and accurate solar radiation predictions based on limited observation data. In this study, a fast and accurate solar radiation nowcasting method is proposed by combining FY-4A satellite data and the McClear clear sky model under the condition of only radiation observation. The results show that the random forest (RF) performed better than the support vector regression (SVR) model and the reference model (Clim-Pers), with the smallest normalized root mean square error (nRMSE) values (between 13.90% and 33.80%), smallest normalized mean absolute error (nMAE) values (between 7.50% and 24.77%), smallest normalized mean bias error (nMBE) values (between −1.17% and 0.7%) and highest R2 values (between 0.76 and 0.95) under different time horizons. In addition, it can be summarized that remote sensing data can significantly improve the radiation forecasting performance and can effectively guarantee the stability of radiation predictions when the time horizon exceeds 60 min. Furthermore, to obtain the optimal operation efficiency, the prediction results were interpreted by introducing the latest SHapley Additive exPlanation (SHAP) method. From the interpretation results, we selected the three key channels of an FY-4A and then made the model lightweight. Compared with the original input model, the new one predicted the results more rapidly. For instance, the lightweight parameter input model needed only 0.3084 s (compared to 0.5591 s for full parameter input) per single data point on average for the 10 min global solar radiation forecast in Yuzhong. Meanwhile, the prediction effect also remained stable and reliable. Overall, the new method showed its advantages in radiation prediction under the condition that only solar radiation observations were available. This is very important for radiation prediction in cities with scarce meteorological observation, and it can provide a reference for the location planning of photovoltaic power stations.

Keywords

• global solar radiation
• forecast
• machine learning models
• SHapley Additive exPlanation