Solar Irradiation Nowcasting Using Local Cloud Coverage Satellite Images for CNN-based Method: A Comprehensive Methodology and a Real Case Study

Electric power production from renewable sources is characterized by intrinsic volatility due to reliance on weather conditions and often causes imbalances between power generation and demand. Therefore, it is becoming clear that a reliable method of predicting solar radiation is becoming crucial for the stability of photovoltaic-based power generation and industrial applications. This study proposes a novel method to forecast solar irradiance in the very short range of 15 minutes. In this paper, a double-input Convolutional Neural Network is proposed that uses satellite cloud images to forecast the Global Horizontal Irradiance with a time horizon of 15 minutes. Additionally, the cloud data is coupled with a representation of physical information of the time of day to increase the accuracy of predictions. A real case study exploiting 10 months of data with a 15-minute resolution is analyzed. With the naïve persistence method as a baseline, a clear improvement across the key metrics has been noted with the proposed methodology. The CNN outperformed persistence during 4 out of 5 test days selected from the test dataset with the highest improvement noted on sunny days with forecast skill equal to 55.2% in the best tested scenario. On the other hand, the worst performance has been observed in difficult weather conditions with forecast skill dropping below 0 to -22.9%.