A new image-processing based approach for solar radiation forecasting

The intermittent and variable nature of the solar source makes it very difficult to use energy efficiently. In order to overcome these problems and benefit from solar energy effectively, many different methods such as solar radiation estimation have been used until today. In this study, a deep learning approach has been developed that predicts future cloud movements by tracking the cloud movements that occur during the day and then performs solar radiation forecasting using the obtained cloud movement forecast and extraterrestrial solar radiation data. In this context, sky images and radiation data collected at specific intervals through the experimental setup established at Afyon Kocatepe University Sun and Wind Application and Research Center are used. Cloud motions in sequential sky images are followed using Shi-Tomasi and Lucas-Kanade methods. Cloud, sky, and sun detections on the images are performed with a hybrid detection approach consisting of red/blue ratio and K-means clustering method. Finally, solar radiation estimates with a resolution of 10 seconds for the time horizon of 5 minutes are performed, and the performance of the approach is tested.