Transfer Learning on the Feature Extractions of Sky Images for Solar Power Production

With the increasing popularity of integrating solar energy into the power system, solar power prediction has recently attracted much interest, where the movement of clouds has a crucial impact on solar irradiance and is the major cause of rapid, violent and irregular fluctuations of solar power production. Meanwhile, it is necessary for solar power prediction to capture these fluctuations several minutes ahead in order to facilitate scheduling and operations that maintain the system stability. Considering such importance of the cloud movement, sky images provided by all-sky cameras consist of important data for solar prediction. However, for solar predictors, raw sky images are usually too large to be directly used as the input data. Instead, the features in the sky images related to solar irradiance should be extracted and fed to the predictors. Hence, in this paper, we propose a transfer learning method to extract features from sky images with a Convolutional Neural Network (CNN) to capture the close relationship between clouds and solar irradiance. To accomplish this goal, we first train a classifier on sky images to determine whether the sun is covered by clouds or not, where the classification accuracy achieved is as high as 97.92%. Then, the spectral and textural features of sky images are further extracted by the regression layers in CNN and a nonlinear relationship between sky images and the solar irradiance is revealed. Experimental results confirm that the proposed method can successfully map the sky images to solar irradiance and have significance for future solar prediction applications.