Sky Image forecasting with Generative Adversarial Networks for cloud coverage prediction

This work focuses on the problem of sky image prediction using Generative Adversarial Networks (GANs). Sky images observation is used for cloud coverage short-term forecasting and therefore by predicting the future frames of a sky image sequence, a better estimate of the available solar resource for renewables could be achieved. For this purpose, a Deep Convolutional Neural Network (CNN) topology is trained with an Adversarial loss in order to produce realistic future frames of the input image sequences. The prediction of future frames is incurred by two different approaches. In the first approach, a network is receiving 4 frames and predicting the next frame and in the second approach, given a sequence of 8 input images a network is predicting the next 8 frames. In order to evaluate the quality of the generated images, a criterion based on cloud coverage is proposed. This cloud coverage metric provides an objective criterion for assessing the quality of the generated sky images, especially in the presence of clouds.