Deep learning for intra-hour solar forecasting with fusion of features extracted from infrared sky images

The increasing penetration of solar energy leaves power grids vulnerable to fluctuations in the solar radiation that reaches the surface of the Earth due to the projection of cloud shadows. Therefore, an intra-hour solar forecasting algorithm is necessary to reduce power instabilities caused by the impact of moving clouds on energy generation. The most accurate intra-hour solar forecasting methods apply convolutional neural networks to a series of visible light sky images. Instead, this investigation uses data acquired by a novel infrared sky imager on a solar tracker, which is capable of maintaining the Sun in the center of the images throughout the day and, at the same time, reducing the scattering effect produced by the Sun's direct radiation. In addition, infrared sky images allow the derivation and extraction of physical cloud features. The cloud dynamics are analyzed in sequences of images to compute the probability of the Sun intercepting air parcels in the sky images (i.e., voxels). The method introduced in this investigation fuses sky condition information from multiple sensors (i.e., pyranometer, sky imager, solar tracker, weather station) and feature sources using a multi-task deep learning architecture based on recurrent neural networks. The proposed deterministic and Bayesian architectures reduce computation time by avoiding convolutional filters. The proposed intra-hour solar forecasting algorithm reached a forecast skill of 18.6% with a forecasting horizon of 8 min. Consequently, the proposed intra-hour solar forecasting method can potentially reduce the operational costs of power grids with high participation of solar energy.