Adaptive image features for intra-hour solar forecasts

Abstract

We introduce a simple and novel technique to extract dynamic features from sky images in order to increase the accuracy of intrahour forecasts for both Global Horizontal Irradiance (GHI) and Direct Normal Irradiance (DNI) values. The proposed methodology is based on a block-matching algorithm that correctly identifies the bulk motion of clouds relative to the position of the Sun in the sky. Adaptive rectangular- and wedge-shaped Regions Of Interest are used to select the image pixels for the new features. The results show an average increase of 6.8% (6.7%) in forecast skill for GHI (DNI) across all horizons tested as measured against a model with global (nonadaptive) image features. Relative to clear-sky persistence, the new model achieves skills ranging from 20% to 30% (22%-35%) for GHI (DNI), among the highest ever reported for these time horizons. An analysis based on Mutual Information and Pearson correlation coefficients between the image features and the training data reveals overall improvements in all metrics. The proposed adaptive method also improves the predictability of the ramp magnitude and direction.