Approaches to improve the accuracy of estimating the diffuse fraction of 1-min resolution global horizontal irradiance using cloud images

Abstract

Obtaining high-precision diffuse irradiance from global horizontal irradiance (GHI) can serve comprehensive and effective data for PV system design, operation and maintenance. This study has incorporated cloud features in an artificial neural network (ANN) model to improve the estimation accuracy of diffuse fraction on 1-min resolution dataset. The cloud features are extracted from ground-based cloud images, including spectrum features, texture features and cloud cover ratio, with image processing algorithms. After data validation, the ANN model which incorporated all the cloud features has achieved a normalized root mean square error (NRMSE) of 17.1 %, representing a 13 % reduction compared to the basic ANN model, we have investigated additional strategies that further optimize the model performance, including cloud classification, weather classification and data averaging, and quantified the effects of the proposed approaches based on actual station data. The data averaging based on proper time scale has brought about 2 % in accuracy improvement; the weather classification and cloud classification have both brought above 10 % of accuracy improvement in some cases but others may deteriorate for some reasons that need to be further investigated, based on this, we have analyzed and summarized the deficiencies in our research and proposed detailed research directions for future endeavors.