Solar irradiance component separation benchmarking: The critical role of dynamically-constrained sky conditions

The decomposition of global horizontal irradiance into its direct and diffuse components is critical in many applications. To guarantee accurate results in practice, the existing separation techniques need to be validated against reference ground measurements from a variety of stations. Here, four versions of the recent GISPLIT model are compared to a strong benchmark constituted from nine leading models of the literature. The validation database includes ≈24 million data points and is constituted of one calendar year of 1-min high-quality data from 118 research-class world stations covering all continents and all five major Köppen-Geiger climates. The results are analyzed with various statistical metrics to be as generalizable and explicative as possible. It is found that even the simpler GISPLIT version reduces the mean site RMSE of the best benchmark model by ≈11 % for the direct component and ≈17 % for the diffuse component. The improvement reaches ≈17 % and ≈25 %, respectively, when using the best GISPLIT version. The improvements are more important in cases of highly variable sky cloudiness, per the CAELUS sky classification scheme. A ranking analysis shows that all four versions of GISPLIT ranked higher than the benchmark models, and that the use of machine learning significantly improves the separation performance. In contrast, only marginal improvements are obtained through preliminary conditioning by Köppen-Geiger climate class. Overall, it is concluded that GISPLITv3, which is not dependent on climate class but makes use of machine learning for the most challenging sky conditions, can be asserted as the new high-performance quasi-universal separation model.