GISPLIT: High-performance global solar irradiance component-separation model dynamically constrained by 1-min sky conditions

Abstract

The separation of global horizontal irradiance (GHI) into its direct and diffuse components is necessary in a variety of applications, most specially in solar energy utilization, where knowledge of direct normal irradiance (DNI) is of paramount importance. Here a novel and efficient model, referred to as GISPLIT, is presented to perform this task accurately, using time series of measured data at 1-min resolution. To better describe the radiative effects of different cloud situations, the model takes advantage of a preliminary classification of the sky conditions into six sky types. An empirical submodel is assigned to each sky class to split GHI into its components, using a limited number of predictors that are related to GHI’s magnitude and variability, and to coincident estimates of the clear-sky irradiance components. Those submodels are trained and validated using rigorously quality-assessed measurements from 120 radiometric stations over all continents and all five major Köppen-Geiger (KG) climate classes, totaling ≈64 million valid data points. Four model versions are evaluated using training data for either all KG climate regions combined or conditioned by KG climate, and either with or without additional support from machine learning. The validation of the four versions suggests that the conditioning by KG climate does not add any significant benefit over the “all-climates” training approach and that, overall, the model version trained with data from all KG climates combined and supported by machine learning generally predicts DNI with the best RMSE results at unseen sites, although with little difference over the other versions.