Site adaptation of satellite-based GHI and DNI using clustered features and combined local–global modeling

Abstract

Estimating solar irradiance over several years is essential for assessing large-scale photovoltaic and heliothermic projects. The usual way to achieve this is to adapt long-term satellite-based solar irradiance estimates for specific locations using short-term ground measurements. In this context, a novel site adaptation method is proposed for global horizontal irradiance (GHI) and direct normal irradiance (DNI) based on a two-level modelling strategy incorporating both clustered and unclustered features. These lead to global and local models, which are evaluated separately and in combination. Input variables are provided by the Copernicus Atmosphere Monitoring Service Radiation Service (CAMS) and ECMWF ERA5-Land. The method is tested at four locations – El Rosal and Salta (Argentina), Petrolina (Brazil), and Gobabeb (Namibia), addressing sites in the Southern Hemisphere with similar climates, but different latitudes and satellite viewing angles. A new non-supervised sky conditions classification method is proposed for the local models using the clear sky index, a variability index, and Kernel Density Estimation. This classification forms an integral part of the site-adaptation procedure by capturing local cloud-irradiance interactions. For GHI at 15-minute resolution, the method improves accuracy relative to CAMS at locations near the edge of the Meteosat Second Generation field of view satellites, with RMSE reductions of 26.2% (El Rosal) and 4.8% (Salta). For DNI in Petrolina and Gobabeb, the local and combination models outperform CAMS model, achieving RMSE reductions of 10.2% and 5.8%, respectively. Combination models present the more accurate results as they use the global, local, and CAMS outputs as input variables.