Artificial neural network based model for retrieval of the direct normal, diffuse horizontal and global horizontal irradiances using SEVIRI images

Abstract

A statistical model for the prediction of the solar irradiance components, utilizing six thermal channels of the SEVIRI instrument (onboard Meteosat Second Generation satellite), is presented. Additional inputs to the model include the solar zenith angle, solar time, day number and eccentricity correction. Treating the cloud-free and cloudy observations separately, the model employs two trained artificial neural network ensembles, one for estimating the direct normal irradiance and the other for estimating the diffuse horizontal irradiance. The global horizontal irradiance is then computed from the model’s outputs. The model has been trained using reference data from three ground measurement stations for the full year of 2010 and tested over two independent stations for the full year of 2009. Over the two independent stations for all sky conditions, the relative root mean square errors for the direct, diffuse and global components are 26.1%, 25.6% and 12.4%, respectively, while the relative mean bias errors are −6%, +3.6% and −2.9%, respectively. The temporal and spatial variations of the direct, diffuse and global components are also presented for three days exhibiting different sky conditions in the year 2009.