A improvement for the surface solar insolation retrieval from geostationary sensor

A successful retrieval of SSI highly depends on how to describe the cloud attenuation since most of clouds have larger spatial and temporal variability and complicated physical character. Moreover, the accuracy of SSI estimation for cloudy condition is substantially lower than for clear sky. This study aims to generate a neural network-based cloud factor retrieval system, which can improve accuracy of SSI estimation for cloudy condition. In this study, multilayer feed-forward (MLF) neural network (NN) was employed with Levenberg-Marquardt back-propagation (LM-BP) and early stopping method to avoid the over-fitting. The number of hidden nodes was determined by using trial and error method since too complicated network was apt to be over-fitting, while a too simple network structure will have difficult training the network. The validation of the estimated SSI using NN-based cloud factor was performed with pyranometer measurement data obtained from 22 meteorological stations over Korea peninsula. This SSI estimation for cloudy condition showed a good agreement with ground-based measurements (RMSE = 66.0 W/m²). This accuracy indicates that the use of NN-based cloud factor leads an improvement for SSI estimation in comparison with use of previous system of cloud factor.