Improved estimation of surface solar insolation using a neural network and MTSAT-1R data

Solar surface insolation (SSI) provides information on how much solar radiance reaches the Earth’s surface at a specified location during the daytime. The amount of insolation reaching the surface is a critical parameter for climate change estimation and numerical weather prediction (NWP). We calculated SSI from MTSAT-1R data using a neural network (NN) model to obtain more accurate results than obtained using empirical and physical methods. The use of retrieved SSI data depends on the accuracy of the output results. Thus, before adding the input parameters to the NN, principal component transformation was performed using the eigenvectors and normalized input data to eliminate data redundancy. An NN model with one hidden layer was then used to simulate SSI using early-stop and Levenberg–Marquardt back-propagation (LMBP) methods. We separated the NN architecture into two parts according to cloudy or clear-sky conditions, which have different processes because of complicated cloud physical characteristics. The SSI estimates from the NN model were compared with pyranometer measurements and showed better agreement with ground-truth values than did estimates obtained using conventional methods, especially under the clear-sky condition.