A deep learning algorithm to estimate hourly global solar radiation from geostationary satellite data

Abstract

To apply deep learning technique for estimating hourly global solar radiation (GSR) from geostationary satellite observations, a hybrid deep network is proposed, relying on convolutional neural network (CNN) to extract spatial pattern from satellite imagery, multi-layer perceptron (MLP) to link the abstract patterns and additional time/location information to target hourly GSR. Its representative advantage lies in the ability to characterize changeable cloud morphology and simulate complex non-linear relationships. The deep network is trained using ground measured GSR values at 90 Chinese radiation stations in 2008 as well as the radiative transfer model simulation at the top of Mt. Everest which serves as constraints of extrapolation for high elevation regions. The extensibility of trained network is validated at 5 independent stations in 2008, yielding an overall coefficient of determination (R2) of 0.82, and at all stations in 2007 along with an R2 of 0.88. Comparative experiments confirm that the combination of spatial pattern and point information can lead to more accurate estimation of hourly GSR, achieving a minimum root mean square error (RMSE) of 84.18 W/m2 (0.30 MJ/m2), 1.92 MJ/m2 and 1.08 MJ/m2 in hourly, daily total and monthly total scales, respectively. Moreover, the deep network is capable of mapping spatially continuous hourly GSR which reflects the regional differences and reproduce the diurnal cycles of solar radiation properly.