Neural networks for estimating surface solar irradiation from satellite images

In this work, we present a MultiLayer Perceptron (MLP) model to estimate ground solar radiation, in terms of Global Horizontal Irradiance (GHI), over a given site and for a specific time. The MLP model generates GHI estimates from the Meteosat 12-channel satellite images centered over the target location, and GHI values in clear-sky conditions over the same location. The dataset includes two years of data, covering 2016 and 2017, with a temporal granularity of 15 min, relative to a set of 16 test sites distributed across Europe, Africa and South America. We populate the training- and test-sets with all available data for the 15 and remaining station, respectively. We test all possible combinations of stations to define training and test sets, demonstrating the generalizability of the presented MLP model over potentially any location included in the Meteosat full-disk image. The estimated GHI values are compared to ground-measured GHI data achieving an overall Root Mean Square Error (RMSE) and Coefficient of determination (R 2 ) of 77.682 W/m 2 and 0.929, respectively, across all locations. Finally, the GHI estimates are set against those generated by the Heliosat-4 method, our benchmark, yielding an overall RMSE improvement of 3 W/m 2 . The experiments show that neural networks produce competitive results with fewer and accessible inputs compared to complex physical models for estimating solar radiation. Furthermore, historical and near real-time GHI estimation enabled by the proposed methodology would help photovoltaic (PV) planners determine the irradiance profile of a site where the deployment of a weather station is precluded.