Effective Prediction of Global Solar Radiation Using Machine Learning Algorithms and Satellite Images

Accurate prediction of Daily Global Solar Radiation (DGSR) is of prime importance for various applications in renewable energy, agriculture or climate studies. This paper presents the effectiveness of using Machine Learning (ML) algorithms and satellite imagery to improve DGSR prediction. The most commonly used ML methods consider as input the use of several meteorological parameters (temperature, wind speed, atmospheric pressure, sunshine duration, etc.) and/or radiometric parameters (aerosol optical thickness, water vapor, etc). In this work, the use of normalized reflectance obtained from satellite images in different channels is studied with the aim of improving the accuracy of solar radiation forecasts. Both ML approaches, ANN (Artificial Neural Network) and SVM (Support Vector Machine), were used as regression models to predict DGSR. The obtained results revealed that the number of parameters used as ML inputs significantly influences the accuracy of daily forecasts of global solar radiation. Furthermore, the results showed superior performance of ANN compared to SVM. Indeed, the errors obtained with the ANN model are an RMSE of 212.21 Wh/m 2 , an NRMSE of 3.46%, a MAPE of 2.85%, an MBE of -7.26Wh/m 2 and R of 0.99 while the SVM model gave respectively 441.95Wh/m 2 , 6.6%, 5.62%, 69.46W h/m 2 and R of 0.98.