Global Horizontal and Direct Normal Solar Irradiance Modeling by the Machine Learning Methods Xgboost and Deep Neural Networks with CNN-LSTM Layers: A Case Study Using the GOES-16 Satellite Imagery

Abstract

Restrictive legislations over the usage of fossil fuels are encouraging the research on clean and renewable energies. As a drawback, renewable energies are characterized by having random behavior, which hampers its integration on the current energy base system. Thus, solar irradiation estimation is important to a better assimilation of the renewable energies into the energy matrix. In this paper, two machine learning (ML) estimation models for global (GHI) and direct normal solar irradiance (DNI) are proposed: the first using XGBoost and the second using a hybrid convolutional neural network (CNN) with a long short-term memory (LSTM), CNN-LSTM. Both models use images from the GOES-16 satellite, taken from the city of Petrolina, Brazil. The results of the proposed models are analyzed against the reference models Copernicus Atmosphere Monitoring Service (CAMS), Solcast and the Physical Solar Model (PSM) provided by the National Solar Radiation Database (NSRDB). For GHI estimation, the PSM model reached the smallest RMSE, equal to 147.23 W/m^2, while the model CNN-LSTM was the best one to estimate DNI with a RMSE equal to 238.22 W/m^2. The proposed model managed to improve RMSE when compared against the benchmarking models by 2.89 % and 1.70 % for CNN-LSTM and XGBoost, respectively.