学术文献库

Knowing the behavior of solar radiation at a geographic location is essential for the use of energy from the sun using photovoltaic systems; however, the number of stations for measuring meteorological parameters and for determining the size of solar fields in remote areas is limited. In this work, images obtained from the GOES-13 satellite were used, from which variables were extracted that could be integrated into datasets from meteorological stations. From this, 3 different models were built, on which the performance of 5 machine learning algorithms in predicting solar radiation was evaluated. The neural networks had the highest performance in the model that integrated the meteorological variables and the variables obtained from the images, according to an analysis carried out using four evaluation metrics; although if the rRMSE is considered, all results obtained were higher than 20%, which classified the performance of the algorithms as fair. In the 2012 dataset, the estimation results according to the metrics MBE, R2, RMSE, and rRMSE corresponded to -0.051, 0.880, 90.99 and 26.7%, respectively. In the 2017 dataset, the results of MBE, R2, RMSE, and rRMSE were -0.146, 0.917, 40.97 and 22.3%, respectively. Although it is possible to calculate solar radiation from satellite images, it is also true that some statistical methods depend on radiation data and sunshine captured by ground-based instruments, which is not always possible given that the number of measurement stations on the surface is limited.