Previsão de irradiação solar através de um modelo de rede neural otimizada com o algoritmo de broyden-fletchergoldfarb-shanno

Abstract

Solar energy is a promissing renewable source of energy in Brazil due to its tropical climate and new solar system installations receive more incentives. In order to evaluate the viability of a solar system, it is important to have access to knowledge about solar radiation conditions throughout the year. However, solar radiation data is not always available. Therefore, the development of solar radiation prediction models based on meteorological variables available is a good alternative to the absence of solar radiation data. In this work a neural network algorithm was developed in order to create and evaluate solar radiation prediction models using meteorological data as inputs. The algorithm is developed using the R free software and maximum temperature, minimum temprature, relative humidity, sunsine hours, extraterrestrial radiation, total preciptation, evaporation, wind speed and cloudiness data from 1974 to 1988. From daily data, a weekly and a monthly series were generated. A neural network was trained on each one of these groups in order to find the best neural network architecture in each case. The neural networks were compared using root mean squared error (RMSE), mean percentage error (MAPE) and coeficient of determination (R²).The analisys shows that the model created for monthly average solar radiation prediction presents the best results.