Um estudo comparativo entre redes neurais profundas e o modelo numérico WRF-ARW para previsão de energia solar a curto prazo

Abstract

The use of renewable energy sources is a big step for the future of a nation, especially Brazil has been highlighted by the large availability of solar energy throughout its territory. However, the presence of clouds and aerosols in the atmosphere can compromise the use of this type of energy in the electrical system, predicting associated meteorological variables a challenging task. In this sense, in view of a sustainable model of energy generation and recent advances in the area of Artificial Intelligence (AI) for the development of predictive models, this work aims to develop a computational model based on deep machine learning (deep learning) to perform the prediction of temperature and global solar radiation, both used to calculate solar energy power and the respective efficiency, and also to make a comparison with the traditional atmospheric model Advanced Research Weather Research and Forecasting (WRF-ARW). The training of the networks was carried out with data from four cities in the central west region of Bahia, recognized for having a high solarimetric potential. For testing purposes, data obtained in the region of the city of Ibotirama in Bahia and Petrolina in Pernambuco were used, without the inclusion of any such data in the training, to verify if the networks were able to correctly represent the atmospheric behavior of the environment under study. The results showed, in general, that a configuration composed of two Multilayer Perceptron (MLP) networks presented the best metric in the calculation of energy power for the Ibotirama region. While a combination of Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) was better for Petrolina, predicting energy power. All networks were built to predict from 1 to 48 hours ahead with the 1 hour gap. A statistical analysis of the final results for both target cities suggest a good acceptance of the use of neural networks for atmospheric predictability, in view of the low computational cost and superior performance for the prediction of solar energy generation, compared to WRF-ARW.