Previsão de variáveis ambientais na Amazônia com uso de redes neurais artificiais do tipo Long Short-Term Memory.
Visualizar/ Abrir
Data
2021-08-25Autor
http://lattes.cnpq.br/2153466045008312
SANTOS, Paulo Guilherme Silva dos
Metadata
Mostrar registro completoResumo
With the expansion of the photovoltaic solar energy industry came the search for improving the efficiency of its systems. The prediction of environmental variables such as temperature and solar irradiance help in making decisions about the use of these systems. In this work, a methodology for the evaluation of Artificial Neural Networks (ANN) of short-long-term memories (LSTM) was used to predict the time series of these two variables. The data used were obtained from two cities located in the Amazon region, two sets of temperature data and a set of solar irradiance data. During the processing of data sets, fundamental characteristics for time series prediction were verified, such as autocorrelation and stationarity, and the division into training-test and validation sets. The architectures used have differences in their number of layers, to analyze the influence of their complexity on their performance. As a result, in data validation, the 3-layer architecture presented a statistically significant difference compared to the 7-layer architecture, for the same number of times. For the LABIC temperature dataset, the RMSE averages of the two architectures were 0.9393°F and 1.4531°F, for 3 and 7 layers, respectively; for the GEDAE temperature dataset the mean RMSE was 1.6499°F and 1.9767°F, for 3 and 7 layers, respectively; and in the solar irradiance dataset we obtained an average RMSE of 170.6649 W/m² and 204.7825 W/m², for 3 and 7 layers, respectively. The methodology used allowed the comparison between the architectures and could be used in the future to evaluate other models of ANNs for forecasting time series.