Modelos de séries temporais e redes neurais na previsão de vazão

Forecasting the hydrological behavior of inflowing rivers into reservoirs of hydroelectric plants is one of the main tools for managing the production of electric power in Brazil. Knowing the future values of a river’s flow is critical when planning hydroelectric systems. Considering such background...

ver descrição completa

Detalhes bibliográficos
Autor: Batista, André Luiz França
Tipo de documento: dissertação
Estado:Versão publicada
Data de publicação:2009
País:Brasil
Recursos:Universidade Federal de Lavras (UFLA)
Repositório:Repositório Institucional da UFLA
Idioma:português
OAI Identifier:oai:repositorio.ufla.br:1/1918
Acesso em linha:https://repositorio.ufla.br/handle/1/1918
Access Level:Acceso aberto
Palavra-chave:CNPQ_NÃO_INFORMADO
Redes neurais artificiais
Vazão fluvial
Séries temporais
Análise de séries temporais
Bacias fluviais - Vazão
Artificial neural networks
River flow
Time series
Modelo SARIMA
Descrição
Resumo:Forecasting the hydrological behavior of inflowing rivers into reservoirs of hydroelectric plants is one of the main tools for managing the production of electric power in Brazil. Knowing the future values of a river’s flow is critical when planning hydroelectric systems. Considering such background, this work aims at investigating two different methods to forecast time series of river flows: Box & Jenkins and Artificial Neural Networks. The data used in this work are the values of average monthly flow of Rio Grande (stream gauge station of Madre de Deus de Minas, MG). The data set consists of 216 observations that were done between January/1990 to December/2007. Models originated from the Box & Jenkins method, as well as models based on the Artificial Neural Networks technique, have been constructed. These models were evaluated according to the EQMP and MAPE criteria in order to select the best models for the studied time series. The statistical model that best suited the data set was a SARIMA(0,1,1)(0,1,2)12. The neural networks model that best adjusted to the data set was an MLP(12,20,1). The selected models were used to forecast future values of the historical series of Rio Grande’s natural flows. A comparative analysis between both techniques used at the prognostication of time series has been done. The results obtained from this comparison have shown that each method can be adequately adjusted to the set of studied observations; however, each technique has both advantages and disadvantages. The Box & Jenkins method has as an advantage the fact that it extracts important information from the time series, such as identification of cycles and trends. This extraction of information from the series does not occur in the Artificial Neural Networks technique, which is a drawback to this technique. In Rio Grande’s flow series, the positive aspect of using Neural Networks was that the obtained prediction values were more accurate than the ones from the statistical models proposed by Box & Jenkins.