Modelagem inteligente para previsão de séries de vazões afluentes

• Ano de defesa: 2016
• Autor(a) principal: Brito, Bethânia Oliveira De
• Orientador(a): Salgado, Ricardo Menezes
• Banca de defesa: Beijo, Luiz Alberto, Ohishi, Takaaki
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Alfenas
• Programa de Pós-Graduação: Programa de Pós-Graduação em Estatística Aplicada e Biometria
• Departamento: Instituto de Ciências Exatas
• País: Brasil
• Palavras-chave em Português: Redes Neurais (Computacão); Algoritmos - Genética; Previsão Hidrológica
• Área do conhecimento CNPq: PROBABILIDADE E ESTATISTICA::PROBABILIDADE E ESTATISTICA APLICADAS
• Link de acesso: https://repositorio.unifal-mg.edu.br/handle/123456789/955

The generation of power is of strategic importance for the economic development of any nation. The ability to generate power is fundamentally linked to the availability of natural resources. The exploitation of such resources should be guided by principles to maximize the benefit provided and minimize the negative impact on the environment and society. The generation of electricity from hydraulic system depends mainly on the water inflow series (TUCCI, 2002). The forecast strategy series streamflow estimates the water flow with the goal of minimizing uncertainties and risks while reducing factors that hinder the planning of hydroelectric energy production. There are several models in the literature for performing streamflow series forecasting. They include: artificial neural networks, genetic programming, and autoregressive models, among others. In this paper, we propose the construction of ensembles - the combination of individual components - in order to improve the performance of forecasts of streamflow rates. We used one database from the Operador Nacional do Sistema Elétrico (ONS) in two plants located in Rio Grande: Água Vermelha and Itutinga. The models that stood out were the artificial neural network (ANN) with the training algorithms Backpropagation (BPM), Gradient Method (GRAD), and genetic programming (GP). The ensemble BPM showed greater efficiency and generalizability. The forecast MAPE of models for dry periods is less than for the wet season. Model results depended upon the characteristics of the plant and the period under study. Making predictions by periods led to minor mistakes when taken throughout the year. After combining the individual components, there was up to a 14% reduction of the average percentage absolute error (MAPE).