Estratégias de previsão multipassos à frente para vazão afluente em bacias hidrográficas de diferentes dinâmicas

• Ano de defesa: 2014
• Autor(a) principal: Samira Domingos Costa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Multipassos; Predição; Hidrologia; Séries temporais; Engenharia elétrica; Bacias hidrográficas
• Link de acesso: http://hdl.handle.net/1843/BUOS-9QHK4A

One of the problems in the time series analysis is to obtain multi-step ahead predictions using models that can reproduced the phenomena under investigation with good accuracy. In order to achieve this goal, not only the algorithm used but also the strategy applied to the problem is of great importance. For example, in the context of hydrology, there is the necessity of producing real flow rate of river basins as part of the strategic plan of electric companies. This work sets as one of its goals to apply and to study prediction methods up to 10 steps ahead for the real problem of defining the flow of CEMIG river basins. To do so, two techniques will be used: AutoRegressive with eXogenous inputs modelling, chosen as a comparison model as it is a classic of specialized literature and is widely used, and the Lazy Learning Algorithm, a tool discussed by (Birattari e Bontempi, 1999; Taieb et al., 2012), and originated from the field of nonlinear dynamics. To optimize the predictions using the ARX models and the Lazy Learning models, five different multistep ahead prediction strategies are discussed: Direct, Recursive, DirRec, MIMO and DIRMO. The real measured data of precipitation and flow used for identification and validation purposes corresponds to two river basins: a small one, from Rosal Hydroelectric Power Station, and a large one, from Três Marias Hydroelectric Power Station. Besides the usual performance metric, MAPE, other typical hydrologic metrics, MAE, MSE, RMSE, PBIAS, RSR, and NSE, will be used to quantify results. The strategy of using different models for different prediction steps is shown to be efficient when applied to real data of a specific river basin, according to the chosen performance index.