O uso da transformada Wavelet na previsão de vazão

• Ano de defesa: 2012
• Autor(a) principal: Freire, Paula Karenina de Macedo Machado
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraí­ba; BR; Engenharia Cívil e Ambiental; Programa de Pós-Graduação em Engenharia Urbana e Ambiental; UFPB
• Palavras-chave em Português: Wavelet; Previsão de vazão; Redes Neurais Artificiais - RNA; Reservatório de Sobradinho; Flow forecasting; Sobradinho reservoir; ENGENHARIAS::ENGENHARIA CIVIL
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/tede/5514

The Brazilian energetic system is mainly based on hydropower, which is highly dependent on the watershed water availability. In order to minimize the risk of failure, which affects the uptake from the water bodies, this system is interconnected. During the dry season, there is less volume stored in the reservoirs, which leads to a lower production of energy. Before the flood period, in order to attend the multiple uses of the water resources, it is necessary to keep an operational volume in the reservoir, which also decreases the water level and has impacts on power generation. In order to make the flood control, the electricity sector forecasts the availability of a waiting volume in the reservoirs, which are capable of receiving parts of the inflows to prevent, at prefixed risk, damage at downstream. It is in this scenario that the problem highlighted in this dissertation arises, the forecasting of inflow to a reservoir, in order to have a judicious allocation of these void spaces in the reservoirs for the flood control. Thus, the main objective of this study is to analyze the use of the wavelet transform to forecast daily inflows in Sobradinho reservoir (Bahia State) seven days ahead, by a wavelet-ANN hybrid system, with the following specific objectives: (a) eliminate the noise present in the observed inflow time series by wavelet analysis, (b) define the optimal level for decomposition of the signals, (c) determine the appropriate mother-wavelet for this type of forecasting with ANN, and (d) carry out simulations with the proposed wavelet-ANN hybrid system and compare the results with the predictions made without the application of wavelet transform. It was used the daily data for the period from January 1931 to December 2010. From the obtained results, it was observed that the wavelet-ANN hybrid system performed better forecasting for seven days ahead than the system using the ANN with the raw data, and the approximation A3 from the discrete mother-wavelet Meyer obtained the best result (R2 = 0.9977; Nash = 0.9954 and RMS = 96.4523 m³/s), whereas the prediction using RNA with raw data obtained the following results: R2 = 0.9481; Nash = 0.8971 and RMS = 456.7712 m³/s, i.e., the RMS decreased almost 80%, while the Nash and R2 coefficients have increased more than 5% and 10%, respectively, when compared with the forecasts using the raw data.