Previsão de vazões afluentes utilizando redes neurais artificiais e ensembles

• Ano de defesa: 2019
• Autor(a) principal: Belotti, Jônatas Trabuco
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Ponta Grossa; Brasil; Programa de Pós-Graduação em Ciência da Computação
• Palavras-chave em Português: Análise de séries temporais; Previsão hidrológica; Redes neurais (Computação); Time-series analysis; Hidrological forecasting; Neural networks (Computer science); CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Ciência da Computação
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/4037

The Brazilian energy matrix is predominantly composed of hydroelectric plants. In this way, it is important to ensure maximum efficiency in the operation of these plants since the direct consequence is a significant impact on the cost of energy production and pricing. Determining the streamflow to a hydroelectric plant is a fundamental step in the efficiency of the operation. Over the years, several linear models, such as Autorregressive, and nonlinear, as Artificial Neural Networks have been used to predict streamflows. In order to improve the existing forecasting techniques, this work accomplished the forecast of monthly streamflows through the use of 2 linear models: Autoregressive and Autoregressive of Moving Averages; 10 Architectures of Artificial Neural Networks: MLP, RBF, ELM, ELM (CR), Elman, Jordan, ESN Jaeger, ESN Jaeger (CR), ESN Ozturk and ESN Ozturk (CR); and 6 Ensembles: Medium, Median, MLP, RBF, ELM and ELM (CR) combiners. The term CR is related to the presence of the regularization coefficient. The tests were carried out of the historical séries of the plants of Água Vermelha, Belo Monte, Ilha Solteira, Paulo Afonso and Tucuruí with forecasts horizons of 1, 3, 6 and 12 steps ahead. In addition, the inputs used by the neural models were selected using the Wrapper method. Also, we proposed and tested 3 forecasting strategies using data from the El Niño and La Niña climatic events, two of which resulted in significant improvements in the performances. We verified that the performance of the neural models were better than the linear models in all the simulations, proving the superiority of the nonlinear predictors. We highlight the ELM as the best predictor.