Avaliação da utilização de redes neurais para previsão do risco de inundação em áreas urbanas na cidade de Curitiba/PR

• Ano de defesa: 2021
• Autor(a) principal: Valle, Robson Felipe do
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Curitiba; Brasil; Programa de Pós-Graduação em Engenharia Civil; UTFPR
• Palavras-chave em Português: Drenagem; Precipitação (Meteorologia) - Previsão; Redes neurais (Computação); Avaliação de riscos ambientais; Inundações - Previsão - Curitiba (PR); Drainage; Precipitation forecasting; Neural networks (Computer science); Environmental risks assessment; Floods - Forecasting - Curitiba (PR); CNPQ::ENGENHARIAS::ENGENHARIA CIVIL; Engenharia Civil
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/26216

Soil impermeabilization, the reduction of green areas, the absence of maintenance in sewers and the lack of palliative measures relating to high levels of precipitation have caused countless cases of floods. Therefore, It is of paramount importance to know the levels of precipitation in order to produce proper urban planning. The city of Curitiba-PR had 13 floods in March 2021, that is, the problem is still recurrent. Thus, a neural network method was applied from 2010 to 2020 in the city of Curitiba, Brazil with the intent of predicting the risk of floods by using precipitation data and numbers of flooded streets throughout the period. The neural network used was Feed Forward, in which the processing occurs from the input layer towards the output layer, without feedback. And to train the network, a standard Bayesian reverse propagation regularization algorithm was used, which minimizes the linear combinations. For the precipitation statistics it was used data from National Meteorological Institute (INMET) and for the flooded streets statistics it was used data from the Civil Defense of Curitiba. It was observed that by the use of seven neurons the coefficient of determination (R²) was significantly higher, producing the number of 0,9829 as the prevision of flooded streets. Twelve neural networks were tested to predict flooded streets in one year, with R² greater than 0.95 in all networks. Therefore, with the use of neural networks it was possible to predict the risk of floods in relation to the precipitation levels.