Modelagem e simulação de um sistema de tratamento de esgoto sanitário usando rede neural artificial

• Ano de defesa: 2015
• Autor(a) principal: Silva Neto, Pedro Nogueira da
• Orientador(a): Vieira, Fernando Fernandes
• Banca de defesa: Sousa, José Tavares de , Barbosa, Enivaldo Santos
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual da Paraíba
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência e Tecnologia Ambiental - PPGCTA
• Departamento: Pró-Reitoria de Pós-Graduação e Pesquisa - PRPGP
• País: Brasil
• Palavras-chave em Português: Tratamento de esgoto; Matéria orgânica; MATLAB; Redes neurais artificiais
• Palavras-chave em Inglês: Organic matter; Sewage treatment
• Área do conhecimento CNPq: ENGENHARIAS::ENGENHARIA SANITARIA
• Link de acesso: http://tede.bc.uepb.edu.br/tede/jspui/handle/tede/2472

The population growth in cities, and in parallel, the intensification of agricultural and livestock activities , in addition to industrial and agro-industrial development are the main causes for the increase in the dumping of sewage , often arranged in lakes , seas and rivers in its raw form or with insufficient treatment. UpflowAnaerobicSludgeBlanket(UASB) reactors appear as an option for treatment of various types of wastewater , widespread and well being applied more than the other , and having as an essential aspect of the process the nature of the active biomass. The Artificial Neural Networks ( ANN ) are a type of mathematical model of the area of Artificial Intelligence are a modeling technique based on input-output relations ( black -box structures) that has received great interest from the scientific community , being carried out research in a multidisciplinary way. Thus, this research is structured in order to model and simulate a UASB reactor in order to evaluate and establish control parameters, using Artificial Neural Networks (ANN). The set of data relating to the 13 analyzed parameters , has been subdivided in three different forms and the use of a hidden layer and two hidden layers with varying number of neurons in amounts of 5, 10 and 15. In addition, the functions activation were modified in order to obtain fluctuations networks. Thus, it was found that when using a single intermediate layer, the best activation function was Logsig and better distribution data is presented when it was combined 80% of the data used for training, 10% for trial and 10% for validation. At the same time, one can conclude that the greater the number of neurons used in the hidden layer, best determination coefficients are found. In this 2 circumstance, the R was 0.9985, which was considered most suitable for predicting the parameters analyzed. When patterned with 2 hidden layers, the results resembled well with the simulation of a hidden layer, since the best results have been obtained for distributing data, 80%, 10%, 10% respectively to training, testing and validation. The best determination coefficient was obtained with the combination of Tansig activation functions in the hidden 2 layer 1 and Logsig in the hidden layer 2, obtivendo an R of 0.9912, considered excellent for predicting the data.