Desenvolvimento de modelo de redes neurais para identificação de eventos com alta concentração de 2- metilisoborneol (MIB) e geosmina no reservatório Jundiaí, no sistema produtor Alto Tietê.
Data
2022-12-16
Autores
Santos Filho, Hélio Vitor dos 
Orientadores
Hanisch, Werner Siegfried
Tipo
Trabalho de conclusão de curso
Título da Revista
ISSN da Revista
Título de Volume
Resumo
Um dos problemas provocados pela eutrofização dos reservatórios utilizados para abastecimento público são as florações de cianobactérias, as quais, além de poderem lançar toxinas nas águas, podem lançar compostos orgânicos tais como o 2-metilisoborneol (MIB) e a geosmina. Esses metabólitos conferem gosto e odor de terra e mofo à água mesmo quando presentes em baixas concentrações, ocasionando uma fácil percepção pelos consumidores. Uma forma de contornar esse problema é melhorar a gestão do monitoramento desses reservatórios valendo-se da utilização de modelos preditivos para a identificação da iminência do surgimento de concentrações relevantes dessas substâncias. À vista disso, esse trabalho teve como objetivo identificar, por meio de análises estatísticas (Análise de Componentes Principais e Correlação de Pearson), em um reservatório tropical utilizado para abastecimento público, quais organismos são os responsáveis pela liberação/produção de MIB e geosmina. Seguido desta análise, foram programadas duas redes neurais artificiais (RNA) previsionais do tipo NARX – modelo autorregressivo não linear com entrada exógena – sendo uma para MIB e outra para geosmina, capazes de predizer, com uma semana de antecedência, as concentrações dos compostos de interesse. Isto posto, no intuito de aplicar as RNA’s obtidas, o trabalho visou modelar uma interface interativa capaz de alocar a programação da rede para uso industrial. A respeito da modelagem das RNA’s, foi utilizada para ambas as redes a técnica de divisão randômica de dados com 70%, 15% e 15% dos dados para treino, validação e teste, respectivamente. Além disso, os algoritmos de treinamento Scaled Conjugate Gradient, Bayesian Regulation Backpropagation e Levenberg-Marquardt foram submetidos aos testes de performance juntamente com a variação dos números de camadas e neurônios ocultos. Os resultados foram extremamente significativos obtendo-se, para a melhor rede de previsão do MIB: erro percentual médio de 24,4% e erro absoluto médio de 7,968 ng/L; e para geosmina: erro percentual médio de 21,1% e erro absoluto médio de 3,353 ng/L. Vale ressaltar que devido ao comportamento caótico da sintetização de 2-metilisoborneol, esse apresentou um desempenho na etapa de testes inferior ao encontrado para a previsão de geosmina, porém, ainda assim, a rede selecionada afigurara-se eficiente para o proposto.
One of the problems caused by the eutrophication of reservoirs used for public supply is cyanobacterial blooms, which, in addition to intoxicate water, can release organic compounds such as 2-methylisoborneol (MIB) and geosmine. These metabolites impart a taste of dirt and an odor of mold to water even in low concentrations, causing an easy perception by consumers. One way to avoid this problem is to improve the monitoring management of these reservoirs using prediction models to identify the imminence of relevant concentrations of these substances. Thus, this work focused on identify, using statistical analysis (Principal Compounds Analysis and Pearson’s Correlation), in a tropical reservoir used for public supply, which organisms are responsible for MIB and geosmine’s production. In addition to this analysis, two predictive artificial neural networks (ANN) of the NARX type – non-linear autoregressive model with exogenous input – were programmed, one for MIB and the other one for geosmine, being capable of predicting, with one week in advance, the concentrations of these two compounds. Therefore, in order to apply the ANN’s obtained, an interactive interface was created capable of allocating the networks for industrial use. Regarding the modeling of ANN’s, the random division technique was used for both networks with 70%, 15% and 15% of the data for training, validation and testing, respectively. In addition, the Scaled Conjugate Gradient, Bayesian Regulation Backpropagation and Levenberg-Marquardt training algorithms were subjected to performance tests at the same time as the number of layers and hidden neurons were variated. The results were extremely significant, obtaining, for the best MIB prediction: mean percentage error of 24.4% and mean absolute error of 7.968 ng/L; and, for geosmine: mean percentage error of 21.1% and mean absolute error of 3.353 ng/L. It is noteworthy that, because of 2-methylisoborneol’s chaotic behavior, the performance for this network, during testing, was inferior to geosmine prediction, however, the network selected is efficient for the purpose of this work.
One of the problems caused by the eutrophication of reservoirs used for public supply is cyanobacterial blooms, which, in addition to intoxicate water, can release organic compounds such as 2-methylisoborneol (MIB) and geosmine. These metabolites impart a taste of dirt and an odor of mold to water even in low concentrations, causing an easy perception by consumers. One way to avoid this problem is to improve the monitoring management of these reservoirs using prediction models to identify the imminence of relevant concentrations of these substances. Thus, this work focused on identify, using statistical analysis (Principal Compounds Analysis and Pearson’s Correlation), in a tropical reservoir used for public supply, which organisms are responsible for MIB and geosmine’s production. In addition to this analysis, two predictive artificial neural networks (ANN) of the NARX type – non-linear autoregressive model with exogenous input – were programmed, one for MIB and the other one for geosmine, being capable of predicting, with one week in advance, the concentrations of these two compounds. Therefore, in order to apply the ANN’s obtained, an interactive interface was created capable of allocating the networks for industrial use. Regarding the modeling of ANN’s, the random division technique was used for both networks with 70%, 15% and 15% of the data for training, validation and testing, respectively. In addition, the Scaled Conjugate Gradient, Bayesian Regulation Backpropagation and Levenberg-Marquardt training algorithms were subjected to performance tests at the same time as the number of layers and hidden neurons were variated. The results were extremely significant, obtaining, for the best MIB prediction: mean percentage error of 24.4% and mean absolute error of 7.968 ng/L; and, for geosmine: mean percentage error of 21.1% and mean absolute error of 3.353 ng/L. It is noteworthy that, because of 2-methylisoborneol’s chaotic behavior, the performance for this network, during testing, was inferior to geosmine prediction, however, the network selected is efficient for the purpose of this work.