Modelagem do processo de tratamento do lixiviado de um aterro sanitário usando processos oxidativos avançados, via redes neurais artificiais

Abstract

Artificial Neural Networks (ANNs) is an alternative approach to traditional statistical methods in mapping a nonlinear system. They are models that have been developed in recent years, and has as a metaphor the functioning of the human brain. These models are based on biological neural networks associated with the parallel processing of the human brain, consisting of interconnected processing elements, called neurons, which are arranged in layers and are responsible for nonlinearity and network memory. This work aims to model a leachate treatment system with advanced oxidative processes using Artificial Neural Networks. The data that were used in this work came from the research developed by Martins (2014). With an experimental system that was installed and monitored in the physical facilities of the Laboratory of the Experimental Station of Biological Treatments of Sanitary Sewage (EXTRABES) of the State University of Paraíba - UEPB. In the treatment of leachate, two types of advanced oxidative processes were used: Photo-Fenton and H2O2 / UV photolysis. For the Photo-Fenton system, the variables concentration of H2SO4, concentration of H2O2, time and rate of ultraviolet radiation were considered. In the H2O2/ UV system H2O2 concentration, time and ultraviolet radiation rate. These parameters that were used to monitor the advanced oxidative processes will be used as input variables in the elaboration of artificial neural networks. For the training and validation of the network, a computational code was used in Matlab software. The best error-mining algorithm was the Levenberg Marquardt with an RNA configuration with a gradient of 2.39.10^-5 and a mean square error (MSE) of 1.10^-6. A hidden layer containing 10 neurons and tangent and purelin sigmoid activation functions were used in the training of these networks. In the H2O2 / UV system the best performance in the training and validation stages was RNA with a hidden layer with 25 neurons presenting a coefficient of determination above 0.97. In the Photo-Fenton system with an architecture with two hidden layers containing in each hidden layer 5 and 25 neurons respectively presented significant results as coefficients of training determination and validation above 0.95.