Desenvolvimento de um modelo em redes neurais artificiais para controle de um reator de fluxo contínuo destinado a eletrofloculação de efluente de matadouro e frigorífico de suínos

Abstract

The processing of meat used in slaughterhouses and refrigerator generates a great amount of effluents which are constituted of protein, fat, salts, among other substances. Thus, this waste has a high pollutant degree with high organic matter content, oils and greases, nutrients and total solids. To avoid environmental and social problems, industries that slaughter and industrialize animal meat treat their effluents, thereby reducing their pollutant burden and complying with environmental legislation. Conventional systems for treatment of this effluent are considered efficient, however these treatment plants occupy large areas and can cause bad smells, causing discomfort to the surrounding area. One technology that uses a small area to treat large volumes of effluents and does not generate odor is electroflocculation. Also known as electrocoagulation and electroflotation, this technique is based on the use of electric current involving electrochemical reactors, in which coagulants are generated in situ by electrolytic oxidation of an appropriate material at the anode. As the whole process to obtain good results is necessary to be controlled, the objective of this work was to implement a control based on Artificial Neural Networks (ANN) in a slaughterhouse and refrigerator effluent treatment system by means of electro-flocculation in a continuous flow reactor. The choice of Central Composite Rotational Design (CCRD) was to cover the entire experimental space using a smaller number of tests. From this design were carried out 11 tests for statistical analysis which validated the mathematical model generated. Based on this model, the ANN training and validation databases that were implemented in Matlab software were generated. Empirical tests defined the ANN architecture based on training performance with the 2-layer configuration being 1 hidden with 11 neurons and 1 in the output layer. The control used feedforward type, but due to the impossibility of reading in-line turbidity the feedback action was provided by the predictive mathematical model. The prediction obtained for the controlled variable provided by ANN compared to that provided by the empirical mathematical model proved to be efficient providing values with errors considered low for this type of experiment that ranged from 0.0043 A to 0.5418 A with a average value of 0.1277 A which demonstrated the validity of this computational tool in controlling processes