Estudo comparativo da regressão linear múltipla e redes neurais artificiais na modelagem do processo de adsorção

Abstract

The textile sector is a branch of activity that has been standing out over the years in Brazil. However, there is a lot of discussion about the consumption of these products, but little is said about the waste generated throughout the production. These residues are often improperly discarded and cause serious damage to the environment, especially to water bodies. Among these residues, it is possible to mention the effluents discarded by the textile industries. The adsorption process can be used as an alternative for the sector's effluent treatment. The purpose of this work is the applicability of mathematical modeling through two methodologies: Regression Analysis that made use of Multiple Linear Regression and Artificial Neural Networks in experimental data of adsorption of textile dyes, as a way of identifying a model that fits to the data and check the adsorption efficiency. Model training was performed in order to fit the best model. Seven variables were used in the application of RLM and ANN, being five control variables: mass, concentration, pH, rotation and time; and two responses: adsorbent mass unit and removal efficiency. The best response of the RLM for the variable adsorbent mass unit was observed in the acid treatment for training with 40% and test with 30% of validation. For training, R_a^2 presented an average of approximately 84% of variation to explain the model, in the test an average value of approximately 84% was obtained for R_a^2. The removal efficiency variable presented the best response in the alkaline treatment with 30% of validation, where an R_a^2 of approximately 58% was obtained for testing, while for training with 30% of validation, an R_a^2 value of 56 was obtained 26%. For the RNA with one layer, the following results were obtained: for the variable adsorbent mass unit, the in-naturra treatment presented an R_a^2 of 99.67%, while for the removal efficiency in the acid treatment, an R_a^2 was obtained. of 99.85% both with 40% of data validation. The RNA with two layers showed an R_a^2(Y1) of 99.73% for the in-natura treatment with 30% validation, while the R_a^2 (Y2) was 99.93% for the acid treatment with 40 % validation. The ANN with one and two layers allows to obtain better values for the variation of the model fit when compared to the RLM.