Aplicação de redes neurais artificiais na modelagem de processos de separação

Abstract

This study investigates the application of Artificial Neural Networks (ANNs) in the modeling of separation processes, with a practical focus on predicting the drying kinetics. The bibliometric analysis, based on data collected from Web of Science and processed using Bibliometrix and VOSviewer, revealed the main trends in the field, highlighting Feedforward and Radial Basis Function (RBF) models as the most commonly used, along with optimization algorithms such as Backpropagation (BP), Levenberg-Marquardt (LM), and Genetic Algorithm (GA). To validate the efficiency of ANNs, models were developed in MATLAB and their predictive capability was compared to traditional and fractional-order models using statistical metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Fit Efficiency (EF). The training process was conducted with experimental moisture content data at 40°C and 70°C, and subsequently, the networks were tested for prediction at 55°C. The results demonstrated that Multilayer Perceptron (MLP) architectures with Tanh and Log-Sigmoid activation functions outperformed conventional models, achieving high precision and better fit to experimental data. Among the developed ANNs, the best models exhibited EF values close to 0.999, RMSE below 0.001, and MSE in the order of 10-6, indicating greater accuracy in moisture content prediction compared to conventional mathematical models. The effectiveness of ANNs in process modeling lies in their ability to learn from data, eliminating the need for complex mathematical equations and dynamically adjusting to different operating conditions. Their application reduces the need for extensive laboratory experiments, optimizes time and costs, and provides a more robust and adaptable approach for forecasting and controlling industrial processes.