Predição de comportamento de coluna de destilação via redes neurais artificiais

Abstract

This study explores the application of Artificial Neural Networks (ANNs) for predicting the top molar fraction in ethanol-water binary systems within fractional distillation columns. Utilizing data generated through rigorous Aspen Plus [trade mark] simulations, a comprehensive database was developed, incorporating variations in temperature, reflux ratios, and stage configurations. A Multilayer Perceptron (MLP) architecture was optimized using the Optuna framework for hyperparameter tuning, achieving high statistical convergence with a coefficient of determination (𝑅 2 ) of 0.998 and mean squared error in the order of 10−5. The findings demonstrate the superior ability of ANNs to capture process non-linearities compared to classical methods, such as the Fenske-Underwood-Gilliland (FUG) correlation and linear regression. Furthermore, the model exhibited high robustness and resilience under instrumental noise levels of up to 5%. The study concludes that the development of this soft-sensor provides a computationally efficient and high-fidelity alternative for real-time monitoring and process optimization in the chemical industry.