Otimização da adsorção de gás natural em tanque de combustível veicular através de redes neurais artificiais e carbonos ativados

Abstract

In this study, the application of artificial intelligence models to predict energy deactivation curves in adsorption beds composed of 17 different types of activated carbons over the operational cycles of a vehicular fuel tank is explored. Using simulated data from the adsorption of methane, ethane, propane, and butane, three methods of characterizing the activated carbons were employed: Langmuir equilibrium parameters, complete pore size distribution (Complete PSD), and representative pore size distribution (Representative PSD). The tank model, coupled with the Ideal Adsorbed Solution Theory (IAST), was iteratively solved in the gPROMS software to obtain data over 200 operational cycles. The Latin Hypercube Sampling (LHS) technique was used to determine efficient combinations of molar fractions of natural gas components used in calculations. Subsequently, Artificial Neural Networks (ANNs) were trained and evaluated to predict the energy deactivation curves of the adsorption beds,

optimizing their architectures through the Particle Swarm Optimization (PSO) methodology. The results indicate that ANNs, trained with simpler characterization data such as Representative PSD, demonstrated comparable or superior performance to more complex models, emerging as an effective approach in predicting this phenomenon. Opting for simpler models not only streamlines the training of ANNs but also provides valuable insights into the influence of different methods of characterizing activated carbons on predictions of variables of interest related to the complex phenomenon of adsorption.