Comparison between different Artificial Neural Network architectures in VLE prediction models using experimental and simulated data (via Modified Raoult's Law) of binary systems
Daniel Gomes Ribeiro1; João Flávio Vieira de Vasconcellos2; Gustavo Mendes Platt2
1 UERJ/CEFET; 2 UERJ/IPRJ
doi:10.20906/CPS/COB-2015-1311
Resumo
In this present work, a comparison between different artificial neural networks (ANN) where made in order to determine the best approach in VLE predictions which would be capable to make correct predictions and avoid iterative methods, required to solve state equations. To do so, it was used a set of VLE data containing the molar fraction of liquid phase of each substance and the pressure of the system. The objective is to determine the molar fractions of vapor phase and temperature. Three training types where compared: using only monobaric experimental data, using only simulated data with various pressure values (obtained via thermodynamic model) and a set of experimental data mixed with monobaric simulated data. Three binary systems where used to evaluate the quality of interpolation given by ANN predictions when compared with experimental data. The results have shown that its possible to replace simulated predictions given by thermodynamic models with correctly trained ANNs. Also, the ANNs were capable to interpolate experimental data with reasonable accuracy.
Palavras-chave: Artificial Neural Network; Wilson Model; Vapor Liquid Equilibrium