Solving vapor-liquid flash problems using artificial neural networks
Jonah P. Poort (ZEF, Student TU Delft)
Mahinder Ramdin (TU Delft - Engineering Thermodynamics)
J. van Kranendonk (ZEF)
Thijs J.H. J. H. Vlugt (TU Delft - Engineering Thermodynamics)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Vapor-liquid phase equilibrium —flash— calculations largely contribute to the total computation time of many process simulations. As a result, process simulations, especially dynamic ones, are limited in the amount of detail that can be included due to simulation time restrictions. In this work, artificial neural networks were investigated as a potentially faster alternative to conventional flash calculation methods. The aim of this study is to extend existing applications of neural networks to fluid phase equilibrium problems by investigating both phase stability and property predictions. Multiple flash types are considered. Classification neural networks were used to determine phase stability, and regression networks were used to make predictions of thermodynamic properties. In addition to well established flash-types such as the pressure-temperature (PT), and pressure-entropy (PS) flashes, neural networks were used to develop two concept flashes: an entropy-volume (SV), and an enthalpy-volume (HV) flash. All neural networks were trained on, and compared to, data generated using the PT-flash from the Thermodynamics for Engineering Applications (TEA) property calculator. Training data was generated for binary water-methanol mixtures over a wide range of pressures and temperatures. Overall phase classification accuracy scores of around 97% were achieved. R
2
scores of property predictions were in the general order of 0.95 and higher. The artificial neural networks showed speed improvements over TEA of up to 35 times for phase classification, and 15 times for property predictions.
help_outline