Conventional and machine learning improved prediction of hydrocarbon density using volume-translation at high-pressure high-temperature conditions

Abstract

ABSTRACT Non-cubic equations of state (nCEOS) are increasingly showing improved performance at predicting volumetric properties of hydrocarbons, nitrogen, and carbon dioxide at high-pressure high-temperature over volume-translation-based cubic equations of state (VT-CEOS). However, since nCEOS are rather complex, a less mathematically complex and more accurate CEOS is desired. Hence, in this study, we have explored different techniques, including conventional (non-linear regression) and machine learning-based approaches (random forest) to predict a more accurate molar volume deviation term of the VT-CEOS. We used an extensive high-pressure and high-temperature PVT dataset ranging from 50 to 150 MPa and 300 – 500 K respectively in this study. The VT was modeled as a function of reduced temperature only as well as reduced temperature and molecular weight/critical pressure of the pure hydrocarbon components. Statistical analyses and graphs displayed high performance of the developed predictive models over existing VT-CEOS models applied to HPHT and PC–SAFT. More specifically, the machine learning model gave 99% accuracy while the accuracy of the conventional approach ranged from 60-98%. To the best of the knowledge of the authors, the application of machine learning to estimating volume-translation based on CEOS for pure hydrocarbon components of natural gas and heavy hydrocarbons is nonexistent. This paper presents the first application of physics-based machine learning and the use of features that honors thermodynamic principles for prediction of hydrocarbon density.