Abstract
No one can deny the ever-increasing importance of oil since it has influenced every aspect of humans' life. One of the most important pressure-volume-temperature (PVT) properties of crude oil, which is needed in a majority of production and reservoir engineering calculations, is the saturation pressure (bubble point pressure (P-b)). Having accurate knowledge about P-b is significant for both academia and industry. This communication concentrates on providing reliable experimental data from constant composition expansion (CCE) test as well as rigorous compositional models to predict saturation pressure of crude oils based on oil composition (H2S, N-2, CO2, C-1 to C7+), reservoir temperature, C7+ specifications (molecular weight and specific gravity). Seven advanced machine learning approaches, namely, decision trees (DTs), random forest (RF), extra trees (ETs), cascade-forward back propagation network (CFBPN), and generalized regression neural networks (GRNN) as well as multilayer perceptron (MLP) and radial basis function (RBF) neural networks were used for modeling. The CFBPN and MLP models were trained by three different training algorithms, namely scaled conjugate gradient (SCG), Bayesian regularization (BR), and Levenberg-Marquardt (LM). The modeling was done based on a databank consisting of 206 data points (130 points were previously published in the literature plus 76 points determined experimentally in this study). The results show that the DT model could provide the most reliable prediction with an average absolute percent relative error (AAPRE) of 4.43%. The efficiency of various equations of state (EoS) and empirical correlations were checked. According to the results, Peng-Robinson (PR) and the correlation developed by Elsharkawy were the most efficient EoS and empirical correlation with AAPRE values of 8.46% and 12.05%, respectively. Then, the sensitivity analysis revealed that the P-b was extremely affected by methane and C7+ mole percent. Finally, the Leverage approach confirmed the validity of the employed data, detecting only 7 points as outliers.
NSTL
Elsevier