A Dl-KF Modeling for Acceleration of Flash Calculations in Phase Equilibrium Using Deep Learning Methods

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Flash calculation is essential for compositional model to determine the composition of the existing hydrocarbon phases and understand the mechanics of fluids flow in reservoirs. Due to its inherent nonlinearity, flash calculation usually turns out to be unstable and high time-consuming in the determination of phase classification and composition concentrations at each grid-block and each time step in compositional simulation. The efficiency and robustness of phase equilibrium calculation is crucial for reservoir simulation. A DL-KF (Deep Learning for K-values & Fugacity Calculation) modeling is proposed to accelerate phase equilibrium calculation using deep learning methods, in which the three-steps neural networks are concluded: ANN-STAB (Artificial Neural Network for Stability Test) model, ANN-KV (Artificial Neural Network for K-values Calculation) model and ANN-FUG (Artificial Neural Network for Fugacity Calculation) model respectively. The ANN-STAB model is generated to test phase stability. When temperature, pressure and feed composition are given, the phase classification is obtained directly with very low computation cost. In the ANN-KV model, initial K-values are determined by trained networks instead of employing Wilson equation employed in traditional flash calculation. Its initial estimations of K-values significantly reduce the number of iterations and avoid converging to incorrect or unphysical solutions. The ANN-FUG model is built to replace the fugacity coefficient calculation in traditional flash calculation algorithms, and simplifies the nonlinear calculation of internal iterative calculation. The proposed DL-KF modeling is validated and compared with the original EOS modeling and three other deep learning methods using two typical hydrocarbon fluid cases. A sample of C3H8CO2-heavy oil systems from Huabei oilfield and a PVT experiment in Tahe oilfield are used to examine the DL-KF modeling. The physical properties of oil sample of Bakken reservoir with CO2 injection are also investigated. These results reveal that the DL-KF method is accurate and efficient for phase equilibrium calculations of reservoir fluids.

This publication has 22 references indexed in Scilit: