TY - JOUR

T1 - Solving the delumping problem using the neural network based algorithm

AU - Arentov, D. O.

AU - Matroshilov, N. O.

AU - Lykhin, P. A.

AU - Usov, E. V.

AU - Kolchanov, B. A.

AU - Kozlov, M. G.

AU - Krylov, A. M.

AU - Taylakov, D. O.

AU - Ulyanov, V. N.

N1 - The work was done in accordance with Ministry of Education and Science of the Russian Federation , FSUS-2022-0020 Project.

PY - 2024/3

Y1 - 2024/3

N2 - During the simulation of multiple producing wells with different PVT-models the problem of mixing of models arises. Usually the composition of each model is compressed (lumped) into pseudo-components and the number of pseudo-components can vary from model to model. In order to mix such fluids correctly they have to be converted to a single standard with fixed number of pure components. The mixing between such fluids reduces to simple summation of corresponding molar fractions of pure components. This paper considers the applied problem of delumping of a compositional fluid model using an ensemble of five identical neural networks and an algorithm created by the authors to find an optimal solution together named Approximator-Predictor pair. Numerical experiments with two laboratory fluid compositions of hydrocarbon mixture are carried out, in which the phase diagrams of original and delumped fluids are compared. The reference phase diagrams and stability tests are calculated using the PVT-module of the “d-Flow” hydraulic simulator. The algorithm produces delumped compositions based on the lumped composition and saturation points. Comparison of phase states at different regions of PT-plane between original and delumped fluids show high accuracy exceeding 98 %.

AB - During the simulation of multiple producing wells with different PVT-models the problem of mixing of models arises. Usually the composition of each model is compressed (lumped) into pseudo-components and the number of pseudo-components can vary from model to model. In order to mix such fluids correctly they have to be converted to a single standard with fixed number of pure components. The mixing between such fluids reduces to simple summation of corresponding molar fractions of pure components. This paper considers the applied problem of delumping of a compositional fluid model using an ensemble of five identical neural networks and an algorithm created by the authors to find an optimal solution together named Approximator-Predictor pair. Numerical experiments with two laboratory fluid compositions of hydrocarbon mixture are carried out, in which the phase diagrams of original and delumped fluids are compared. The reference phase diagrams and stability tests are calculated using the PVT-module of the “d-Flow” hydraulic simulator. The algorithm produces delumped compositions based on the lumped composition and saturation points. Comparison of phase states at different regions of PT-plane between original and delumped fluids show high accuracy exceeding 98 %.

KW - Delumping

KW - Lumping

KW - Multiphase fluid properties calculation

KW - Neural networks

KW - Phase envelopes

UR - https://www.scopus.com/record/display.uri?eid=2-s2.0-85184009106&origin=inward&txGid=53032a84cf0f4b7ce3787d7cf4ef55d1

UR - https://www.mendeley.com/catalogue/7b96f507-596a-3269-a211-8e5c0397a789/

U2 - 10.1016/j.geoen.2023.212622

DO - 10.1016/j.geoen.2023.212622

M3 - Article

VL - 234

JO - Geoenergy Science and Engineering

JF - Geoenergy Science and Engineering

SN - 2949-8910

M1 - 212622

ER -