Open Access

ARTICLE

An Integrated Multi-Scale Modeling Framework for Gas Entrainment Prediction in Coalbed Methane Production Systems

Qin Zhao¹, Yuxin Wang¹, Gang Chen¹, Hui Zhang¹, Lei Wang¹, Mulin Zhou¹, Songfei Zhang¹, Yu Weng^2,*

1 Petrochina Coalbed Methane Co., Ltd., Beijing, China
2 College of Pipeline Engineering, Xi’an Shiyou University, Xi’an, China

* Corresponding Author: Yu Weng. Email:

Fluid Dynamics & Materials Processing 2026, 22(6), 9 https://doi.org/10.32604/fdmp.2026.083174

Received 30 March 2026; Accepted 08 June 2026; Issue published 30 June 2026

Abstract

This study presents an integrated multi-scale framework for predicting gas entrainment and flow behavior in coalbed methane production systems under gas-liquid two-phase flow conditions. The approach combines three-dimensional computational fluid dynamics simulations, reduced-order modeling, and machine-learning-based prediction to achieve both high physical fidelity and computational efficiency. Such an improved strategy stems from a specific need. As coalbed methane extraction increasingly encounters complex multiphase flow conditions, accurate characterization of gas entrainment has become essential for improving production stability and optimizing downstream gathering and separation systems. In practice, the flow entering rod pumps frequently deviates from the conventional assumption of a purely aqueous phase, leading to strongly coupled gas-liquid interactions that are difficult to capture using traditional modeling approaches. At the same time, the computational cost associated with full three-dimensional simulations limits their applicability to real-time field operations. The present approach is introduced to map the three-dimensional flow dynamics onto a one-dimensional numerical manifold while retaining the dominant physical characteristics of the system. Building upon this reduced representation, a Stacking ensemble learning framework is developed using Support Vector Regression, Back-Propagation Neural Networks, and Random Forests as base learners, combined through an Adaptive Neuro-Fuzzy Inference System meta-learner for rapid field-scale prediction. The results show that the suction effect generated by the rod pump is the dominant mechanism governing gas entrainment. Notably, the proposed prediction model achieves high predictive accuracy, with a root mean square error of 1.75 L/min, a mean absolute error of 0.82 L/min, and a coefficient of determination of 0.98. Furthermore, analysis of the surface pipeline network indicates that terminal flow rates are characterized by low-frequency pulsations that promote stable gas-liquid interface formation within downstream separators.

---

Keywords

---