Supervised Machine Learning Mode for Predicting Gas-Liquid Flow Patterns in Upward Inclined Pipe

Abstract

Accurate identification of gas-liquid two-phase flow patterns during oil and gas drilling is critical to analyzing bottom hole pressure, detecting overflows in time, and preventing blowout accidents. Since the gas-liquid two-phase flow has deformable interfaces, resulting in complex gas-liquid two-phase flow patterns, the existing gas-liquid two-phase flow patterns are limited in width in terms of pipe diameter and incline, leading to adaptation problems in experimental flow patterns and mechanistic models. Machine learning methods provide potential tools for solving gas-liquid two-phase flow pattern identification. In this paper, a sample database with 5879 data points was established by reviewing and organizing existing literature focusing on normal pressure and temperature, and air-water experimental conditions to provide a data-preparation for the relationship between gas and liquid velocities, pipe diameter and incline characteristics and flow pattern objectives. Four machine learning models, including K-Nearest Neighbor, Naïve Bayes, Decision Tree and Random Forest, were investigated, and each model was trained and tested using a sample database to reveal the performance of four types of supervised machine learning methods, representing similarity, probability, inductive inference and ensemble-learning principles, for gas-liquid two-phase flow pattern recognition, and the prediction accuracy was 0.86, Naïve Bayes is 0.56, Decision Tree is 0.89 and Random Forest 0.97. Comprehensive analysis of each model confusion matrix shows that the machine learning method has the best recognition of dispersed bubble flow, better recognition of slug flow, and the worst recognition of churn flow among the nine flow patterns which proves the controversial nature of the mechanism model in the transition from slug flow to churn flow. This paper uses experimental data as model input features, making the machine learning-based gas-liquid two-phase flow pattern identification model meaningful for practical engineering applications, and also demonstrating the feasibility of using supervised machine learning methods for gas-liquid two-phase flow pattern identification at normal pressure and temperature, wide-range of pipe diameter and incline.