Prediction of liquid holdup and pressure drop in gas-liquid two-phase flow based on integrating mechanism analysis and data mining

Abstract

Gas-liquid two-phase flow widely exists in gas transportation pipelines, and the prediction of liquid holdup and pressure drop is crucial for the safety and efficiency of pipelines. Due to the complexity of multiphase flow, the mechanism model is difficult to accurately and quickly predict the liquid holdup and pressure drop. Machine learning (ML) models have high accuracy, but the lack of physical meaning limits their application. In order to fully leverage the advantages of mechanism models and machine learning models, a new method using physically guided neural networks (PGNN) to predict pressure drop and liquid holdup is proposed. Based on the flow mechanism analysis of different flow patterns, the structure of PGNN is designed according to the calculation process of the mechanism model. Key physical intermediate variables such as shear stress and friction coefficient are added to the model as neurons. Physical constraints in gas-liquid two-phase flow are added to the loss function to give physical significance to neurons. 1390 publicly available experimental data are collected for model training and testing. Due to the uneven distribution of experimental data, the local outlier factor (LOF) algorithm is used to screen outlier of data. By comparing PGNN with other ML models, empirical models, and semi empirical models, it is proved that integrating mechanism into the ML model improves the accuracy and physical consistency. This research presents a fresh outlook on the investigation of the flow dynamics in gas-liquid two-phase systems.