Flow pattern identification of gas-liquid two-phase flow based on integrating mechanism analysis and data mining

Abstract

The flow pattern identification of gas-liquid two-phase flow is very important to gas field gathering pipeline. Due to the complexity of gas-liquid two-phase flow, it is difficult to accurately and quickly identify the flow pattern by mechanism model. Machine learning (ML) model has high accuracy, but poor generalization and physical significance. In order to improve the deficiencies of ML model and mechanism model, a new physically guided neural network (PGNN) is proposed. Dimensionless numbers with physical significance are used as the input of PGNN. The structure of PGNN is designed according to the physical modeling process of flow pattern identification. The physical intermediate variables that affect flow pattern transformation are added to the loss function. 1812 experimental data in two literatures are used for training and testing different models, and 88 experimental data in one literature are used as an additional test dataset to evaluate the generalization of different models. The average accuracy of PGNN on test set and additional test data set is 95.2% and 95.9%, which are higher than other models. This study proves that the integration of mechanism can improve the accuracy and generalization of ML model, and provides a new perspective for the study of gas-liquid two-phase flow.