Prediction of Two-Phase Flow Regimes in Vertical Pipes Using Tree-Based Ensemble Models

Abstract

Abstract The multi-phase fluid transfer pattern in vertical flow through pipelines is a significant parameter to be predetermined for predicting the pressure gradient, liquid holdup, and other flow properties. In the present study, the prediction of two-phase flow patterns in vertical pipes using ensemble machine-learning classification models is presented. For this purpose, ensemble machine learning techniques including boosting, bagging, and random forest have been applied. A decision tree-based classifier is proposed, such as Random trees (RT), J48, reduced-error pruning decision trees (REPT), logistic model trees (LMT), and decision trees with naive Bayes (NBT), to predict flow regimes. Datasets consisting of more than 2250 data points were used to develop the ensemble models. The importance of attributes for different models was investigated based on a dataset consisting of 1088 data points. Feature selection was performed by applying six different optimization methods. For this task, training, and cross-validation were used. To check the performance of the classifier, a learning curve is used to determine the optimal number of training data points to use. The performance of the algorithm is evaluated based on the metrics of classification accuracy, confusion matrix, precision, recall, F1-score, and the PRC area. The boosting approach and random forest classifiers have higher prediction accuracy compared with the other ensemble methods. AdaBoost, LogitBoost, and MultiBoosting algorithms were applied as boosting approaches. Multiposting has a better performance compared with the other two techniques. The random forests provided a high level of performance. Its average precision, recall, and F1 scores are 0.957, 0.958, and 0.949, respectively. It is concluded that comparing the results of single classifiers, the ensemble algorithm performed better than the single model. As such, the accuracy rate of the prediction of flow regimes can be increased to 96%. This study presents a robust and improved technique as an alternative method for the prediction of two-phase flow regimes in vertical flow with high accuracy, low effort, and lower costs. The developed models provide satisfactory and adequate results under different conditions.