Improving Two-Phase Flow Pattern Classification Accuracy in Horizontal Pipes Using Ensemble Techniques

Abstract This study explored the application of ensemble machine-learning models to predict two-phase flow patterns in horizontal pipes. Ensemble techniques including boosting, bagging, and random forests (RF) were emplyed. A novel decision-tree classifier was developed by combining Random Trees (RT), J48, Reduced-Error Pruning Decision Trees (REPT), Logistic Model Trees (LMT), and Naive Bayes (NBT) algorithms, specifically designed to predict flow regimes. The ensemble models were built using approximately 2380 experimental data points. Feature selection involved the application of six optimization methods, with training and cross-validation used for optimal selection. Performance evaluation utilized metrics like classification accuracy, precision, recall, confusion matrix, a measure of predictive performance, F-Measure, and Precision-Recall Curve (PRC) area. The study's results indicate that boosting and RF classifiers demonstrate superior prediction accuracy compared to other ensemble algorithms. Particularly, the RF model exhibited exceptional performance across various evaluation metrics, emphasizing the efficacy of ensemble algorithms over single classifiers for tree-based models in predicting flow regimes with an impressive accuracy rate of 91%. The findings underscore the consistent outperformance of ensemble models, which integrate multiple classifiers, over individual classifiers regarding prediction accuracy. This underscores the benefits of leveraging diverse models to enhance accuracy and robustness in flow pattern classification. The analysis also highlights the significant influence of volumetric liquid velocity and gas velocity as key features in determining the models’ performance. This research presents a robust and refined ensemble approach, offering a cost-effective and highly accurate method for predicting two-phase flow regimes in horizontal conduits under various operating conditions.