Multiphase fluid transfer in pipelines is an extensive process, starting with pipeline design, maintenance (e.g. corrosion-control), and mitigating operational issues (slugging and fluid management). There are physics-based and CFD (Computational Fluid Dynamics) models that predict vital information supporting pipeline design and operation from wells to the GOSP. The proposed model is independent from mechanistic correlations and assumptions, relying completely on measurements data from a multiphase flow database to make final predictions.
The objective of this study is to assess the predictability of multiphase flow regimes in horizontal pipes via supervised machine learning (ML) classification techniques with: liquid and gas velocity, liquid and gas viscosity, liquid and gas density, surface tension, pipe diameter, and absolute roughness. The data is obtained from the Stanford Multiphase Flow Data (SMFD). The selected dataset size is 2,254 points with 9 input variables describing the fluid flow and pipe, and the output is the flow regime type. The study compares the performance of five algorithms: decision tree, random forest, logistic regression, support vector machine, and neural-network (multi-layer perceptron). The evaluation metrics are based on accuracy (F-1 score), and efficiency (training runtime).
After the running the model with all three random data splits, the average F-1 accuracy score and training time were evaluated. The accuracies were consistent across data-splits (standard deviations <0.02). The best performing algorithms for classifying flow regimes using the SMFD dataset were Random Forest, then Decision Tree. They are high in accuracy (86% for the Decision Tree, and 89% for Random Forest), and low in training-time (less than 0.005 seconds). They would perform well with a larger training set due to their flexibility, high efficiency, and ability to capture trends relating the variables to flow regimes.