Supervised Machine Learning in Predicting Multiphase Flow Regimes in Horizontal Pipes

Abstract

Abstract The petroleum industry has been undergoing transitional changes that were mandated by the high competition in the business. Making decisions and engineering approaches grew from being trial- and-error, educated guesses, well-studied designs, to finally creating advanced and well developed models and simulations. Yet, the development of approaches has been and will always be rapidly ongoing. In order to stay competitive, the approaches must be constantly improving. This study was performed as a proof-of-concept to assess and prove the practicality of using supervised machine learning (ML) techniques to predict multiphase flow regimes in horizontal pipes. The flow is comprised of air, water, and oil. The input features used were water cut (the percentage of water), gas superficial velocity, and liquid superficial velocity. The predicted output was one of six possible flow regimes. The algorithms assessed in the study were Decision Tree, Random Forest, Logistic Regression, Support Vector Machine (SVM), and Neural Network Multi-Layer Perceptron (MLP). According to the results, the best candidate for the dataset is to use the random forest algorithm with a high accuracy of 90.8% and low training time (0.13 seconds) in case of increasing the size of the data and features. All the predictive algorithms can be easily improved in accuracy by increasing experience, either increasing the size of the dataset or addition of relevant features. The trained model and logic can be applied in industry by automating flow control or installing prediction and mitigation systems in pipelines and field operations.