Multiphase Flow Pattern Recognition in Horizontal and Upward Gas-Liquid Flow Using Support Vector Machine Models

Abstract

Abstract Multiphase flow occurs in wellbores during the production of oil and gas. Depending on the physical forces and interactions acting on different phases, there can be various phase distributions in the pipes, known as flow patterns or flow regimes, such as bubble flow, slug flow, annular mist flow, and stratified flow. Because multiphase flow pressure gradients change significantly with different flow patterns, the flow pattern prediction is usually the first step before any pressure drop estimation is performed. Moreover, in gas production wells, flow regime prediction can help engineers to determine the continuous phase to deal with liquid loading problems. Many efforts, including correlation fitting, fluid dynamic calculation, and back-propagation neural network models, have been used to match experimental observations, which are usually presented as flow regime maps. However, there are often mismatches or errors between the prediction results and the experimental data. To avoid such matching errors, this study applies Support Vector Machine (SVM) models to directly represent the measured experimental data. If the assumption is made that there is no error in the experimental data, the SVM models always give correct output results. An SVM model is a mathematical model that is popularly used for pattern classification and nonlinear regression. For producing oil and gas wells, horizontal and upward multiphase flow is studied in this paper. Experimental data was collected from literature and other sources in order to train the SVM models. Different flow regimes are divided by the boundaries created by the trained models. The model prediction results are plotted in 3-D plots, which provide a clear visualization of how the well inclination angle affects the flow regime transition. The SVM models also perform interpolation approximation to predict the flow regimes at various inclination angles where no experiments have been conducted. Well trained SVM models can be conveniently used and easily combined with pressure loss correlations to calculate pressure drops in wellbores. Finally, an approach using the trained SVM models to deal with liquid loading problems in gas production wells is presented.