A Data-Driven Model for Multiphase Flow in Pipes

Abstract

Steady-state models commonly used to determine pressure gradient in wells are either empirical correlations or are dependent on empirical parameters correlated to specific flow conditions such as pressures, temperatures, fluid types, and pipe inclinations. The empirical nature of these models leads to limitations in predicting cases in a wide range of conditions. Established models may be delivering inaccurate results when applied at conditions beyond those which they were derived from. For instance, models available in the literature have shown to have limitations when used in large diameter pipes and high-velocity scenarios. This work aims to evaluate different flow models with field and laboratory data and propose a data-driven modeling technique to determine liquid holdup and pressure gradient for a wide range of flowing scenarios. This methodology can be implemented either using drift-flux concepts or direct estimation of pressure gradient through data analytics techniques. A fully automated flow loop was designed and built to demonstrate how data can be generated in real-time to cover a wide range of pipe inclinations, and different liquid and gas velocities, without the need of flow regime estimation to accurately predict liquid holdup and pressure gradient. The results from this work were compared with drift-flux and empirical models from the literature. Liquid holdup prediction using the data-driven approach resulted in errors at least 15% lower than the other models including in the comparison. While the estimation of pressure gradient based on existing data provided errors on average 10% lower than the other models evaluated. This study showed that empirical and mechanistic models have limitations when are applied in flow conditions beyond their range of application. The drift-flux approach showed improvements but is not feasible to determine the pressure gradient for a wide range of flow velocities. It was evident that the application of dimensionless numbers that account for the dynamics of two-phase flow to determine pressure gradient has the potential to improve multiphase flow modeling. The integration of automated flow loops, as the one built in this work, can improve data-driven models like the one proposed in this work.