A Machine Learning Approach to Predict the Pressure Gradient of Different Oil-Water Flow Patterns in a Horizontal Wellbore
Conference Paper
Overview
Identity
Other
View All
Overview
abstract
AbstractFluid pressure gradient in a wellbore plays a significant role to efficiently transport between source and separator facilities. The mixture of two immiscible fluids manifests in various flow patterns such as stratified, dispersed, intermittent, and annular flow, which can significantly influence the fluids pressure gradient. However, previous studies have only used limited flow patterns when developing their data-driven model. The aim of this study is to develop a uniform data-driven model using machine-learning (ML) algorithms that can accurately predict the pressure gradient for the oil-water flow with two stratified and seven dispersed flow patterns in a horizontal wellbore. Two different machine-learning algorithms, Artificial Neural Network (ANN) and Random Forest (RF), were employed to predict the pressure gradients. A total of 662 experimental points from nine different flow patterns were extracted from five sources that include twelve variables for different physical properties of oil-water, wellbores surface roughness, and input diameter. The variables are entrance length to diameter ratio, oil and water viscosity, density, velocity, and surface tension, between oil and water surface tension, surface roughness, input diameter, and flow pattern. The algorithms performance was evaluated using median absolute percentage error (MdAPE) and root mean squared error (RMSE). A repeated train-test split strategy was used where the final MdAPE and RMSE were computed from the average of all repetitions. The MdAPE and RMSE for the prediction of pressure gradients are 13.89% and 0.138 kPa/m using RF and 12.17% and 0.088 kPa/m using ANN, respectively. The ML algorithms ability to model the pressure gradient is demonstrated using measured vs. predicted analysis where the experimental data points are mostly located in close proximity of the diagonal line, indicating a suitable generalization of the models. Comparing the performance between RF and ANN shows that the latter algorithms prediction accuracy is significantly better (p>0.01).
