Leveraging Machine Learning for Predicting Gas Separation Efficiency of a Downhole Separator

Abstract

Abstract Artificial lift systems, specifically Electrical Submersible Pumps and Sucker Rod Pumps, often face operational challenges due to high Gas-Oil Ratio, leading to premature tool failure and increased non-productive time. These issues underscore the necessity for efficient gas separation at upstream of the artificial lift pump to ensure uninterrupted operations. Downhole gas separators, situated upstream of the artificial lift pump, are designed to segregate gas from the fluid, reducing downtime. This study aims to predict the gas separation efficiency of the downhole gas separator using experimental data from a centrifugal separator and testing the generated models on experimental data for a gravity separator. To predict the separation efficiency of the downhole separator, multiple machine learning models were evaluated, and a comparative analysis was performed. The experimental test data obtained for the centrifugal packer-type separator was used for model training and experimental test data for the gravity separator was used for model testing (blind test). To conduct experiments, multiphase flow setup (31ft. (9.4m) long horizontal and 27ft. (8.2m) tall vertical section of multiphase flow setup) was used. In this study, seven regression methods, including multilinear (MLR), random forest (RF), support vector machine (SVM), ridge, lasso, k-nearest neighbor (KNN) and XGBoost, were examined to assess their ability to effectively predict the gas separation efficiency. Several performance metrics, including Root mean square error (RMSE), Mean absolute percentage error (MAPE) and R-squared were evaluated to select the efficient model. In-depth exploratory data analysis and feature engineering were performed to identify significant input parameters for model training. The scatterplot and correlation plot indicated the response variable were mostly affected by inlet liquid and gas volume flow. Therefore, liquid and gas volume flow at inlet were selected as two independent parameters to predict gas volume flow per minute at the outlet (GVFO) as the response variable. After data scaling, a comparative analysis of the seven incorporated regression models revealed that, the Random Forest model to be the most efficient for predicting GVFO, closely followed by the KNN model. The Random Forest model exhibits R-squared values of 96% and RMSE value of 112. Hyperparameter tuning and tenfold cross-validation ensured optimal values for tuning parameters and prevented oveffitting, respectively. This developed machine learning workflow enables the prediction of the gas separation efficiency of different type of downhole gas separator by using experimental test data from a centrifugal downhole separator. It provides critical findings through a comparative analysis of seven regression models. This research offers valuable insights that can aid in optimizing the performance of artificial lift systems.