Real-Time Estimation of Production Rates in Gas Condensate Wells Using a Machine Learning Model

Abstract

Abstract For efficient reservoir management, production allocation, and surface facility optimization, recovery rates in gas condensate wells must be continuously and accurately estimated. Conventional measurement systems, like well testing separators and multiphase flow meters, are costly, need upkeep, and are frequently not available at every well location. With the help of readily measured surface parameters for gas-condensate wells, this study proposes a machine learning method for the real-time estimation of gas, condensate, and water production rates. Three separate extreme gradient boosting (XGBoost) regression models were created, one for each target phase, using a dataset of about 33,000 records from producing gas-condensate wells. To preserve model generalization, an extensive feature engineering was used, including computed parameters like flow energy, differential pressure ratios, and temperature-pressure products. Using RandomizedSearchCV with 10-fold cross-validation, hyperparameter tuning was carried out, with RMSE serving as the main optimization metric. With low mean absolute errors, the models obtained R2 values of 0.999 for gas, 0.996 for condensate, and 0.985 for water on the test set. Both manual and batch data input were supported by the trained models, which were then integrated with decline curve analysis for production forecasting in an internal Streamlit-based gas wells virtual meter application. Excellent agreement between predicted and measured rates was shown in a field application, confirming the method's viability as an affordable and scalable substitute for conventional flow measurement systems.