Production Forecasting in Gas Lifted Wells using Interpretable Machine Learning Techniques

Abstract Traditional techniques, such as decline curve analysis (DCA) and material balance, are industry-accepted standards for evaluating reservoir performance and estimating developed reserves. However, while these methods are reliable for estimating well productivity under natural flow conditions, accurate evaluation of gas lifted wells can be challenging due to the additional energy injected to augment mthe natural flow. Although attempts have been made to model gas lifted wells using machine learning (ML) techniques, these models provide little or no interpretability, thereby offering no clues to the underlying physical interactions behind their predictions. To address this issue, an empirical correlation was developed to predict gas lifted oil production rates using interpretable non-linear regression ML models. Production data from four wells were obtained from literature, encompassing production periods both with and without gas lift. A Python program was written to train the ML models. A ridge regression model was trained on production data from three of the wells using this program. The coefficients and intercepts of the resulting model were extracted to generate an empirical correlation and its predictive performance was then statistically analyzed. The performance of the derived empirical correlation was validated using production data from a new well, exhibiting a correlation coefficient of 0.96 and a root mean squared error of 27.55 STB. Analysis of predicted production rates under both natural and gas lifted flow conditions indicated that proper utilization of gas lift can increase developed reserves under natural flow by as much as 200%. The predicted production life of the well also increased significantly from 28 months under natural flow to 48 months with gas lift. A non-linear empirical correlation has been developed and tested to predict the oil production rate in gas lifted wells. With slight modifications, this correlation and the workflow implemented here can be applied to other wells, providing a practical tool for forecasting oil production rates not only in gas lifted scenarios but for other artificial lift methods as well.