Production Forecasting and Optimization of Gas Lift Operations Using a Dual-Track Machine Learning Framework

Abstract

As reservoirs mature and natural driving energy declines, artificial lift systems become essential for sustaining hydrocarbon production. Gas lift, which operates by injecting high-pressure gas to reduce the hydrostatic gradient and increase inflow from the reservoir, is among the most extensively used methods for this purpose. Its performance, however, is sensitive to a complex interplay of injection conditions, reservoir properties, well architecture, and multiphase flow behavior that varies across wells. Production forecasting in gas-lifted wells is therefore challenging, and while data-driven machine learning models have shown strong predictive capability for this task, their ability to generalize across wells with distinct operating behaviors remains limited and unexamined. This study presents Track A of a dual-track machine learning framework, applying data-driven models to real field data from 11 horizontal gas-lifted wells completed in the Eagle Ford Shale to evaluate predictive performance, diagnose generalization behavior, and interpret model decision patterns. Three supervised regression algorithms, specifically ANN, Random Forest, and XGBoost, were trained to forecast daily oil production rate using 15 input features spanning operational, pressure, architectural, and reservoir variables. Model performance was evaluated under both a conventional 80/20 random train-test split and a strict Leave-One-Well-Out (LOWO) cross-validation framework. Distributional diagnostics and exponential smoothing were applied to investigate the origins of generalization behavior. SHAP analysis was used to examine feature attribution and interpret algorithm-dependent model decisions. Under the 80/20 split, all three models achieved strong predictive performance with R2 values of 0.960, 0.953, and 0.946 for ANN, Random Forest, and XGBoost, respectively. Under LOWO evaluation, performance degraded, with average R2 of -1.65, -0.20, and -0.62. Wells operating in distinct production regimes not well represented in the training data, experienced complete performance collapse. Signal conditioning reduced absolute prediction errors but did not resolve cross-well generalization, indicating that behavioral heterogeneity across wells is a primary contributor to the observed limitations. SHAP analysis revealed algorithm-dependent attribution patterns, with ANN driven by reservoir-state variables and Random Forest by wellbore and completion variables, directly explaining the differential generalization behavior under LOWO. The results confirm that the available field dataset does not fully represent the operating space required for reliable cross-well generalization. The nonlinear and behaviorally diverse nature of gas lift systems across wells motivates a complementary physics-based simulation track, capable of generating structured operating scenarios beyond what field data alone can provide, directly addressing the coverage and behavioral gaps identified in Track A.