Well-aware stacking ensemble machine learning framework for oil production prediction in gas-lift wells: leakage-aware validation in the Halewah Field

Abstract Lift-gas decisions in the mature Halewah Field, Yemen, require fast, well-specific oil-rate predictions. PIPESIM models of wells H-06 and H-09, each calibrated to a stabilized field operating point, were evaluated at 350 operating scenarios in each well. Six learners and 22 ensemble configurations were compared, including WellAware_Stack, one six-learner stack per well. On a 105-record holdout, WellAware_Stack reached R² = 0.996 (RMSE 42.0 STB/d), against 0.949 (155.1 STB/d) for the best pooled stack. Repeated scenario-grouped nested cross-validation of these two finalists gave R² = 0.994 and 0.857, exposing a cross-well information path in row-wise splits. A controlled comparison attributed the gain to well information rather than stacking: supplying it reduced the error of every model family five- to tenfold, a well-specific Gaussian process was more accurate than WellAware_Stack (RMSE 23.6 against 44.9 STB/d), and standardizing the Gaussian-process target stabilized the stack. On 60 fresh PIPESIM scenarios simulated after the models were frozen, WellAware_Stack reached R² = 0.995 and 0.999 for H-06 and H-09, and the well-specific Gaussian process 0.999 for both; at the calibrated operating points their errors were 10–17% and below 4%, respectively. The study quantifies surrogate-to-simulator fidelity and provides a leakage-aware protocol for paired simulation data.