Accurately predicting Flowing Bottom-Hole Pressure (FBHP) is critical for optimizing oil and gas production. Existing predictive methods often rely on oversimplified or complex, yet computationally expensive, models that fail to capture the intrinsic nonlinearities of well dynamics, leading to inaccurate predictions and potential economic losses. This paper introduces a three-layer heterogeneous stacking ensemble model to address the latter challenge. In particular, the key novelty of the developed work is a hierarchical architecture that integrates five distinct Machine Learning (ML) base learners, two meta-learners, and a final super-learner, i.e. , an additional meta-model that combines the outputs of the meta-learners to capture complex, non-linear relationships in the data. When evaluated on a field dataset (total dataset samples N = 795 ; test set samples N = 199 ), the proposed Super Learner Stacking model (ST-S) demonstrated superior predictive performance on the independent test set, achieving R-squared ( R 2 ) = 0 . 857 ± 0 . 006 and Root Mean Squared Error (RMSE) = 146 . 382 ± 2 . 806 . In addition, the ST-S model outperformed all individual models and simpler stacking ensembles reported in the article. As a result, the developed ST-S model provides a robust, data-driven tool for FBHP prediction, achieving high predictive accuracy without resorting to computationally expensive methods, thereby supporting improved well management and production optimization. • Accurate FBHP prediction using heterogeneous stacking ensemble • Super learner architecture improves FBHP prediction accuracy. • Ensemble outperforms individual models for oil well optimization. • Uncertainty analysis allows assessing model reliability and robustness. • Shapley analysis highlight key features influencing FBHP.
Campos et al. (2026) studied this question.