Integration of AI for Predicting Water Production and Enhancing Reservoir Management in Naturally Fractured Reservoirs

Abstract Naturally Fractured Reservoirs (NFRs) with strong bottom aquifers and extensive fractures often experience significant water channeling, driven by highly conductive fracture networks and matrix interactions. This dual water channelling behaviour—early fracture-driven breakthrough followed by sustained matrix flow—creates challenges in managing water production and optimizing reservoir performance. This study introduces a novel Machine Learning (ML) predictive framework to model water channelling and accurately forecast water production trends in NFRs. The study's objectives were achieved through simulation scenarios using a dual-porosity dual-permeability model representing diverse fracture-matrix configurations. Water cut (WC) responses were analysed for key parameters, including fracture permeability (5–3000 md), matrix permeability (2–200 md), fracture spacing (100–500 ft), storativity ratio (0.002–0.2), mobility ratio (0.5–5), production rate (100–10000 bbl/d), matrix anisotropy ratio (0.1–0.5), and perforated interval (0.2–0.8), with inter-porosity flow coefficients ranging from 10−3 to 10−8. To efficiently design the experiment, the Box-Behnken design was used, significantly reducing the number of simulation runs to 117 runs while preserving the ability to capture the model’s nonlinearities. We also developed a single logistic correlation that accurately represents the relationship between dimensionless WC (WCd) and dimensionless time (td). The derived correlation demonstrates a strong agreement with the data, as indicated by a high coefficient of determination (R² = 0.9759) and a low Root Mean Square Error (RMSE = 0.0515), indicating excellent predictive accuracy. To further improve accuracy and better capture the behaviour of water channelling in NFRs, different ML models were tested. Models including decision trees, bagged trees ensemble, Gaussian Process Regression (GPR), least squares regression kernel, linear regression, neural networks, Support Vector Machines (SVM), were trained and validated using simulation data. High-performing models were further tested with additional simulation data to ensure accuracy and applicability. The bagged trees ensemble model performed excellently with an RMSE of 0.02811 and an R² of 0.99. GPR also showed high predictive accuracy, with an RMSE of 0.07591 and an R² of 0.95. Neural networks performed well, achieving an R² of 0.97 on the test set with an RMSE of 0.06230, though its slightly higher RMSE indicated minor inconsistencies in prediction accuracy. In contrast, regression-based models like linear regression and least squares regression kernel struggled to capture the complex nonlinear relationships in the data. Linear regression produced a high RMSE of 0.30832 with a low R² of 0.16, while the least squares regression kernel model had an RMSE of 0.17893 and an R² of 0.72, showing moderate but still limited accuracy. The Support Vector Machine model performed the worst, with an RMSE of 0.35696 and a negative R² of -0.12, indicating poor generalization and an inability to learn meaningful trends from the dataset. Given these results, ensemble models emerged as the best model, outperforming all others in both accuracy and computational efficiency, therefore it was further optimized, achieving an R2 of 0.99 and an RMSE of 0.025788. This AI framework represents a significant advancement in water breakthrough and production forecasting, enabling operators to optimize reservoir management strategies. By accurately capturing water channelling phenomena, the developed models provide actionable insights into water control, even in the absence of detailed reservoir data or under tight decision-making timelines. The integration of AI offers practicing engineers a smarter, more efficient tool for optimizing reservoir performance and addressing water production challenges effectively.