AbdulQoyum Adegoke Olowookere

Accurate prediction of reservoir production remains a critical challenge due to the complex interplay of multiphase flow behavior, temporal dynamics, and reservoir heterogeneity. This study proposes a Hybrid Physics-Informed Dynamic Productivity Index (HPD-PI) model that integrates classical inflow performance relationships with data-driven temporal features to enhance production forecasting. The model extends the conventional productivity index formulation by incorporating water-cut and gas-oil ratio corrections, dynamic production trends, and temporal memory through lagged production terms. A comprehensive dataset comprising production, pressure, and multiphase-flow parameters was used to evaluate the proposed approach. A baseline Random Forest model was first developed for comparison, achieving a coefficient of determination (R ² ) of 0.8138. However, analysis revealed that the model was predominantly driven by temporal dependencies with limited physical interpretability. In contrast, the proposed Hybrid HPD-PI model achieved superior performance, with an R ² of 0.9466, root mean square error (RMSE) of 71.01, and mean absolute error (MAE) of 38.78. The improvement is attributed to the integration of temporal memory, multiphase flow corrections, and dynamic behavior within a physically consistent framework. The results demonstrate that the hybrid model effectively captures both short-term fluctuations and long-term production trends while maintaining interpretability. The study highlights the importance of combining physical principles with data-driven techniques for robust reservoir modeling. The proposed HPD-PI framework provides a reliable tool for production forecasting, reservoir management, and decision-making, with potential applicability to a wide range of reservoir systems.

Accurate oil-rate forecasting is essential for effective reservoir management, yet it is challenged by nonlinear flow behavior, operational variability, and data uncertainty. Conventional physics-based approaches, such as decline-curve analysis, provide interpretability but lack short-term responsiveness, while purely data-driven machine-learning models, although accurate, may violate physical production trends. This study addresses this gap by developing a physics-guided hybrid forecasting framework. A dataset containing 22 engineered features representing pressure dynamics, temperature behavior, choke performance, production characteristics, and temporal attributes was used to train Linear Regression, Random Forest, Gradient Boosting, and Support Vector Regression models. An Arps decline-curve model was fitted to capture long-term reservoir behavior, and a hybrid model was constructed by anchoring machine-learning predictions to physics-based decline trends. Model performance was evaluated using MAE, RMSE, R2, and cross-validation. Results indicate that ensemble machine-learning models outperformed linear and kernel-based methods, with Gradient Boosting achieving an RMSE of 54.15 STB/day and an R2 of 0.9984. The proposed hybrid model further improved predictive accuracy, reducing RMSE to 53.83 STB/day and increasing R2 to 0.9986, demonstrating enhanced stability and physical consistency. The study concludes that integrating decline-curve physics with machine learning yields more reliable oil-rate forecasts. Future work should extend the framework to field-scale applications and incorporate uncertainty quantification to support production-optimization decisions.

This study presents a leakage‐aware machine learning framework for predicting minimum horizontal stress ( σ _h min) using structured geomechanical and fracture‐related parameters. A dataset comprising 21,499 records from approximately 200 horizontal wells in the Marcellus Shale was preprocessed using a strictly leakage‐controlled pipeline, including feature refinement, outlier capping, and training‐fold–based transformations. Three ensemble models—random forest (RF), gradient boosting regressor (GBR), and extreme gradient boosting (XGBoost)—along with a stacking regressor and artificial neural network (ANN), were evaluated using root mean squared error (RMSE), mean absolute error (MAE), and coefficient of determination ( R ² ). The RF model achieved the best baseline performance (RMSE = 6.5846, R ² = 0.9988), while gradient boosting showed improved performance after tuning (RMSE ≈ 6.59). The stacking model delivered competitive results (RMSE = 7.4963), whereas the ANN showed lower performance (RMSE = 55.5008), indicating limited suitability for structured tabular data. A scenario‐based evaluation using Nigerian reservoir data demonstrated reasonable pr3332edictive consistency but was not treated as external validation due to data limitations. The results confirm that leakage‐aware ensemble learning provides a robust and physically consistent approach for predicting minimum horizontal stress, with potential application in reservoir characterization and hydraulic fracturing design.