Application of supervised machine learning and unsupervised data compression models for pore pressure prediction employing drilling, petrophysical, and well log data.

Accurate determination of pore pressure is critical in the design of wells, determining a safe range of mud properties, and estimating the required mud weight to ensure wellbore stability. Conventional techniques for forecasting pore pressure, such as the Eaton, Bower, or compressibility methods, have certain constraints. These methods depend on empirical relationships and constants that can differ between basins. This study proposes an effective data-driven approach that utilizes machine learning algorithms to forecast reservoir pore pressure. A total of five machine learning algorithms, namely multivariable regression (MVR), polynomial regression (PR), random forest (RF), CatBoost regression, and multilayer perception (MLP), are applied in this research. Hybrid stacking modeling is employed for the first time to forecast pore pressure and to improve the accuracy and robustness of the results by combining different methodologies. Principal component analysis is also utilized (PCA) to extract features, hence expediting the entire process by reducing dimensionality. To accomplish the objectives, 1811 recordings are selected from the Volve Field, situated approximately 200 km west of Stavanger, Norway. These recordings encompass depth data; well logs, including NPHI, GR, DT, RD, RHOB, RS, and RT; drilling activities, specifically ROP; and petrophysical parameters, including BVW, K, PHIF, SW, and VCL. Pore pressure is used as the output level to generate data-driven models. 70% of the dataset is used for training the machine learning models, while the remaining 30% is reserved for testing the models to evaluate their performance and generalization capability. Data standardization is conducted to ensure that the utilized data is statistically well-distributed, devoid of measurement mistakes, and impervious to instrumental noise. Regression metrics, such as mean MAE, R, Adjusted R RMSE, MinE, and MaxE are employed to evaluate the efficacy of the models. The results suggest that the stacking model, which integrates CatBoost and Random Forest (RF) as base models and Polynomial Regression (PR) as the meta-model, achieves an R of 0.9846, an adjusted R of 0.9842, MAE of 11.20 and an RMSE of 22.747 on the testing dataset. This makes it the most accurate model for pore pressure prediction, followed closely by CatBoost. The MVR, exhibiting an R of 0.896 and an RMSE of 57.931, is the least efficient model. A thorough comparison of all analyzed models indicates that the algorithms, ranked by performance metrics, are Stack_2, CatBoost, Stack_1, RF, PR, Stack_3, MLP, and MVR. Hybrid stacking improves performance even without hyperparameter tuning. PCA significantly speeds up the entire process by lowering the number of dimensions, hence enhancing the cost-effectiveness of the procedure. Using a few petrophysical, drilling, and well log data, the methodology presented in this work can help engineers and researchers quickly and precisely determine the reservoir pore pressure, validating the safe and cost-effective drilling operations.