Heliyon (Jun 2024)
Advanced machine learning approaches for predicting permeability in reservoir pay zones based on core analyses
Abstract
Permeability is the most important petrophysical characteristic for determining how fluids pass through reservoir rocks. This study aims to develop and assess intelligent computer-based models for predicting permeability. The research focuses on three novel models—Decision Tree, Bagging Tree, and Extra Trees—while also investigating previously applied techniques such as random forest, support vector regressor (SVR), and multiple variable regression (MVR). The primary dataset consists of 197 data points from a heterogeneous petroleum reservoir in the Jeanne d'Arc Basin, including laboratory-derived permeability (K), oil saturation (SO), water saturation (SW), grain density (ρgr), porosity (φ), and depth. The most effective machine learning models are identified by a thorough analysis that makes use of a variety of statistical metrics, such as the coefficient of the determinant (R2), mean squared error (MSE), mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), maximum error (maxE), and minimum error (minE). Additionally, core features are ranked based on their importance in permeability modeling. This study deviates from conventional approaches by proposing an efficient means of forecasting permeability, reducing reliance on labor-intensive and time-consuming laboratory work. The findings reveal that MVR is unsuitable for permeability prediction, with all developed models outperforming it. Extra Trees emerges as the most accurate model, with an R2 of 0.976, while random forest and bagging tree exhibit slightly lower R2 values of 0.961 and 0.964, respectively. The ranking of these algorithms based on performance criteria is as follows: extra trees, bagging tree, random forest, SVR, decision tree, and MVR. The study also presents a detailed analysis of the impact of input parameters, highlighting porosity (φ) and water saturation (SW) as the most influential, while grain density (ρgr), oil saturation (SO), and depth are considered less important. This study contributes to the petroleum industry's knowledge by showcasing the inadequacy of MVR and highlighting the superior performance of machine learning models, particularly Extra Trees. The proposed models employed in this study can help engineers and researchers determine reservoir permeability quickly and accurately by using a few core attributes, reducing the dependency on resource-intensive and time-consuming laboratory work.
