Abstract
Conventional methods, represented by multivariate linear/nonlinear regression, empirical formulas, and petrophysical models, struggle to adequately capture the complex high-dimensional nonlinear relationships between logging responses and reservoir parameters, resulting in inherent limitations in shale reservoir type predictions. This work focuses on deep shale reservoirs from the Upper Ordovician Wufeng Formation (O(3) w) and the Lower Silurian Longmaxi Formation, Member 1, Submember 1 (S(1) l (1) (1)) in the western Chongqing area, Sichuan Basin. Two schemes that identify deep shale reservoir types based on Light Gradient Boosting Machine (LightGBM) algorithm were established, followed by rigorous selection of the optimal scheme. Subsequently, the SHapley Additive exPlanations (SHAP) algorithm was implemented to quantitatively evaluate the importance of logging curves. The optimized model was ultimately deployed for comprehensive reservoir grading evaluation in the target area. The results demonstrate that compared to regression-based scheme, classification-based scheme not only substantially reduces model complexity for deep shale reservoir identification but also achieves remarkable improvements in computational efficiency and predictive performance. For the classification-based scheme, weighted precision (WP) and weighted recall (WR) metrics on the testing data set attained 90.5% and 90.4%, respectively, outperforming those of the regression-based scheme at 85.9% and 86.1%, respectively. Feature importance analysis via SHAP revealed that compensated density (DEN), gamma ray (GR), and compensated neutron (CNL), exert the most significant influence on type I and III reservoir identification, while DEN, acoustic transit time (AC), and GR dominate type II reservoir identification. The dependence plots elucidate complex nonlinear relationships between logging responses and reservoir classification outcomes. The grading evaluation in the target area indicates that type I reservoirs predominantly occur in the Upper O(3) w and bed 1 of S(1) l (1) (1) (S(1) l (1) (1)-1). The Bayesian-optimized LightGBM algorithm enables efficient and accurate identification of deep shale reservoir types, providing a novel approach for grading evaluation of deep shale reservoirs.