Continental shale oil is widespread in many basins worldwide and constitutes an important potential incremental source of unconventional oil, making reservoir evaluation of such plays particularly significant. Total Organic Carbon (TOC) is a key parameter for evaluating oil potential in continental shale reservoirs. Based on conventional well logs, this study proposed a hybrid stacking ensemble method tthat integrated Extreme Gradient Boosting (XGBoost), Random Forest (RF), Least-Squares Support Vector Regression (LS-SVR), and an Improved Gated Recurrent Unit (IGRU). Input features included lithology and reservoir-layer priors, continuous lithology digitalization, and multiple overlapping well-log response features, together with amplitude–phase attributes extracted by the Time-Parameterized Complex-Valued Overlapping Convolutional Architecture (TP-CVOCA) and integrated petrophysical representations from Principal Component Analysis (PCA). Heuristic optimization algorithms were employed for feature selection and hyperparameter tuning, and Regularized Mixup (RegMix) was applied for data augmentation. The IGRU and TP-CVOCA modules were designed to handle non-uniform time-step issues. Based on 2,374 core-measured TOC samples from the northern Songliao Basin, seven comparative experiments were conducted. The proposed method achieved R² of 0.8184 ± 0.0301 for intra-well prediction and 0.7132 ± 0.0590 for cross-well prediction.
Lü et al. (Sat,) studied this question.