A Deep Learning Method for Oil–Water Layer Identification Based on SMOTE-ENN Resampling

To address the challenges of strong reservoir heterogeneity, complex relationships between logging parameters and fluid responses, and sample imbalance in the Ordos Basin, this study proposes an intelligent oil–water layer identification method based on deep learning. Focusing on Block M, six core feature parameters were precisely selected using the Pearson correlation coefficient method. The SMOTE-ENN hybrid sampling method was employed to resolve class imbalance and enhance dataset quality. Four identification models (GBDT, XGBoost, CatBoost, and Random Forest) were established and comprehensively evaluated. The results demonstrate that the GBDT model achieved a remarkable 99.3% accuracy on the test set, with its gradient boosting framework effectively mining sensitive features of oil–water layers, significantly outperforming other models. The SMOTE-ENN balanced samples improved average accuracy across multiple models by 2.6%. In practical application, the SMOTE-ENN optimized GBDT model achieved 99.5% identification accuracy when deployed in Block K, validating the method's effectiveness and reliability. This approach overcomes limitations such as redundant feature interference and class bias through feature optimization and hybrid sampling, providing an efficient and precise technical solution for complex reservoir identification. The methodology demonstrates GBDT's superior capability in feature mining compared to XGBoost and CatBoost, particularly in handling heterogeneous reservoirs with imbalanced data distributions.