Prediction of Lithofacies in Heterogeneous Shale Reservoirs Based on a Robust Stacking Machine Learning Model

The lithofacies of a reservoir contain key information such as rock lithology, sedimentary structures, and mineral composition. Accurate prediction of shale reservoir lithofacies is crucial for identifying sweet spots for oil and gas development. However, obtaining shale lithofacies through core sampling during drilling is challenging, and the accuracy of traditional logging curve intersection methods is insufficient. To efficiently and accurately predict shale lithofacies, this study proposes a hybrid model called Stacking, which combines four classifiers: Random Forest, HistGradient Boosting, Extreme Gradient Boosting, and Categorical Boosting. The model employs the Grid Search Method to automatically search for optimal hyperparameters, using the four classifiers as base learners. The predictions from these base learners are then used as new features, and a Logistic Regression model serves as the final meta-classifier for prediction. A total of 3323 data points were collected from six wells to train and test the model, with the final performance evaluated on two blind wells that were not involved in the training process. The results indicate that the stacking model accurately predicts shale lithofacies, achieving an Accuracy, Recall, Precision, and F1 Score of 0.9587, 0.959, 0.9587, and 0.9587, respectively, on the training set. This achievement provides technical support for reservoir evaluation and sweet spot prediction in oil and gas exploration.