The method of turbidite reservoir prediction based on improved Stacking ensemble model

Seismic attributes contain a wealth of reservoir information, and the integration of various seismic attributes can enhance the accuracy of reservoir prediction. Due to the complex and heterogeneous underground geological structures, the suitable fusion algorithms vary not only among different oil fields but also at different locations or layers within the same oil field. Therefore, there is an urgent need to explore a multi-algorithm ensemble approach for attribute fusion to improve the generalization capability of seismic attribute integration methods. To improve the accuracy of reservoir prediction, an improved Stacking ensemble model-based method for predicting turbidite reservoirs has been proposed. Firstly, well log seismic attributes are optimized based on correlation analysis and unsupervised clustering techniques to construct a relationship model between seismic attributes and the thickness of turbidite reservoirs, reducing the ambiguity of seismic attributes. Then, hyperparameter optimization of the model is conducted using Optuna, and several types of models with good application effects and significant differences in the field of reservoir prediction are selected as the base learners of the Stacking ensemble model based on root mean square error (RMSE), mean absolute error (MAE), and correlation analysis. Finally, corresponding weights are assigned to the prediction results based on the test accuracy of the base learners, and the original dataset is also included in the meta-learner training, enabling the meta-learner to learn the implicit relationship between the original and new training sets, thereby enhancing the model's predictive performance. This method is applied to the prediction of turbidite reservoirs in the NZ Subsag, and the results show that compared with single prediction models and traditional Stacking ensemble models, the improved Stacking ensemble model significantly reduces the root mean square error in the prediction of turbidite reservoir thickness, and the correlation coefficient between the integrated attributes and sand thickness reaches 0.92, proving that the method has good application prospects.