Abstract
Currently, the common step of intelligent learning application in reservoir characterization is to study its performance in intelligent computing directly. When the specific task is relatively simple and the adequate learning samples are not beyond reach, the model established by the intelligent learning algorithm can often achieve satisfying results. However, small labeled samples are usually the majority in reservoir characterization, making the method with stronger learning ability more likely to fall into the trap of noise in data, which further results in the poor or unstable performance of the model in practice. In this paper, we propose a learning method that aggregates fc-Nearest Neighbors (fcNN), Decision Tree (DT), Random Forest (RF), and XGBoost by stacked generalization (Stacking) to obtain a more robust model. We adopted this model to classify the lithofacies of a real well log data and compared it with the models established by the above four learning methods and the Soft Voting ensemble, through a statistical test measured by Fi score. The results show that compared with other methods, the proposed method can establish a more robust model with higher prediction accuracy for logging lithofacies classification with limited data.
NSTL