Stacking is an ensemble machine learning method designed to improve overall performance by combining the predictions of multiple base learners. The core idea of Stacking is to use the output of different base learners as input and make final predictions through a meta-learner, allowing the model to benefit from the strengths of various base learners and adapt better to complex data patterns. Predicting pore structures in low permeability is a sophisticated nonlinear modeling issue affected by multiple factors, including sedimentary and diagenetic effects on the structure. An individual model to predict pore structure is not satisfactory. In this paper, we applied a fusion of random forest and Adaboost models as base learners, which emerges as the cornerstone of this research, showcasing remarkable versatility and integration of diverse learner strengths. The Silurian system in Tazhong area features a vast expanse of lithological reservoirs with low abundance and high heterogeneity. Capillary pressure plays a crucial role in the distribution of these less permeable reservoirs by pore structure type influence. Aiming to predict the pore structure for the low-permeability reservoir in the field, which is characterized by low abundance and high heterogeneity, the study strategically integrates geological theory into specialized databases and harnesses machine learning methods. The resulting machine learning modeling dataset leverages geological factors and logging response characteristics to strengthen the ensemble machine learning stacking model. This sophisticated model not only achieves exceptional accuracy but also undergoes rigorous validation by analyzing predicted pore structure types alongside production data. The model's average accuracy was 79.4%, which practically hit the test set's accuracy threshold, meaning that the model was in an optimal state. In essence, the proficiency of the stacking model in predicting pore structure types serves as a transformative force, offering significant advancements in hydrocarbon distribution predictions and paving the way for more streamlined and effective future EOR activities.
Read full abstract