Machine learning framework for evaluating fracturing-flooding effectiveness: From prediction to decision recommendations

Abstract

Fracturing-flooding is an emerging technique designed to enhance oil recovery in waterflooding reservoirs, however, its effectiveness remains difficult to predict due to limited theoretical foundations and the lack of reliable methods. This paper proposed a novel machine learning framework for predicting, analyzing and optimizing fracturing-flooding effectiveness. Data from 498 fracturing-flooding well groups within an oilfield were collected, encompassing 3 categories of indicators: reservoir geological parameters, fracturing-flooding parameters, and effectiveness outcomes. 9 machine learning models were screened, and cumulative SHapley Additive exPlanations (SHAP) values were utilized to identify both key and risk factors. The established model was integrated with SHAP analysis through a visualized output, providing a clearer interpretation of the factors influencing the predictions of fracturing-flooding effectiveness. Experimental results show that the eXtreme Gradient Boosting (XGBoost) model outperformed other models, achieving an 81.5% accuracy during cross-validation and an AUC value of 0.9. On the test set, the model sustained nearly 80% accuracy with an AUC of 0.89. The key factors identified include oil production before fracturing-flooding (OP), water production before fracturing-flooding (WP), the number of oil wells in the well group (NWG), and reservoir thickness. Among these, WP was identified as the primary risk factor contributing to potential misjudgment. Through the application of the framework, the effectiveness of fracturing-flooding operations improved by 16.7%. This research provides a robust foundation for optimizing fracturing-flooding strategies to enhance oil recovery and contribute to a more efficient energy supply in oilfield.