A Fusion Model for Predicting the Minimum Miscible Pressure of a CO₂–Oil System Based on a Classification Sampling Strategy

Accurate prediction of minimum miscible pressure (MMP) in CO2–oil systems is critical for optimizing CO2 flooding in enhanced oil recovery (EOR). MMP data sets, derived from a wide variety of oils, exhibit a range from 6.55 to 40.5 MPa, with complex reservoirs (e.g., ultralow permeability and heavy oil) often exhibiting higher MMP values. These data sets are sparse and unbalanced due to the need for high-pressure phase experiments, which are both costly and dangerous. Conventional machine learning models perform poorly for these data sets. To address this issue, this paper proposes a K-means algorithm combined with a mixed sampling (KMMS) strategy, integrated into a fusion model designed for high MMP prediction in CO2–oil systems. The KMMS approach clusters similar samples to mitigate data unbalance. The fusion model improves model robustness by combining diverse base learners, thereby surpassing the limitations of standalone machine learning models. The findings reveal that this method outperforms individual machine learning models in terms of accuracy and generalizability. Specifically, it achieves an average mean absolute error (MAE) of 1.08, a root-mean-square error (RMSE) of 1.53, and a coefficient of determination (R2) of 0.955, calculated as the mean performance. In particular, the method improved the prediction accuracy by an average of 12.35%. Furthermore, SHapley Additive exPlanations (SHAP) are employed to evaluate feature contributions across clustered classes, validating the model’s interpretability and effectiveness. This model effectively provides accurate, efficient, and cost-effective predictions and holds a strong potential for application in energy and environmental engineering.