A Method for Evaluating the Suitability of CO2 Injection in Oil Reservoirs Based on Multi-model Coupled Machine Learning Algorithm

Both indoor and on-site pilot experiments have demonstrated the enormous potential of CO2 injection to improve crude oil recovery, but there is currently no consensus on how to evaluate reservoirs suitable for CO2 injection development. Most traditional methods rely on developing subjective screening conditions; some conditions are contradictory in practical applications. In addition, the nested relationships between reservoir characteristics, fluid properties, and development stages also result in low accuracy in the prediction results obtained by an independent judgment of various parameters. This study constructed a database of actual development effects of CO2 injection in oil reservoirs, considering a total of 19 parameters such as reservoir characteristics, fluid physical properties, and development stages. The missing data was optimized using the K-nearest neighbor (KNN) algorithm, which provided a possibility for the practical application of multi-parameter evaluation in the field. On this basis, based on support vector machine (SVM), classification and regression tree (CART), KNN, random forest (RFC), optimized distributed gradient enhancement library (XGBoost), and the principle of minority obeying the majority, the problem of overfitting and distortion of a traditional single algorithm is avoided, and a fast, accurate and intelligent screening model of reservoir CO2 injection suitability is established. The results show that the model accuracy of the above five algorithms is 0.910, 0.854, 0.809, 0.865, and 0.921, respectively. The accuracy of the proposed mixed model is 0.933, confirming the accuracy of the CO2 injection suitability screening method established in this study. In addition, the main factors affecting the suitability of CO2 development are ranked as follows: crude oil viscosity, reservoir heterogeneity, reservoir permeability, reservoir pressure, original reservoir pressure, current oil saturation, reservoir depth, surface crude oil density, etc. The degree of miscibility of the CO2 crude oil system should not be the primary criterion for real-time CO2 flooding. The composite machine learning method proposed by this research institute is flexible and accurate, which can avoid misjudgment caused by fixed indicator ranges. It is of great significance to achieve rapid screening of CO2 development reservoir suitability under the dual carbon strategy goal.