Minimum Miscibility Pressure Of Carbon Dioxide In Crude Oil Based On Machine Learning Method

CO₂ injection flooding can not only sequester CO₂ but also improve oil and gas recovery.Therefore,this method has been widely concerned and valued by scholars in the industry.The difference between miscible flooding and immiscible flooding is very large,so the minimum miscible pressure（MMP）parameter,which determines whether miscibility can be achieved in the oil displacement process,is particularly critical.Although the conventional slim tube experiment is accurate to determine MMP,the test efficiency is low.Most of the existing empirical formulas have low accuracy and great limitations.Machine learning is one of the best methods for predicting MMP due to its convenience,speed and accuracy.However,most of the current researches based on machine learning methods have problems such as inadequate screening of main control factors and inadequate model verification.Therefore,in order to prove and solve these problems,this thesis is based on 147 sets of MMP data.Research on MMP model is carried out from the perspectives of main control factor screening,model establishment,model verification and model application:（1）In this thesis,methods such as grey correlation,Pearson correlation coefficient,Kendall correlation coefficient,Spearman correlation coefficient,variance characteristic method,univariate linear regression method and tree screening were used to screen and evaluate the main control factors,which proved that the screening method of the main control factors in the current study had defects.（2）Based on Support Vector Machine,Neural Network,Ordinary Least Square method,Stochastic Gradient Descent,Decision Tree,Random Forest,Bayes,K Nearest Neighbor eight common machine learning algorithms to establish the model.Conventional prediction accuracy verification method was used to verify the eight models,and good evaluation results were obtained.（3）Learning curve method and single-factor control variable analysis method were used for additional validation.Three defect models were screened out by learning curve verification method.The single-factor control variable analysis method screened out two additional defect models,and ranked the excellence of five defect models,which proved the limitations of the current model verification method,leading to the potential defects of the current model.（4）The three optimal models of Ordinary Least Squares,Bayes and Neural Network optimized in this thesis are used to predict the actual reservoir MMP.At the same time,the MMP of the reservoir is measured by the slim tube experiment method.The predicted value is close to the measured value,and the prediction effect is better.