Data Driven Prediction of the Minimum Miscibility Pressure (MMP) Between Mixtures of Oil and Gas Using Deep Learning

Abstract Knowing the minimum miscibility pressure (MMP) between different oil and gas compositions is important to predict reservoir performance for gas-based injection as a secondary gas flood or tertiary technique such as water alternating gas (WAG). Machine Learning (ML) has been used widely and has been proven efficient in estimating these properties. In this work, the development of ML as well as commonly used algorithms in predicting bubble point pressure and oil formation volume factor is reviewed. Just a few studies are found before 2000. From 2001 to 2010, the use of ML increased steadily. However, a sharp augmentation in number of articles is observed from 2011 up to now. More than that, Artificial Neural Networks (ANN) is the most employed algorithm with 23 applications out of 38 studied papers. In addition, for the first time, deep learning- multiple fully connected networks algorithm is implemented to predict the MMP for oil and gas through 250 datasets covering a wide range of CO2 concentration from 0 to 100% in the injected gas. The wide range of CO2 concentrations is to cover all modes of gas injection from a pure CO2 flood to CO2 being negligibly present when injecting a sweet gas. The model is then optimized using Early Stopping and K-Fold Cross Validation techniques, showing the average result of k splitting data sets. The eight input parameters are as follows: reservoir temperature, oil characteristics (molecular weight, ratio of volatile components, and intermediate components), and gas characteristics (mole percentage of CO2, Cl, N2, H2S, C2+). The proposed model is compared with other Machine Learning Techniques such as Decision Tree and Random Forest Regression. The results show that reservoir temperature, the amount of CO2 and Cl in the gas source were the parameters to affect MMP the most significantly. The presence of CO2 in the gas stream will lower the MMP significantly. The Deep Learning model obtained an R2 = 0.96 and a Root Mean Square Error (RMSE) of 5.4%. Through Early Stopping technique, the proposed model reach the R2 result of 0.97 in 7 epochs. An R2 value of 0.954 was found using K-Fold Cross Validation technique, resulting in a good model generated by five folds data set. The model built by Deep Learning algorithm was more accurate than these ones built by Decision Tree and Random Forest Regression, which had an R2 value below 0.9 and RMSE larger than 10%. This work goes beyond other prior research by adding a ‘stopping point’ concept, increasing the overall performance of the methods for general applications, and considering the full range of CO2 in the gas stream.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.