Enhanced machine learning—ensemble method for estimation of oil formation volume factor at reservoir conditions

Since the oil formation volume factor (B(o)) is crucial for various calculations in petroleum engineering, such as estimating original oil in place, fluid flow in the porous reservoir medium, and production from wells, this parameter is predicted using conventional methods including experimental tests, correlations, Equations of State, and artificial intelligence models. As a substitute to conventional black oil methods, the compositional oil method has been recently used for accurately predicting the oil formation volume factor. Although oil composition is essential for estimating this parameter, it is time-consuming and cost-intensive to obtain through laboratory analysis. Therefore, the input parameter of dissolved gas in oil has been used as a representative of the amount of light components in oil, which is an effective factor in determining oil volume changes, along with other parameters, including pressure, API gravity, and reservoir temperature. This study created machine learning models utilizing Gradient Boosting Decision Tree (GBDT) techniques, which also incorporated Extreme Gradient Boosting (XGBoost), GradientBoosting, and CatBoost. A comparison of the results with recent correlations and machine learning methods adopting a compositional approach by implementing tree-based bagging methods: Extra Trees (ETs), Random Forest (RF), and Decision Trees (DTs), is then performed. Statistical and graphical indicators demonstrate that the XGBoost model outperforms the other models in estimating the B(o) parameter across the reservoir pressure region (above and below bubble point pressure); the new method has significantly improved the accuracy of the compositional method, as the average absolute relative deviation is now only 0.2598%, which is four times lower than the previous (compositional approach) error rate. The findings of this study can be used for precise prediction of the volumetric properties of hydrocarbon reservoir fluids without the need for conducting routine laboratory analyses by only employing wellhead data.