Estimating Pressure-Volume-Temperature Properties of Crude Oil Systems Using Boosted Decision Tree Regression

Machine learning has been successfully implemented in the estimation of reservoir fluid properties, competing with the empirical correlations used in this field. One of the most commonly used modeling schemes is the artificial neural network, which is known for its black-box problem. This study offers a different modeling approach that overcomes this limitation. The model provides accurate estimations and facilitates a deeper understanding of the key input parameters and their importance to the estimated results. It uses a boosted decision tree regression (BDTR) predictive modeling scheme to estimate the bubble point pressure (Pb) and oil formation volume factor at the bubble point pressure (Bob) as a function of oil and gas specific gravities, solution gas-oil ratio, and reservoir temperature. The built BDTR model exhibits higher accuracy and performance than previous machine learning models and the most commonly used empirical correlations for estimating Pb and Bob. The results indicate the higher efficacy of the developed model integrated with an imputation pre-processing step compared with the most commonly used empirical correlations in estimating Pb. This model brings significant predictive capability and versatility to datasets with multiple missing input features.