Abstract
Machine learning has been successfully implemented for the past 20 years in estimating reservoir fluid properties competing with the empirical correlations. One of the most commonly utilized modelling schemes is the artificial neural network which is known for its black-box problem that does not show the steps taken to reach the final estimation of the fluid properties. This study offers a different modeling approach that overcomes the limitations of the current implemented modeling scheme providing users with better predications and deeper understanding of the key input parameters in modeling. The proposed model predicts the bubble point pressure (Pb) and the oil formation volume factor at bubble point pressure (Bob) as a function of oil and gas specific gravity, solution gas-oil ratio, and reservoir temperature by a boosted decision tree regression (BDTR) predictive modeling scheme. The K-means clustering algorithm is performed as a preprocessing step based on the Pressure-Volume-Temperature (PVT) input features to increase the prediction accuracy. In addition, the predictive power of the built K-means clustered BDTR model implemented in this study is compared against the most commonly used empirical correlations, the ANNs, and the standalone BDTR model. Moreover, the feature importance of predicting Pb and Bob is discussed. The universal dataset used in building the predictive model consists of 5200 experimentally derived data points representing worldwide crude oils covering a wide range of geographical regions. The built BDTR model is more accurate and it outperforms the most commonly used empirical correlations and the previous machine learning models in predicting Pb and Bob in terms of the average absolute percent relative error. Furthermore, the proposed model can be integrated into simulators and it can also be applied towards predicating other oil and gas properties, in the gas-liquid two-phase flow pattern identification, and in predicting rock properties. As an interpretable approach in predicting the PVT properties of crude oils, the proposed model can be used as an alternative modeling scheme in PVT characterization where the importance of the input features can heuristically and accurately be determined. This can be applied towards preventive maintenance and anomaly detection studies where prediction decisions can be further investigated by the interpretable representation of the decision trees. In addition, it is the most accurate model to date in predicting the bubble point pressure and the oil formation volume factor at bubble point pressure of crude oils.
