Reliability Assessment of PVT-Properties of Reservoir Fluids on the Basis of a Probabilistic Mixture Model of Student's Distributions

Abstract

Abstract Substantiation of the component composition and PVT-properties (Pressure, Volume, Temperature) of reservoir fluids is one of the most crucial conditions for increasing the reliability of reserves calculation and efficiency of oil and gas field development. The main tasks of this article are abnormal samples definition and a prediction of some PVT-properties form other ones. In machine learning this tasks names as the anomaly detection and the regression analysis. To solving this task, the model based mixtures of multidimensional Student's distributions was developed, training of which using variational Bayesian inference. The developed model was tested on data base containing more than 3200 samples of reservoir fluids. The article compares the proposed method by the quality of the predictions with other machine learning methods, for example random forest, artificial neural networks and SVM regression. Using these models, the quality of PVT-sample parameters prediction was significantly higher than the proposed method of probabilistic mixture of Student's t-distributions.