Determination of bubble point pressure and oil formation volume factor: Extra trees compared with LSSVM-CSA hybrid and ANFIS models

Abstract

Successful field development relies on effective reservoir management, which, in turn, is, to a great extent, influenced by the knowledge of reservoir fluid properties and phase behaviour. Laboratory experiments, known as Pressure-Volume-Temperature (PVT) tests, are the most reliable sources of detailed information on reservoir fluid properties required for different recovery stages and methods. However, these laboratory experiments are costly and time-consuming. Also, the experimental data of reservoir fluids are not readily available in many cases, particularly in early stages of field development when such data are required for estimation of the in-place fluid volumes and/or design of enhanced recovery processes. For these reasons, the development of models for prediction of reservoir fluid properties, especially the bubble point pressure (Pb) and oil formation volume factor (Bo), has been an active area of research since decades ago. In this paper, we have used the effective machine learning approaches of Extremely Randomized Trees, also called Extra Trees (ET), Least Square Support Vector Machine (LSSVM) hybridized with the Coupled Simulated Annealing (CSA) algorithm, and Adaptive Network-based Fuzzy Inference System (ANFIS) to develop models for prediction of Bo and Pb. An extensive dataset of the essential reservoir and fluid properties comprised of initial solution gas-oil ratio (Rsi), gas specific gravity (γg), oil API gravity (API), and reservoir temperature (T), in addition to the target parameters, has been used to introduce the models. The main focus of this study is on the ET modelling method which we have used as a robust tool to predict the fluid properties of interest with high accuracy over wide ranges of the input features. The other two modelling techniques, i.e. LSSVM-CSA hybrid and ANFIS, have been used in this work to provide a basis for comparison and evaluation of the models. The predictions of the developed models greatly match the real data. Through detailed model evaluation and error analysis, it is shown that the ET model outperforms the other models in terms of accuracy and robustness (with very small mean absolute percentage errors of 0.099% and 0.917% for Bo and Pb predictions, respectively, over the total dataset), although all the developed models exhibit strong performance and comparably accurate predictions. Moreover, the developed ET models show a strong correlation between the predicted data and the target values with R2 > 0.99 in almost all cases over the training, test, and total datasets. The findings of this study can help to better understand the relative importance of the different reservoir fluid parameters typically used for the prediction of Bo and Pb. Based on this work, with the ET approach, Pb and Rsi have the highest importance in determining Bo and Pb, respectively, while temperature has the lowest significance in both cases. To the best of our knowledge, this work is the first Bo and Pb modelling attempt using the ET method. Additionally, we have applied the feature selection capability of the tree-based estimators to develop a new ET model for prediction of Pb using only two parameters (Rsi and γg); an outcome that is of particular importance in case of the lack of data on other parameters which are required in almost all the other Pb predictive tools and correlations reported in the literature.