Estimating oil\u2013gas ratio for volatile oil and gas condensate reservoirs: artificial neural network, support vector machines and functional network approach

Abstract

Laboratory experiment ideally, is the main method to obtain PVT properties of the oil and gas reservoir fluids. The alternative two methods widely used when laboratory experiments are not available are: equation of state (EOS) and empirical PVT correlations. The EOS requires lots of numerical computations based on identifying the full compositions of the reservoir fluids properties. The measurement and calculation of these properties are very expensive and time consuming. On the other hand, using of PVT correlations which are based on easily measured field data such as reservoir pressure and temperature, and gas and oil density is reliable and more economic. In this work, three artificial intelligence (AI) technique models were developed to predict the oil–gas ratio (Rv) for volatile oil and gas condensate reservoirs. Thirteen actual reservoir fluid samples (five volatile oils and eight gas condensates) covering a wide range of fluid behavior and characteristics were used. Whitson and Torp three parameters EOS were used to generate modified black oil (MBO) PVT properties that were used as a data set for model development. The MBO PVT data points were extracted for each sample using commercial PVT software at five different separator conditions. The nature of the input data was studied showing that data type is clustered. In addition, the correlations between the input parameters were checked. This preprocessed is helpful in selecting the best method to deal with the input parameters that will be fed to the developed models. According to this analysis and since the input parameters have different ranges, normalization of these parameters is vital to improve the accuracy of the models and to get the solution quickly and efficiently. Results showed that taking the log for the input parameters is the best among the other normalization techniques. The AI techniques that have been implemented in this research are; Artificial Neural Network (ANN) models, Functional Networks (FNs) and Support Vector Machines (SVMs). Models developed based on these techniques used 17,941 data points and a ratio of 70% for training, 15% for validation, and 15% for testing. To develop these models, three Matlab codes were written for each tool where the provided input data in excel format were read and prepressed before implementation. Results obtained using these techniques showed that the ANN model predicted Rv with an average square correlation coefficient of 0.9999 and an average relative error of 0.15% while FNs predicted Rv with an average correlation coefficient of 0.9635 and an average relative error of 27.6%. It was noted that SVMs gave the best results with an average correlation coefficient of 0.9990 and an average relative error of 0.12%. The results concluded that ANN and SVM accurately predicted such data since this type of data are clustered and these two models can handle this kind of data. The newly developed models depend only on easily obtainable parameters in the field and can have varied applications when typical lab PVT reports are not available.