A Study And Application Of Support Vector Machine For Prediction Of PVT Properties Of Crude-oil

The PVT properties of crude-oil, such as bubble point pressure, formation volume factor, dissolved gas-oil ratio, reservoir temperature, crude-oil API gravity and gas relative density, played a key role in the calculation of reserves of the oil and gas reservoirs as well as for the identification of the reservoir characteristics. Typically, in order to determine the PVT properties of crude-oil, the samples which collected form cores or surface, are used to calculate the PVT properties which are difficult to obtain. However, this process spends a lot of money, and the costs are also high. Therefore, the researchers deduced many empirical formulas which are used to calculate the PVT properties of crude-oil. At present, there are a lot of empirical formulas in the oil and gas industry. Because of the traditional experience formulas are only suitable for specific property reservoir. Therefore, the empirical formulas are not suitable for calculating of PVT properties of all types of crude-oil.Artificial Neural Networks (ANN) can learn the model of data by self-adjusting its parameters. After trained, ANN can accurately match with the expected data, and it can be quickly and accurately predict the unknown sample data. It can be said that one of the most important feature of ANN is its ability to discovery the rules and patterns in the data which the general observation and some standard statistical methods can not find. Especially, when dealing with the nonlinear problems, this characteristics of ANN is obvious. Many researchers recognized that through the use of ANN, we can establish the appropriate model which can accurately predict the PVT properties of crude-oil in the petroleum engineering.However, ANN suffers from the complicated model structure, parameter selection difficult, prone to over-fit and low accuracy problems. But the Support Vector Machine (SVM) can solve the problems existed in ANN. SVM is one of the effective algorithms in machine learning and data mining. With the advantages of simple model structure selection, fast processing speed, and high learning precision, it is widely used in handling classification and regression problems. Therefore, based on these advantages of SVM, it is introduced in this paper to predict the PVT properties of crude-oil of China. Then it analyzes the basis theoretical of SVM, the statistical learning theory, the SVM algorithm and the theory of SVM for regression (SVR). At the same time, SVR algorithm is analyzed with the results of the BP algorithm in ANN for predicting the oil formation volume factor under bubble point pressure to validate the feasibility and effectiveness of SVR in application of obtaining the PVT properties of crude-oil in China. The experimental results show that the SVM model has a good prediction effect as well as better value of practical application. The designed SVR model can also accurately identify the oil formation volume factor under bubble point pressure in a smaller error.