This study presents a universal neural network based models for the prediction of PVT properties of crude oil samples obtained from all over the world. The data, on which the network was trained, contains 5200 experimentally obtained PVT data sets of different crude oil and gas mixtures from all over the world. They were collected from major producing oil fields in North and South America, North Sea, South East Asia, Middle East and Africa. This represents the largest data set ever collected to be used in developing PVT models. An additional 234 PVT data sets were used to investigate the effectiveness of the neural network models to predict outputs from inputs that were not used during the training process. The neural network model is able to predict the solution gas-oil-ratio and the oil formation-volume-factor as a function of the bubble-point pressure, the gas relative density, the oil specific gravity, and the reservoir temperature. The neural network models were developed using back propagation with momentum for error minimization to obtain the most accurate PVT models. A detailed comparison between the results predicted by the neural network models and those predicted by other correlations are presented for these crude oil samples. This study shows that artificial neural networks, once successfully trained, are excellent reliable predictive tools for estimating crude oil PVT properties better than available correlations. These neural network PVT models can be easily incorporated into reservoir simulators and production optimization software.


Empirical correlations for predicting reservoir fluid properties have been used in evaluating newly discovered formations, studying fluid recoveries, designing production equipment and surface facilities, planning future production and economics.

In 1949, Katz introduced the first correlation to predict oil formation volume factor for Mid-Continent US crudes. Since then, several correlations for the prediction of crude oil properties from various locations worldwide have been presented in the literature. Ideally, the PVT properties such as bubble point pressure, gas-oil ratio, and oil formation volume factor are measured on collected bottom hole samples or recombined surface samples. In some occasions, experimentally measured PVT data are not available because adequate samples cannot be obtained or the production horizon does not warrant the expense of detailed reservoir fluid studies. In these cases, field measured data such as reservoir pressure, temperature, crude oil API gravity and gas specific gravity are used to estimate the PVT properties using these empirical correlations. Local PVT correlations for a particular field or region can also be used to check the accuracy of the PVT report for a given crude from the same field or the region.

The accuracy of the well-known empirical PVT correlations such as Standing, Vasquez and Beggs, and Glas and the recently developed ones has been the subject of numerous studies. All these studies indicated that these correlations are not accurate to be generalized to predict crude oil properties from various locations. All the correlations mentioned above were developed using conventional regression methods, which may not give reliable results. Artificial neural networks, on the other hand, were shown to have excellent and reliable predictive capabilities. The objective of this paper is two fold:

  1. to develop a universal network model using large collection of crude oil properties representing oils from different oil fields in the world to predict the PVT properties of various crude oil systems;

  2. is to compare the accuracy of the neural network model to several published correlations.

In recent years, the application of artificial neural networks to petroleum engineering problems has been the subject of much study.

P. 619^

This content is only available via PDF.
You can access this article if you purchase or spend a download.