Artificial Neural Networks Models for Predicting PVT Properties of Oil Field Brines

摘要

Knowledge of chemical and physical properties of formationrnwater is very important in various reservoir engineeringrncomputations especially in water flooding and production.rnIdeally, those data should be obtained experimentally. Onrnsome occasions, these data are not either available or reliable;rnthen, empirically derived correlations are used to predict brinernPVT properties. These correlations offer a handy and anrnacceptable approximation of formation water properties.rnHowever, the success of such correlations in predictionrndepends mainly on the range of data at which they werernoriginally developed. These correlations were developedrnusing linear, non-linear, multiple regression or graphicalrntechniques.rnRecently, researchers utilized artificial neural networksrn(ANN) to develop more accurate oil PVT correlations. Therndeveloped models outperformed the existing correlations.rnHowever, there is no similar research done so far to utilize thernpower of ANN in developing similar models for formationrnwaters. In the present study, two new models were developedrnto predict different brine properties. The first model predictsrnbrine density, formation volume factor (FVF), and isothermalrncompressibility as a function of pressure, temperature andrnsalinity. The second model is developed to predict brinernviscosity as a function of temperature and salinity only. Anrnattempt was made to develop a comprehensive model tornpredict all properties in terms of pressure, temperature andrnsalinity. The results were satisfactory for all other propertiesrnexcept for viscosity. This was attributed to the fact thatrnviscosity depends only on temperature and salinity. Thernmodels were developed using 1040 published data sets. Theserndata were divided into three groups: training, cross-validationrnand testing. Radial Basis Functions (RBF) and Multi-layerrnPreceptor (MLP) neural networks were utilized in this study.rnTrend tests were performed to ensure that the developedrnmodel would follow the physical laws. Results show that therndeveloped models outperform the published correlations inrnterms of absolute average percent relative error, correlationrncoefficient and standard deviation.