Artificial neural networks models for predicting PVT properties of oil field brines

Knowledge of chemical and physical properties of formation water is very important in various reservoir engineering computations especially in water flooding and production. Ideally, those data should be obtained experimentally. On some occasions, these data are not either available or reliable; then, empirically derived correlations are used to predict brine PVT properties. These correlations offer a handy and an acceptable approximation of formation water properties. However, the success of such correlations in prediction depends mainly on the range of data at which they were originally developed. These correlations were developed using linear, non-linear, multiple regression or graphical techniques. Recently, researchers utilized artificial neural networks (ANN) to develop more accurate oil PVT correlations. The developed models outperformed the existing correlations. However, there is no similar research done so far to utilize the power of ANN in developing similar models for formation waters. In the present study, two new models were developed to predict different brine properties. The first model predicts brine density, formation volume factor (FVF), and isothermal compressibility as a function of pressure, temperature and salinity. The second model is developed to predict brine viscosity as a function of temperature and salinity only. An attempt was made to develop a comprehensive model to predict all properties in terms of pressure, temperature and salinity. The results were satisfactory for all other properties except for viscosity. This was attributed to the fact that viscosity depends only on temperature and salinity. The models were developed using 1040 published data sets. These data were divided into three groups: training, cross-validation and testing. Radial Basis Functions (RBF) and Multi-layer Preceptor (MLP) neural networks were utilized in this study. Trend tests were performed to ensure that the developed model would follow the physical laws. Results show that the developed models outperform the published correlations in terms of absolute average percent relative error, correlation coefficient and standard deviation.