Modeling of Oil Viscosity for Southern Iraqi Reservoirs using Neural Network Method

The calculation of the oil density is more complex due to a wide range of pressuresand temperatures, which are always determined by specific conditions, pressure andtemperature. Therefore, the calculations that depend on oil components are moreaccurate and easier in finding such kind of requirements. The analyses of twenty liveoil samples are utilized. The three parameters Peng Robinson equation of state istuned to get match between measured and calculated oil viscosity. The Lohrenz-Bray-Clark (LBC) viscosity calculation technique is adopted to calculate the viscosity of oilfrom the given composition, pressure and temperature for 20 samples. The tunedequation of state is used to generate oil viscosity values for a range of temperature andpressure extends from the reservoir to surface conditions.The generated viscosity data is utilized in the neural network tool (NN) to get fittingmodel correlates the viscosity of oil with composition, pressure and temperature. Theresulted error and the correlation coefficient of the model constructed are close to 0and 1 respectively. The NN model is also tested with data that are not used in set upthe model. The results proved the validity of the model. Moreover, the model’soutcomes demonstrate its superiority to selected empirical correlations.


Introduction:
Viscosity is a very important parameter that governs the flow of fluids either in a porous media or through transporting pipes. Sometimes, the measurement of oil viscosity is costly especially when the oil has dissolved gas. Therefore, a number of correlations [2][3][4][5][6][7][8][9] have been developed to provide alternative tool for getting the viscosity of specific hydrocarbon at the certain conditions. The most common correlation is the Lohrenz-Bray-Clark [1] (LBC). LBC is an efficient tool for viscosity estimation for the operating time management. Therefore, it is usually used in reservoir simulation where the execution time is a principal factor.
In this work, 20 live sample analyses are implemented. The data has been provided as The above-mentioned data is used to tune an equation of state using using the principle of the corresponding state for computing the oil viscosity. The viscosity is calculated by LBC correlation, which is then used to create viscosity data used in developing the model. The oil viscosity model is the ultimate target of the current project. A nonparametric regression method is elected to correlate the data. The neural network [10] technique is selected for creating the model.

Viscosity matching:
In this work, the composition of 20 Iraqi oil samples is considered. The samples were taken from several southern Iraqi oil fields, reservoirs, and wells. Before estimating the oil viscosity from its composition, temperature, and pressure, the experimental data should be fitted with the results of the equation of state. In the tuning of the equation of state, its variables are adjusted well to make its results close enough to the experimental test. The used oil properties in tuning phase are the saturation pressure and oil properties at pressure range from the atmospheric pressure to the initial reservoir pressure. The good match between real PVT data and calculated data is necessary to generate a large set of viscosity values in different temperatures.
This set of values is used in neural network fitting tool to get a good viscosity correlation.
Matching is done with PVTi software by doing regression for the equation of states parameters such as critical pressure, critical temperature, and the acentric factor for all components. Matching between calculated and real data is tested by the root mean square (RMS) of these values. The smallest RMS is found for the best match. Table   These values are used in neural network function to get the viscosity model.

Viscosity Model
The nonparametric regression tool, neural network, is proved to produce good fitting models even if the network is a simple one. Consequently, this technique is adopted in the current work. Neural network function needs 3 sets of data; training data, validation data, and test data. Training data are the set of data that would be used to been made with Bayesian Regularization [12,13] back propagation, which takes longer time but may be better for challenging problems, like existence of many independent variables and their large amount and differences of them .While the third model has been made with scaled conjugate gradient [14] back propagation, which uses less memory and so it is suitable in low memory situations.
The difference between the three neural network models is presented in Figures (2 to   10).  The training stops when the validation data reach the best performance. Figure (9) shows that in model 2, no validation data has been inserted; that means, the training has stopped when the maximum epochs have reached the maximum default number of epochs (1000) and the improvement still continuous. However, it is concluded that few epochs are required to reach the state of no improvement since the MSE approaches to a very small value.

Viscosity Comparison:
Viscosity of three wells could be represented in this work. One of these three wells wasn't inserted in the neural network while developing the model (the one unnamed well). The viscosity has been calculated with the three models and shown in tables (2 to 4) with the root mean square error (RMS).
Data of the same mentioned three wells has been inserted in PVTp software to calculate their viscosity with three other empirical correlations which are Beal et al.
[2], Beggs [3], Petrosky [9]. The results of these three correlations have been compared with the result of the three models and shown in Tables (2 to 4).
In Tables 2 to 4      2. The calculated viscosity was found very close to the measured viscosity due to the perfect match exhibited by the new model.