Abstract
Oil viscosity is used in any fluid transport calculation in both subsurface and surface conditions.It is possible to determine oil viscosity from laboratory measurements or empirical correlations.However,laboratory measurements are not always available;and empirical correlations suffer from low accuracy.This work implements data mining algorithms to suggest a new correlation for oil viscosity calculation in a wide pressure range from subsurface to surface conditions.First,a scatter plot matrix is applied to analyze 1950 PVT experimental data points from Iranian oil reservoirs.Therefore,the most correlated parameters for predicting oil viscosity are determined.Next,75% of data points are randomly selected to train the models.The remaining data,i.e.,25% of data points,are utilized to investigate the accuracy of the developed correlation.Then,a symbolic regression analysis is performed in all pressure ranges,i.e.,dead oil viscosity,bubble point oil viscosity,below and above the bubble point pressure.Finally,a new oil viscosity correlation is proposed.The statistical and graphical evaluations reveal that the new correlation outperformed the previously proposed correlations by lowering average absolute errors.It can be concluded that the presented correlation improves the prediction accuracy in all pressure ranges.Consequently,it is inferred from the results that machine learning could provide a highly accurate prediction for oil viscosity in all pressure regions.Overall,the proposed correlation could be used to calculate oil viscosity in all pressure ranges with reasonable accuracy.
NSTL