Experimental measurement and compositional modeling of crude oil viscosity at reservoir conditions

Abstract

In the present study, experimental and modeling investigations were performed and combined to implement trustworthy paradigms to predict the viscosity value under different circumstances and a wide variety of conditions. The experimental approach was conducted on a considerable number of Iranian crude samples using a Rolling Ball viscometer. Accordingly, more than 1000 experimental points were gained. These latter were utilized as a databank in the modeling approach which included many advanced soft computing techniques, namely radial basis function (RBF) neural network, multilayer perceptron (MLP), support vector regression (SVR), adaptive neuro-fuzzy inference system (ANFIS), decision trees (DTs) and random forest (RF). When performing the modeling tasks using these techniques, two distinct cases were considered: the first includes all available parameters as inputs such as pressure, temperature, API°, Mw of C12+ and the mole fractions till C11; whereas in the second case, a grouping scheme was considered to reduce the number of fractions. The obtained results revealed that DTs for the first case is the best implemented model with an overall average absolute relative deviation (AARD) of 3.379%. In addition, the comparison results with the preexisting approaches showed the superiority of the newly proposed model.