Comparing classical and neural regression techniques in modeling crude oil viscosity

Abstract

The importance of crude oil viscosity makes its accurate determination necessary for reservoir performance calculations, evaluation of hydrocarbon reserves, planning thermal methods of enhanced oil recovery, and designing production equipment and pipelines. Viscosity data are also involved in several dimensionless parameters to calculate flow regimes, friction factors and pressure gradients in multiphase flow problems. Numerous research efforts have been directed towards the development of viscosity models that are capable of accurately predicting crude oil viscosity as a function of production data, and/or composition of well stream fluids, if available, using equation of State. Since fluid compositions are not always available, most of the efforts were focused on developing viscosity correlations using classical regression techniques.The study presents, for the first time, a comparison among several models developed using both classical regression techniques (CRT) and neural regression techniques (NRT). These models are developed in this study from viscosity data collected from different oil fields. The models have also been tested using another collection of viscosity data that was not used before in the development phase. Results show that viscosity models developed using NRT were more accurate than viscosity models developed using CRT. Based on this comparison, a viscosity model is therefore presented, which uses stock-tank oil API gravity, gas gravity, pressure(s), and temperature(s) to predict crude oil viscosity. The model was developed using General Regression Neural Network algorithm.