An artificial neural network model to predict gas viscosity

Viscosity of gas mixtures is one of the most crucialgoverning parameters which affect oil and natural gasflow in reservoirs and exploitation equipment. Ideally,viscosity must be determined experimentally in thelaboratory on actual fluid samples. However, in theabsence of experimentally measured data, due to itsdifficulty or when invalid samples are available,parameters would be simulated by differentmathematical models. Existing models needexperimentally obtained gas components’ mole fraction.This study presents an artificial neural network modelwhich predicts the hydrocarbon and non-hydrocarbongases viscosity by only three input parameters:temperature, pressure and gas molecular weight. Theprediction procedure in this study was carried out using alarge data base containing ۲۴۴۵ experimental data in awide range of temperatures (۱۰۰-۴۶۰ °F), pressure (۱۴.۷-۱۰۰۰۰ psia) and molecular weights (۱۶.۰۴-۵۳.۹۲ Ib/Ibmole).A network with two hidden layers including ۱۲neurons in the first hidden layer and ۱۰ neurons in thesecond hidden layer shows the least error. Activationfunctions of this network are sigmoid function. Theaverage absolute relative error (AARE) of the networkfor test data is ۳.۲۷%. A comparison study between thismodel and five other models is done. Not only inputparameters of this new model are much easier to beobtained, but also the proposed artificial neural networkmodel can predict gas viscosity more accurately.

کلیدواژه ها:

Artificial Neural networks ، gas viscosity ، Empirical correlation