Application of Artificial Neural Network in Prediction of Fluid Properties of CO2-Decane System
J. Foroozesh and M. Seyyedi
Event name: IOR 2015 - 18th European Symposium on Improved Oil Recovery
Session: Poster Session
Publication date: 14 April 2015
Info: Extended abstract, PDF ( 1.31Mb )
Price: € 20
CO2-based injection methods have been favourably applying in oil fields due to simultaneous enhanced oil recovery (EOR) and CO2 storage potential. Various strategies of CO2 injection including miscible CO2 injection, water alternating gas (WAG) and carbonated water injection have been considered by petroleum engineers. During injection of CO2 into the oil reservoirs, CO2 migrates into the the oil phase which this results in higher oil recovery. This is mainly due to improvement of oil mobility by changing of oil properties including density and viscosity. Therefore, accurate calculation of fluid properties of mixture plays an important role in compositional simulation of CO2-EOR techniques. Equations of state (EOS) are normally used for description of fluid properties; however, CO2 fluid at usual reservoir conditions is at supercritical state and most available EOSs show poor capability in predicting the properties of fluids near or above the critical conditions. Artificial neural network (ANN) has shown a good potential for modelling the engineering systems. ANN is an effective algorithm which finds a relationship between input and output data during a learning process. ANN can be used to predict the fluid properties of supercritical CO2-oil mixture. In this study ANN is used to estimate the density and viscosity of CO2-decane mixture from 310 to 403 K and 7 to 30 MPa. The data are collected from the measured data available in the literature. The measured data are the density and viscosity of CO2-decane mixture as a function of temperature, pressure and CO2 mole fraction. Multilayer perceptron (MLP) network is used and trained to find a model which can accurately predict the density and viscosity of CO2-decane mixture at different pressures, temperatures and compositions. Mean squared error (MSE) is checked to evaluate the performance of the designed network. Results show that a trained ANN can predict the fluid properties of CO2-decane mixture properly. Furthermore, ANN is used and some density and viscosity data for CO2-decane mixture are generated. Moreover in this study the potential of ANN for the CO2-decane system is considered, however, it can be extended to study other real oil- CO2 systems if some experimental data be available.