Please use this identifier to cite or link to this item: http://hdl.handle.net/20.500.12323/4676
Title: Estimation of PVT Properties Using Artificial Neural Networks and Comparison of Results with Experimental Data
Authors: Mehrizadeh, Masoud
Keywords: neuron
neural network
error backpropagation
neural network application
PVT properties
supersaturated reservoirs
Issue Date: 2020
Publisher: Khazar University Press
Citation: Khazar Journal of Science and Technology
Series/Report no.: Vol. 4;№ 1
Abstract: The importance of pressure, volume and temperature (PVT) properties, bubble pressure, gas-to-oil solubility ratio and oil volume factor have made it necessary to precisely determine these properties for the calculation of reservoir performance. In the absence of laboratory measurements to determine the PVT properties of crude oil, two methods, used commonly, include the equations of state and the experimental relations of PVT. The equation of state is based on the information concerned with the fluid composition details of the reservoir, which is very expensive and time-consuming to determine. Whereas, PVT relationships are based on data obtained from easily measured ground layers. These data include reservoir pressure, reservoir temperature, and the specific gravity of oil and gas. Recent Studies show that artificial neural networks have a great ability to predict the PVT properties. Due to the training capability in neural networks, these networks were rapidly applied in engineering and were widely used in petroleum engineering. Estimation of porosity and permeability of reservoirs, prediction of outflow generated by oil wells, estimation of oil recovery, prediction of asphaltene deposition and estimation of PVT properties are the most important applications of artificial neural networks in petroleum engineering. By preparing and collecting more than 1000 PVT data related to southern Iran reservoirs, 577 data were selected to be used in the project and were randomly divided into two parts, 486 data for network training and 91 data for testing. The three-layer structure (one hidden layer with 6 neurons) was selected as the best structure and the batch backpropagation training method as the best learning algorithm. The results of the network were in a good agreement with experimental data, which the average relative error using training set in estimation of the volume factor and densities of oil were 0.557 and 0.509% respectively and using test data were 1.032 and 1.104% respectively.
URI: http://hdl.handle.net/20.500.12323/4676
ISSN: 2520-6133
Appears in Collections:2020, Vol. 4, № 1



Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.