Page 260 - PVT Property Correlations

Artificial Neural Network Models for PVT Properties Chapter | 10 229

ARTIFICIAL NEURAL NETWORK MODELS FOR PREDICTING
PVT PROPERTIES

Many attempts to predict PVT properties using ANN models are documented
in the petroleum literature. Investigators tried different ANN models to predict
properties such as equilibrium constants (K values), oil PVT properties, dew
point pressure, and gas z-factor. Tables 10.1a and 10.1b summarize many of
the attempts at predicting different PVT properties using ANN models.
Habiballah et al. (1996) used neural network with back propagation and
scaled conjugate gradient optimization algorithm to predict K values for
fourteen multicomponent hydrocarbon mixtures. They used a database of
more than 3000 points and constructed two ANN to predict two parameters
that can be used to calculate the K value. The input nodes to the two ANN
models included the values of pressure and temperature. The two networks
had two hidden layers each.
Gharbi and Elsharkawy (1997) developed neural-network-based models
for the prediction of PVT properties of crude oils from the Middle East.
They used back propagation technique with momentum. Their dataset con-
tained nearly 500 experimentally obtained data of bubble point pressures and
oil formation volume factors at the bubble point.
Elsharkawy (1998) introduced ANN models to calculate formation vol-
ume factor, solution gas oil ratio, oil viscosity, saturated oil density, under-
saturated oil compressibility, and evolved gas gravity. Elsharkawy used back
propagation technique with radial basis transformation function. Four input
nodes were assumed for reservoir pressure, temperature, stock tank oil grav-
ity, and separator gas gravity. The accuracy of the proposed network models
to predict PVT properties for crude oils and gas systems was compared to
that of numerous published PVT correlations. The ANN models were found
to be superior.
Elsharkawy and Gharbi (2001) compared neural network models with
models constructed according to classical regression techniques in order to
estimate oil viscosity using crude oils from Kuwait. Results of their work
show that viscosity models developed using ANN were more accurate than
viscosity models developed using regression techniques. Their ANN model
had four input nodes, one hidden layer, and one output layer Training and
testing data were around 700 points.
Hajizadeh (2007) introduced new models for viscosity prediction using
back propagation neural network. He used a genetic algorithm technique to
optimize the weights of the neural network. The dataset consisted of 89 sam-
ples from Iranian oil PVT reports.
Dutta and Gupta (2010) expanded the work on Hajizadeh (2007) and
developed ANN models for estimating bubble point pressure, solution
gas oil ratio, saturated and undersaturated oil formation volume factor, and
saturated and undersaturated oil viscosity for Indian crude oils. They used
back propagation technique with genetic algorithm for optimization.