Page 267 - PVT Property Correlations

Artificial Neural Network Models for PVT Properties Chapter | 10 233

Hemmati-Sarapardeh et al. (2014) analyzed a large database and evalu-
ated three of the most accurate correlations for each region for comparison
with ANN models. They used dead oil viscosity, viscosity below bubble
point, viscosity at bubble point, and undersaturated oil viscosity properties in
their comparison study. They developed four ANN models for Iranian oil
reservoirs.
Ahmadi et al. (2015) developed an ANN model for prediction of bubble
point pressure using back propagation technique. They used particle swarm
optimization (PSO) algorithm for error minimization in calculating the net-
work weights. Their dataset from the literature included 123 records from
Farasat et al. (2013).
Osman et al. (2001) presented an ANN model for prediction of bubble
point pressure and formation volume factor at the bubble point. They used
803 published data points from the Middle East, Malaysia, Colombia, and
Gulf of Mexico reservoirs. They used back propagation technique and one
hidden layer of five nodes. The inputs were the four parameters routinely
used for each network.
Al-Marhoun and Osman (2002) presented another study that included
new models to predict bubble point pressure and formation volume factor at
bubble point pressure. They used 293 data sets collected from Saudi reser-
voirs. The two ANN models used four inputs, one hidden layer with five
nodes, and one output node for each model. They compared their ANN mod-
els with published PVT correlations and concluded that the ANN models
were superior.
Gonza ´lez et al. (2003) used neural network models to predict dew point
pressures in retrograde gas reservoirs. Their model predicted dew point pres-
sure with accuracy of 8.74%. Their input layer nodes included data for tem-
perature, hydrocarbon, and nonhydrocarbon compositions, molecular weight,
and specific gravity of heptane plus fraction. The study used back propaga-
tion technique and one hidden layer.
Oloso et al. (2009) used a new approach for predicting a curve of oil vis-
cosity and solution gas oil ratio. In this approach, PVT properties can be
predicted over the entire pressure range (rather than at specified pressure
points). The models were built by use of support vector regression and func-
tional network with ANN. Smooth curves for the PVT properties were
produced.
Al-Shammasi (2001) conducted a study on neural network models for
estimation of bubble point pressure and oil formation volume factor at and
below bubble point using global data sets. The usual four inputs and two hid-
den layers were used. Performance of ANN models was compared with pub-
lished correlations for bubble point pressure and formation volume factor.
Alimadadi et al. (2011) used a committee machine type ANN to build
two models for oil formation volume factor and oil density. The input para-
meters were component mole %, solution gas oil ratio, bubble point