Page 256 - PVT Property Correlations

Chapter 10

Artificial Neural Network

Models for PVT Properties

In the last few decades, artificial neural network (ANN) models have been
used in many applications in the petroleum industry (Mohaghegh et al.,
2011; Cranganu et al., 2015). ANN models have been developed for PVT
properties prediction; model recognition in well test analysis; identification
of faults in sucker-rod pumping systems; log response estimation; reservoir
rock properties prediction; multiphase flow pressure drop calculation; identi-
fication of infill drilling locations in unconventional reservoirs; selection of
optimum hydraulic fracturing parameters, and many more applications.
ANN applications in prediction of PVT properties have received signifi-
cant attention in recent years. In several investigations, the capabilities of
ANN have been exploited to predict PVT properties using different data sets
collected from PVT reports and/or the literature. As with the PVT correla-
tions, ANN models can predict the PVT properties within the limits of the
data used to develop them.

BASICS OF ARTIFICIAL NEURAL NETWORK MODELS

Neural network calculations use two calculation passes: (1) forward pass for
prediction purposes and (2) backward pass for training purposes. The ANN
model structure consists of several components: (1) input nodes; (2) hidden
node(s); (3) activation function for the hidden node(s); (4) transformation
function for the output node(s); (5) output node(s); (6) objective function;
(7) optimization algorithm; and (8) training algorithm.
Fig. 10.1 illustrates the eight components of the neural network and the
internal communication paths between components. X1, X2, X3, X4, ...,Xn
represent the input values, while a1, a2, a3, a4, ...,an represent the output
values of the first hidden layer. Weights w1,1; w1,2; ...;w1,n; w2,1; w2,2;
...; w2,n; ... are assigned for the connections between each input node and
the first hidden layer nodes. The output values of the first hidden layer are
usually calculated as the summation of the connected input node values with
each multiplied times its weight. The values of the first hidden layer are then
treated with activation function (function “F”). h1, h2, h3, h4, ...,hn values
represent the output of the activation function for the first hidden layer. The

PVT Property Correlations. DOI: https://doi.org/10.1016/B978-0-12-812572-4.00010-2
© 2018 Elsevier Inc. All rights reserved. 225