Sunteți pe pagina 1din 12

Societyof PetroleufnEn@reere

Fractured Reservoir Characterization


A.+;f;e;-l
biilblai
Al

A. Ouenes, SPE, S. Richardson, W.W.

Copyright

and Performance Forecasting Using Geomechanics and

ln+nll;-am~n

II ILGIII~CIIQC

1995, Society of Petroleum

Weiss, SPE, New Mexico Petroleum Recovery Research Center

rely on a qualitative description of the fractures. This is


achieved by using mainly structure properties,l2 seismic
velocity anisotropy observed with shear or S-waves, and
more recently compression or P-waves.4 However, a reservoir engineer struggling to numerically simulate a fractured reservoir needs more than just the location of sweet
spots.

Engineers, Inc.

This paper was prepared for presentation at the SPE Annual Technical Conference & Exhibition held in Dallas, TX, u.S.A., 22-25, October 1995.
This paper was selected fw presentation by an SPE Program Committee following review
Contents of the paof infmmation contained in n batract submitted by the author(s).
per: s presented, have not been reviewed by the Society of Petroleum Engineers and are
uthor( s). The material, as. presented, does not necessarilyy
subjected to correction by the
reflect any position of the Society of Petroleum Engineers, Its officers, or members. Papers
presented at SPE meetings re subject to publication review by Editorial Committees of the
Society of Petroleum Enqmeers. Permission to co
is restricted to an abstract of not mom
than 300 words. Illustrations ma not be copied. %
bstractshwcldcontain compicuous
acknowledgement of where nd ~ whom the pape, was Presented. Write Librarian, SPE,
P.O. Box 833836, Richardson, TX 75083, U. S. A., fax 01-214-952-9435.

The objective of this paper is to provide a reservoir


description methodology that leads to a computer input
file for a fractured reservoir simulator which can be used
for ncwfnrma.ncc=
Thic
mmthnrlnlnuv
r-licm
nm
r -.----.-.
ww ~Q~~c@@.
- . ..
. ..- . . ...--.
~
. v..
..

Abstract
A new approach in fractured reservoir characterization

the use of geomechanical concepts derived from reservoir


structure and artificial intelligence (AI) tools.

and simulation that integrates geomechanics, geology, and


reservoir engineering is proposed and illustrated with actual oil reservoirs. This approach uses a neural network
to find the relationship between, reservoir structure, bed
thickness and the well performance used as an indicator
of fracture intensity. Once the relation established, the
neural network can be used to forecast primary production, or for mapping the reservoir fracture intensity. The
resulting fracture intensity distribution can be used to represent the subsurface fracture network. Using the fracture
intensity map and fracture network, directional fracture
permeabilities and fracture pore volume can be estimated
via a history matching process where only two parameters
are adjusted.

Neural

networks

During the past few years, the petroleum industry enthusiastically supported the concept of integrated systems.
Integration of everything is everywhere. Ilom a reservoir
engineering point of view, the concept of integration is
a necessity not fashion. The necessity exists because of
the scarcity of reservoir information and the wide range of
scales over which this information is measured. Therefore,
a reliable reservoir description must somehow integrate all
the existing information at all the scales. The application
of stochastic global optimization methods, e.g. simulated
annealing, in reservoir description provided new tools for
achieving a certain level of integration. However, stochastic global optimization methods were developed in an artificial intelligence context and are more than just simple
mathematical optimization methods, as believed by some
users. Within the artificial intelligence framework, other
tools exist and can be used to integrate various information into a complex reservoir model. The most practical of
these integration tools can be found in neurocomputing.

Introduction
Conventional reservoir simulation has benefited from important research during the last few years.
The use
of geostatistics is slowly moving from the production of
grayscale maps with dubious value and multi-million cell
reahzations with little practical value to useful input data
for reservoir simulators. Although there is still much to be
done before these geostatistical reaJzations will be able to
reproduce the past performance of a reservoir, the recent
trend shows clearly that major advances have been made
in conventional reservoir description. On the other hand,
naturally fractured reservoir (NFR) characterization has
not enjoyed a similar benefit from any major research effort. Until this work, there is no quantitative methodology
to fill the NFR simulator gridblock

There are various ways of looking at a neural network.


The most common application is a pattern recognition tool
where from a given amount of known information, a neural network is able to be trained to recognize some patterns. In this case, the output of the neural network is
very often a binary variable where a value O means NO,
and a value 1 means YES. Such tools are routinely used
by financial institutions to decide on the outcome of our

Most of the current fractured reservoir characterization

425

.
FRACTURED
FORECASTING

RESERVOIR CHARACTERIZATION
AND PERFORMANCE
USING GEOMECHANICS AND ARTIFICIAL INTELLIGENCE

A neural network

credit card or mortgage application. Scientists and engineers found other uses for neural networks, the most com%on is the equation maker. In this case, the scientist or
engineer may have experimental or field data for a given
system with N inputs, (zi Ii = 1,2, ., N) and M outputs,
c.-A
-,,+-,,+.
omm he
(~vj~j = ~,~,...
h,f} n-tk +h- ;=-,,+.
,
lVL

J .

l-JUblA

Ui&G

LILyUUO

-AU

UL4Uy

LL..

b-.

rameters tiecting the fracture intensity. Within the same


philosophy, Lisle2 also uses the product of two main curvatures (Gaussian curvature) to detect fractures. However,
the bed thickness and lithology have an important impact
on rock fracturing, and therefore need to be included in a
quantitative fracture model. With at least three major parameters (structure curvatures, thickness, and lithelogy),
it becomes impossible for humans to find a fracture modei
that may correlate all these parameters to fracture intensity as indicated by log data or well production. In this
situation, a neural net would be a perfect tool for finding
the fracture model able to integrate all parameters affecting rock fracturing.

Suppose we have a system with N inputs (z~Ii =


1,2, -.. ,NJ,
- and M- modei outputs (yjlj = 1,2,.. ., M).
Additionally, we consider that there are L training data
records, (zil,yj~li
= 1,2, .o, N;j
= 1,2, O, M;1 =
1,2,... , L), available. The backprop neural net uses the
L training patterns to adjust the weight matrix, W~i~j
through an optimization procedure called training. For
the lth training pattern, the N inputs xil are fed-forward
from the input layer, through all the hidden layers, and
Which
finally the neural net provides its output Yjl (~),
is different from the known target output Yjl measured
in
the laboratory or field. The training process consists of
estimating the weights w~i ~j that minimize the quadratic
error:

[YjJ-~l(W)]2

. . . . . . . . . . . . . . . . . . .

characterization

In our early work,10 we estimated fracture intensity


distribution in a carbonate reservoir using a generalized
curvature method. The major idea behind this early work
is to construct a quantitative fracture model using structural properties. Extending Murraysl original idea of using structure curvature, we used two curvatures CZ= and
Cvy to derive a fracture model. The determination of this
fracture model was time-consuming for a 30 well reser.--:.
lX...&,.,...--.,. ,.- .. UUL
n... ...m+
.. . ..l.. .. . . ..,-.-: JAA..
AAa r..
vuu.
r UI IJXI IIN1 e, dy
VGUULm w CLc uuumum
GUI a= pa-

The process of determining these mathematical models


is described as training. The basic idea is to provide to
the neural net L training records (patterns) that will direct
the adjustment of the neural net parameters represented
hv
matrix W..,, n, . The reader is referred to neural
-J -a wei~ht
.. --m--- --------nets literature78 for a complete description of these AI
tools. The following paragraph gives a brief description
necessary for understanding the next section. Furthermore, we will address only backpropagation (or backprop)
neural networks which we use for the fracture mapping
application.

1=1

for fracture

In reservoir description, a neural network can be viewed


as an integration tool where various inputs of different nature and scale can be correlated to some crucial reservoir
property. In this framework, the neural net is more than a
crude (black iox,!~ it is a practical tooi to integrate various
types of data into a reservoir model. A recent example of
this good integration ability can be found in the porosity
estimation using seismic attributes.g With this integration
ability in mind, we consider a naturally fractured oil or gas
reservoir and try to characterize the areal distribution of
fracture intensity.

measured in the laboratory or field. The neural network


can be used to find a mathematical model that relates the
N inputs, z, to the M outputs, y. In other words, the neural network becomes a multivariate regression tool. However, the advantage of the neural network model over conventional multivariate regression methods is its ability to
mimic complex non-linear models. This ability has been
illustrated by Ouenes et aL6 where the complex partial
differential equations describing two-phase flow in porous
media could be reproduced with a neural network.

~(~)
=ff

SPE 30572

Neural network outputs. As described earlier, a neural


network is a mathematical model representing a relationship between a set of inputs x and a set of outputs y.
Because we are interested in describing fracture intensity,
the question that arises is what is the best way to describe
fracture intensity, at reservoir simulator scale? Additionally, we would like to develop a fracture mapping approach
that will not create another problem, i.e. upscaling. We
have found that at the reservoir simulator scale, the best

...(l)

j=l

., -J tn
v A-sr-rih-----

wav

fract.m~
.----

imkPnsit,v
-.------J

is tn
., --- nrcdnction
~_- -.. -..
-J.- use
--- well

data. These data can either be the maximum oil rate in


the early life of the well, the initial potential (IP) of the
well, or the cumulative production at a given time in the
life of the reservoir. This well performance is strongly dependent on the fracture network around the wellbore and
provides information at a reasonable scale represented by
the drainage radius. However, when choosing well performance as an indicator of fracture intensity, one may
have also to take reservoir pressure into account which
also affects the well performance. Since reservoir pressure
is rarely measured, we use the field cumulative oil pro-

The error E backpropagates from the output layer to


the input layer and is used to adjust the matrix weight
Wn,nj. We use a hybrid optimization algorithm6 that
combines simulated annealing and gradient methods to
estimate and adjust the weight matrix.
After successful training, the neural network can be
used for testing the outcome of some inputs not included
in the training process. There are various practical SSpects related to the training and testing stages which will
be described within the fractured reservoir application.

426

.
A. OUENES, S. RICHARDSON,

SPE 30572

duction to account for the change in pressure. In other


words, for each new well drilled during the development
of a field, we compute the total cumulative oil production
of the entire field and use this number as an indication of
the reservoir pressure at the time the new well was drilled.
This reservoir pressure indicator is added as an input to
the neural network because it affects the output which is
the well performance. In addition to this undesirable
input which is not related to fractures, the neural network
will need other inputs that are related to rock fracturing.
nl-. -l
lx euriu

-..--1.

Ile Lwurli

Z..

A-

lIlp ULS.

TL..

:--..4.

c,...&l.,. -,....1

..,.4

1 lle lup u m 1(J1 I,llC Iltzul cm llC b-

work must be the parameters which are related to fracture


intensity. We consider three major factors affecting the
rock fracturing: reservoir structure, bed thickness, and
lithology. These factors must be represented by one or
more quantitative variables that should be available over
the entire area of the reservoir.
Evaluating reservoir structure may seem an easy task.
However, this fractured reservoir description approach requires a detailed and accurate structure, not simply a
are
only
two
possibilihand-drawn contour
----.-- -rnZLp, There
ties for obtaining a meaningful structure: 3-D seismic or
interpolation using well log data. If 3-D seismic is available, then the major changes in the structure can be seen
clearly either on a structure obtained simply from the
travel timel 1 or after a time-depth conversion. At this
stage, it is important to realize that the most important
information needed in the structure map is the areal variation in depth, not the depth itself. Therefore, a time-depth
conversion may not dramatically change the structure contraat from one point to another.
Unfortunately, 3-D seismic is not available for all fractured reservoirs. In this case, one must rely on tops obtained from well logs and use a mapping method to generate the entire reservoir structure. With the effect of
structure on rock fracturing being accounted for through
curvatures and slopes at each location in the reservoir, it
is important to use a smooth mapping method. A noisy
structure will lead to extremely high curvatures (second
derivative of the tops) that are the result of the noise
and may have no relation to fracturing. A smooth mapping method such as kriging can be used provided that a
reasonable and meaningful variogram is available. After
trying various approaches to generate the reservoir structure, we have found a new approach that can be very useful
for mapping tops especially for folded structures.12
MS.
1 ms approach relies on the assumption that reservoir
depth, D, of a folded structure obeys the Laplace equation:
A2
.-
=

AZ
+~=o

n
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

(2)

In terms of curvatures, this assumption means that at


each location the curvature in the z direction, CC== ~,
and the curvature in the y direction, CYV= @

are equal

W.W. WEISS

and have opposite signs. This assum;}!ix can he seen in


many structures where hyperbolic (scddle; folded surfaces,
can be found. For reservoirs where most c$ th structure
is made of elliptical (dome) folded surfaces this assumption is not valid and another mapping approach must be
considered. If the observed folding appears to h{ hyperbolic, the Laplace equation can be used with the known
depths at the well locations to estimate the entire top of
the structure. We use a finite element method to solve the
Laplace equation and take advantage of the sparse matrix
(resuiting from the finite eiement approach! soivers to inTwin+.
;,.
2 Ffl. z! ..m-n?-lrl.
nrl
9
+r...mn,
+,+.-.VOaow,,n;v
C+v,,
#.+,,v* p.,.,..!.
I&lpwcmc
lGSCL
V. * 0..
Ue u.=
1.
u 1. .!... .,.,
U
,.
u
simple PC.
With the resulting structure map, or using seismic
travel-time,l 1 the discrete curvatures Cl~ and Cyu are
computed at each grid location. Howevex, for a complete
characterization of the structural changes, we have found
that cross curvatures C=v and Cy. taken in the d@pnal directions (Fig. 1) yield additional information and
therefore could be used as inputs for the neural network.
Furthermore, the slopes Sr = ~, SY = ~, and the cross
Aiacrnnal
rlirertion<
clnmoQtnlron
.pwu
w.../.. .in. . the
. . . . --------. . . -.. -..-, wh~n
.. ----- ~~~~~ M ~IE
puts, allow a faster and superior training of the neural
network. In summary, the geomechanical effect of structure on fracturing is taken into account by using the depth,
four slopes, and four curvatures.
The bed thickness is also another important parameter
to be included as an input. For a fractured reservoir where
bed thickness varies significantly, this input is crucial because it affects the neural network output in two ways.
When considering production as an indicator of fracture
intensity, a high oil or gas rate can reflect a very thick pay
zone, a highly fractured zone, or a very thick and highly
fractured zone. Therefore, a correct fracture model must
include thickness as an input. On the other hand, bed
thickness is intimately related to fracturing. In general,
very thick beds have less fractures than thin beds.13 Based
on these two considerations, bed thickness appears to be
an important input for fractured reservoirs. The problem
of obtaining bed thickness over the entire reservoir is similar to the interpolation of structure points. Mapping or
3-D seismic can be used, and the comments made for the
structure mapping also apply to thickness.
The last parameter that haa an impact on rock fracturing is lithology. Thk parameter would be an important
input to consider if reservoir lithology varies areally. If so,
geostatistics applied to an indicator variable could provide
a good estimation of the lithology distribution whkh then
can be included as input to the neural net. Given these
different inputs, the neural network architecture may now
be designed.
Neural network architecture.
A backprop net is a system that links a series of basic blocks called nodes. These
nodes, represented by a set NO = {nl, m, ... nn}, are
connected in a certain way to create the artiltecture (or

..

--,

,ni-m

rlvk
I Umcu
FORECASTING

m.crm
IAID
rscxnifvln

CUADAr7CD17ATlf_l~

Lnf-lnmL

, L,,, &n, ,.

USING GEOMECHANICS

- .nmim
,!sopp.1.~

topoiogy), whereas the Coniiedhl betmeefi tiVOfide~ n;


and nj is characterized by a weight W~i nj. Each node
sums the inputs multiplied by the weights, then applies a
nonlinear function (sigmoid function) to produce the node
output. The network architecture includes the input layer,
the oiitpiit layer, aiid a certain number of hidden !ayers
(Fig. 2). Each layer contains a finite number of nodes
connected to other layer nodes. These nodes can be represented in practice by a few lines of software or a few
electronic components in the hardware. The choice of this
titiltecture is important being that it directly tiects the
performance of the neural network and its ability to predict correct values.

. ..

m nf
.

ac.t~lal
-----

rcwmwnirs
. v.
, -.--,

~h~S

]~X~~~

CJQeS

IIOt

efit

Using the architecture shown in Fig. 2, three actual


oil reservoirs were considered for the application of this
new approach. The first unit, referred to as Unit A is
located in a large naturally fractured reservoir in Western Texas. The study included 16 sections area of thk
unit. For Unit A, the fracture intensity was represented
by the cumulative oil produced at each well for 78 wells
drilled prior to 1960. In Unit A, we are interested in
mapping the fracture intensity of an approximately 900ft
-I-l
.. ---~
L..-,. Cm -r, +:,w. ..,h;
t.h m mill
.~z-l.
wum sumey
SWIUSLAJUC lurlllGUJullw~UU.L
w ~ ... mofor
. ~.-. tn.= ac
Formation AA. For Unit A, little information was available at the time of the study. Besides the estimated individual cumulative production of 78 wells, the tops of
Formation AA and a 3-D seismic survey was available.
a.mra
ma Aonsit.v
p~~ &@CtiOll k
For ~TR~~ .4, +hn
s. u
. . .-~
--...
J of w~]i~

of this methodology.
Therefore, the proper architecture
was found by testing various models, and it appears that
a neural network with two hidden layers (Fig. 2) was sufficient for the modeling of the relation between the fracture
nravimlqlv
d~S~l~SS~(j= ~,ach IK)de
:-+a-.:+..
nrl +ho
ULUGAKJA
J w...ti imn,,+~
. ..~
y.

SPE 30572

INTELLIGENCE

When training a neural network with known data, it is


helpful to know the limiting value of the error E at which
the training could be considered sufficient. The training
process starts by assigning random weights which will lead
to an initial error Eini. At each iteration (epoch), the neural network adjusts the weights and the neural network
error, E, is slowly reduced. The training could be stopped
at a certain error Eji. and the obtained weights could
be considered as representative of the non-linear model
sought by the user. The training process could be stopped
when the neural network fracture intensity F1l reasonably
match the actual fracture intensity filfor all the L wells
used for the training. A bad match of the well data indicates insufficient training, and a very close match of all
the data may indicate over-training, which needs to be
avoided. The number of iterations required to obtain an
acceptable match will be discussed and illustrated with
three different fractured reservoirs.

However, in the fracture mapping application discussed


in this paper and its companion paper,l 1 the amount of
available data is not sufficient to find the optimal architecture in an objective manner. The lack of data is a con. . --.:---- dc,l,
. . ..+.
...1
..,.A...
rn: far
+hn
ammlioa+ifim
ual
1 Cmzk v Um
.1
lIC,
oppAG-...
Seq-uerlce
of Sdeumg

..

p~RFQRM~3N~~

and one may have to live with 25 to 30 training patterns


obtained from the very limited information available at
the wells. Although, 25 patterns may seem very small
number to train a neural network, we have found that it
is .sufficient
. ... . . . . to derive a fracture model.

There are two major aspects to consider when designing the architecture. First, complex non-linear models require more than one hidden layer, many nodes, and therefore many weights to estimate during the training process. Second, the training of a neural network is a multidimensional optimization problem very rich in local minima. Hence, an increase in the number of weights or node
connections reduces the chance of approaching a global
optimum required for reasonable training. In thk tradeoff between reducing the connections for better training,
and increasing the connections for better modeling of the
non-linear behavior, there are very limited objective tools
that can assist in the design of the optimal architecture.
Fuzzy logic combined with a special neural network can be
used to find the optimal architecture,6 if sufficient training
patterns are available.

~~~

AND ARTIFICIAL

---J

of the input layer is connected to all the nodes of the first


hidden layer. For 11 inputs, there are 121 connections between the input layer and the first hidden layer. It is very
-.-.4-.,... . -..1:--+--= .$*v
~nrl
irl INWid
IIW WUI k aPPtJU-&lu.JD
UL +ha
...= Arc+
... . . -.u
cOnimIOTI
second hidden layers to be fully connected, which means
121 additional connections. This common procedure is not
necessary and the user should first try one-to-one connections between the first and second hidden layer as shown
in Fig. 2. In the fracture mapping application, we have
found that the 11 connections between the first and the
second hidden layers were sufficient which means 110 (12111=1 10) less weights to estimate during the training.

WDA = 4.9 wells jsq.mile. Another unit, referred to aa


Unit B encompasses 22 wells that produced on primary
before water injection into formation AA was initiated.
a,ro;lmh]m
at
KS WdlQ
1r----+ :-- n-A A .-,--lO~S
Of tJK3 Lwiillamwll
WGLG

Neural network training


In most neural network applications, there is a wealth of

For the West Texas Unit B and New Mexico YDU, the
structure of the considered fractured formation was obtained by using the log data and the mapping procedure
that assumes the Laplace equation described in a previous

-v-au

S,

U.

u-

. ..

..-

cated in 4 sections. The average well density per section


in Unit B is WDB = 13.7 weUs/sq.mile. The third example is a carbonate reservoir located in southeast New
Mexico and referred to as the Young Deep Unit (YDU).
Our early work10 on fractured reservoir characterization
was applied to YDU. Production history and log data of
30 wells were available at the YDU whkh spans over an
area of about 2 sections. The weii density per section is
WDYDU = 15 wells/sq.mile.

data that can be used for training, and most probably a


reliable generalization of the neural network. In fracture

428

.
A. OUENES, S. RICHARDSON,

SPE 30572

For Unit B, data for 22 wells were available. Two sets


of 18 wells were used for training and each time 4 different
wells were removed from the training patterns. After completion of the training based on 18 wells, the performance
of the 4 wells removed from training are predicted. Figs.
7 and 8 show the actual 1P versus the predicted 1P. The
empty circles indicate the 18 wells used for the training,
and the filled circles are the predicted performance of the
4 wells for each data set. For both sets, the error on the
predicted performances varied between 11% and 50%. In
other words, the largest error was made when the neural
network predicted an IP=200, when the actual IP=400.
Considering the very small number of available wells used
for training, these predictions can be considered remarkable. Unfortunately, this prediction tool cannot be used
after fluids are injected in the fractured reservoir. The
change in pressure cannot be included easily in the neural network model. For this reason, we developed another
procedure to obtain a fracture intensity map that can be
used for reservoir simulation.

In such cases, the user may stop the training at a high


error El. Using the weights associated with this error, a
fracture intensity map can be obtained as described in the
next sections. If the training is continued to a lower error
.&, the new fracture intensity will be different. For the
YDU, Fig. 3 shows the fracture intensity map for an error
El = 0.1, and Fig. 4 shows the map for an error Ez = 0.08.
Although there are some local differences, the main trends
..
.:- .L
- --- lU1
r-. ALA---- -------:.--2 t- . ...*....IJIle Salle
IJlle IIWu U 1U1 a . me
(L11 Veu u Wb Lu e
1txuaul
network from the Fig. 3 fracture intensity map is shown
in Fig. 5 for the high error El. For the lowest error Ez
when using the fracture intensity map shown in the Fig.
4 and the weighting method ciescrilleci in the companion
paper,ll the fracture network obtained is shown in Fig.
6. Notice that despite the difference in the training error, the two fracture networks are almost the same. This
indicates that for both the errors, an acceptable training
was achieved. However, the subsurface fracture network
is unknown and other means must be used to evaluate
the quality of the training. A simple approach to verify if
generalization was achieved, is to test patterns which were
not included in the training process.

network

Primary performance
forecasting.
If we consider a set
of 22 wells with their known performance and use only 18
wells for the training, the testing process consists of predicting the performance of the 4 wells not included in the
training. By doing so, we are simply predicting the performance of an infill well. As an illustration, this testing
procedure is applied to Unit B.

section. Given the reasonable well density of Unit B and


YDU, an interpolation method provides a good estimation
of the structure. The fracture intensity at each well location was represented by the Idial Potential (1P) for Unit
B, and by the maximum oil rate for YDU. For both reservoirs, the correlation between neural net inputs and the
considered fracture intensity was distinctly evident. Thk
can be inferred because in less than 3000 iterations the
neural net converged to a very small error and achieved
very good training. Using the scarce well information in
r,. ,
urm A, an interpolated structure~~ was derived. When
using this structure map to create the neural network inputs, more than 5 million iterations were required to reach
an acceptable error. Thus, if a good structure map is available and the fracture intensity data are correct, the training will be short. For verification purposes, the neural
network used for Unit A was trained with a structure map
derived from 3-D seismicll and resulted in faster training.
Therefore, slow training indicates poor data quality.

Neural

W.W. WEISS

Mapping
the fracture
network. When choosing the
well performance as an indicator of fracture intensity,. we
include the cumulative oil production as an indicator of
----_.._:_ --------:- Al
-- :--..4T- -A L--.--Jr-- -.-.
,.1.
WUI us, lU1 Ixbul
L-esel-vull- pmssul-e
111 LJlt! Illpul)s . 111 Ullllel
new well drilled at a certain time, corresponds a reservoir pressure indicator. This ind:cator is simply the totxd
cumulative oil produced by all the previous wells. When
the neural network is used to predict the performance of
each gridblock at the initial reseruoir pressurv, we have a
tool that can drill fictitious wells, in each gridblock, at
the initial pressure. In thh situation, the predicted performance is only tiected by the fracture network. Therefore,
the predicted value represents the fracture intensity of the
considered gridblock. Applying the same process for all
the gridblock provides a complete description of the axeal
distribution of fracture intensity. Figs. 3 and 4 show the
fracture intensity map for the YDU and similar maps can
be found for Unit A and Unit B in the companion paper.ll
These fracture intensity maps can be further used to derive
a fracture network using the weighting method described
in the companion paper. Briefly speakhg, this method
uses at each gridblock the fracture intensity of the eight
neighboring gridblock to find the orientation of the fracture. When focusing only on the most intense fractures
that affect fluid flow in the reservoir, the fracture network
derived from Fig. 4 for YDU is shown in Fig. 9. ThE
fracture network clearly shows the shear fractures associated with a folded surface. This type of fractured reservoir

testing

Before using the neural network for any prediction, one


must confirm the ability of the neural network to reproduce correctly the non-linear model that may exist
between the various inputs and the fracture intensity.
Thk validation process is mandatory and allows the user
to evaluate the goodness of the obtained mathematical
model. In the application to fractured reservoirs, there
are two ways to test a neural network. The trained model
can be used to predict the performance of an infill well
(well deleted from the training data set), and/or obtain
the fracture intensity map over the entire reservoir.

429

.
FRACTURED
FORECASTING

RESERVOIR CHARACTERIZATION
AND PERFORMANCE
USING GEOMECHANICS AND ARTIFICIAL INTELLIGENCE

is the most complex and difficult to simulate for reasons


described in the next section.

Estimation
volume

of fracture

permeability

As discussed in the companion paper,ll we are focusing


on two types of fractures: regional and structure related.
In the regional fractures, two major fracture directions can
be seen by different means. The principal fractures along
the maximum horizontal stress direction, and another set
of fractures orthogonal to the major regional fractures. In
this situation, a reservoir grid that has its z-axis along
the regional fractures, and its y-axis in the perpendicular direction is sufficient to represent the fluid flow in the
fractures. Unfortunately, structure-related fractures are
not perpendicular and do not keep the same direction in
the field. Often these shear fractures, shown in Fig. 9,
form an angle of 60 degrees and are found along the folding. In other words, the fluid flow in the fractures can
occur in any direction.
When using a finite difference
reservoir simulator, only two major flow directions can
be tolerated. Therefore, the major difficulty in simulating structure-related fractured reservoirs is the inability
of the simulators to account for preferential flow paths in
any direction.

and pore

The objective of this paper is to develop a methodology to


fill reservoir simulator gridblock. This objective includes
history matching, which we view as an integral part of
reservoir characterization, and not a downstream application disconnected from the characterization effort. This
methodology relies on three components: 1) a fracture intensity map obtained with a neural network or using seismic attributes 1, 2) a fracture network derived from the
fracture map using the weighting methodl 1, and 3) an inverse modeling procedure using a reservoir simulator for
automatic history matching.
The basic idea behind this methodology was first described in earlier work10 where we assume that directional
permeabilities, KZ and KY, as well as fracture pore volume PV are a function of the fracture intensity, FI. The
functions considered are linear and simple. For example,
the permeability in the x direction is given by:
Kz=C1x

There are two solutions for this problem. The first solution, is to use a very fine grid, and the second solution is
to use full permeability tensors. Because we are primarily
concerned with full-field simulation that includes history
matching, we try to avoid a very fine arezd grid systein.
Hence, we prefer to deal with full permeability tensors.

FI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...(3)

and the pore volume of the fractures is given by:


PV=

C2XFI

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (4)
Diagonal
versus full permeability
tensors. ?vfost
reservoir simulation studies use a diagonal tensor for permeability where at most two directional permeabilities KZ
.~ v ... eoa;~arc~ Tha
z-.n~a-norrpaahility
~e~Ams.~=u
A ..e u. y-. . AU- . . . .

Where Cl and C2 are two constants to be estimated


during the history matching. Since the fracture intensity
is available over the entire reservoir, the estimation of fracture permeability K. and pore volume PV in each gridblock is reduced to the estimation of two simple parame+rm.a
v. C?.
w~

.and
...

(!.
VA.

SPE 30572

Thcmofnr~
A ... -.V.

V,

t.hiq
... -

annrnach
-rr.
--.,.

- t.riwa
----

to
.-

al~u ~.g -=

b A...ub&tiu.

and Kvz have been introduced by Leung17, White and


Horne18, and Samierlg for different applications. For fractured reservoir simulation, these cross-permeability terms
may be the best tools to represent the flow in all the directions as it occurs in a fold-related fracture system. Notice
at this stage that we did not discuss the dual-porosity or
single-porosity fracture representation. When concerned
with fluid flow in the fractures, the choice of dual or singlei+.r
i.
n
+ im nnr+
n n+
.Ho~~~~~,
~~~~
&Q&~
&~Q-m~~
-fiwfl
w..
I.Ju. .AUJ
..3 ..0 .,..y.

int.ema,te
-.- -o----

all the available field data to produce a fracture intensity


map which will ease the history matching task by reducing
the number of tuning parameters. Furthermore, geomechanical and geological constraints can be honored in the
fracture intensity map which will lead to a realistic reservoir model.
The estimation parameters Cl and C2 can be found
by manual history matching, or by using an automatic
history matching procedure, which will not encounter major difficulties, since the number of estimation parameters
is very small. The same philosophy of fractured reservoir characterization based on inverse modeling was used
by Long et aL14, Doughty et al.ls, and Datta Gupta et
al 16 to obtain fracture distributions.
However, in their
investigations, 14-16 the obtained fracture intensity map is
a result of a pure numerical inversion, and geomechanical and geological input were not imposed to the model.
In our work, the emphasis is given to the integration of
geomechanical, geological, and geophysicalll information
in the fracture model through a neural network. Despite,
thk integration effort there are other problems that need
to be addressed when simulating actual oil or gas fractured
reservoirs.

important when considering the fluid transfers between


matrix and fractures.
For illustration purposes, we will use a diagonal permeability tensor and a single-porosity model to simulate the
performance of the YDU reservoir. The purpose of this
choice is to illustrate the limited ability of a diagonal permeability tensor to reproduce correctly the fine details of
the production history at different wells. The advantages
of a full tensor will be illustrated in future papers.
The YDU reservoir model is described with two layers. The upper layer represents the matrix and the lower
layer represents the fractures. Each layer contains 759
gridblock (33 x 23) of 350 ft. The details concerning the
reservoir and its characteristics can be found in a previous
paper.10 Using the fracture intensity map shown in Fig. 4
and the Eqs. 3 and 4, the inverse model history-matched

430

A. OUENES, S. RICHARDSON,

SPE 30572

.
months of 29 weiis production history. in contrast
~/.r_m~h:l;t,r
~-;m,,+w,,m.,
$-m
+h,.,
+k - /.-. -1.,
..,/. -1. 10 +L,,
+a
LUG pa Lllcaulli
by alllclubl
Upy
lauI,w I,llG
IJ1lC Gal Ly WU1 A

tial Potential (1P), or cumulative production.

tor, which was considered constant over the field, was not
included in the estimation parameters. The permeability
in the g direction Ky was computed directly by using Kz
and an anisotropy map obtained from the fracture network
shown in Fig. 6.
At each gridblock, the fracture orientation was available. In terms of fluid flow, this fracture orientation indicates the direction of flow. Unfortunately, when using a
diagonal tensor there are only two possible flow directions.
In this situation we have only three possible permeability
anisotropy factors (Fig. 10). When the fracture direction
has an angle around O degrees we assume an anisotropy
factor AN = 0.01 and:

3. The geomechanical effects on fracturing can be represented by the slopes and curvatures of the reservoir
structure
4. Using a reliable structure and its geometrical properties, bed thickness, and cumulative production as
inputs, and well performance as output, a neural network can find easily the complex relation that exists
between the inputs and output.
5. Primary performance forecasting can be obtained
simply by testing the neural network.

XKZ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (5)

mL:._
-.-.. .:- .:11
iiSSUIIIJ)LIUIl Will
1111S

r----

l_-

a----

101-(X? L11(5 II(JW

.L

111 TJ163 ~

(h?dkln

:-.---:.-.
- mm
-- L-L. -? --l L-.--L,
11-i3ULUL-e UILf31SILy
map
ue outameu
uy rmmng
6. nA r-------the neural network at initial reservoir pressure.

tfl

honor the fracture orientation. When the fracture direction is around 90 degrees, the assumption is:
Kv=loox

7. Using a fracture intensity and anisotropy map, the


number of history matching parameters was reduced
to two.

Kz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...(6)

And Finally when the fracture orientation is around 45


degrees we use:
KY=K

2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

8. The use of diagonal permeability tensors when simulating structure-related fractured reservoirs do not
allow a correct representation of the actual fluid flow,
and a iuii permeability tensor is needed.

(7)

Since the fracture orientation is available over the entire


field (Fig. 6), an anisotropy map can be obtained at each
gridblock and Ku computed directly by using Kz.
After less than a 100 iterations, using the automatic
history matching algorithm, 10 the production history waa
matched (Fig. 11). Waterflooding was initiated in the
early life of the reservoir and different injectors were used
to maintain the reservoir pressure. The water response of
different wells is directly related to the fracture networks.
In the considered grid, most of the producing wells shown
w-ith an
injWO~
h
tb
x Or ~
in Fig. 11 ~-~ Kit dignd
direction. Since, the grids are large (350ft) and most the
producers-injectors are along the diagonal, we realize inconsistent water breakthrough. For some wells, the breakthrough is late (well 1 and well 12) for others it is early
mmrl mwvl,,.or
..- Q cM.~Awu
.1i,-rm,w-1 rmu~
al,-mcr
(we]! ~). Wham
. . . . . . . tha
. ..ti imicmtnm
.. AJ./v . W.u
y. u Utiti. m

Nomenclature
AN=
C.z =
C~Y=
C.V =
Cyz =
Cl =
Cz =
D =
E =
IP =
FI =
K==
KY =
L=
N =

the principal directions, we reach a better match as seen


on the last months of well 15 and well 24. These examples
illustrate the necessity of including cross-permeabilities in
the fractured reservoir model.

n =

M =
PV=
S. =
SY =
W =
WD =
Zil =

Conclusions
Based on the considered application,
the following conclusions:

2. The fracture intensity can be represented by a well


. . -..:..
Un
-:1 1-.-4,.
T-:
.-.--$----..,.:- A:-.+-- au(,ll
. ..-1. da
lU1 lI1allLc
lllUILGLW1
lllCLU1llU1ll
am,
lnlpm

15(I

KV=O.OI

W.W. WEISS

we have arrived at

1. A neural network can integrate geological, geomechanical, and reservoir engineering information into
a single fractured reservoir model.

permeability anisotropy factor


structure curvature in the z axis
structure curvature in the y axis
structure curvature in the diagonal
structure curvature in the diagonal
history matching estimation parameter
history matching estimation parameter
reservoir depth
neural network training error
well Initial Potential
fracture intensity
fracture permeability in the z axis
fracture permeability in the y direction
number of training patterns
number of neural network inputs
neural network node
number of neural network outputs
fracture pore volume
structure slope in the x axis
structure slope in the g axis
neural network weights
well density per section
ith neural network input for pattern 1

Yji = jth actual target output for pattern 1


Yjl = jth neural network output for pattern

431

.
FRACTURED
FORECASTING

RESERVOIR CHARACTERIZATION
AND PERFORMANCE
USING GEOMECHANICS AND ARTIFICIAL INTELLIGENCE

SPE 30572

Drilling, paper SPE/DOE 27799 presented at


the 1994 Symposium on Improved Oil Recovery,
Tulsa, Apr. 18-20.

Acknowledgments
The authors would like to thank the State of New Mexico for supporting this research, and HEYCO for providing the field data. Appreciation is extended to Adwait
Chawath6 and Mark Valenzuela for reviewing the paper.

[11] Zellou, A. M., Ouenes, A., Banik, A.: Improved


fractured reservoir characterization using neural
networks, geomechanics, and 3-D seismic, paper
SPE 30722 to be presented at the 1995 Annual
Technical Conference and Exhibition, Dallas, 2225 October.

References
Quantitative fracture study,
[1] Murray, G. H.:
Sanish Pool. Fracture-controlled production,
AAPG reprint Series 21.
[2] Lisle, R. J.: Detection of zones of abnormal
strains in structures using Gaussian curvature
analysis, AAPG Bull. (December 1994) V. 78,
No 12, 1811-1819.

[12] Richardson,

S., Generalization of the curvature


method and application of neural networks to
fractured reservoir chamcterization, M.S. Thesis,
New Mexico Tech, 1995.

[3] Lynn, H., and Thomsen, L.: Reflection sheardata collected near the principal axes of azimuthal anisotropy, Geophysics, (Feb 1990) 55,
147-156.

[13] McQuillan, H.: Small-Scale Fracture Density in

Asmari Formation of Southwest Iran and Its Relation to Bed Thickness and Structural Setting,
AAPG Bull. (1973) V. 47 No 12, 2367-2385.

[4] Lynn, H.B., Bates, R., Layman, M., and Jones,


M.: Natural fracture characterization using Pwave reflection data, VSP borehole imaging, and
in-situ stress field determination, paper SPE
A*v,-.F.-. .-.. .3 -L .~- *nnr 17mn n_ -1... na ---dmm

presemea

me

lVVO

m_ n

n,umy

[14] Long, J., Doughty, C., Hestir, K., and Martel, S.:

(iMnAOl:mfiua+nwn~finm,,a .mJ &_P+,,Vo~ naCOriYiWuGA.&,5


..s.=L ~u8AtiI.L.
u.- a. -.=. -U A.,---voirs with Inverse Methods Based on Iterated
Function Systems, Reservoir Characterization
111, B. Linville (cd.), PenWell Publishing Co.,
Tulsa, OK (1993).

Nmun-

tain Regional/Low-Permeability
Reservoirs Symposium, Denver Co, 20-22 March.

[15] Doughty,

C., Long, J., Hestir, K., and Benson, S.: Hydrologic characterization of hetero-. 1.-.,. --.-1- -,AL - :-------

[5] Ouenes, A., Bhagavan, S., Bunge, P.H., and


Travis, B.: Application of simulated annealing
and other global optimization methods to reservoir description: myths and realities, paper SPE
28415 presented at the 1994 Annual Technical
Conference and ExMbition, New Orleans, 25-28
Sept.

-A

geneous

f&olo~C

IIK!ala WITXI WI

lIIVeHe

L-J

IWSLIIUU

based on iterated function systems, Wat. Res.


Res. (June 1994) V. 30, No 60, 1721-1745.
[16] Datta Gupta, A., Vasco, D.W., and Long, J. C.S.:

Detailed Characterization of a Fractured Limestone Formation Using Stochastic Inverse Approaches, paper SPE/DOE 27744 presented at
the 1994 symposium on improved oil recovery,
Tulsa, April 17-20.

[6] Ouenes, A., Doddi, R., Lin, Y., and Cunningham, G.: A new approach combining neural
networks and simulated annealing for solving
petroleum inverse problems, paper presented at
the 1994 European Conference on the mathematics of oil recovery, Rtwos, Norway, June 7-10.

[17] Leung, W. F.:

A tensor model for anisotropic


and heterogeneous reservoirs with variable direction permeability, paper SPE 15134 presented
at the 1986 Western regional meeting, Oakland,
CA.

[7] Wasserman, P.D. Neural Computing:


Theory
and Practice , Van Nostrand Reinhold Co., New
York City (1989).
[8] Lippman, R.P.: An introduction to computing
with neural nets, IEEE ASSP Magazine (April
1987), 4-22.

[18] White,

C. D., and Home, R. N.: Computing


absolute transmissibility in the presence of tie
scale heterogeneity: paper SPE 16011 presented
at the 1987 SPE Symposium on Reservoir Simulation, San Antonio, TX.

[9] Schultz, P.S., Ronen, S., Hattori, M., Mantran,


P., and Corbett, C.: Seismic-guided estimation
of reservoir properties, paper SPE 28386 presented at the 1994 Annual Technical Conference
and Exhibition, New Orleans, 25-28 Sept.

[19] Samier, P.: Finite element method for calculat-

ing transmissibilities in N-point difference equations using a non-diagonal permeability tensor:


paper presented at the 1990 European conference
on the mathematics of oil recovery, Paris.

[10] Ouenes, A. et al.: A New Method to Characterize Fractured Reservoirs: Application to Infill

432

.
A. OUENES, S. RICHARDSON,

SPE 30572

W.W. WEISS

cxx

Fig. lFour directional curvatures.


Input
Layer

Hidden
Layer

Hidden
Layer

Surface Depth

First Derivative (X-dir)


First Derivative (Y-dir.)
First Derivative (FD-dir)

First Derivative (BD-dir)

Second Derivative (X-dir)

Fracture Index

Second Derivative (Y-dir)


Second Derivative (FD-dir)
Second Derivative (BD-dir)
Thickness
Cumulative Production

==EY

Fig. 2Neural network architecture

433

.
FRACTURED
FORECASTING

10

RESERVOIR CHARACTERIZATION
AND PERFORMANCE
USING GEOMECHANICS AND ARTIFICIAL INTELLIGENCE

SPE 30572

22

22

I
Fr%iwe Intensity

Fr%lure Intensity

Fig. 4YDU Fracture Intensity Map for E=O.08.

Fig. 3YDU Fracture Intensity Map for E=O.1.


25 [

r< ...

--

!___

J*_=,

\.-./!

.,,

-,

.-./!!
...,,,

!,//..,,,

! ///-.,,,.
,? ..-.,,

North

\ }

Jcg2&

-z.,l--

-. ..

20
1

-~

,. -./,11
// /-.,,,
\ ---.
\\\,./,
./ .,,, -- .\\~
. ...,,,.
t
--\\\\>,
%
I

North

15 -

10 -

lIIJ{?wl

5 - ,1,1,
11,,,,,18

.5~
o

,,,
,,,

,,
,

. .\-.

,,J,

10

15

20

25

30

01
0

35

l&f//

1.;$:$$

10

15

~<m

7
L7i
Y

/xl

20

25

,
\,

30

35

Fig. 6YDU Fracture Network for E=O.08.

Fig. 5YDU Fracture Network for E=O.1.

434

I
40

.
A. OUENES, S. RICHARDSON,

S~E 30572

600

70+3

11

W.W. WEISS

500

400

200

100

100

t
I

~ml

200

400

300
Pmdiclad W.

5CCI

61X

100

200

30s
Pmdicmd

Fig. 7Neural network training and testing for data set


1.

400

600

wl

700

I.P.

Fig. 8Neural network training and testing for data set


2.

~,,
ky = 100 k~

/ North

20 -

15 5

10 -

5-

O(
o

\,
5

10

15

20

25

30

35

Fig. 9High intensity fractures in YDU.

I
40

Fig. 10-Permeability

anisotropy for a diagonal tensor.

.
FRACTURED
FORECASTING

12

SPE 30572

RESERVOIR CHARACTERIZATION
AND PERFORMANCE
USING GEOMECHANICS AND ARTIFICIAL INTELLIGENCE

Well

Well

11

simulation

1.

Month9

Well

LA
.
m

50

6070

#12

miWlmllo1201m

w3wI

Menttm

Well

Well#15

l,mo

#24

0~-mmmm

..

-.

90

lm
110
Mor12hB

120

mo

140

lsn

I
160

Fig. n-Water

110

history matchkg.

436

120

130

140
MOtdlm

1s0

S-ar putea să vă placă și