You are on page 1of 12

n Society of Petroleum Engineers

SPE 30572

Fractured Reservoir Characterization and Performance Forecasting Using Geomechanics and


Artificial Intelligence
A. Ouenes, SPE, S. Richardson, W.W. Weiss, SPE, New Mexico Petroleum Recovery Research Center

Copyright 1995, Society of Petroleum Engineers, Inc. rely on a qualitative description of the fractures. This is
This paper was prepared for presentation at the SPE Annual Technical Conference & Ex- achieved by using mainly structure properties,1-2 seismic
hibition held in Dallas, TX, U.S.A" 22-25, October 1995.
velocity anisotropy 3 observed with shear or S-waves, and
This paper was selected for presentation by an SPE Program Committee following review
of information contained in an abstract submitted by the author(s). Contents of the pa- more recently compression or P-waves. 4 However, a reser-
per, as presented, have not been reviewed by the Society of Petroleum Engineers and are
subjected to correction by the author(s). The material, as presented, does not necessarily voir engineer struggling to numerically simulate a frac-
reflect any position of the Society of Petroleum Engineers, its officers, or members. Papers tured reservoir needs more than just the location of "sweet
presented at SPE meetinl\s are subject to publication review by Editorial Committees of the
Society of Petroleum Engineers. Permission to copy is restricted to an abstract of not more spots."
than 300 words. Illustrations may not be copied. The abstract should contain conspicuous
acknowledgement of where and by whom the paper 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
Abstract file for a fractured reservoir simulator which can be used
A new approach in fractured reservoir characterization for performance forecasting. This methodology relies on
and simulation that integrates geomechanics, geology, and the use of geomechanical concepts derived from reservoir
reservoir engineering is proposed and illustrated with ac- structure and artificial intelligence (AI) tools.
tual oil reservoirs. This approach uses a neural network
to find the relationship between, reservoir structure, bed Neural networks
thickness and the well performance used as an indicator During the past few years, the petroleum industry enthu-
of fracture intensity. Once the relation established, the siastically supported the concept of "integrated systems."
neural network can be used to forecast primary produc- Integration of everything is everywhere. From a reservoir
tion, or for mapping the reservoir fracture intensity. The engineering point of view, the concept of integration is
resulting fracture intensity distribution can be used to rep- a necessity not fashion. The necessity exists because of
resent the subsurface fracture network. Using the fracture the scarcity of reservoir information and the wide range of
intensity map and fracture network, directional fracture scales over which this information is measured. Therefore,
permeabilities and fracture pore volume can be estimated a reliable reservoir description must somehow integrate all
via a history matching process where only two parameters the existing information at all the scales. The application
are adjusted. of stochastic global optimization methods, e.g. simulated
annealing, in reservoir description 5 provided new tools for
Introduction achieving a certain level of integration. However, stochas-
Conventional reservoir simulation has benefited from im- tic global optimization methods were developed in an ar-
portant research during the last few years. The use tificial intelligence context and are more than just simple
of geostatistics is slowly moving from the production of mathematical optimization methods, as believed by some
"grayscale maps" with dubious value and multi-million cell users. Within the artificial intelligence framework, other
realizations with little practical value to useful input data tools exist and can be used to integrate various informa-
for reservoir simulators. Although there is still much to be tion into a complex reservoir model. The most practical of
done before these geostatistical realizations will be able to these integration tools can be found in neurocomputing.
reproduce the past performance of a reservoir, the recent There are various ways of looking at a neural network.
trend shows clearly that major advances have been made The most common application is a pattern recognition tool
in conventional reservoir description. On the other hand, where from a given amount of known information, a neu-
naturally fractured reservoir (NFR) characterization has ral network is able to be trained to recognize some pat-
not enjoyed a similar benefit from any major research ef- terns. In this case, the output of the neural network is
fort. Until this work, there is no quantitative methodology very often a binary variable where a value 0 means NO,
to "fill the NFR simulator gridblocks" and a value 1 means YES. Such tools are routinely used
Most of the current fractured reservoir characterization by financial institutions to decide on the outcome of our

425
FRACTURED RESERVOIR CHARACTERIZATION AND PERFORMANCE
2
FORECASTING USING GEOMECHANICS AND ARTIFICIAL INTELLIGENCE
SPE 30572

credit card or mortgage application. Scientists and engi- A neural network for fracture characterization
neers found other uses for neural networks, the most com- In reservoir description, a neural network can be viewed
'-mon is the "equation maker." In this case, the scientist or as an integration tool where various inputs of different na-
engineer may have experimental or field data for a given ture and scale can be correlated to some crucial reservoir
system with N inputs, (xiii = 1,2,···, N) and M outputs, property. In this framework, the neural net is more than a
(Yjlj = 1,2,···, M). Both the inputs and outputs can be crude "black box," it is a practical tool to integrate various
measured in the laboratory or field. The neural network types of data into a reservoir model. A recent example of
can be used to find a mathematical model that relates the this good integration ability can be found in the porosity
N inputs, x, to the M outputs, y. In other words, the neu- estimation using seismic attributes. 9 With this integration
ral network becomes a multivariate regression tool. How- ability in mind, we consider a naturally fractured oil or gas
ever, the advantage of the neural network model over con- reservoir and try to characterize the areal distribution of
ventional multivariate regression methods is its ability to fracture intensity.
mimic complex non-linear models. This ability has been
In our early work,10 we estimated fracture intensity
illustrated by Quenes et al. 6 where the complex partial
distribution in a carbonate reservoir using a generalized
differential equations describing two-phase flow in porous
curvature method. The major idea behind this early work
media could be reproduced with a neural network.
is to construct a quantitative fracture model using struc-
The process of determining these mathematical models tural properties. Extending Murray's1 original idea of us-
is described as training. The basic idea is to provide to ing structure curvature, we used two curvatures Gxx and
the neural net L training records (patterns) that will direct Gyy to derive a fracture model. The determination of this
the adjustment of the neural net parameters represented fracture model was time-consuming for a 30 well reser-
by a weight matrix Wnm ;. The reader is referred to neural voir. Furthermore, only curvatures were considered as pa-
nets literature7 - 8 for a complete description of these AI rameters affecting the fracture intensity. Within the same
tools. The following paragraph gives a brief description philosophy, Lisle2 also uses the product of two main curva-
necessary for understanding the next section. Further- tures (Gaussian curvature) to detect fractures. However,
more, we will address only backpropagation (or backprop) the bed thickness and lithology have an important impact
neural networks which we use for the fracture mapping on rock fracturing, and therefore need to be included in a
application. quantitative fracture model. With at least three major pa-
Suppose we have a system with N inputs (xiii = rameters (structure curvatures, thickness, and lith(")logy),
1,2, .. ·,N), and M model outputs (Yjlj = 1,2, .. ·,M). it becomes impossible for humans to find a fracture model
Additionally, we consider that there are L training data that may correlate all these parameters to fracture inten-
records, (xil,Yjlli = 1,2, .. ·,N;j = 1,2, .. ·,M;l = sity as indicated by log data or well production. In this
1" 2 ... , L) ,.available
... .. The backprop neural net uses the situation, a neural net would be a perfect tool for finding
L training patterns to adjust the weight matrix, Wn;n; the fracture model able to integrate all parameters affect-
through an optimization procedure called training. For ing rock fracturing.
the lth training pattern, the N inputs Xil are fed-forward
from the input layer, through all the hidden layers, and Neural network outputs. As described earlier, a neural
finally the neural net provides its output Yjl (W), which network is a mathematical model representing a relation-
is different from the known target output Yjl measured in ship between a set of inputs x and a set of outputs y.
the laboratory or field. The training process consists of Because we are interested in describing fracture intensity,
estimating the weights Wn;n; that minimize the quadratic the question that arises is what is the best way to describe
error: fracture intensity, at reservoir simulator scale? Addition-
ally, we would like to develop a fracture mapping approach
L M that will not create another problem, i.e. upscaling. We
E(W) = LL[Yjl- Y}1(W)]2 (1) have found that at the reservoir simulator scale, the best
1=1 j=1 way to describe fracture intensity is to use well production
data. These data can either be the maximum oil rate in
The error E backpropagates from the output layer to the early life of the well, the initial potential (IP) of the
the input layer and is used to adjust the matrix weight well, or the cumulative production at a given time in the
W nm ;. We use a hybrid optimization algorithm 6 that life of the reservoir. This well performance is strongly de-
combines simulated annealing and gradient methods to pendent on the fracture network around the wellbore and
estimate and adjust the weight matrix. provides information at a reasonable scale represented by
After successful training, the neural network can be the drainage radius. However, when choosing well per-
used for testing the outcome of some inputs not included formance as an indicator of fracture intensity, one may
in the training process. There are various practical as- have also to take reservoir pressure into account which
pects related to the training and testing stages which will also affects the well performance. Since reservoir pressure
be described within the fractured reservoir application. is rarely measured, we use the field cumulative oil pro-

426
SPE 30572 A. OUENES, S. RICHARDSON, W.w. WEISS 3

duction to account for the change in pressure. In other and have opposite signs. This 3,fisumptioD. can be seen in
words, for each new well drilled during the development many structures where hyperbolic (saddle) folded surfaces,
of a field, we compute the total cumulative oil production can be found. For reservoirs where most of the structure
of the entire field and use this number as an indication of is made of elliptical (dome) folded surfaces this assump-
the reservoir pressure at the time the new well was drilled. tion is not valid and another mapping approach must be
This reservoir pressure indicator is added as an input to considered. If the observed folding appearR to he hyper-
the neural network because it affects the output which is bolic, the Laplace equation can be used with the known
the well performance. In addition to this "undesirable" depths at the well locations to estimate the entire top of
input which is not related to fractures, the neural network the structure. We use a finite element method to solve the
will need other inputs that are related to rock fracturing. Laplace equation and take advantage of the sparse matrix
(resulting from the finite element approach) solvers to in-
Neural network inputs. The inputs for the neural net- terpolate reservoir structure points in a lpw seconds on a
work must be the parameters which are related to fracture simple PC.
intensity. We consider three major factors affecting the With the resulting structure map, or using seismic
rock fracturing: reservoir structure, bed thickness, and travel-time,l1 the discrete curvatures C:r,r and C yy are
lithology. These factors must be represented by one or computed at each grid location. Howevpr, for a complete
more quantitative variables that should be available over characterization of the structural changes, we have found
the entire area of the reservoir. that cross curvatures G",y and Cy ", taken in the diago-
Evaluating reservoir structure may seem an easy task. nal directions (Fig. 1) yield additional information and
However, this fractured reservoir description approach re- therefore could be used as inputs for the neural network.
quires a detailed and accurate structure, not simply a Furthermore, the slopes S'" = ~~, Sy = !!Jf" and the cross
hand-drawn contour map. There are only two possibili- slopes taken in the diagonal directions, when added as in-
ties for obtaining a meaningful structure: 3-D seismic or puts, allow a faster and superior training of the neural
interpolation using well log data. If 3-D seismic is avail- network. In summary, the geomechanical effect of struc-
able, then the major changes in the structure can be seen ture on fracturing is taken into account by using the depth,
clearly either on a structure obtained simply from the four slopes, and four curvatures.
travel time l1 or after a time-depth conversion. At this The bed thickness is also another important parameter
stage, it is important to realize that the most important to be included as an input. For a fractured reservoir where
information needed in the structure map is the areal varia- bed thickness varies significantly, this input is crucial be-
tion in depth, not the depth itself. Therefore, a time-depth cause it affects the neural network output in two ways.
conversion may not dramatically change the structure con- When considering production as an indicator of fracture
trast from one point to another. intensity, a high oil or gas rate can reflect a very thick pay
Unfortunately, 3-D seismic is not available for all frac- zone, a highly fractured zone, or a very thick and highly
tured reservoirs. In this case, one must rely on tops ob- fractured zone. Therefore, a correct fracture model must
tained from well logs and use a mapping method to gen- include thickness as an input. On the other hand, bed
erate the entire reservoir structure. With the effect of thickness is intimately related to fracturing. In general,
structure on rock fracturing being accounted for through very thick beds have less fractures than thin beds. 13 Based
curvatures and slopes at each location in the reservoir, it on these two considerations, bed thickness appears to be
is important to use a smooth mapping method. A "noisy" an important input for fractured reservoirs. The problem
structure will lead to extremely high curvatures (second of obtaining bed thickness over the entire reservoir is sim-
derivative of the tops) that are the result of the "noise" ilar to the interpolation of structure points. Mapping or
and may have no relation to fracturing. A smooth map- 3-D seismic can be used, and the comments made for the
ping method such as kriging can be used provided that a structure mapping also apply to thickness.
reasonable and meaningful variogram is available. After
The last parameter that has an impact on rock frac-
trying various approaches to generate the reservoir struc-
turing is lithology. This parameter would be an important
ture, we have found a new approach that can be very useful
input to consider if reservoir lithology varies areally. If so,
for mapping tops especially for folded structures. 12
geostatistics applied to an indicator variable could provide
This approach relies on the assumption that reservoir
a good estimation of the lithology distribution which then
depth, D, of a folded structure obeys the Laplace equa-
can be included as input to the neural net. Given these
tion:
different inputs, the neural network architecture may now
2 2 be designed.
-88xD2 + -88 D2 = 0 (2)
y
Neural network architecture. A backprop net is a sys-
In terms of curvatures, this assumption means that at tem that links a series of basic blocks called nodes. These
each location the curvature in the x direction, G",,,, = ~, nodes, represented by a set NO = {nl,n2,''',nn}, are
and the curvature in the y direction, Gyy = ~:1? are equal connected in a certain way to create the architecture (or

427
FRACTURED RESERVOIR CHARACTERIZATION AND PERFORMANCE
4 SPE 30572
FORECASTING USING GEOMECHANICS AND ARTIFICIAL INTELLIGENCE

topology), whereas the connection between two nodes ni mapping of actual reservoirs, this luxury does not exist
and nj is characterized by a weight Wninj' Each node and one may have to live with 25 to 30 training patterns
sums the inputs multiplied by the weights, then applies a obtained from the very limited information available at
nonlinear function (sigmoid function) to produce the node the wells. Although, 25 patterns may seem very small
output. The network architecture includes the input layer, number to train a neural network, we have found that it
the output layer, and a certain number of hidden layers is sufficient to derive a fracture model.
(Fig. 2). Each layer contains a finite number of nodes When training a.neural network with known data, it is
connected to other layer nodes. These nodes can be rep- helpful to know the limiting value of the error E at which
resented in practice by a few lines of software. or a few the training could be considered sufficient. The training
electronic components in the hardware. The choice of this process starts by assigning random weights which will lead
architecture is important being that it directly affects the to an initial error Eini. At each iteration (epoch), the neu-
performance of the neural network and its ability to pre- ral network adjusts the weights and the neural network
dict correct values. error, E, is slowly reduced. The training could be stopped
There are two major aspects to consider when design- at a certain error Efin and the obtained weights could
ing the architecture. First, complex non-linear models re- be considered as representative of the. non-linear model
quire more than one hidden layer, many nodes, and there- sought by the user. The training process could be stopped
fore many weights to estimate during the training pro- when the neural network fracture intensity FIt reasonably
cess. Second, the training of a neural network is a multi- match the actual fracture intensity fil for all the L wells
dimensional optimization problem very rich in local min- used for the training. A bad match of the well data in-
ima. Hence, an increase in the number of weights or node dicates insufficient training, and a very close match of all
connections reduces the chance of approaching a global the data may indicate over-training, which needs to be
optimum required for reasonable training. In this trade- avoided. The number of iterations required to obtain. an
off between r~ducing the connections for better training, acceptable match will be discl,1ssed and illustrated with
and increasing the connections for better modeling of the three different fractured res~rvoirs.
non-linear behavior, there are very limited objective tools Using the architecture shown in Fig. 2, three actual
that can assist in the design of the optimal architecture. oil reservoirs were considered for the application of this
Fuzzy logic combined with a special neural network can be new approach. The first unit, referred to as Unit A is
used to find the optimal architecture,6 if sufficient training located in a large naturally fractured reservoir in West-
patterns are available. ern· Texas. The study included 16 sections area of this
However, in the fracture mapping application discussed unit. For Unit A, the fracture intensity was represented
in this paper and its companion paper, 11 the amount of by the cumulative oil produced. at each well for 78 wells
available data is not sufficient to find the optimal archi- drilled prior to 1960. In Unit A, we are interested in
tecture in an objective manner. The lack of data isa con-· mapping the fracture intensity of an approximately 900ft
sequence of selecting actual reservoirs for the application thidc shaley sandstone formation which we will refer to as
of this methodology. Therefore, the proper architecture Formation AA. For Unit A, little information was avail-
was found by testing various models, and it appears that able at the time of the study. Besides the estimated in-
a neural network with two hidden layers (Fig. 2) was suffi- dividual cumulative production of 78 wells, the tops of
cient for the modeling of the relation between the fracture Formation AA and a 3-D seismic survey was available.
intensity and the inputs previously discussed. Each node For Unit A, the average density of wells per section is
of the input layer is comiected to all the nodes of the first WDA = 4.9 wells/sq.mile. Another unit, referred to as
hidden layer. For 11 inputs, there are 121 connections be- Unit B encompasses 22 wells that produced on primary
tween the input layer and the first hidden layer. It is very before water injection into formation AA was initiated.
common in neural network applications for the first and Tops of the .formation AA were available at 55 wells lo-
second hidden layers to be fully connected, which means cated in 4 sections. The average well. density per section
121 additional connections. This common procedure is not in Unit B is WDB = 13.7 wells/sq.mile. The third ex-
necessary and the user should first try one-to-one conneC- ample is a carbonate reservoir located in southeast New
tions between the first and second hidden layer as shown Mexico and referred to as the Young Deep Unit (YDU).
in Fig. 2. In the fracture mapping application, we have Our early work lo on fractured reservoir characterization
found that the 11 connections between the first and the was applied to YDU. Production history and log data of
secon'd hidden layers were sufficient which means 110 (121- 30 wells were available at the YDU which spans over an
11=110) less weights to estimate during the training. area of about 2 sections. The well density per section is
i
WDYDU = 15 wells/sq.mile.
Neural network training For the West Texas Unit B and New Mexico YDU, the
In most neural network applications, there is a wealth of structure of the considered fractured formation was ob-
data that can be used for training, and most probably a tained by using the log data and the mapping procedure
reliable generalization of the neural network. In fracture that assumes the Laplace equation described in a previous

428
SPE 30572 A. OUENES, S. RICHARDSON, W.w. WEISS 5

section. Given the reasonable well density of Unit Band Primary performance forecasting. If we consider a set
YDU, an interpolation method provides a good estimation of 22 wells with their known performance and use only 18
of the structure. The fracture intensity at each well loca- wells for the training, the testing process consists of pre-
tion was represented by the Initial Potential (IP) for Unit dicting the performance of the 4 wells not included in the
B, and by the maximum oil rate for YDU. For both reser- training. By doing so, we are simply predicting the per-
voirs, the correlation between neural net inputs and the formance of an infill well. As an illustration, this testing
considered fracture intensity was distinctly evident. This procedure is applied to Unit B.
can be inferred because in less than 3000 iterations the For Unit B, data for 22 wells were available. Two sets
neural net converged to a very small error and achieved of 18 wells were used for training and each time 4 different
very good training. Using the scarce well information in wells were removed from the training patterns. After com-
Unit A, an interpolated structure ll was derived. When
pletion of the training based on 18 wells, the performance
using this structure map to create the neural network in-
of the 4 wells removed from training are predicted. Figs.
puts, more than 5 million iterations were required to reach 7 and 8 show the actual IP versus the predicted IP. The
an acceptable error. Thus, if a good structure map is avail-
empty circles indicate the 18 wells used for the training,
able and the fracture intensity data are correct, the train-
and the filled circles are the predicted performance of the
ing will be short. For verification purposes, the neural 4 wells for each data set. For both sets, the error on the
network used for Unit A was trained with a structure map predicted performances varied between 11% and 50%. In
derived from 3-D seismic ll and resulted in faster training. other words, the largest error was made when the neural
Therefore, slow training indicates poor data quality.
network predicted an IP=200, when the actual IP=400.
Considering the very small number of available wells used
In such cases, the user may stop the training at a high
for training, these predictions can be considered remark-
error E l . Using the weights associated with this error, a
able. Unfortunately, this prediction tool cannot be used
fracture intensity map can be obtained as described in the
after fluids are injected in the fractured reservoir. The
next sections. If the training is continued to a lower error
change in pressure cannot be included easily in the neu-
E z , the new fracture intensity will be different. For the
ral network model. For this reason, we developed another
YDU, Fig. 3 shows the fracture intensity map for an error
procedure to obtain a fracture intensity map that can be
E l = 0.1, and Fig. 4 shows the map for an error E z = 0.08.
used for reservoir simulation.
Although there are some local differences, the main trends
remain the same for the two errors. The derived fracture
network from the Fig. 3 fracture intensity map is shown Mapping the fracture network. When choosing the
in Fig. 5 for the high error E l . For the lowest error E z well performance as an indicator of fracture intensity" we
when using the fracture intensity map shown in the Fig. include the cumulative oil production as an indicator of
4 and the weighting method described in the companion reservoir pressure in the inputs. In other words, for each
paper,ll the fracture network obtained is shown in Fig. new well drilled at a certain time, corresponds a reser-
6. Notice that despite the difference in the training er- voir pressure indicator. This indicator is simply the total
ror, the two fracture networks are almost the same. This cumulative oil produced by all the previous wells. When
indicates that for both the errors, an acceptable training the neural network is used to predict the performance of
was achieved. However, the subsurface fracture network each gridblock at the initial reservoir pressure, we have a
is unknown and other means must be used to evaluate tool that can "drill" fictitious wells, in each gridblock, at
the quality of the training. A simple approach to verify if the initial pressure. In this situation, the predicted perfor-
generalization was achieved, is to test patterns which were mance is only affected by the fracture network. Therefore,
not included in the training process. the predicted value represents the fracture intensity of the
considered gridblock. Applying the same process for all
the gridblocks provides a complete description of the areal
Neural network testing distribution of fracture intensity. Figs. 3 and 4 show the
fracture intensity map for the YDU and similar maps can
Before using the neural network for any prediction, one be found for Unit A and Unit B in the companion paper.!l
must confirm the ability of the neural network to re- These fracture intensity maps can be further used to derive
produce correctly the non-linear model that may exist a fracture network using the weighting method described
between the various inputs and the fracture intensity. in the companion paper. Briefly speaking, this method
This validation process is mandatory and allows the user uses at each gridblock the fracture intensity of the eight
to evaluate the goodness of the obtained mathematical neighboring gridblocks to find the orientation of the frac-
model. In the application to fractured reservoirs, there ture. When focusing only on the most intense fractures
are two ways to test a neural network. The trained model that affect fluid flow in the reservoir, the fracture network
can be used to predict the performance of an infill well derived from Fig. 4 for YDU is shown in Fig. 9. This
(well deleted from the training data set), and/or obtain fracture network clearly shows the shear fractures associ-
the fracture intensity map over the entire reservoir. ated with a folded surface. This type of fractured reservoir

429
FRACTURED RESERVOIR CHARACTERIZATION AND PERFORMANCE
6
FORECASTING USING GEOMECHANICS AND ARTIFICIAL INTELLIGENCE
SPE 30572

is the most complex and difficult to simulate for reasons As discussed in the companion paper, 11 we are focusing
described in the next section. on two types of fractures: regional and structure related.
In the "regional fractures, two major fracture directions can
Estimation of fracture permeability and pore be seen by different means. The principal fractures along
volume the maximum horizontal stress direction, and another set
The objective of this paper is to develop a methodology to of fractures orthogonal to the major regional fractures. In
fill reservoir simulator gridblocks. This objective includes this situation, a reservoir grid that has its x-axis along
history matching, which we view as an integral part of the regional fractures, and its y-axis in the perpendicu-
reservoir characterization, and not a downstream applica- lar direction is sufficient to represent the fluid flow in the
tion disconnected from the characterization effort. This fractures. Unfortunately, structure-related fractures are
methodology relies on three components: 1) a fracture in- not perpendicular and do not keep the same direction in
tensity map obtained with a neural network or using seis- the field. Often these shear fractures, shown in Fig. 9,
mic attributes 11 , 2) a fracture network derived from the form an angle of 60 degrees and are found along the fold-
fracture map using the weighting method 11 , and 3) an in- ing. In other words, the fluid flow in the fractures can
verse modeling procedure using a reservoir simulator for occur in any direction. When using a finite difference
automatic history matching. reservoir simulator, only two major flow directions can
The basic idea behind this methodology was first de- be tolerated. Therefore, the major difficulty in simulat-
scribed in earlier work 10 where we assume that directional ing structure-related fractured reservoirs is the inability
permeabilities, K x and K y , as well as fracture pore vol- of tp,e simulators to account for preferential flow paths in
ume PV are a function of the fracture intensity, Fl. The any direction.
functions considered are linear and simple. For example, There are two solutions for this problem. The first so-
the permeability in the x direction is given by: lution, is to use a very fine grid, and the second solution is
to use full permeability tensors. Because we are primarily
!(x = 01 X Fl (3) concerned with full-field simulation that includes history
and the pore volume of the fractures is given by: matching, we try to avoid a very fine areal grid system.
Hence, we prefer to deal with full permeability tensors.
PV=02 xF1 (4)
Where 0 1 and O2 are two constants to be estimated Diagonal versus full permeability tensors. 'Most
during the history matching. Since the fracture intensity reservoir simulation studies use a diagonal tensor for per-
is available over the entire reservoir, the estimation of frac- meability where at most two directional permeabilities K x
ture permeability K x and pore volume PV in each grid~ and K y are considered. The cross-permeability terms K xy
block is reduced to the estimation of two simple parame- and K yx have been introduced by Leung 17 , White and
ters 0 1 and O2 . Therefore, this approach tries to integrate Horne 18 , and Samier 19 for different applications. For frac-
all the available field data to produce a fracture intensity tured reservoir simulation, these cross-permeability terms
map which will ease the history matching task by reducing may be the best tools to represent the flow in all the direc-
the number of "tuning parameters." Furthermore, geome- tions as it occurs in a fold-related fracture system. Notice
chanical and geological constraints can be honored in the at this stage that we did not discuss the dual-porosity or
fracture intensity map which will lead to a realistic reser- single-porosity fracture representation. When concerned
voir model. with fluid flow in the fractures, the choice of dual or single-
The estimation parameters 0 1 and O2 can be found porosity is not important. However, this choice becomes
by manual history matching, or by using an automatic important when considering the fluid transfers between
history matching procedure, which will not encounter ma- matrix and fractures.
jor difficulties, since the number of estimation parameters For illustration purposes, we will use a diagonal perme-
is very small. The same philosophy of fractured reser- ability tensor and a single-porosity model to simulate the
voir characterization based on inverse modeling was used performance of the YDU reservoir. The purpose of this
by Long et al. 14 , Doughty et al. 15 , and Datta Gupta et choice is to illustrate the limited ability of a diagonal per-
al. 16 to obtain fracture distributions. However, in their meability tensor to reproduce correctly the fine details of
investigations,14-16 the obtained fracture intensity map is the production history at different wells. The advantages
a result of a pure numerical inversion, and geomechani- of a full tensor will be illustrated in future papers.
cal and geological input were not imposed to the model. The YDU reservoir model is described with two lay-
In our work, the emphasis is given to the integration of ers. The upper layer represents the matrix and the lower
geomechanical, geological, and geophysical 11 information layer represents the fractures. Each layer contains 759
in the fracture model through a neural network. Despite, gridblocks (33 x 23) of 350 ft. The details concerning the
this integration effort there are other problems that need reservoir and its characteristics can be found in a previous
to be addressed when simulating actual oil or gas fractured paper.l° Using the fracture intensity map shown in Fig. 4
reservoirs. and the Eqs. 3 and 4, the inverse model history-matched

430
SPE 30572 A, OUENES, S, RICHARDSON, W,W, WEISS 7

150 months of 29 wells' production history. In contrast 2. The fracture intensity can be represented by a well
to the the early work lO the permeability anisotropy fac- performance indicator such as maximum oil rate, Ini-
tor, which was considered constant over the field, was not tial Potential (IP), or cumulative production.
included in the estimation parameters. The permeability
in the y direction K y was computed directly by using K x 3. The geomechanical effects on fracturing can be rep-
and an anisotropy map obtained from the fracture network resented by the slopes and curvatures of the reservoir
shown in Fig. 6. structure
At each gridblock, the fracture orientation was avail-
able. In terms of fluid flow, this fracture orientation indi- 4. Using a reliable structure and its geometrical prop-
cates the direction of flow. Unfortunately, when using a erties, bed thickness, and cumulative production as
diagonal tensor there are only two possible flow directions. inputs, and well performance as output, a neural net-
In this situation we have only three possible permeability work can find easily the complex relation that exists
anisotropy factors (Fig. 10). When the fracture direction between the inputs and output.
has an angle around 0 degrees we assume an anisotropy
factor AN = 0.01 and:
5. Primary performance forecasting can be obtained
Ky = 0.01 x Kx ...................•.•.....•.•..•.... (5) simply by testing the neural network.

This assumption will force the flow in the x direction to 6. A fracture intensity map can be obtained by testing
honor the fracture orientation. When the fracture direc- the neural network at initial reservoir pressure.
tion is around 90 degrees, the assumption is:
7. Using a fracture intensity and anisotropy map, the
K y = 100 x K x ...••.•.•....•............•.......•.•. (6) number of history matching parameters was reduced
to two.
And Finally when the fracture orientation is around 45
degrees we use:
8. The use of diagonal permeability tensors when sim-
Ky = Kx ......•.......•.•..........•................ (7) ulating structure-related fractured reservoirs do not
allow a correct representation of the actual fluid flow,
Since the fracture orientation is available over the entire and a full permeability tensor is needed.
field (Fig. 6), an anisotropy map can be obtained at each
gridblock and K y computed directly by using K x • Nomenclature
After less than a 100 iterations, using the automatic
AN = permeability anisotropy factor
history matching algorithm,lO the production history was
C xx = structure curvature in the x axis
matched (Fig. 11). Waterflooding was initiated in the
C yy = structure curvature in the y axis
early life of the reservoir and different injectors were used
C xy = structure curvature in the diagonal
to maintain the reservoir pressure. The water response of
C yx = structure curvature in the diagonal
different wells is directly related to the fracture networks,
C 1 = history matching estimation parameter
In the considered grid, most of the producing wells shown
C2 = history matching estimation parameter
in Fig. 11 are not aligned with an injector in the x or y
D = reservoir depth
direction. Since, the grids are large (350ft) and most the
E = neural network training error
producers-injectors are along the diagonal, we realize in-
I P = well Initial Potential
consistent water breakthrough. For some wells, the break-
F I = fracture intensity
through is late (well 1 and well 12) for others it is early
K x = fracture permeability in the x axis
(well 3). When the injector and producer are aligned along
K y = fracture permeability in the y direction
the principal directions, we reach a better match as seen
L = number of training patterns
on the last months of well 15 and well 24. These examples
N = number of neural network inputs
illustrate the necessity of including cross-permeabilities in
n = neural network node
the fractured reservoir model.
M = number of neural network outputs
PV = fracture pore volume
Conclusions Sx = structure slope in the x axis
Based on the considered application, we have arrived at Sy = structure slope in the y axis
the following conclusions: W = neural network weights
W D = well density per section
1. A neural network can integrate geological, geome·· Xi! = ith neural network input for pattern I
chanical, and reservoir engineering information into yjl = jth actual target output for pattern I
a single fractured reservoir model. YJI = jth neural network output for pattern

431
FRACTURED RESERVOIR CHARACTERIZATION AND PERFORMANCE
8
FORECASTING USING GEOMECHANICS AND ARTIFICIAL INTELLIGENCE
SPE 30572

Acknowledgments Drilling," paper SPE/DOE 27799 presented at


The authors would like to thank the State of New Mex- the 1994 Symposium on Improved Oil Recovery,
ico for supporting this research, and HEYCO for provid- Tulsa, Apr. 18-20.
ing the field data. Appreciation is extended to Adwait
Chawathe and Mark Valenzuela for reviewing the paper. [11] Zellou, A.M., Ouenes, A., Banik, A.: "Improved
fractured reservoir characterization using neural
References networks, geomechanics, and 3-D seismic," paper
[1] Murray, G.H.: "Quantitative fracture study, SPE 30722 to be presented at the 1995 Annual
Sanish Pool. Fracture-controlled production," Technical Conference and Exhibition, Dallas, 22-
AAPG reprint Series 21. 25 October.

[2] Lisle, R.J.: "Detection of zones of abnormal [12] Richardson, S., Generalization of the curvature
strains in structures using Gaussian curvature method and application of neural networks to
analysis," AAPG Bull. (December 1994) V. 78, fractured reservoir characterization, M.S. Thesis,
No 12, 1811-1819. New Mexico Tech, 1995.
[3] Lynn, H., and Thomsen, L.: "Reflection shear- [13] McQuillan, H.: "Small-Scale Fracture Density in
data collected near the principal axes of az- Asmari Formation of Southwest Iran and Its Re-
imuthal anisotropy," Geophysics, (Feb 1990) 55, lation to Bed Thickness and Structural Setting,"
147-156. AAPG Bull. (1973) V. 47 No 12, 2367-2385.
[4] Lynn, H.B., Bates, R., Layman, M., and Jones, [14] Long, J., Doughty, C., Hestir, K., and Martel, S.:
M.: "Natural fracture characterization using P- "Modeling Heterogeneous and Fractured Reser-
wave reflection data, VSP borehole imaging, and voirs with Inverse Methods Based on Iterated
in-situ stress field determination;" paper SPE Function Systems," Reservoir Characterization
29595 presented at the 1995 SPE Rocky Moun- III, B. Linville (ed.), PenWell Publishing Co.,
tain Regional/Low-Permeability Reservoirs Sym- Tulsa, OK (1993).
posium, Denver Co, 20-22 March.
[5] Ouenes, A., Bhagavan, S., Bunge, P.R., and [15] Doughty, C., Long, J., Hestir, K., and Ben-
Travis, B.: "Application of simulated annealing son, S.: "Hydrologic characterization of hetero-
and other global optimization methods to reser- geneous geologic media with an inverse method
voir description: myths and realities," paper SPE based on iterated function systems," Wat. Res.
28415 presented at the 1994 Annual Technical Res. (June 1994) V. 30, No 60,1721-1745.
Conference .and Exhibition, New Orleans, 25-28
[16] Datta Gupta, A., Vasco, D.W., and Long, J.C.S.:
. Sept.
"Detailed Characterization ·of a Fractured Lime-
[6] Ouenes, A., Doddi, R., Lin, Y, and Cunning- stone Formation Using Stochastic Inverse Ap-
ham, 0.: "A new approach combining neural proaches," paper SPE/DOE 27744 presented at
.networks and simulated annealing for solving the 1994 symposium on improved oil recovery,
petroleum inverse problems," paper presented at Tulsa, April 17-20.
the 1994 European Conference on the mathemat-
ics of oil recovery, R0ros, Norway, June 7-10. [17] Leung, W.F.: "A tensor model for anisotropic
and heterogeneous reservoirs with variable direc-
[7] Wasserman, P.D. Neural Computing: Theory tion permeability," paper SPE 15134 presented
and Practice, Van Nostrand Reinhold Co., New at the 1986 Western regional meeting, Oakland,
York City (1989). CA.
[8] Lippman, R.P.: ~'An introduction to computing [18] White, C.D., and Horne, R.N.: "Computing
with neural nets," IEEE ASSP Magazine (April absolute transmissibility in the presence of fine
1987), 4-22. scale heterogeneity," paper SPE 16011 presented
[9] Schultz, P.S., Ronen, S., Hattori, M., Mantran, at the 1987 SPE Symposium on Reservoir Simu-
P., and Corbett, C.: "Seismic-guided estimation lation, San Antonio, TX.
of reservoir properties," paper SPE 28386 pre-
sented at the 1994 Annual Technical Conference [19] Samier, P.: "Finite element method for calculat-
and Exhibition, New Orleans, 25-28 Sept. ing transmissibilities in N-point difference equa-
tio~s using a non-diagonal permeability tensor,"
[10] Ouenes, A. et al.: "A New Method to Charac- paper presented at the 1990 European conference
terize Fractured Reservoirs: Application to Infill on the mathematics of oil recovery, Paris.

432
SPE ::l0572 A. OUENES, S. RICHARDSON, W.w. WEISS 9

y Cx y

~----~-----..... CXX

L....-- _ ..... X

Fig. I-Four directional curvatures.

Input Hidden Hidden


Layer Layer Layer
Surface Depth
First Derivative (X-dir)
First Derivative (Y-dir.) Output
First Derivative (FD-dir) Layer
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

Fig. 2-Neural network architecture.

433
FRACTURED RESERVOIR CHARACTERIZATION AND PERFORMANCE
10
FORECASTING USING GEOMECHANICS AND ARTIFICIAL INTELLIGENCE
SPE 30572

9900

22 22

7425

6188

4950

3712

2475

1238 1238

o o I I
Fracture Intensity Fracture Intensity

Fig. 3-YDU Fracture Intensity Map for E=O.1. Fig. 4-YDU Fracture Intensity Map for E=0.08.
25 -----,,-----,,---,-----,,-------,,-----,,-----,,----, 25 ~-_,--_,--_,--_,--____r--____r--_.

20
~rth
20 ,. I I

15
15

10

10

5
o 1/1 I
I I I I I

./ I

_5'-----'----'-------'----"----"------'----'-----' °0~---5'----1LO---1L5---2LO-
o 5 10 15 20 25 30 35 40 30 35 40

Fig. 5-YDU Fracture Network for E=O.1. Fig. 6-YDU Fracture Network for E=0.08.

434
SPE 30572 A. OUENES, S. RICHARDSON, W.W. WEISS 11

600 700
a
500
o 600

500
400

0: Training ~ 400 - Training


..l
a a
~
~
300 Testing
• J 300
Testing

-
200
III
a 200
a
100 100
a

o
o 100 I 200 300 400 500 600 100 200 300 400 500 600 700
Predicted I. P. Predicted I.P.

Fig. 7-Neural network training and testing for data set Fig. 8-Neural network training and testing for data set
1. 2.

\ ,
\, ky = 100 kJl'
" " ,/
,
ky=kx

ky = 0.01 kx
JL
8

40

Fig. 9-High intensity fractures in YDU. Fig. lO--Permeability anisotropy for a diagonal tensor.

435
FRACTURED RESERVOIR CHARACTERIZATION AND PERFORMANCE
12
FORECASTING USING GEOMECHANICS AND ARTIFICIAL INTELLIGENCE
SPE 30572

Well Well
20,000 15,000

'UJ
;[
15,000

simulation
field data
III
,.• 1., 10,000

~ 10,000 E
lii .,,~ .,
~

~ ~ 5,000
5,000

0 0
0 10 20 50 0 130
Months

Well #12
25,000 r-----------------,

simulation
20,000
• field data
~
.,
;[ 15,000

e
j 10,000

5,000

oL...... .~~~~w..I
30 40 50 60 70 80 90 100 110 120 130 140 150 160
Months

Well # 15 Well #24


5,000 10,000 r;:::======:;--------,
simulation simulation
4,000 6,000
• field data • field data
'UJ
;[ 3,000 ••
~
S 2,000 j
~

4,000
~
1,000 2,000


~10 120 130 140 150 180
Months

Fig. ll-Water history matching.

436

You might also like