SPE 75223 Using Statistics To Evaluate A History Match

Download as pdf or txt
Download as pdf or txt
You are on page 1of 21

SPE 75223

Using Statistics To Evaluate A History Match


*

David Uldrich, SPE, BP Kuwait ; Saad Matar, Kuwait Oil Company; Hugh Miller, ChevronTexaco Kuwait
Copyright 2002, Society of Petroleum Engineers Inc.
This paper was prepared for presentation at the SPE/DOE Improved Oil Recovery Symposium
held in Tulsa, Oklahoma, 1317 April 2002.
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 paper, as
presented, have not been reviewed by the Society of Petroleum Engineers and are subject to
correction by the author(s). The material, as presented, does not necessarily reflect any
position of the Society of Petroleum Engineers, its officers, or members. Papers presented at
SPE meetings are subject to publication review by Editorial Committees of the Society of
Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this paper
for commercial purposes without the written consent of the Society of Petroleum Engineers is
prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300
words; illustrations may not be copied. The abstract must contain conspicuous
acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O.
Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.

Abstract
Three methods are proposed for quickly evaluating the history
match of a numerical simulation to actual reservoir
performance. All of the methods rely on computing a set of
deviation values, each of which is defined to be a calculated
simulator result minus the corresponding surveillance
measurement value.
For any particular type of surveillance data, such as rates,
watercuts, or gas-oil ratios, the deviation values can be
grouped by well, by area, or combining all measurements in
the database. The first two proposed methods rely on simple
graphical presentations of each group of deviation values to
show how well the simulation results match the surveillance
data. Plotting together the results from more than one
simulation run allows a quick comparison of the match for
each run, which is useful during the history match process.
The third method converts each deviation value to a
quantity called Match Factor, which is a relative measure of
the confidence that the simulator actually reproduced the
particular reservoir performance at the time the surveillance
measurement was made. Weighted-average Match Factors can
reveal the degree of match by well, by area, and by data type.
These techniques are especially valuable when matching
reservoirs with a large volume of surveillance data. They can
help focus the history matching process by identifying areas
less well matched. They can identify when the history
matching process is not significantly improving the match and
can stop.

Now with Phillips Alaska, Inc.

Introduction
One of the more challenging aspects of history matching large
numerical simulators is assessing how well the simulator
results match observed field behavior. Unfortunately, there are
few objective measures of the degree of match readily
available.
The traditional approach is to plot the observed data values
versus time, along with the corresponding simulator output,
and visually assess how well the simulator reproduced the
measured values1. The quality of this evaluation can vary,
depending on the experience and judgment of the simulation
engineer. It can also be very time consuming, especially for
fields with many wells and years of surveillance data.
This paper describes a two-stage approach, developed by
BP Kuwait for Kuwait Oil Company, and validated on large
models of giant reservoirs in Kuwait. In the first stage,
observed surveillance data values are directly compared to the
corresponding predicted values extracted from simulator
output. Each pair of observed and predicted values defines a
deviation value. Groups of deviation values are presented in
two graphical ways, showing the overall degree of match for
the type of data in that group for that run. Plotting together
results from two or more simulation runs can be used to
quickly compare the matches for the runs.
In the second stage of the analysis, each deviation is
converted to a Match Factor value, which represents the
confidence level that the simulated result matches the actual
field behavior represented by the observed surveillance value.
This conversion takes into account the inherent uncertainty of
the field measurement technique, and the limits of the
simulator calculation. Plotting on a map the Match Factors
averaged by well quickly shows where a model is better
matched compared with other areas. This helps guide where
changes in the reservoir description should be made in
subsequent history match runs.
Averaging the Match Factors for all values in one data
type quickly provides a measure of the overall degree of
match for that data type for that simulation run. Comparing
this average to that from other runs, shows whether the history
match is improving or deteriorating. If the average Match
Factor value has stabilized, the useful end of the history match
process may have been reached (unless a significant alteration
in the reservoir description is made).

D. ULDRICH, S. MATAR, AND H. MILLER

This method can also prove valuable when doing


sensitivity analysis on simulators which have already been
history matched. Different reservoir descriptions which
give substantially the same degree of match are equally
likely to describe the actual reservoir. Forecasts based on
these different reservoir descriptions are also equally probable,
and can help to define the uncertainty range around
official forecasts.
Theory And Definitions
The Standard Approach To History Match Evaluation
Fundamentally, history match evaluation relies on comparing
a set of surveillance measurements, gathered over the active
life of a field, to the corresponding values predicted by a
dynamic reservoir simulator. Typically, in oil fields, the types
of data gathered are well rates (production and injection),
watercuts, gas-oil ratios (GOR), static bottomhole pressures
(SBHP), and possibly zone pressures using wireline tools (for
example, RFT), and water saturations using pulsed neutron
tools (for example, TDT). Other data types can be monitored
as well, such as tracer concentrations or production and
injection pressures and profiles. In gas fields, the surveillance
data gathered will be similar. In best practice, numerical
simulators are built with a set of outputs in mind, including
those to be compared with observed data in history
matching. Thus, for each observed surveillance measurement,
there will be a corresponding simulator value. The
accumulated number of such observations defines the size of
the history match database.
The quality of the measured data must be considered. If
possible, inaccurate measurements should be corrected. But
often the only quality control option is to decide whether to
include a measurement in the history match dataset exactly as
it was reported, or to remove it altogether. Implicit in
removing observations from the dataset is the conclusion that
the value does not meet a minimum confidence level that it
represents the actual performance of the reservoir at the time it
was measured. For example, a GOR measurement might be
rejected if it is far below solution GOR in an under-saturated
oil reservoir.
However, surveillance data often give qualitative
information about reservoir performance, which could benefit
the history match by guiding the results. In the example of the
erroneously-low GOR value, it might still qualitatively
indicate that the well is producing no free gas. Completely
ignoring that information because it is not accurate enough
may reduce the reliability of the final version of the simulation
(the Match Factor method discussed later offers a means by
which qualitative or imprecise data may provide some benefit
to the history matching process).
Standard practice when evaluating a history match is to
plot the observed values for a particular data type versus time,
along with the corresponding simulator result. Figure 1 shows
a schematic example of such a plot. The observed values are
plotted as Xs on the date of measurement. The predicted
values from the simulation are plotted as a step-function, in

SPE 75223

which each step represents the period of a time step in


the simulation.
Typically, these plots are scanned by the simulation
engineer, looking for how closely the simulator results
approach the surveillance data, and whether the two sets of
data show similar trends. Ideally, the predicted results would
exactly match the field measurements, but it is common for
matches to be assessed on the basis of matching a trend or
change in trend in the surveillance data. Any final conclusions
as to how well the simulation matches the actual field
performance are often a function of the judgment and
experience of the simulation engineer.
Definition Of Deviation Value
In the history match process, there should be a predicted value
calculated by the reservoir simulator corresponding to each
measurement of field performance made as part of a
surveillance program. Extracting the appropriate value from
the simulation output that corresponds in time with the
observed value is best handled by a computer program
designed to process the output from the specific simulator
being used in the history matching process. Once extracted,
the simulator value is used to calculate the deviation value,
defined here to be the calculated simulator value minus the
measured surveillance value. Note that deviations can be
positive or negative.
History Match Evaluation Using The Deviation
Distribution Plot
The purpose of the Deviation Distribution Plot is to represent
the distribution of deviation values for a specific group of
observations. All of the observations in the group must be of
the same data type (e.g. only watercuts, or only static
bottomhole pressures), but they may then be grouped by well
or by area, or they may represent the entire set of deviations
for that data type.
One traditional way of presenting a distribution of these
deviation values is through a histogram in which the total
range of deviations is divided into bins of equal range, and
the fraction of the total number of deviation values which land
in each bin (here called the relative frequency) are plotted
versus the deviation value at the middle of the bin range. This
is illustrated by the Relative Frequency Plot in Figure 2, which
shows a group of static bottomhole pressure deviation values
which have been sorted into bins 25 psi wide. In this example,
for instance, 18% of the values have deviations between 25
and 0 psi.
Although the example in this plot shows a reasonably welldeveloped bell-shaped curve (normal distribution), the shape
is very much a function of the bin size. To avoid any
erroneous conclusions due to inappropriate bin size selection,
we propose presenting the data in a different way: the
Deviation Distribution Plot. First, all of the deviation values
are sorted from smallest to largest (most negative to most
positive). Then, each value is assigned a relative frequency
equal to its fraction of the total number of values. For
example, if there are 20 deviations in a group, then each

SPE 75223

USING STATISTICS TO EVALUATE A HISTORY MATCH

deviation represents 5% (or 1/20) of the total. In the Deviation


Distribution Plot, the sum of the relative frequencies of all
deviations from the minimum value up to a particular
deviation (here called the cumulative frequency) is plotted
versus that deviation value.
The data used to make Figure 2 have been replotted in this
way to produce the line labeled Case A in the Deviation
Distribution Plot in Figure 3. The minimum deviation value is
plotted at zero on the cumulative frequency axis because none
of the deviation values are less than that minimum value.
Similarly, the median deviation value plots at 50% on the
cumulative frequency axis because half of the deviation values
are less than this value. Strictly speaking, this is true only for
groups with an even number of values, but the point at which
the curve crosses the 50%ile line on the cumulative frequency
axis can be used as a measure of median value, even if no
point is actually plotted there. By convention, the maximum
deviation value is plotted at 100% on the cumulative
frequency axis, even though this is not strictly consistent with
the definition of cumulative frequency (this convention is
suffers most with small sets of deviation values).
Because some datasets can contain thousands of deviation
values, we choose to plot deviation values interpolated at a set
number of equally-spaced points along the cumulative
frequency axis, typically every 5%, from 0% to 100%. It has
been found that results displayed at this interval (such as in
Figure 3) adequately define the shape of the curve.
Inspection of the Case A curve in Figure 3 reveals some
important features. The median deviation value (at 50% on the
cumulative frequency axis; marked by the square symbol) is
an indication of any systematic bias in the simulation
prediction. If the median deviation value is very positive, the
simulator is over-predicting values for the data type group. A
negative median deviation value indicates that the simulator is
under-predicting values.
In the Case A curve in Figure 3, the median deviation
value is very close to zero. Although the median deviation
value for Case B (triangle symbol) is just as close to zero
deviation as Case A, its curve shape shows greater departure
from the vertical line at zero deviation. Thus, overall, it is not
as well matched as Case A. The match for Case C is worse
than the other two cases because of 1) the systematic underprediction indicated by its large negative median deviation
value (circle symbol), and 2) the greater departure of its curve
from the vertical zero deviation line. If these cases had been
generated as part of a history matching exercise, it would be
easy to select Case A as the best matched case (which would
then be used as the basis for new cases).
Ideally, all deviations values would fall on the vertical line
at zero deviation in this plot. In practice, this is almost never
the case. The practical goal in history matching then becomes
reducing the area between the curve and that vertical line. At
present, we are not proposing any definitive standard for
declaring a simulator sufficiently matched based on the
departure of this curve from the zero deviation value line, but
it easy to visually compare one case to another using this plot,

selecting the case with the median value and the overall curve
shape closest to zero deviation.
As the examples of Cases A, B and C show, by plotting
results from two simulation runs in a single Deviation
Distribution Plot (such as successive cases in the history
match process, or different sensitivity cases after history
matching), it quickly becomes clear whether one case is better
matched than the other. If there is no clearly superior case
during history matching, it may be an indication that the
process has reached its useful end. Finding more than one
reservoir description that matches past performance equally
well (such as when running sensitivity cases after history
matching), allows the simulation engineer to assess the impact
of reservoir description uncertainty on future predictions.
Future predictions of rates and recoveries from simulators
equally matched to past performance should be equally
probable, which is useful when defining the uncertainty range
for those future predictions.
History Match Evaluation Using The Deviation Band Plot
Finding positive or negative bias in the match is important in
comparing two history match cases and deciding what changes
in the reservoir description to make next in the history match.
But evaluating the total amount of deviation, regardless of
whether it is positive or negative, can be useful as well. This is
equivalent to repeating the method used to build the Deviation
Distribution Plot, this time using the absolute value of the
deviations instead of the actual values.
A convenient way to visualize this concept is to picture
each of the observation values plotted in Figure 1 with a
vertical bar defining the same deviation range above and
below each observed surveillance value. At each step in the
deviation band analysis, the length of the bars defines a
deviation band above and below the observed data, in the
way illustrated by Figure 4. For any particular deviation band
width (length of the bars), a certain number of the predicted
values from the simulation will fall within the deviation band.
If the width of the deviation band is zero, only those predicted
values which perfectly match the observed values will fall
within the deviation band. As the width of the deviation band
increases, a larger and larger percentage of the predicted
values fall within the band. When the band width equals the
maximum of the absolute values of all of the deviations, the
band will enclose 100% of all of the predicted values.
A Deviation Band Plot presents the relationship between
the width of the deviation band and the cumulative percentage
of the predicted values which fall within the band. Case A in
Figure 5 shows the same static BHP deviation data from
Figure 3, replotted in the Deviation Band Plot format.
Although usually described in a different way from that
presented here, the concept behind the deviation band has
often been used by simulation engineers to specify a
quantifiable requirement for declaring a simulator sufficiently
history matched. The requirement is usually expressed as a
target: to match a certain percentage of observations within a
certain range (positive or negative). In this example, 82% of
the predicted values are within 100 psi of the observed

D. ULDRICH, S. MATAR, AND H. MILLER

values (square symbol). The specific target chosen is often


based on the past experiences of the simulation engineer, and
acknowledges the fact that it is practically impossible to match
all of the surveillance data perfectly.
Like the Deviation Distribution Plot described above, the
Deviation Band Plot can be used to compare the match for a
single data type from two separate cases. In Figure 5, the
dashed line shows that Case D is a much worse match than
Case A because its curve is significantly farther from the
vertical line at zero deviation (Y axis).
Assuming that the distribution of deviations approximates
a normal, bell-shaped distribution (as Figure 2 suggests it
might), the deviation band width which encompasses about
67% of the data can be considered to be the apparent standard
deviation (ASD) of the distribution. In Case A, that deviation
band is about 60 psi, as indicated by where the horizontal
arrow at 67% cumulative frequency intersects the Case A
curve. The ASD for Case D is much larger than for Case A,
because the same horizontal arrow intersects the Case D curve
at about 500 psi. Thus the ASD value can be used as a single
measure of a match, and as a quick comparison between
different simulation runs (but only for the same data type).
The Deviation Band Plot, unlike the Deviation Distribution
Plot described above, does not have the advantage of showing
whether a systematic bias exists, either positive or negative.
For that, the Deviation Distribution Plot must be used.
Nevertheless, the benefit of the Deviation Band Plot comes
from its ability to demonstrate the intrinsic degree of match,
and to express that degree of match with a single number: the
apparent standard deviation, ASD. As with the previous plot,
combining results from more than one simulation allows for a
quick comparison of the relative match.
History Match Evaluation Using Match Factor
A perfectly history-matched model would provide only zero
deviations. But in practice, that almost never happens. This is
due to a combination of 1) the inherent inaccuracy in the
surveillance measurements, 2) the simplifications of the
process physics in the simulator, and 3) the highly underspecified reservoir description. Therefore, if the surveillance
program is properly designed and executed, and an
appropriately rigorous simulator chosen, the main source for
non-zero deviations is more likely to be the difference
between the reservoir description in the simulator (both fluid
and rock properties) and the properties of the actual reservoir.
Simulations which more closely reproduce past behavior are
presumed to contain a more realistic reservoir description,
which will in turn provide more reliable predictions of
future performance.
History matching a model then becomes an exercise in
reducing deviations by modifying the reservoir description in
the simulator, consistent with sound geologic and reservoir
engineering principles and constraints. The Match Factor
method proposed here integrates the concepts of measurement
uncertainty, simulator limitations, and relative data importance
into a single indicator of confidence in the match, which, in
turn, indicates the confidence that the reservoir description in

SPE 75223

the simulator adequately represents the actual reservoir. Past


attempts to develop automated history matching procedures
have relied on minimizing an objective function based on a
power law function of deviations and weighting factors2-5. The
Match Factor method proposed here is not proposed for that
purpose, but it does use deviation results and weighting factors
to assess the degree of the match.
Definition Of Match Factor
The Match Factor is designed to approximate the confidence
level that a particular predicted value matches the actual field
behavior at the time of the measurement. When the Match
Factor is at its maximum value, it indicates that the simulator
prediction is as matched to the surveillance value as it
effectively can be. Although the deviation may not be zero,
there would be no benefit in moving the deviation value closer
to zero.
Figure 6 shows an example of how a Match Factor curve
might be defined for a watercut observation (expressed as a
decimal fraction). In this case, any deviation value between
-0.05 and +0.05 would be assigned the maximum
Match Factor value, 100. This is equivalent to saying that
the uncertainty in the combination of watercut measurement
method and simulator precision is about 0.05. Forcing a
simulator to match a surveillance measurement more
precisely than its inherent uncertainty does not improve the
history match.
Just as there is typically a range of deviation values at the
maximum match confidence level, there also are deviations
beyond which there is no confidence that the predicted value
represents in any way the actual field behavior. In the example
in Figure 6, those limits are 0.20 (labeled Zero Confidence
Limits). This is interpreted to mean that any deviation less
than 0.20 or greater than +0.20 indicates definitively no
match between the predicted and observed values. Beyond
those limits, the Match Factor is assigned the minimum Match
Factor value of zero. Setting these limits of zero confidence
depends on the type and quality of the data.
The use of 100 as the maximum value is completely
arbitrary, but has been found to be convenient. The goal is to
be consistent (that is, use the same maximum value for all
Match Factors), so that when they are averaged together, if the
result is the maximum Match Factor value, then all
observations are as matched as they can be, within the
uncertainty limits of the data and the tools. Any average
less than the maximum value indicates that some of
the observations are not as well matched as they could be,
and should be the focus for subsequent history
matching modifications.
To sensibly relate Match Factor to deviation value, it is
necessary to understand how the surveillance measurement
was made. This is especially important when setting a
reasonable width range for the portion of the Match Factor
curve at its maximum value (the plateau). Even so, there is
no theoretical requirement that a Match Factor curve even
have a plateau at the maximum value. A single point at the
zero deviation line might be more appropriate in some cases, if

SPE 75223

USING STATISTICS TO EVALUATE A HISTORY MATCH

the measurement technique is extremely accurate and the


simulator very rigorously precise. There is also no requirement
that the Match Factor curve be symmetric around the zero
deviation value point. And although more complex Match
Factor curve shapes could be defined (possibly using more
line segments, or even curved segments), our experience to
date suggests that the trapezoidal shape in the example figure
is sufficient.
If the surveillance measurement technique changes with
time, the shape of the Match Factor curve (especially the
width of the maximum value plateau) may have to change
as well. For instance, early in the producing life of an oil well,
watercut might be measured by assaying salt content in the
produced fluids. If the formation water is very saline, it would
only take a very small water content to increase the salt
content dramatically, and, thus, the uncertainty in the watercut
measurement would be very small. After water breakthrough,
watercut might be measured with a 3-phase separator, and the
uncertainty in the measurement could be much larger. The
Match Factor curve assigned to the early watercut data should
have a much narrower deviation range at the maximum value
than that assigned to the late-time watercut data.
By manipulating the shape of the Match Factor curve, this
analysis can provide benefits to history matching, even when
the surveillance data are valid only in a qualitative sense.
Sometimes, the information provided by the surveillance data
can only be used to set limits on the predicted values coming
from the simulator, rather than defining a precise target value
to match. In these cases, the Match Factor method is much
better suited to quickly assessing whether the simulator results
are within those limits than the deviation-based methods
described before. Appendix A gives a series of related
examples to demonstrate how Match Factors should be used in
this manner.
The Match Factor method is a logical extension of the
Deviation Band concepts discussed above. Figure 7 illustrates
this graphically (as compared with Figure 4). Instead of a
series of vertical bars of equal length centered on each
observed value, which expand or contract in unison to
measure the distribution of deviations, the Match Factor
method applies a more complex but unchanging bar to each
observed datum. On top of each observed value, a vertical bar
is plotted, and then subdivided into three segments. The
central segment (thick line) represents the deviation range
above and below each observed datum value within which a
simulator result would be assigned the maximum Match
Factor value. The other segments (thin lines), above and below
the central segment, extend out to the zero confidence limits.
Unlike the bars in the Deviation Band method, these bars do
not expand and contract to describe the distribution of
deviations. Instead, the Match Factor value is determined by
whether the predicted value falls within the central segment of
the bar (maximum Match Factor value), one of the extension
segments (an intermediate Match Factor) or outside all of the
segments (zero Match Factor value). Note that different
observed values can be assigned Match Factor bars of
different segment lengths, depending on the method used at

the time to make the measurement (such as in the previous


watercut measurement example). As Figure 7 shows, the
goal for history matching should be to adjust simulator
output until the predicted values fall anywhere within the
range of maximum Match Factor values, not to match the
observation exactly.
History Match Evaluation Using Average Match Factors
One advantage that working with Match Factors has over
using deviation values is that they can be arithmetically
averaged to produce a single number which indicates a relative
confidence of match. If the average equals the maximum
Match Factor value, then the simulator is as matched as it can
be, for all practical purposes, and the history match process is
complete. Averages below the maximum value indicate the
relative degree of match (compared to other data types in the
same simulation run, or to the same data types in other
simulation runs). In theory, a single average Match Factor
value could be calculated for the entire observation dataset,
but this has not been found to be as useful to the history
matching exercise as averaging by well and by data type.
An important option in the Match Factor method is the use
of weighting factors when averaging individual Match Factor
values. Relative weighting factors can be used to emphasize
data during critical history match periods, such as matching
water breakthrough times in the water production data. Setting
all weighting factors to zero except during the water
breakthrough period produces an average Match Factor for the
breakthrough period only. Comparing that average to the
unweighted average indicates whether the water breakthrough
period data have been matched to actual performance better or
worse than the water production data as a whole.
Weighting factors can also be used to remove a bias in the
average Match Factor towards values in periods of very
frequent data measurement. Appendix B describes two ways
to use weighting factors to remove that bias.
The benefit obtained from the average Match Factor value
depends on which group of Match Factors have been averaged
together. Typical groupings would be by well or by data type.
Well average Match Factors can be plotted as a bubble map,
showing the locations within the reservoir which are better
matched than others (see Example 3 below). During the
history match process, it is very useful to identify area trends
in the match, to see if a new geologic model is required in
some areas to explain the performance. Comparing bubble
maps from subsequent history match simulation runs indicates
whether the match in these areas is improving significantly or
not. If it is not, the history match process may have reached its
practical end.
Extraction Of Simulator Results
The single most difficult aspect of applying the statistical
techniques proposed here is the extraction of the relevant
results from the simulation. Typically, this involves building a
computer program to post-process the text files generated by a
simulator, searching for specific data values at particular times
in the simulation. Because each simulator outputs its results in

D. ULDRICH, S. MATAR, AND H. MILLER

its own unique format and style, a different post-processor is


required for each simulator.
One advantage this analysis does offer is the fact that all of
the surveillance data used in the analysis are processed only
once, at the beginning of the evaluation process. The
production/injection rate data, pressure data, saturation data,
etc. are collected and stored (usually in a text file), along with
the Match Factor curve descriptions and weighting factors for
each observation. Once completed, these data remain
unchanged throughout the evaluation process (history
matching, sensitivity analysis, etc.). After each new simulation
run finishes, the post-processor program goes back to the same
pre-processed surveillance data. We have found that even for
very large surveillance databases (exceeding 60,000
observations), UNIX-based post-processors can return results
in less than five minutes. These turnaround times were
verified on two of the most widely used commercial
simulation software packages.
Evaluating A History Match For A Large Kuwait Oil
Field
In 2000, Kuwait Oil Company was engaged in building a new
full-field simulator for a large undersaturated oil reservoir.
The previous simulator for this field was being replaced for
several reasons, including the recognition that data gathered
by more recent drilling supported updating the previous
reservoir description. There was a need to demonstrate that the
new model was at least as reliable as the previous version.
One crucial part of satisfying that goal required comparing the
history match in the new model to that in the old. The
evaluation techniques described in this paper were developed
for that purpose.
The oil field in question had been in production for well
over 20 years. With over 150 wells drilled during that time,
there was a very large volume of production data available to
match to simulator results. Because of this large surveillance
database, traditional history match evaluation methods would
have been very cumbersome and highly subjective. The
techniques proposed here were used instead.
Although no fluid injection had yet been implemented in
this field, there was a significant amount of water movement
due to aquifer influx noted in openhole electric logs and
cased-hole pulsed neutron logs. In addition, ever since
produced water handling facilities had been installed, some
areas showed high water production. Matching water
production in the simulator, then, was a key objective in the
history match to verify water movement.
Reservoir pressure had been principally monitored through
the years by measuring bottomhole pressures in producers
shut-in for 24-48 hours. In addition, openhole repeat formation
testing tools had been occasionally used to measure pressures
in newly-drilled wells. Although reservoir pressure had
declined since the start of production, pressure depletion was
not yet sufficient to take the oil column below the initial
bubblepoint pressure and cause significant amounts of free gas
production. However, small variations in producing GOR

SPE 75223

were expected, caused by areal and vertical variation in


oil composition.
Evaluating the history match of both versions of the fullfield simulation of this reservoir to its past performance
provided the following examples of the value of the statistical
evaluation methods proposed here.
Example 1: Deviation Distribution Analysis Of Static
Bottomhole Pressure By Well
At one point in the history match, the Deviation Distribution
Plot for the static bottomhole pressure data type looked like
the Case A curve in Figure 3. With a median deviation very
near 0 psi, the overall reservoir pressure in the simulator was
well matched to the observed static pressure data.
However, the individual Deviation Distribution Plots for a
group of wells in one area showed another trend. The
curves in Figure 8 show the deviation distributions
of the static bottomhole pressure data from 12 wells.
Although the field-wide match showed neither positive
nor negative bias, the individual wells separated into two
groups: ten of the wells showed negative median deviations
(under-predicted pressures), while two wells (Wells 04 and
08) showed significant positive median deviations (overpredicted pressures).
When it was recognized that all of the wells in the first
group were located high on structure, while the two outlier
wells were downdip (close to the OWC), it became apparent
that the simulator was allowing the down-structure aquifer
(the only pressure source in this depletion drive reservoir) to
over-support the downdip wells, while restricting pressure
support to the updip wells. This suggested that the actual
effective permeability in the reservoir was higher than
that assigned to the simulator. Thus, being able to quickly
and easily compare the deviation results by well was
valuable in first identifying a history match trend, and then
using that information to suggest modifications to the
reservoir description.
Example 2: Deviation Band Analysis Of Misleading
Watercut Data
Sometimes, it is so easy to match the surveillance data that the
important information carried by it is obscured. For example,
for many years in this large Kuwait oil field, there were no
facilities to handle the highly saline produced water, nor was
there any treatment available to remove salt from the produced
fluids. Consequently, whenever a producer began to produce
even 0.001 watercut, it was shut-in immediately (and another
producer was brought on line to maintain oil production).
Consequently, matching water production in this period
was an exercise in suppressing water production in the
simulator. Figure 9 shows that the simulator was doing a good
job of matching most of the watercut data. Over 40% of the
reported watercut values were matched exactly (zero
deviation). The ASD for this simulation was only about 0.04,
which most simulation engineers would probably accept as a
reasonable watercut match.

SPE 75223

USING STATISTICS TO EVALUATE A HISTORY MATCH

Unfortunately, most reservoir simulators can be forced to


produce no water, at least for a certain period of time,
although it may require very draconian alterations to the
reservoir description if the water flow in the simulator is not
similar to that in the reservoir. Simply turning off water
production by any means calls into question the reliability of
the future predictions that are made using that altered reservoir
description, which is, after all, the reason for building the
simulator in the first place.
The key point in this example is that the simulation
engineer must be aware of the constraints placed on operations
and surveillance before drawing conclusions. In this case, in
the period before wells were allowed to produce any water, the
most valuable indicator of water movement in the field was
not watercut, but the date on which a producer was shut-in due
to salt content. Inspecting the simulator output for indications
of water influx near a producer on its shut-in date (such as
increased mobile water saturation in the connected grid
blocks) would have been a more reliable indicator of a good
match to historic performance than matching the watercut
data, no matter how good that match.
Example 3: Match Factor Analysis With A Watercut
Bubble Map
Since water influx represented the major source of reservoir
energy in this reservoir, minimizing water production while
maximizing oil production was a key reservoir management
goal for this large Kuwait oil field. As a result, the full-field
simulation needed to capture actual water movement as
realistically as possible. The simulator was history matched to
watercut performance (after the installation of produced water
handling facilities) to verify water movements. It was
important, then, to identify which regions had poor versus
good watercut match. The Match Factor method was used for
this purpose.
First, a Match Factor curve shape was defined for the
watercut observations (Figure 6). Second, watercut results
were extracted from each simulation run, and Match Factor
values calculated from the deviation results (no weighting
factors were used). Finally, the resulting Match Factor values
were arithmetically averaged by well, and the averages plotted
on a bubble map. Figure 10 shows the results at the end of the
history match process.
The map shows that wells in some large areas of the field
are well matched (smaller circles indicating average Match
Factors near 100). This is principally true in the northern area.
However, in other areas, such as in the southern end of the
field, many wells have significantly lower average
Match Factors (larger circles). If the history matching process
had continued, these areas would be the focus for further
reservoir description modifications aimed at improving the
watercut match.
The map also shows some isolated problem wells. These
are wells with low average Match Factors in areas dominated
by wells with high Match Factors. Special attention should be
paid to these wells, because they may indicate either a
mechanical problem in the well (for example, poor cement

isolation from bottom water) or a localized thief zone.


Remedial action may be needed to control abnormal water
production. As this example shows, this approach to
evaluating history matches can quickly identify anomalous
well behavior, possibly leading to corrective action that adds
value to the project, as well as to the history match.
Example 4: Match Factor Analysis Of Well Rate Controls
A special application of the Match Factor method is evaluating
the control of well rates during history match simulations.
Because well rates are usually known with some confidence,
wells are often simulated at their historic total reservoir rates.
For a producer in an undersaturated oil reservoir, this means
control by specifying total liquid production rate (oil plus
water rate). For a dry gas producer, this would be total voidage
rate (gas plus water rate). A producer in a depleted oil
reservoir would be controlled by total voidage rate, based on
reservoir volumes of produced oil, free gas and water.
It is important to remember that the specified well rates are
only targets, and the simulator may not be able to operate the
wells at those rates, even when the flowing bottomhole
pressure limits are greatly relaxed (set very low for producers,
or very high for injectors). It becomes important to know
whether or not the wells are operating at target rates when
evaluating the history match. The Match Factor method allows
a quick evaluation of this situation.
In this analysis, the target rate is treated as the observed
value. A Match Factor curve (Figure 11) is defined based only
on the inherent precision of the simulator to report well rates.
For instance, when wells are operating at their rate target, the
simulator will normally report their total liquid production
rates (TLPRs) within 1 bpd of that target. The Match Factor
curve assigns a maximum value when the well is operating on
target, and zero whenever the well is off target rate by more
than 2 bpd.
Ideally, when Match Factor analysis is run on the target
versus predicted rates, all of the deviations would be less than
1 bpd, and the well average Match Factors would all be 100.
If any wells fail to make target rate at any time in the history
match period, then the well average Match Factor will fall
below the maximum value. This will quickly become apparent
when plotting average Match Factor by well versus
cumulative percent of wells (Figure 12). This plot is built in
the same way as the Deviation Distribution Plot: the average
Match Factor values by well are sorted from smallest to
largest, and plotted versus the cumulative percentage of the
total well set that each well represents. In this example, 75%
of the wells have an average Match Factor of 100, meaning
that they operate at target rate 100% of the time. This means
that 25% of the wells failed to operate at target rate sometime
during the history match period in the simulation.
If the target rates are specified at roughly the same
frequency throughout the history match period (as they are in
this case), the average Match Factor value is a good estimate
of the percentage of a wells active life that it is operating at
target rate. With that in mind, only about 5% of the wells in
this example are true problem wells, operating below target

D. ULDRICH, S. MATAR, AND H. MILLER

rate more than half of the time (Match Factor less than 50).
Plotting these average match factors on a bubble map will
show whether the problem is well-specific (problem
wells widely scattered amongst other wells), or regional
(problem wells grouped into one or more regions). Finding a
regional trend would suggest there may be a problem with the
reservoir description, causing insufficient reservoir pressure
around the producers in some areas. In this way, Match Factor
analysis of the well control rate performance can provide
valuable data to assess a history match and guide reservoir
description modifications.
Example 5: Match Factor Analysis Of The Overall Match
(Comparing New To Previous Full-Field Simulation)
Because Match Factors are a relative measure of the degree of
history match, they can be compared with Match Factor values
from other data types directly. A bar chart is a convenient way
to summarize the history match evaluation, by presenting the
average Match Factors for each of the data types in the
surveillance dataset. When comparing the match for two
simulator runs, bars representing the average Match Factors
for each data type are plotted side by side.
For the large Kuwait oil field being evaluated here, the
surveillance dataset (for both the new and previous
simulations) included Total Liquid Production Rate (TLPR),
watercut (WC), gas-oil ratio (GOR) static bottomhole pressure
(SBHP), and RFT pressure (RFT). For each data type, Match
Factor curves were developed, using the concepts described
above. Since both simulator versions used essentially the
same surveillance dataset, the same Match Factor curves
were used to evaluate both. Weighting factors were not used in
this example.
The resulting averaged Match Factors for the previous and
new simulator runs (after each were history matched) are
plotted in Figure 13. The TLPR match is very good, as would
be expected for the data type used to control well rates in the
simulation. The watercut is also well matched in both cases,
reflecting the time and effort devoted to it during both history
matching efforts. The GOR match is slightly improved in the
new simulation, a result of improvements in the description of
fluid properties.
The SBHP Match Factor value is somewhat lower in the
new simulator than in the previous version, although
considerable effort was spent on matching that data type
(because of the importance of reproducing the proper level of
aquifer support on predicting future recovery). Although this
drop in average Match Factor is not a large enough drop to
invalidate the new simulation, it indicates the need to review
the methods used to build the reservoir description before this
simulator is updated again.
Although the average RFT Match Factor is very consistent
for both versions of the simulation, it has the lowest value of
the five data types. Part of the reason for this is that the RFT
pressure measurement is one of the most accurate and least
ambiguous surveillance measurements made. The depth is
defined by wireline techniques to within a few feet. The
reservoir pressure is typically measured by a very accurate

SPE 75223

gauge to within a few psi. Pressure transients around the tool


are typically very small, especially when compared to those
generated around active production or injection wells (which
must be accounted for with sophisticated transient analysis
techniques). Thus, the pressure reported in an RFT survey is
inherently more accurate than most other measurements in the
surveillance database. Consequently, the Match Factor curve
used on RFT pressures has a much narrower plateau width
at the maximum Match Factor value than that used on the
SBHP data, making it much harder for the simulator to
produce high Match Factor values. Despite this difficulty,
Match Factor analysis is still a valid comparison of the RFT
match between the previous and new simulations, especially in
a relative sense.
Limitations Of The Proposed Methods
Although the methods described in this paper have proven to
be useful tools in the evaluation of the history matching of
large full-field models, it is important to keep in mind their
limitations as well:
1. These methods do not eliminate the need for the
traditional evaluations techniques applied to history match
evaluation, such as plotting simulator results and
surveillance data together. These methods help the
process by allowing the simulation engineer to quickly
identify and focus on the most important problem areas.
2. The methods proposed here give only a relative
comparison of the degree of match between two or more
simulation runs on simulations of the same reservoir
matched to the same (or very similar) observed
surveillance data. No specific degree of match (for
example, a particular Match Factor value) is proposed
here as a minimum requirement for declaring a
model sufficiently matched, except for the case in which
all deviations are within the range of maximum Match
Factor value.
4. No direct correlation has been established between the
Match Factor value and the probability that the model is
correctly simulating the recovery process in the reservoir.
It is reasonable to presume, however, that models with
higher resulting Match Factors will yield more probable
predictions of future behavior. This is consistent with the
basic principle of history matching.
5. The methods presented here do not readily lend
themselves to assessing the degree of match as time
progresses in the history match simulation. The deviationbased methods need data grouped into statistically
significant numbers of deviations to work well, which
makes deviation-by-deviation evaluation impossible. One
useful technique is using weighting factors to isolate and
calculate average Match Factors only for surveillance
data near the end of the history match period.
This focuses the analysis effort on improving the
transition between simulating the past (history match) and
the future (forecasting).

SPE 75223

USING STATISTICS TO EVALUATE A HISTORY MATCH

Conclusions
In this paper, we have proposed using three simple statistical
methods for evaluating the degree to which a simulation is
history matched to past performance, as defined by the
surveillance data collected in the field. The goal is to quickly
extract simulator results and generate graphical displays which
can be used to 1) guide subsequent history matching iterations,
2) determine when further history matching no longer adds
value, and 3) compare how closely two or more simulations
match the surveillance data.
All three methods rely on extracting specific results from a
simulation run and calculating the deviation of that value from
the corresponding surveillance observation value. The
Deviation Distribution Plot shows whether the simulator is
over or under-predicting surveillance data of a particular type,
and the general degree of that match. The Deviation Band Plot
also demonstrates the degree of match and provides a
quantitative measurement of the match: the apparent standard
deviation (ASD). In Match Factor analysis, each deviation is
converted to a relative measure of the confidence that the
simulator actually reproduced the particular reservoir
performance at the time the surveillance measurement was
made. Arithmetically averaging Match Factors by well
allows bubble maps to be created that quickly show which
areas are better matched. Bar charts of Match Factors
averaged by data type summarize the entire history match
evaluation for one or more simulation runs. Weighting factors
are used to emphasize the impact of key Match Factor values
on the average.
Because of the volume of data involved in history
matching simulators with large surveillance databases, a
simulator-specific computer program is required in order to
extract the required data from the simulators output files.
However, once the extraction routine is constructed, the
analyses proposed here can be carried out in a matter of
minutes following the completion of a simulation run. That
makes it well suited to the iterative nature of history matching.
In addition, the ability to evaluate the match well-by-well,
focuses the process on those areas which need it the most.
Acknowledgements
The authors wish to thank the management of Kuwait Oil
Company for their support during the development and
implementation of the techniques described in this paper.
Specifically, our thanks go to Mr. Ali Al-Shammari, the
General Superintendent of the West Kuwait group in the
Fields Development Department of the Exploration And
Development Division in the Kuwait Oil Company (KOC).
Also, our thanks go to Mr. Jaber Al-Hunaif and Mr. Fahad AlMehdadi, the Superintendents responsible for the original
development work and implementation.
In addition, the authors also wish to thank the management
of BP Kuwait and Chevron Kuwait for their support. Finally,
from among the numerous colleagues in KOC, BP Kuwait and
Chevron Kuwait who contributed their input and suggestions,
we wish to draw special attention to the support provided by
Dr. Robert Merrill and Alan Clark, both KOC secondees from

BP Kuwait, who provided valuable ideas, and the benefit of


their many years of simulation experience.
Nomenclature
ASD
Cumulative Frequency
Data Type
Deviation
GOR
Match Factor
Relative Frequency
RFT
SBHP
Sorw
Swi
Swmax
TDT

Apparent Standard Deviation


Percentage Of Dataset With Value
Less Than A Limit
Category Of Surveillance
Measurement
Predicted Value Observed Value
Gas-Oil Ratio
Relative Confidence Of Match
(based on deviation value)
Percentage Of Dataset With A
Value Between A Pair Of Limits
Repeat Formation Tester
Static Bottomhole Pressure
Residual Oil Saturation
Initial Water Saturation
Maximum Water Saturation
Thermal Decay Time Pulsed
Neutron Log

References
1.

2.

3.

4.

5.

Williams, M. A., Keating, J. F., and Barghouty, M. F.: The


Stratigraphic Method: A Structured Approach to History
Matching Complex Simulation Models, Paper SPE 38014
presented at the 1997 SPE Reservoir Simulation Symposium,
Dallas, Texas, June 8-11.
Grussaute, Th., and Gouel, P.: Computer Aided History
Matching of a Real Field Case, Paper SPE 50642 presented at
the 1998 SPE European Petroleum Conference, The Hague, The
Netherlands, October 20-22.
MacMillan, D. J., Pletcher, J. L., and Bourgeois, S. A.:
Practical Tools To Assist History Matching, Paper SPE 51888
presented at the 1999 SPE Reservoir Simulation Symposium,
Houston, Texas, February 14-17.
Portella, R. C. M., and Prais, F.: Use of Automatic History
Matching and Geostatistical Simulation to Improve Production
Forecast, Paper SPE 53976 presented at the 1999 SPE Latin
American and Caribbean Petroleum Engineering Conference,
Caracas, Venezuela, April 21-23.
Schulze-Riegert, R. W., Axmann, J. K., Haase, O., Rian, D. T.,
and You, Y. -L.: Optimization Methods for History Matching
of Complex Reservoirs, Paper SPE 66393 presented at the
2001 SPE Reservoir Simulation Symposium, Houston, Texas,
February 11-14.

Appendix A: Establishing Match Factor Curves For


Qualitative Data (TDT Example)
On occasion, surveillance data will have a large inherent
uncertainty, to the point of being only qualitative, but still
have valuable information regarding the behavior of the
reservoir. Attempting to match the measured value exactly is
not justified, but discarding the observations altogether may
mean ignoring important indications of reservoir processes.
One example is the use of TDT water saturation logs.
Sometimes, they can qualitatively indicate water movement in
a reservoir, even though water saturation values may not be

10

D. ULDRICH, S. MATAR, AND H. MILLER

calculated with acceptable accuracy. In these cases, the Match


Factor method can use the qualitative information
extracted from a TDT survey to determine the degree of match
of the simulation results to the limits imposed by the
TDT interpretation.
In this example, each TDT survey is partitioned into
sections corresponding to the layers in the simulator. Within
each layer, the average water saturation is calculated using
whatever assumptions are appropriate (accuracy is not the
critical issue). Then the Match Factor curve is defined for that
single observation based on the degree of water encroachment
which is observed, using the case which most closely
approaches one of the following examples.
Note: the variable used in this example to define the Match
Factor curves is the grid block water saturation, varying
between Swi and Swmax = 1-Sorw. Once the Match Factor
curve is established in this way for a particular grid block, it
must be converted back to a deviation basis before the analysis
can be run on the simulator results. This requires knowing the
Swi and Swmax values assigned to each grid block in the
TDT survey, which can be extracted if necessary from the
simulator initialization.
Case 1: Interval Completely Unswept
In this case, the character of the TDT log suggests that very
little mobile water has displaced the original hydrocarbon
phase. This situation might look like the TDT log in the left
half of Figure A 1. The goal of the history matching process in
this case is for the simulator to predict a water saturation in the
corresponding grid block that is very close to Swi (which is
defined at simulator initialization). The Match Factor curve on
the right of Figure A 1 accomplishes this goal. The width of
the maximum Match Factor value range should be
approximately the uncertainty in the TDT saturation
calculation method. The Match Factor curve falls rapidly from
the maximum value to zero as block Sw increases. Thus,
if the simulator calculates any significant water influx, the
Match Factor value will quickly drop from its maximum to
close to zero.
Case 2: Interval Completely Swept
In this case, the TDT response is interpreted to mean that the
interval is close to being completely swept by water. The
objective for the simulation is to calculate an average water
saturation close to the residual oil condition (Swmax in this
case). Figure A 2 presents both an example TDT log (on the
left) and Match Factor curve (on the right) for this situation.
This time, the range for the maximum Match Factor value
should be wider than for Case 1 above because of the greater
uncertainty in the actual water saturation (that is, whether the
zone is actually at completely residual conditions).
Conversely, the TDT log character leaves little confidence
that the actual water saturation could be anywhere close to
Swi (that is, that the interval could be completely unswept).
Therefore, the Match Factor value assigned to a simulator
result near Swi must be zero. As the simulator water saturation
climbs above some intermediate simulator result (set to

SPE 75223

midway between Swi and Swmax in this example), the Match


factor value begins to increase from zero towards its
maximum value. Experience to date shows that rigorously
defining these zero confidence points does not significantly
change the final interpretation of the Match Factor results. The
key is to set the maximum Match factor value for any
simulator result near Swmax, and a minimum value for any
saturation close to Swi.
Case 3: Water Contact Within Layer
One task that TDTs do well is identifying water contacts
within layers (such as the example in Figure A 3). Above the
contact, the water saturation is interpreted to be very close to
Swi. Below the contact, the layer appears to be well swept,
approaching residual oil saturation. Unless the simulation is
run in vertical equilibrium mode, the only information the
simulator has concerning a possible contact inside a grid block
is a calculated Sw intermediate between Swi and Swmax.
The Match Factor curve in Figure A 3 shows the
recommended way to handle this situation. The average water
saturation for the entire interval covered by the grid block is
approximated from the TDT log. Around this estimated value,
a wide range is defined for the maximum Match Factor value,
taking into account the combined uncertainty in the TDT Sw
calculation and the residual conditions below the contact. The
zero confidence limits are set at or near the Sw limits for the
grid block (Swi and Swmax). In this way, the assigned Match
Factor value will be maximum when the simulator predicts an
intermediate Sw, and zero when it predicts no water influx
(near Swi) or residual conditions (near Swmax). In this way,
the qualitative limits imposed by the TDT results (that is, the
grid block Sw is greater than Swi and less than Swmax) have
been translated into a quantifiable test of the simulator results.
Case 4: Complex Water Saturation Distribution
Even more complex water saturation situations can be
imagined within the single model layer interval. One objective
in designing a layering scheme for a simulator is to capture as
many of the independent flow units as possible in separate
layers, but that is not always practical, especially when using
vertical upscaling. Water fingering through thin, high perm
layers can easily be missed when building a model, but appear
readily in a TDT log. Assuming that the layering scheme
cannot be changed to reflect the new flow unit behavior, the
existing model layer must attempt to predict the gross flow of
the fluids.
As the example in Figure A 4 shows, there could be a
transition zone as well as a contact in the interval defined by a
model layer. The recommended method for evaluating the
match is very similar to that in the previous case. Again, the
intent is to assign a zero Match Factor value to predicted water
saturations near initial or residual conditions, and maximum
match factor to Sw near the calculated average in the layer.
The difference in this case would be to increase the width of
the range of maximum Match Factors values to reflect the
increased uncertainty of the actual Sw in the reservoir.
Simulators showing any intermediate water saturation

SPE 75223

USING STATISTICS TO EVALUATE A HISTORY MATCH

condition will be assigned relatively high Match Factor values,


indicating that the simulator results are within the limits
defined by the TDT results.
Summary
This series of examples illustrate how the Match Factor
method can make use of the qualitative information often
available from imprecise or qualitative surveillance data. It
shows that there are practical limits to the process of matching
simulator results to field measurements. There is no value in
matching surveillance results exactly. Conversely, there is
value in applying even qualitative constraints to the simulation
results. This is a benefit often lost by ignoring imprecise
measurements altogether.
Appendix B: Using Weighting Factors To Remove
Sample Frequency Bias
One common application for the use of weighting factors
when averaging Match Factor values is removing the bias
caused by measuring surveillance data at very different
frequencies during the history match period. Two methods are
proposed here, depending on whether the data type is defined
1) only when a well is active, or 2) at any time in the history
match period. The following two examples illustrate these two
proposals. In both cases, the weighting factors are based on
the time period in which the individual surveillance
measurement best represents reservoir behavior.
Case 1: Data Types Defined Only When Wells Are Active
Figure B 1 shows an example for a data type which is defined
only when a well is active. This data type might be watercuts,
GORs, or well rates. In the example, a production well is
active for two periods in the history match period, and shut-in
between these two active periods. During the first producing
time, surveillance measurements are made four times (marked
by the star symbols labeled 1-4). During this period, the
simulator executes three time steps, represented by the step
function curve in Figure B 1.
The weighting factor assigned to each surveillance
measurement is the segment of the producing time period in
which the observation occurred. The boundaries for these time
segments are set at the midpoint times between the
observations. For observations at the beginning and end of the
producing periods, the segments are restricted to the producing
period. In this example, the weighting factor for the
surveillance measurement 1 is the time period starting at the
beginning of Producing Time 1 and ending at a time half way
between the times for measurements 1 and 2. For
measurements 2 and 3, the weighting factors are defined
purely by the midpoint times with their closest neighbor
measurements. The weighting factor period for measurement 4
starts at the time midway between measurements 3 and 4, and
ends at the end of Producing Time 1. The duration of time
periods T1-T4 in Figure B 1 illustrate the weighting factors
that should be assigned to measurements 1-4. Note that since
weighting factors are used in a relative sense, they may be

11

expressed in any unit of time that is convenient, as long as it


is consistent.
In Producing Time 2, the surveillance measurements are
made less frequently than in Producing Time 1. Unless
weighting factors are used, the average Match Factor value for
this well will be dominated by the values in Producing Time 1.
To offset this effect, weighting factors for measurements 5 and
6 are calculated in the same manner as the earlier
measurements, and are illustrated by the time segments T5
and T6.
Case 2: Data Types Defined Throughout The History
Match Period
Some data types can be measured at any time during the
history match period, regardless of whether a well is active or
not. Typically, these are observations of reservoir conditions,
such as static bottomhole pressures, saturations, and RFT
pressures, rather than well conditions (although the
observations are usually tied to a particular well, in order to
associate it with a location in the reservoir).
The method proposed for assigning weighting factors to
these surveillance measurements is the same as in Case 1,
except that the time period segments are not restricted to the
intervals of well activity. Figure B 2 presents an example for a
data type measured in the same well as in Figure B 1 (note the
same pre-production and shut-in time periods), but the
corresponding simulator output is defined and calculated
throughout the history match period. In fact, in this example,
measurements 1 and 5 were made during periods when the
well was shut-in.
In this case, measurements 2-4 have a higher sampling
frequency than 1, 5 or 6. Consequently, they are assigned
smaller weighting factors. Measurement 5 receives a large
weighting factor only because the next measurement, 6, occurs
so long after it. The simulation engineer should review this
automatic assignment of weighting factors to verify that the
measurement is truly representative of the entire period. In
some cases, it may be appropriate to impose a maximum
weighting factor, say 365 days, so that very infrequent
measurements do not bias the result unrealistically because of
the very large weighting factors they might be assigned.
Similarly, a minimum weighting factor (for instance, 1 day)
might also be needed if some of the measurements are
made simultaneously.

12

D. ULDRICH, S. MATAR, AND H. MILLER

SPE 75223

Data Type Value

Observed
Values

Predicted
Values
Deviations

Time

Relative Frequency (%)

Figure 1: Example History Match Plot Showing Deviations

25
20

Bin Size: 25 psia

15
10
5
0
-300

-200

-100

100

SBHP Deviation (psia)


Figure 2: Example Relative Frequency Plot For Static BHP

200

300

Cumulative Frequency
(%)

SP3E 7522

USING STATISTICS TO EVALUATE A HISTORY MATCH

100
90
Case C
80
70
60
50
40
30
20
10
0
-800 -600 -400 -200

13

Case B

Case A

200 400 600 800

SBHP Deviation (psia)


Figure 3: Example Deviation Distribution Plot For Static BHP

DATA TYPE VALUE

Observed
Values

Predicted
Values
Deviation
Band Bars

TIME
Figure 4: Example History Match Plot Showing Deviation Band

D. ULDRICH, S. MATAR, AND H. MILLER

Cumulative Frequency
(%)

14

100
90
80
70
60
50
40
30
20
10
0

SPE 75223

Case A

Case D
0

100

200

300

400

500

600

700

SBHP Deviation Band (psia)


Figure 5: Example Deviation Band Plot Comparing Alternate Cases

Match Factor

100
90
80

Maximum
Value

70
60
50
40
30
20
10
0
-0.3

Zero
Confidence
Limits

-0.2

-0.1

0.0

0.1

0.2

Watercut Deviation (fraction)


Figure 6: Example Match Factor Curve For Watercut

0.3

USING STATISTICS TO EVALUATE A HISTORY MATCH

15

Observed
Values

DATA TYPE VALUE

SP3E 7522

Predicted
Values
Match
Factor
Bars

TIME

Cumulative Frequency
(%)

Figure 7: Example History Match Plot Showing Match Factor Bars

100
90
80
70
60
50
40
30
20
10
0
-600 -400 -200

Well 01
Well 02
Well 03
Well 04
Well 05
Well 06
Well 07
Well 08
Well 09
Well 10
Well 11
Well 12

200 400 600

SBHP Deviation (psia)


Figure 8: Example Deviation Distribution Plot Comparing Several Wells

D. ULDRICH, S. MATAR, AND H. MILLER

Cumulative Frequency
(%)

16

100
90
80
70
60
50
40
30
20
10
0

SPE 75223

40% Of Data At
Zero Deviation

0.0

0.2

0.4

0.6

0.8

1.0

Watercut Deviation Band (fraction)


Figure 9: Example Deviation Band Plot For Restricted Watercut Case

Well Location

Problem Well

MF 95-100

Well
Matched
Area

MF 90-95
MF 80-90
MF 60-80
MF 30-60
Poorly
Matched
Area

MF 0-30

Figure 10: Example Bubble Map Showing Average Watercut Match Factors By Well

SP3E 7522

USING STATISTICS TO EVALUATE A HISTORY MATCH

17

Match Factor

100
90
80
70
60
50
40
30
20
10
0
-4

-3

-2

-1

TLPR Deviation (bpd)


Cumulative % Of Wells

Figure 11: Match Factor Curve For Total Liquid Production Rate Example

100
90
80
70
60
50
40
30
20
10
0

75% Of Wells
At Maximum
Match Factor

5% Of Wells
At Very Low
Match Factor

10

20

30

40

50

60

70

80

90 100

Average Match Factor (By Well)


Figure 12: Distribution Plot For Average Well Match Factors (Total Liquid Production Rate Example)

Average Match Factor

18

D. ULDRICH, S. MATAR, AND H. MILLER

SPE 75223

98 99

100

83 85
75

80

81
66

60

53

52 51

40

Previous
New

20
0
TLPR

WC

GOR

SBHP

RFT

Data Type
Figure 13: Average Match Factor Summary Bar Chart

Completely
Interval
Completely
Unswept:
Unswept:
Swmax
Sorw

TDT
SwSw

Swi

Match Factor
Match
Factor

Model Layer

100

Figure A 1: TDT Example #1 Interval Completely Unswept

Swi

Block Sw

Sorw
Swmax

SP3E 7522

USING STATISTICS TO EVALUATE A HISTORY MATCH

Completely
Interval
Completely
Swept: Swept:
Swmax
Sorw

TDT
SwSw

Swi

Model Layer

Match Factor
Match
Factor

100

Swi

Block Sw

Sorw
Swmax

Figure A 2: TDT Example #2 Interval Completely Swept

Contact
In Layer:
Water
Contact
Within Interval :
Swmax
Sorw

TDT
SwSw

Average Sw
From Log

Swi

New OWC

Figure A 3: TDT Example #3 Water Contact Within Interval

Match Factor
Match
Factor

Model Layer

100

Swi

Block Sw

Sorw
Swmax

19

20

D. ULDRICH, S. MATAR, AND H. MILLER

SPE 75223

TransitionWater
Complex
Zone:Saturation Distribution:
Swmax
Sorw

TDT
SwSw

Average Sw
From Log

Swi

Match Factor
Match
Factor

Model Layer

100

Transition
Zone

New OWC

Swi

Block Sw

Sorw
Swmax

Figure A 4: TDT Example #4 Complex Water Saturation Distribution

Data Type Value

T1

T2

T3

T4

T5

T6

5
4

1
PreDrill
Period

6
Shut-In
Period

Producing Time 1

Producing Time 2

Figure B 1: Weighting Factor Example #1 For Data Types Defined Only When Wells Active

Time

SP3E 7522

USING STATISTICS TO EVALUATE A HISTORY MATCH

Data Type Value

T1

T2

T3

T4

T5

2
1

T6

4
PreDrill
Period

21

5
Shut-In
Period

History Match Simulation Time

Time

Figure B 2: Weighting Factor Example #2 For Data Types Defined Throughout History Match Period

You might also like