SPE 77597

Identification of Successful Practices in Hydraulic Fracturing Using Intelligent Data

Mining Tools; Application to the Codell Formation in the DJ Basin
Mohaghegh, S.D., Popa, A., Gaskari, R., Ameri, S., West Virginia University, and Wolhart, S. L., Pinnacle Technologies
Copyright 2002, Society of Petroleum Engineers Inc.
This paper was prepared for presentation at the SPE Annual Technical Conference and
Exhibition held in San Antonio, Texas, 29 September2 October 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
In a detail data mining study about 150 wells that have been
completed in the Codell formation, DJ Basin, have been
analyzed to identify successful practices in hydraulic
fracturing. The Codell formation is a low permeability
sandstone within the Wattenburg field in the DJ Basin of
Colorado. Since 1997 over 1500 Codell wells have been
restimulated. As part of a Gas Research Institute restimulation
project 150 wells were studied to optimize candidate selection
and identify successful practices.
Hydraulic fracturing is an economic way of increasing gas
well productivity. Hydraulic fracturing is routinely performed
on many gas wells in fields that contain hundreds of wells.
During the process of hydraulically fracturing gas wells over
many years, companies usually record the relevant data on
methods and materials in a database. These databases usually
include general information such as date of the job, Service
Company performing the job, fluid type and amount, proppant
type and amount, and pump rate. Sometimes more detail
information may be available such as breakers, additives,
amount of nitrogen, and ISIP to name a few.
These data are usually of little use in complex 3-D hydraulic
fracture simulators. These models require additional and more
detailed information. On the other hand, the collected data
contain valuable information that can be processed using
virtual intelligence tools. The process covered in this paper
takes the above-mentioned data and couples it with general
information from each well (things like latitude, longitude and
elevation), any information available from log analysis and
production data. The conclusion of the analysis is a set of

successful practices that has been implemented in a particular

field and recommendations on how to precede with further
hydraulic fracture jobs.
In this paper the results of applying this process to about 150
Codell wells during the GRI sponsored project is presented.
This process provides an important step toward constructing a
comprehensive set of methods and processes for data mining,
knowledge discovery, and data-knowledge fusion from data
sets in oil and gas industry.
Introduction
Patina Oil and Gas has been very active in the DJ basin in
recent years. They have been one of the most active operators
in the United States in identifying and restimulationg tigh gas
sand wells. Patina has over 3,400 producing wells in the basin,
and has restimulated over 230 Niobrara/Codell completions so
far. Furthermore, it is estimated that the results they are
achieving in terms of incremental recoveries are up to 60%
better than other operators.
Studies and analysis such as the one being presented in this
paper has the potential to help operators like Patina Oil & Gas
to increase their chance of success even to a higher
percentage. It also has the potential to help other operators in
increasing their chances of success in DJ Basin or any other
locations throughout the North America. This stuy is probably
one of the most comprehensive analyses of its kind ever to be
performed on a set of wells in the United States.
In this technical paper the authors intention is to introduce
this new and novel methodology in its entirety and present as
much of the results as the page limitations of this paper
allows. Please note that due to the comprehensive nature of
this methodology many of the topics cannot be discussed in
much detail. It is our intention to introduce these topics in
much more detail in series of upcoming technical papers.
Methodology
The process of Successful Practices Identification using
state-of-the-art data mining, knowledge discovery and dataknowledge fusion techniques1 includes the following five
steps. In order to comprehensively cover the theoretical

S. MOHAGHEGH, A. POPA, R. GASKARI, S. AMERI & S. WOLHART

background of each step involved in this process, a separate

paper may be needed for each. Some of the ideas have been
introduced in the past and are referred to in the references.
Detail on other topics will be the subject of future papers. In
this article the goal is to provide a view of the methodology as
a whole, and therefore, authors will simply introduce and give
a brief explanation of each of the topics to clarify their role in
the process.
Step One: Data Quality Control
The process starts with a thorough quality control of the data
set. During this process the outliers and their nature (are they
really valid data elements or are they the result of human error
either in measurement or in recording the values?) as well as
missing data are identified, and using advanced intelligent
techniques,2,3,4 the data set is repaired. It is important to note
that the repair of the data set at this stage of the analysis is
aimed at rescuing the remaining data elements in a particular
record (a data record here is referred to a row of the data
matrix) that includes many features (features are the columns
of the data matrix). The goal is not to magically find the
piece of the missing data or substitute the outlier with the
correct value. The goal is merely to put the best possible value
in place of the missing data or the outlier that would allow the
analysis to continue without losing the information content
that exists in the rest of the features in a particular data record.
Moreover, a new and novel methodology has been developed
in order to identify and eliminate erroneous data records from
the data set. These techniques, verification of their accuracy
and how they are implemented are subjects of a future paper.
Step Two: Fuzzy Combinatorial Analysis
The second step of the process is a complete Fuzzy
Combinatorial Analysis FCA that examines each feature in
the data set in order to identify its influence on the process
outcome5. The process outcome is a feature (usually a
production indicator such as cumulative production, 5 year
cum., 10 year cum., best 12 months of production, etc.) in the
data set that is designated to identify the success of the
practices in a field. For example if 5 year cumulative gas
production is selected to be the process outcome, then a high 5
year cum. would indicate performance of good practices for
that particular well. During that Fuzzy Combinatorial
Analysis each features influence is examined on the process
outcome both individually and in combination with other
features. This is due to the fact that influence of a particular
feature (say a fracturing fluid) on the process outcome may be
altered once it is combined with the effects of other features
(say specific additives) that are present in the process.
Therefore it is important to perform the analysis in a
combinatorial fashion (hence the name combinatorial analysis)
in order to reveal the true influence of the features present in
the data set on the process outcome.
A note of caution is in order here. Many commercial, off-theshelf neural network software applications claim to identify

SPE 77597

the influence of features on the output once a neural network

model is build for a data set, and many practitioners in our
industry have been using them as the true influence of
parameters on the output. These products simply use the
summation of the weights connected to a particular input
neuron in order to achieve this. Authors believe that this is a
gross simplification of a complex problem, and does not
provide an accurate account of the influence of each feature
and therefore should not be used as such. This method simply
is an artifact of the modeling process and in the event of
changing the architecture or learning algorithms of the
modeling process the influence of features may be altered.
Step Three: Intelligent Production Data Analysis
The third step in the Successful Practices Identification
methodology is a process called Intelligent Production Data
Analysis IPDA. The word intelligent in the above phrase
refers to the use of intelligent systems techniques in the
production data analysis process. During the IPDA process
production data is used to identify a series of Production
Indicators that would represent the state of the production
from a particular field in time and space. The time and space
representation is aimed at capturing the depletion and pressure
decline in the field as new wells are drilled and put into
production at different rates. The dynamic nature of this
analysis (simultaneous analysis of the data in four dimensions
x, y, z, and t) allows the user to identify the sweet spots as
well as bad (unproductive) spots in a field. Such analysis
would prove quite valuable during field development
strategies that include infill drilling programs and candidate
selection for stimulation, restimulation and workovers.
Step Four: Neural Model Building
Next step (Step 4) in the process calls for building a predictive
neural model based on the available data. This step has been
coved in detail in the several prior papers and will not be
repeated here4,6-9.
Step Five: Successful Practices Analysis
Once a representative neural model is successfully trained,
calibrated and verified, the process of Successful Practices
Identification using state-of-the-art data mining, knowledge
discovery and data-knowledge fusion techniques is concluded
with a three stage analysis. The three stage analysis combines
the neural model with Monte Carlo simulation, genetic
algorithms search and optimization routines3, and fuzzy set
theory2 to identify the successful practices on a single well
basis, on groups of wells basis, and on a field wide basis.
During the single well analysis, each well is thoroughly
analyzed in order to identify the sensitivity of that particular
well to different operational conditions. This analysis can
identify the distance of actual practices on that well from the
successful practices that could have been performed on that
well. The larger the distance the higher would be the potential

SPE 77597

IDENTIFICATION OF SUCCESSFUL PRACTICES IN HYDRAULIC FRACTURING USING

INTELLIGENT DATA MINING TOOLS; APPLICATION TO THE CODELL FORMATION IN THE DJ-BASIN

of a good response from the well to future activities such as

stimulation, restimulation and workover.
During the analysis of successful practices on groups of wells,
two different kinds of grouping (clustering) procedures are
used. First, the grouping of wells in the field is performed on
the basis of their location (geology), using the information
provided during the step three analyses. Then anther set of
grouping of the wells are performed based on specific
attributes identified during the step two analyses. This
grouping is done using either hard K-Mean clustering and/or
fuzzy C-Mean clustering procedures, dictated by the data set.
Upon clustering of the wells successful practices for each
cluster is identified using the neural model in conjunction with
genetic algorithms. As a new well is introduced to the process
for analysis, it is first identified to which group or cluster it
belongs and then accordingly, the successful practices for that
particular group are recommended for the new well.
The last stage is the identification of the field-wide successful
practices. In this stage all the wells in the field are treated as
one group and the successful practices are identified for the
entire filed operation.
Results and Discussions
The original database for stimulation of Codell wells in the DJ
basin needed considerable quality control in order to remove
erroneous records including records that did not belong to that
database and had to be removed. It is interesting to note that
several data records were identified by our data quality control
procedure as suspect. These findings were communicated with
the engineers in the operating company. After further
investigation the engineer at the operating company confirmed
that some of the records identified should not have been
included in the database and their inclusion was the result of
an oversight. This was a confirmation of the accuracy of the
new and novel methodology that has been developed for the
data quality control procedure.
Once the quality control of the data set was completed, Fuzzy
Combinatorial Analysis (FCA) was performed. The analysis
was performed for the combination of up to five features.
Table 1 shows the result of FCA. As shown in this table the
most influential parameter is Flow Back Volume followed
by Codell Gas-Ft which is a reservoir quality parameter
identifying the product of gas saturation and the net pay for
the Codell formation. The next two influential parameters are
related to the water quality used for the restimulation frac jobs.
These are water bicarbonate content and the water
peak viscosity.
The parameters in Table 1 can be divided into six different
categories. These categories are Location, Reservoir Quality,
Original Completion, Original Frac Treatment, Second Frac
Treatment, and Production. Using the Fuzzy Combinatorial
Analysis (FCA) average scores for each of the categories,
one can identify the importance of each category on the

process outcome. It should be noted that the process outcome

in this analysis is the Actual Post-Restimulation Peak
Production which will only be referred to as Actual Peak.
Figure 1 shows the influence of each category on the Actual
Peak. From this figure it is clear that Reservoir Quality
followed by Well Location (which is an indication of the
geology) are clearly the most influential categories in the DJ
basin. It is also interesting to note that original frac treatment
plays a more important role than the restimulation treatment in
identifying the success of a restimulation process. Please note
that in this figure the lower the FCA value (shorter bars) the
higher the influence of the parameter, or category.
Details on two of the parameters from Table 1 are presented in
Figures 2 and 3. Each figure provides three graphs. These
graphs include the conventional regression analysis, the
conventional frequency distribution and finally the fuzzy
combinatorial analysis curve for each of the parameters.
Figure 2 represents the details on the most influential
parameter Flow Back Volume. It shows a well behaved
frequency distribution of the values of this parameter in the
data set that depicts a normal distribution slightly skewed to
the left toward the lower volumes, slightly peaking around 225
bbls. The actual peak of the frequency distribution is around
500 bbls. The regression plot shows that no significant outliers
can be immediately identified. Please note that the regression
plot provides a R2 of about 0.065 that mainly identifies
no correlation.
Next, looking at the FCA plot one can see a smooth and
continuous curve that simply means that the values of this
parameter in the dataset are behaving normally. There are
three regions in this plot that need to be identified. First is the
region in blue shade (Region B from 380 to 650 bbls). This is
the region that has provided the best consistent results in this
process. The area indicated by the shade of green (Region A
from 80 to 240 bbls) is the region with the least desirable
outcome. The rest of the plot with no shades (or white shades)
is consisted of several regions of this parameter that seem to
show sensitivity to other parameters in the data set. This
simply means that in Flow Back Volumes between 250 380
bbls and those above 650 bbls, importance and role of other
parameters determines the degree of success of the process. In
other words, regions in shades of white are the regions of high
sensitivity to other parameters. In order to identify the process
outcome in these areas one need to build a neural model of the
process and study each well separately. This is performed
during the first stage of Step five. Figures 3 is a similar graphs
for Peak Viscosity.
Much more can be discussed and shown as the result of the
Fuzzy Combinatorial Analysis but to keep this paper short
the detail of the findings of this analysis will be the subject of
a future paper.

S. MOHAGHEGH, A. POPA, R. GASKARI, S. AMERI & S. WOLHART

SPE 77597

Upon completion of the first two steps the stage is set for the
third step or the Intelligent Production Data Analysis
IPDA. In this analysis the goal is to assist the operating
company in their future field development strategies by
identifying the locations of the field that would be a prime
candidate for further investment (infill drilling, stimulation,
restimulation or workovers) as well as identifying the portions
of the field in which no new investments are recommended.
Please keep in mind that all this is done based on the
successful practices of the operating company in the past.
Figure 4 shows the identified regions in this field. Region
identified by number 1 is the sweet spot. New investment in
this part of the field is recommended. The conclusion of this
analysis is that new investment in this part of the field (infill
drilling, stimulation, restimulation or workovers) has a high
probability of providing high returns. Region 2, as designated
in this graph would be the next best region in the field. Region
3 in the figure would be the region in which new investment is
not recommended based on this analysis. Many times
identification of bad spots in a field is just as important as the
good ones. Table 2 shows the characteristics of each region.
It can be seen that regional segmentation is supported by
Post-Restimulation Actual Peak Production, and the
reservoir quality identifier which is the product of gas
saturation and Codell net pay. Values of both these parameters
reduce as we go from one region to the next.

Figure 5 through 7 show the results of training, calibration and

verification of the Pre-Restimulation neural model. Figure 5
shows the results of the training data set. The training data set
has a R2 of 0.783 and a correlation coefficient of 0.901. Figure
6 shows the results of the calibration data set. The calibration
data set has a R2 of 0.821 and a correlation coefficient of
0.907. Figure 7 shows the results of the verification data set.
The verification data set has a R2 of 0.516 and a correlation
coefficient of 0.809.

The fourth step in this analysis is the neural network model

building. In order to facilitate a thorough analysis of the wells
in DJ basin it was decided that two neural models should be
constructed, trained, calibrated and verified. The first model
only uses the parameters prior to the restimulation process
(Pre-Restimulation Neural Model) and the second model uses
parameters related to the restimulation process (PostRestimulation Neural Model). This is an effort in order to
study the possibility of identification of key issues that
affect the ultimate process outcome (Post-Restimulation
Actual Peak.)

This analysis can guide the operator and the Service Company
on what has worked in the past on a particular well and what
has failed to work. It can provide a valuable roadmap for
design and implementation of restimulation procedures for
such a well. Furthermore, coupling the results from this stage
with the similar study on the same well using the PostRestimulation Neural Model can provide much more insight
into the future practices that should be implemented on this or
similar wells.

In this paper the results and analysis of the first neural network
(Pre-Restimulation Neural Model) will be discussed. The
results of the second network may be presented in a future
paper. The Pre-Restimulation neural model was developed
using eight inputs and one output (Post-Restimulation Actual
Peak). Following parameters were used during the neural
model building as the input to the neural network.
1.
2.
3.
4.
5.
6.
7.
8.

Well Location Latitude

Well Location Longitude
Reservoir Quality Codel porosity-net pay product
Reservoir Quality Nibrara porosity-net pay product
Completion Number of perforations in Codel
Original Stimulation Amount of sand used as
proppant (Mlbs)
Original Stimulation Amount of fluid used (Mgal)
Result of cluster Analysis Entropy

Last step in this analysis is the Successful Practices

Identification. As was mentioned in the previous section the
Successful Practices Identification is performed in three
stages. During the stage one the analysis is performed on a
single well basis. Please note that only the results of the
Successful Practices Identification for the first neural
network (Pre-Restimulation Neural Model) are being
presented here and the concentration will be on three
parameters, one completion parameter and two parameters
from the original stimulation performed on each well. These
parameters are the number of perforations in the Codell as the
completion parameter, amount of sand (proppant) in thousands
of pounds, and amount of fluid in thousands of gallons as the
original stimulation parameters. The goal is to identify if the
effect of the original completion and stimulation on the
restimulation outcome.

The analysis on each well is performed and presented in two

ways in order to allow the operator and the Service Company
to take full advantage of the process. First the analysis is
performed and presented in the form of a series of twodimensional graphs. In these Cartesian graphs each of the
selected parameters is plotted against the process outcome.
The 2-D graph displays the response of an average
restimulation procedure as a function of different values of the
parameter being studied. In each plot the dependent variable
(parameter being studied) is swept throughout its range from
the minimum to the maximum values present in the dataset.
Then the the dependent variable, being the process outcome
(post-restimulation actual peak in Barrels of Oil Equivalent
BOE) is calculated and plotted as a function of different
values of dependent variable. This is done by dividing the
range of the independent variable by 100 and running the
model 100 times once per each value of the independent
variable. The results of the model are then plotted in a
2-D graph.

SPE 77597

IDENTIFICATION OF SUCCESSFUL PRACTICES IN HYDRAULIC FRACTURING USING

INTELLIGENT DATA MINING TOOLS; APPLICATION TO THE CODELL FORMATION IN THE DJ-BASIN

Figure 8 shows this analysis performed for well Bohlender 85. There are three plots in this figure for three independent
variables, namely, original 20/40 sand, which is the amount of
the proppant (Mlbs) pumped during the original stimulation
job, original fluid, which is the amount of fluid (Mgal)
pumped during the original stimulation job, and finally total
number of perforations in Codell.
As it can be seen in this figure, in the case of the original fluid
a maximum can be clearly seen at about 1,364 BOE, while in
the other two cases a minimum value can be observed for
each. The minimum value in the case of proppant is about 970
BOE and in the case of number of Codell perforations is about
1100 BOE. Since the post-restimulation actual peak for this
well is about 1,344 BOE, the conclusion might be that in the
case of this particular well the amount of the original fluid
used during the stimulation was at its optimum or close to its
optimum value. On the other hand by increasing the amount of
proppant or the number of perforations, one might be able to
increase the restimulation performance. But it has to be noted
that there is not much one can do about the independent
variables being studied in this model, since we cannot go back
and change the original stimulation jobs specifications. But
are there lessons that can be learned from this for the future
wells that will be drilled in this area? May be we find out that
once a well has been damaged (or stimulated in a way that it
would hampers it future restimulation performance) it would
be an indication to stay away from such wells (for
restimulation purposes) and invest on other wells.
The straight black lines in each of the plots in Figure 8 show
the actual value of the independent variable that was used in
the original completion and stimulation. This shows that
although much could not have been done in the case of
amount of fluid, but there definitely was room for
improvement as far as the amount of proppant and the number
of perforations was concerned.
The second part of the single well analysis includes threedimensional sensitivity analysis. The three-dimensional
sensitivity analysis is, in nature, very much similar to the 2-D
analysis. During the three-dimensional sensitivity analysis two
independent variables are selected. Then as the value of one of
the variables are changed in sequence from its minimum to its
maximum value in ten steps, at each step the value of the
second variable is swept for the entire range. The result is a
three-dimensional graph that shows the sensitivity of the postrestimulation actual peak in Barrels of Oil Equivalent to both
variables simultaneously. Figure 9 shows three of these plots
for the well Bohlender 8-5 examining original sand, original
fluid and number of perforations in Codell.
Last but not least for the single well analysis is the Monte
Carlo simulation. During the Monte Carlo simulation analysis
for each well, first the independent variables are selected.
Then the probability distribution for each parameter is

identified along with the corresponding parameters. Finally

the model is executed several times in order to generate a
distribution of the process outcome that is the postrestimulation actual peak in this case.
Table 4 shows the characteristics of the Monte Carlo
simulation that was preformed for a particular well. These are
the same three variables that were selected and studied for
well Bohlender 8-5. Uniform probability distribution was
selected for all three variables. The Monte Carlo simulation
was performed for 5000 runs. The results for well Rumsey
10-4 are shown in Figures 10 and 11. In Figure 10 the
frequency distribution is shown. As you can see the frequency
distribution demonstrates a sort of bi-modal characteristics
around 1,195 BOE and 2,095 BOE.
Figure 11 shows the results of the Monte Carlo simulation for
the same well in the form of cumulative probability
distribution. The actual value of the post-restimulation actual
peak for this well was 2,100 BOE. This means that there is not
much that could have been done for this well to enhance its
post-restimulation actual peak above and beyond the achieved
value. The green line in Figures 10 and 11 identifies the actual
value of the post-restimulation actual peak for this well.
Stage two of the Successful Practices Identification
performs the study on groups of wells. The groups are
identified in two fashions. First using the geological separation
as performed in step three (Figure 4.) In such cases a new well
will be subject to successful practices recommendation based
on where it is located in the field. The second way of grouping
the wells is using either hard K-Mean cluster analysis or fuzzy
C-Mean cluster analysis. The cluster analysis is performed
based on the information provided in the step two (Fuzzy
Combinatorial Analysis) of this methodology. Cluster analysis
by nature does not provide a unique solution. In order to
identify the best possible clusters (groups of wells) a novel
method was developed to couple the cluster analysis routines
with a robust search and optimization routine in order to
converge to the best clustering of the wells based on the
identified features. The objective function of the search and
optimization algorithm was chosen to be the minimum
average entropy associated with each cluster as well as with
the entire data set. The entropy is defined as an indication of
amount of information content (as oppose to noise) in a
data record.
Once the grouping and the clustering of the wells are
completed, a genetic algorithm routine, using the developed
neural network as the objective function, will identify the
successful practices for each of the parameters involved in the
process. The results are presented in the form of
characteristics tables identifying the general behavior of the
optimized frac jobs for each group.

S. MOHAGHEGH, A. POPA, R. GASKARI, S. AMERI & S. WOLHART

In stage three of the Successful Practices Identification the

same procedure (as explained in the above paragraph) is
performed for all the wells in the field, with results being
presented to the user in the form of a characteristics table.
Table 5 and 6 are examples of such Successful Practices
characteristic tables. The next and final step would be to
deduce rules of thumb or recommendations based upon these
characteristics tables. Following are examples of such
conclusions from field-wide Successful Practices
conclusions for amount of proppant and peak viscosity.
Amount of Proppant:
The dominant trend is Linear Descending about 83% of the
wells. Four more cases present the same trend but are defined
by a second order polynomial. Ten cases present the opposite
trend. The dominant trend for amount of proppant (Table 5)
would imply that high Actual Peak would be achieved if low
amounts of sand were used in treatment. The trend indicates
that placing more sand in the fracture will not bring significant
increase in Actual Peak.
Peak Viscosity:
The dominant trend is Half-Parabolic Descending about
53% of the wells. However, another 31 cases present the same
Descending trend but are defined by a second order
polynomial. This would bring the same type of trend to
approximately 95%. The dominant trend implies that high
Actual Peak is achieved for low viscosity of the frac fluids.
The trend indicates that low viscosity fluids are preferred to
higher viscosity fluids. This somehow agrees with the trends
identified above in the amount of proppant analysis. Indeed,
for low proppant concentrations there is no need for high
viscosity fluids. Furthermore, high viscosity fluids would
produce high friction when pumped at high rate.
In conclusion of this section it is important to note that in the
case of field-wide Successful Practices Identification (as
well as the analysis for groups of wells but at a lesser extent)
the law of averages will play an important role. Therefore, it is
recommended that all three stages of analysis, namely single
well analysis, groups of wells analysis, and field-wide analysis
should be performed and the final recommendations and
decision making process should use the results of all three
stages simultaneously. Moreover, once the relevant graphs
from the stage one and the characteristics tables from stages
two and three are generated, it is suggested that the final
recommendations, rules of thumb and future action items
should be decided upon by a team of experts in the particular
field (hydraulic fracturing in this case) together with engineers
from the operating company that are intimately familiar with
the intricacies of the particular filed being studied.
Conclusions
A new and novel methodology for comprehensive Successful
Practices Identification of the oil and gas operations has been
developed. This methodology uses state-of-the-art data

SPE 77597

mining, knowledge discovery and data-knowledge fusion

techniques inspired by virtual intelligence paradigms. The new
methodology was applied to hydraulic fracturing program of
about 150 wells in the Codell formation in DJ Basin, Rock
Mountain region. This methodology has the potential to open
new doors in comprehensive and detail field studies as an
alternative to field wide simulation and modeling that
sometimes can be prohibitive due to high cost and lack of
proper information and data. If enough information (data) is
available and cost is not an issue, field-wide simulation and
modeling would provide valuable detail that can prove to be
crucial in large scale operations. Even in such cases, studies
like the one presented in this paper can be performed on a
relatively low cost and can provide much needed insight at a
short time to the engineers and field managers.
Acknowledgement
The authors wish to thank Dr. Rober Siegfried and Mr. Iraj
Salehi of Gas Technology Institute, and Mr. David Hill of
Ticora Geosciences for their help and support throughout the
life of this project. We also wish to thank Mr. Donald Shaw
for his help and insightful observations during this project. We
would like to extend our gratitude toward Patina Oil & Gas
Corporation for providing the data set.
References:
1.

Mohaghegh, S., Gaskari, R., Popa, A., Ameri, S., and

Wolhart, S.: "Identifying Best Practices in Hydraulic
Fracturing Using Virtual Intelligence Techniques", SPE
72385, Proceedings, 2001 SPE Eastern Regional
Conference and Exhibition, October 17-19, North Canton,
Ohio.

2.

Mohaghegh, S.: Virtual Intelligence Applications in

Petroleum Engineering: Part 3 Fuzzy Logic. Journal of
Petroleum Technology, Distinguished Author Series,
November 2000, pp 82-87.

3.

Mohaghegh, S.: Virtual Intelligence Applications in

Petroleum Engineering: Part 2 Evolutionary Computing.
Journal of Petroleum Technology, Distinguished Author
Series, October 2000, pp 40-46.

4.

Mohaghegh, S.: Virtual Intelligence Applications in

Petroleum Engineering: Part 1 Artificial Neural
Networks. Journal of Petroleum Technology,
Distinguished Author Series, September 2000, pp 64-73.

5.

Mohaghegh, S., Mohamad, K., Popa, A. S. and Ameri, S.:

"Performance Drivers in Restimulation of Gas Storage
Wells. SPE Reservoir Evaluation & Engineering
Journal, December 2001, pp 536-543.

6.

Mohaghegh, S., Goddard, C., Popa, A., Ameri, S., and

Bhuiyan, M.: " Reservoir Characterization Through
Synthetic Logs", SPE 65675, Proceedings, 2000 SPE
Eastern Regional Conference and Exhibition, October 1719, Morgantown, West Virginia.

SPE 77597

IDENTIFICATION OF SUCCESSFUL PRACTICES IN HYDRAULIC FRACTURING USING

INTELLIGENT DATA MINING TOOLS; APPLICATION TO THE CODELL FORMATION IN THE DJ-BASIN

7.

Mohaghegh, S., Reeves, S., and Hill D.: " Development of

an Intelligent Systems Approach to Restimulation
Candidate Selection", SPE 59767, Proceedings, SPE Gas
Technology Symposium, Calgary, Alberta, April 2000.

8.

Mohaghegh, S., Koperna, G., Popa, A. S. and Hill, D. G.:

"Reducing the Cost of Field-Scale Log Analysis Using
Virtual Intelligence Techniques", SPE 57454, Proceedings,
1999 SPE Eastern Regional Conference and Exhibition,
October 21-22, Charleston, West Virginia.

9.

Mohaghegh, S., Popa, A. S. and Ameri, S.: "Intelligent

Systems Can Design Optimum Fracturing Jobs", SPE
57433, Proceedings, 1999 SPE Eastern Regional
Conference and Exhibition, October 21-22, Charleston,
West Virginia.

S. MOHAGHEGH, A. POPA, R. GASKARI, S. AMERI & S. WOLHART

Rank
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

Feature
FCA Value Rank
Feature
FCA Value
Flowback Volbbl
0
22 Frac Type
2.2848
CO -Phi-H
0.5811
23 No-CO-Perfs
2.303
Bicarbonate ppm
0.6666
24 Chloride ppm
2.3298
Peak Visc
0.7486
25 NI- Perfed-H
2.3302
Lat
0.7734
26 Water pHLab
2.3665
Orig20/40 Sand-Mlbs
0.9214
27 Pre-Refrac Mcfd
2.3956
Long
1.1
28 Cum MMcf
2.4009
Refrac Date
1.1934
29 Water Source
2.4018
ViscShear 100-30Min
1.3324
30 Iron ppm
2.4351
TotHardness ppm
1.518
31 MGAL
2.496
Calcium ppm
1.6692
32 TotalPerfs
2.5045
AvgRate BPM
1.7415
33 Sulfate ppm
2.5164
Est-Ult- GOR
1.7706
34 New Perfs
2.552
No-NI -Perfs
1.7863
35 Sodium ppm
2.6039
AvgPsi
1.8438
36 Magnesium ppm
2.6108
ViscShear 100-5Min
1.9401
37 ViscShear 100-0Min
2.6649
Top CO Perf
1.9819
38 Pre- FracISDP
2.7127
TDSolid ppm
2.0084
39 TestedPH
2.8066
MMcf
2.0777
40 Post- FracISDP
2.8256
Orig Fluid-Mgal
2.0855
41 Mlb20-40
2.8907
DOFP
2.2451
42 Communication
2.9554

Table 1. Ranking of all the parameters in the dataset from most to least influential parameter.
The lower the FCA values the higher the rank.

Table 2. Statistics on the selected regions during the IPDA.

Rsquare
Correlation Coefficient

Training
0.783
0.901

Calibration
0.821
0.907

Verification
0.516
0.809

Table 3. Statistics of the Pre-Restimulation neural network model.

SPE 77597

SPE 77597

IDENTIFICATION OF SUCCESSFUL PRACTICES IN HYDRAULIC FRACTURING USING

INTELLIGENT DATA MINING TOOLS; APPLICATION TO THE CODELL FORMATION IN THE DJ-BASIN

Variable
No. of Perforation in Codel (Number)
Original 20-40 Sand pumped (Mlbs)
Original Fluid Pumped (Mgal)

Distribution
Uniform
Uniform
Uniform

Minimum
4
85.5
44.5

Maximum
80
600
147.6

Table 4. Characteristics of the variables used in the Monte Carlo simulations.

Table 5. Characteristics table for Amount of Proppant in field-wide study (Stage Three) of the Best Practices Identification.

Table 6. Characteristics table for Peak Viscosity in field-wide study (Stage Three) of the Best Practices Identification.

10

S. MOHAGHEGH, A. POPA, R. GASKARI, S. AMERI & S. WOLHART

Figure 1. Influence of different categories on the post-restimulation actual peak.

Figure 2. Conventional statistics and FCA results for FlowBack Volume, Ranked #1.

SPE 77597

SPE 77597

IDENTIFICATION OF SUCCESSFUL PRACTICES IN HYDRAULIC FRACTURING USING

INTELLIGENT DATA MINING TOOLS; APPLICATION TO THE CODELL FORMATION IN THE DJ-BASIN

Figure 3. Conventional statistics and FCA results for Peak Viscosity, Ranked #4.

11

12

S. MOHAGHEGH, A. POPA, R. GASKARI, S. AMERI & S. WOLHART

Figure 4. Graphical representation of the IPDA analysis.

SPE 77597

SPE 77597

IDENTIFICATION OF SUCCESSFUL PRACTICES IN HYDRAULIC FRACTURING USING

INTELLIGENT DATA MINING TOOLS; APPLICATION TO THE CODELL FORMATION IN THE DJ-BASIN

13

Figure 5. Actual and predicted Post-Restimulation Actual Peak for the 70 wells in the training dataset. The Pre-Restimulation neural model.

Figure 6. Actual and predicted Post-Restimulation Actual Peak for the 9 wells in the calibration dataset. The Pre-Restimulation neural model.

Figure 7. Actual and predicted Post-Restimulation Actual Peak for the 9 wells in the verification dataset. The Pre-Restimulation neural model.

14

S. MOHAGHEGH, A. POPA, R. GASKARI, S. AMERI & S. WOLHART

SPE 77597

Figure 8. Sensitivity of the post-restimulation actual peak (after an average restimulation job) to different values of three parameters being
studied, namely, amount of sand, amount of fluid, and number of Codel perforations for well Bohlender 8-5.

Figure 9. Sensitivity of the post-restimulation actual peak (after an average restimulation job) to different values of three parameters being
studied (two variable at a time), namely, amount of sand, amount of fluid, and number of Codell perforations for well Bohlender 8-5.

SPE 77597

IDENTIFICATION OF SUCCESSFUL PRACTICES IN HYDRAULIC FRACTURING USING

INTELLIGENT DATA MINING TOOLS; APPLICATION TO THE CODELL FORMATION IN THE DJ-BASIN

Figure 10. Results of Monte Carlo Simulation in the form of frequency distribution for well Rumsey 10-4.

Figure 11. Results of Monte Carlo Simulation in the form of Cumulative probability distribution for well Rumsey 10-4.

15