Tech 101
Artificial Intelligence and Data Mining:
Enabling Technology for Smart Fields
Shahab D. Mohaghegh, Intelligent Solutions and West Virginia University
S
mart completions let engineers
intervene with details of wells’
operations from a distance.
Smart wells transmit nearly
continuous (real-time) data
streams (pressure, flow rate, etc.) to
the remote office, providing immediate
feedback on the consequences of
recent decisions made and actions
taken. Smart fields include multiple
smart wells providing the possibility
of managing the entire reservoir
remotely and in real time. Our industry
is now on the eve of making real-time
reservoir management (RTRM) a reality.
Artificial intelligence and data mining
(AI&DM) is one of the key enabling
technologies for RTRM. AI&DM enables
us to process, model, and use real-time
data streams, build accurate prototypes
of sophisticated reservoir-simulation
models that can respond to changes
in model input in real time to help us
make crucial reservoir management
decisions and close the loop on highfrequency feedback to the reservoir
model for making RTRM a reality.
In short, the contribution of
AI&DM technology to smart fields can
be summarized in two items: (a) It
provides the technological capabilities to
automatically and autonomously handle
high-frequency data streams received
from permanent downhole gauges
(data cleansing, data summarization,
pattern recognition, adaptive online
modeling, etc.); (b) It provides the
modeling framework and workflows
that allows the existing reservoirsimulation models to be run in real
time, thus making RTRM possible.
RT R M
Reservoir management is defi ned as
the practical science of developing
a hydrocarbon field in a manner
that maximizes ultimate recovery.
RTRM involves performing reservoir
engineering analysis in real time to
support frequent field-management
decisions, and using near-real-time
feedback on the consequences from the
reservoir/well to assess management
14
F IG. 1
Schematic diagram of the closed-loop RTRM.
decisions and reservoir performance.
Therefore, RTRM is the enabling
technology for the emerging smart
fields. It is the closed-loop process
during which the reservoir model is
continuously updated by the information/
feedback received from the field
(by means of high-frequency data
streams) that are the consequence of
the decisions made and implemented,
based on the reservoir model.
The ultimate benefit of the smart
field depends on the degree of our
success in building and implementing
RTRM. In other words, the value of highfrequency data streams is realized once
we are able to use them in effectively
updating the reservoir model and
subsequently using the model to make
field-operational decisions. Therefore,
the key to moving toward successful
smart-field operation is to be able
to perform the following steps:
1. Acquire, process, and
analyze real-time, continuous
data streams from the wells.
These real-time pressure and
rate data provide indications of
reservoir reaction to the operational
decisions made by engineers
using the reservoir model.
2. Use the real-time data as
feedback to the reservoir model.
By analyzing the high-frequency
data in the context of the reservoir
model, engineers can compare the actual
reservoir/well response with the model
predictions. Furthermore, the model can
be updated (modified) to compensate for
the discrepancies between its predictions
and reservoir/well responses.
3. Make operational decisions
based upon model simulations.
Make new operational decisions
by comprehensively analyzing the
solution space of the reservoir model and
make predictions on possible response
from the reservoir/well. Implement
the decisions. (Among the possible
decisions is to continue operations as is.)
4. Go to Step No. 1.
Perform all these analyses while
taking into account and quantifying
the uncertainties associated with the
reservoir model.
To accomplish steps 2 and 3, the
reservoir model must have the capability
of analyzing multiple scenarios in real
time (or near real-time) and provide
real-time responses to changes to the
model input or potential modifications
that can be made to the well operation.
The reservoir/well responses to the
modifications are reflected in highfrequency (real-time) data streams.
Fig. 1 is a schematic diagram of
the closed-loop RTRM concept.
SURROGATE RESERVO IR MOD EL
The surrogate reservoir model (SRM)
has been developed in response to the
need for RTRM and to make it a reality
3.00
Rate-relaxation period
14
2.50
50
Rate-relaxation period
45
2.50
40
12
2.00
1.50
1.00
30
1.50
8
Cumulative oil
Cumulative water
Water cut
35
2.00
10
6
25
Cumulative oil
Cumulative water
Water cut
1.00
20
15
4
0.50
0.00
Jun-03 Jan-04
Aug-04 Feb-05 Sep-05 Mar-06 Oct-06
F IG. 2
10
0.50
2
Water Cut %
16
Cluster 1
5
0
Apr-07 Nov-07 Jun-08
0.00
Nov-01
Mar-03
Aug-04
Dec-05
0
Apr-07
Results of lifting rate restriction on typical wells from clusters 1 and 5.
(Mohaghegh et al. 2009; Mohaghegh et
al. 2006a; Mohaghegh 2006; Mohaghegh
et al. 2006b). An SRM is developed using
state-of-the-art capabilities in AI&DM.
AI&DM is a collection of complementary
analytical tools that attempt to mimic life
when solving nonlinear, complex, and
dynamic problems. AI&DM consists
of, but is not limited to, analytical
techniques such as artificial neural
networks (Mohaghegh 2000a), genetic
optimization (Mohaghegh 2000b), and
fuzzy logic (Mohaghegh 2000c).
An SRM is an accurate replica of
a complex reservoir-simulation model
that may include tens or hundreds of
wells. SRM runs provide results, such as
wells’ pressure and production profiles
or pressure and saturation distribution
throughout the reservoir, in real time
The SRM is developed using a
unique series of data-generation,
manipulation, compilation, and
management techniques. These
techniques are designed to take
maximum advantage of characteristics of
artificial neural networks complemented
with fuzzy set theory. Upon completion
of the modeling process and validation,
an SRM can accurately replicate the
results generated by highly sophisticated
reservoir-simulation models in response
to changes made to the model input,
in fractions of a second. The fact
that thousands of SRM runs can be
performed in seconds makes uncertainty
analysis a practical task. This allows
the uncertainty band associated with
any decisions to be identified quickly.
The SRM has been successfully field
tested. In a recent study performed on
a giant oil field in the Middle East, an
SRM was developed to replicate the
existing simulation model of the field
that was developed using a commercial
Upon completion of tens of millions
of SRM runs (equivalent to tens of
millions of simulation runs) the 165
wells in the field were divided into
five clusters, on the basis of rising
water-cut risk. Recommendations for
rate relaxation were made for wells
in clusters 1 and 2. Furthermore, it
was predicted that these wells would
produce small amounts of water and
large amounts of incremental oil over the
next 25 years. On the other hand, more
than 100 wells in clusters 4 and 5 were
predicted to produce large amounts of
water, if rate restrictions were lifted.
With completion of the study, rate
restriction was lifted from 20 wells. These
wells were selected from among all
the clusters to provide a representative
spatial distribution of the reservoir
and examine the accuracy of the SRM
predictions. After more than 2½ years of
production, the results were analyzed.
As can be seen clearly from Fig. 2
simulator. Consisting of approximately
a million grid blocks, the computing
time required for a single run of the
existing simulation model is 10 hours on
a cluster of 12 parallel central processing
units. Upon development, calibration,
and validation of the SRM that could
successfully and accurately replicate
the results of the simulation model, tens
of millions of SRM runs were performed
to comprehensively explore the
reservoir model’s solution space so that
a successful field-development strategy
could be established. The objective
was to increase oil production from the
field by relaxing the rate restriction
imposed on wells. The key was to identify
those wells that would not suffer from
high water cuts once a rate-relaxation
program was initiated. The SRM had
to take into account and quantify the
uncertainties associated with the
geological model, while accomplishing
the objectives of this project.
110
100
Water cut, %
90
Difference in max. water cut (%) after rate relaxation
Water cut, %
Cumulative Production, million bbl
3.00
70
50
25
30
10
-10
10
-8
1
F IG. 3
-4
4
2
3
Cluster No. Assigned by ISI Study
5
Percentage increase in maximum water cut,
normalized for all wells in clusters 1–5.
Vol. 5 // No. 3 // 2009
15
F IG. 4
Autonomous cleansing and summarization of high-frequency data and
online, adaptive modeling of high-frequency data streams. Data is from
an oil-producing carbonate reservoir in the Middle East.
to predict thier behavior and detect,
model, and verify hypotheses about
drive mechanisms by means of
continuously modeling volumetric
and material-balance estimates
• Diagnostic analysis to detect
interwell connectivity and
reservoir compartmentalization
that may be present
• Communication with the SRM (sending
feedback and receiving instructions)
Fig. 4 shows an example of
real-time data preprocessing and
summarization (left) and real-time
adaptive modeling (right) applied
to a well in the Middle East.
CONCLUSION
(similar results observed from all other
wells in the corresponding clusters are
summarized in Fig. 3), wells in clusters
1 and 2 produced large amounts of
incremental oil, while water production
declined. The opposite effect was
observed in wells that were classified
in clusters 4 and 5, as predicted by
the SRM. Fig. 3 shows the maximum
incremental water cut normalized for
all wells in each of the clusters. It is
clear from this figure that in accordance
with the SRM’s predictions, water cut
decreased in wells classified in clusters
1 and 2 while it increased significantly
in wells classified in clusters 4 and 5.
Results from this study, as well
as other similar studies, demonstrate
the robustness of SRM technology. An
SRM can be used to develop replicas
of sophisticated and large reservoirsimulation models that then can be used
to drive the main engine of RTRM.
I N TE LL I G EN T REAL - T I M E
DATA AN ALY S I S
The high-frequency (real-time) data that
is collected from permanent downhole
gauges and transmitted to be stored in
data historians is usually unusable in its
raw form. It needs to be cleansed and
summarized and prepared (processed)
for use in reservoir-engineering
studies. The high-frequency data need
to be denoised, the outliers must be
identified and removed, existing trends
and patterns need to be identified,
and data need to be summarized so
that maximum useful information can
be preserved, while using the least
amount of data storage. Most important,
all these tasks need to be performed
reliably in real time (at the same speed,
or faster, than data are received)
automatically and autonomously, without
16
the supervision of an engineer.
Furthermore, the intelligent realtime data analyzer needs to be capable
of taking maximum advantage of the
information content of the high-frequency
data. AI&DM-based intelligent data
analysis of the high-frequency data
streams includes two main components:
high-frequency data management
(HFDM) and high-frequency data
analysis (HFDA). HFDM performs
(autonomously and in real time) data
preparations and preprocessing. This
process prepares the data for reservoirengineering analysis and includes data
denoising, outlier removal, pattern
recognition and summarization, and,
fi nally, data preparation for SRM.
HFDA performs state-ofthe-art reservoir-engineering
analysis using the prepared and
preprocessed high-frequency data
streams. The analyses include
• Detection, isolation, and analysis
of pressure-transient data used
for continuous monitoring of
well and reservoir behavior
• Empirical modeling using adaptive
technology that learns data behavior
and continuously modifies itself to
match and model the observed data,
AI&DM contributes to all aspects of
the E&P industry to help increase
efficiency and productivity (Tapias et
al. 2001; Saputelli et al. 2002; Zangl and
Oberwinkler 2004; Finol and Buitrago
2002; deJonge and Stundner 2002).
AI&DM provides a key component of the
emerging smart field. At a June 2009 SPE
forum on this topic in Colorado Springs,
Colorado, AI&DM’s contributions
to drilling, production, reservoir
characterization, reservoir modeling,
reservoir management, and assetportfolio management were thoroughly
discussed by numerous industry leaders.
All agreed that the role of AI&DM in the
E&P industry can only increase as more
and more applications are implemented.
It is hard to imagine that smart fi elds
can be implemented without extensive
use of AI&DM as an integral part of
a complete system that also would
include many other components and
technologies. Therefore, including
the fundamentals and applications
of AI&DM in the training of our
workforce (either prehiring through
the universities or post-hiring through
short courses) can contribute to higher
effi ciency and more productivity. TWA
References continued on page 19
Shahab D. Mohaghegh is professor of petroleum and natural
gas engineering at West Virginia University and founder and
president of Intelligent Solutions. A pioneer in the applications
of AI&DM in the E&P industry, he has 18 years of experience in
this area and has published more than 100 technical papers and
articles on the subject. Mohaghegh has been an SPE Distinguished
Speaker and Distinguished Author. He is an associate editor
of several peer-reviewed journals and has recently cochaired
SPE’s forum on artificial intelligence. Mohaghegh holds BS
and MS degrees in natural gas engineering from Texas A&I University and a PhD
in petroleum and natural gas engineering from Pennsylvania State University.
Tech 101... continued from page 16
References
deJonge, G.J. and Stundner, M. 2002. How Routine
Reservoir Surveillance with Neural Networks
and Simplified Reservoir Models Can Convert
Data Into Information. Paper SPE 78334-MS
presented at the European Petroleum Conference,
Aberdeen, 29–31 October. doi: 10.2118/78334-MS.
Finol, J. and Buitrago, S. 2002. Model Identification
with Fuzzy-Optimization Techniques in Geological
Data Mining. Paper SPE 78340-MS presented at
the European Petroleum Conference, Aberdeen,
29–31 October. doi: 10.2118/78340-MS.
Mohaghegh, S.D. 2000a. Virtual Intelligence
Applications in Petroleum Engineering:
Part 1—Artificial Neural Networks.
J. Pet. Tech. 52(9): 64–73. SPE 101474-MS.
Mohaghegh, S.D. 2000b. Virtual Intelligence
Applications in Petroleum Engineering:
Part 2—Evolutionary Computing.
J. Pet. Tech. 52(10): 40–46. SPE-61925-MS.
Mohaghegh, S.D. 2000c. Virtual Intelligence
Applications in Petroleum Engineering:
Part 3—Fuzzy Logic. J. Pet. Tech.
52(11): 82–87. SPE-62415-MS.
Mohaghegh, S.D. 2006. Quantifying Uncertainties
Associated With Reservoir Simulation Studies
Using Surrogate Reservoir Models. Proc., SPE
Annual Technical Conference and Exhibition,
San Antonio, Texas, 24–27 September. SPE
102492-MS. doi: 10.2118/102492-MS.
Mohaghegh, S.D., Hafez, H., Gaskari, R., Haajizadeh,
M., and Kenawy, M. 2006b. Uncertainty
Analysis of a Giant Oil Field in the Middle
East Using Surrogate Reservoir Model. Proc.,
Abu Dhabi International Petroleum Exhibition
and Conference, Abu Dhabi, 5–8 November.
SPE 101474-MS. doi: 10.2118/101474-MS.
Mohaghegh, S.D., Modavi, A., Hafez, H.H.,
Haajizadeh, M., Kenawy, M., and Guruswamy,
S. 2006a. Development of Surrogate Reservoir
Models (SRM) For Fast-Track Analysis of
Complex Reservoirs. Proc., SPE Intelligent Energy
Conference and Exhibition, Amsterdam 11–13
April. SPE 99667-MS. doi: 10.2118/99667-MS.
Mohaghegh, S.D., Modavi, A., Hafez, H.,
Haajizadeh, M., and Guruswamy, S. 2009.
Development of Surrogate Reservoir Model
(SRM) for Fast-Track Analysis of a Complex
Reservoir. Intl. J. Oil, Gas and Coal Tech. 2(1):
2–23. doi: 10.1504/IJOGCT.2009.023627.
Saputelli, L., Malki, H., Canelon, J., and Nikolaou,
M. 2002. A Critical Overview of Artificial
Neural Network Applications in the Context of
Continuous Oil Field Optimization. Paper SPE
77703-MS presented at the SPE Annual Technical
Conference and Exhibition, San Antonio, Texas,
29 September–2 October. doi: 10.2118/77703-MS.
Tapias, O., Soto, C.P., Sandoval, J., Perez, H.H.,
and Bejarano, A. 2001. Reservoir and Artificial
Intelligence Techniques for Data Analysis.
Paper SPE 68743-MS presented at the SPE Asia
Pacific Oil and Gas Conference and Exhibition,
Jakarta, 17-19 April. doi: 10.2118/68743-MS.
Zangl, G., and Oberwinkler, C.P. 2004.
Predictive Data-Mining Techniques for
Production Optimization. Paper SPE 90372MS presented at the SPE Annual Technical
Conference and Exhibition, Houston, 26–29
September. doi: 10.2118/90372-MS.
Vol. 5 // No. 3 // 2009
19