|
Reducing Risks and Costs Through Technology
by Karl Lang,
Hart/IRI Fuels Information Services
Excerpts in PTTC
Network News, 4th Quarter 2001
Two Department of Energy (DOE) initiatives managed by the National
Energy Technology Laboratory (NETL) have generated more than 50 new ideas
for domestic independent producers seeking to reduce the risk of
investment or reduce the costs of operating older oil and gas fields on
the margin of profitability. The Technology
Development with Independents Program has provided funds for 46
projects since its inception, with industry contributions to these
projects providing more than 70% of the total required investment. The Stripper
Gas Well Program has enabled five more projects specifically focused
on low productivity gas wells. Each of these projects demonstrates
collaboration among government, producers, industry experts, and in some
cases, the academic community.
In this article we take a look at five of these projects that focus on
reduction. That is, reducing the
risks of decision-making for investments designed to increase production, reducing
the manpower required to manage production, and reducing
the operating costs of marginal production. All of this effort is commonly
directed towards the goal of keeping domestic wells profitably on stream
for as long a period as possible. The information provided here has been
summarized from a series of workshops on these and other similar projects,
presented by the DOE and PTTC at locations around the country. More
information on these workshops is available at
www.pttc.org/.
Project 1. Reducing Risk When Deciding to
Re-Establish a Waterflood
The questions facing the operator here were not uncommon ones: Does it
make economic sense to re-establish a discontinued waterflood in an older
field that is approaching its economic limit? How can I reduce the risk of
making the wrong decision? In this case the larger problem behind these
questions was even more common: How to carry out an engineering analysis
that requires characterizing a reservoir and predicting its behavior with
very limited reservoir data. In this case, several new techniques were
employed to extract the most value from the limited amount of data
available, making it possible to quantify the economic outcome of various
operating options. The end result? The operator chose to re-establish the
waterflood and the field’s subsequent performance (over 12 months) is
holding reasonably well to the forecast.
The 520-acre Shugart Queen Sand Penasco (SQSP) Unit oil property in the
northeast corner of Eddy County, New Mexico, was first developed in 1939
and placed on waterflood in 1974. Water injection was suspended in 1994. A
total of 14 wellbores had been drilled through the Queen Sand (at 3,350
feet) with 11 of those wells recording production from that reservoir. At
the time the study was initiated (January 1999), production from the five
remaining producing wells during the prior year had ranged between 10 and
100 barrels per month (0.3 to 3.3 BOPD) and about 300 barrels of water per
month (9.9 BWPD). The Unit was facing shut-in and abandonment unless
additional recovery potential could be identified.
The operator, MNA Enterprises, believed that the property had remaining
waterflood oil potential because the water injection program had been
abruptly and, it appeared, prematurely discontinued by the previous
operator. The secondary to primary recovery ratio for the SQSP Unit was
only 0.7, while other Queen Sand waterflood units located in the same
Artesia-Vacuum trend had posted an average S/P recovery ratio of 1.2. The
operator felt that the water injection pattern in place had not provided
the necessary sweep to efficiently recover the oil in the project area.
A waterflood simulation study was
initiated to help define options for re-establishing the waterflood. The
study plan included three phases: data acquisition and reservoir
characterization, production history matching using DOE’s BOAST III
reservoir simulation model, and performance prediction under different
operating scenarios. Unfortunately, as is often the case in older fields,
only one 60 year-old electric log and four gamma ray-neutron count rate
perforating logs were available to characterize the geology of the 11
wells that initially produced oil from the Unit. More information was
needed to characterize the reservoir.
The first step was to expand the
study area beyond the 520-acre Unit to include all wells drilled in the
four contiguous sections. This resulted in 47 well locations, of which 9
had a suite of modern logs, 10 had gamma ray and neutron count rate logs,
and 17 had gamma ray logs. There was also one well with 10 feet of core
through the producing zone. The challenge was to find a way to use the
modern logs available on nearby wells to generate a map of reservoir
properties suitable for reservoir modeling across the Unit.
The strategy for accomplishing this
had five steps:
-
Convert the modern logs
to gamma ray-neutron count rate format.
-
Train a neural network to “see” modern log cross plot porosity
given GR and NCR logs.
-
Correlate porosity with permeability using core
measurements.
-
Correlate permeability with water saturation using core
measurements.
-
Map the three properties using geostatistical algorithms.
The maps that resulted from this strategy were used to define input
parameters for the reservoir simulation model used to history match both
primary and secondary production history for the Unit. The final reservoir
model developed from that history match was then used to evaluate
different operational scenarios.
The model results indicated that re-initiating water injection at 300
BWPD with the existing waterflood pattern (Base Case) would result in
additional secondary oil recovery at a projected rate that would be
economically attractive. The Base Case model forecast that 7,000
incremental barrels of oil (double the estimated remaining oil recovery
under then-current operations) could be recovered over a five-year period.
Additional model results indicated that realignment of the
production-injection pattern by converting one existing producer to an
injection well and converting one injector to a producing well would
increase the waterflood sweep efficiency. At a water injection rate of 300
BWPD over a ten-year period, this could result in the recovery of 36,000
incremental barrels of oil over and above waterflood recovery from the
existing flood pattern. Further, adding another producing well could
result in the recovery of another 15,000 incremental barrels of oil. This
was a total of 51,000 incremental barrels over and above the Base Case and
54,500 barrels over and above then-current operations.
The model also indicated that
if an additional source of injection water could be developed, such that
the water injection rate could be doubled to 600 BWPD, the projected oil
production rates could be doubled for each scenario.
In addition, a deeper, undeveloped interval, the Penrose zone, was
identified during the Queen sand log evaluation. Preliminary analysis
indicated that the Penrose zone has primary and secondary recovery
potential and a possible combination Queen sand/Penrose zone development
location was identified within the Unit.
The study prompted the re-initiation of the waterflood by MNA
Enterprises late last year, the operator of the Unit. MNA cleaned out a
few wells and did some chemical treatments to help reduce a scale and iron
deposit problem when they began reinjecting. The property was sold in
August 2001 for reasons unrelated to performance. After the sale the new
owner, Vintage Drilling Co. of Loco Hills, NM, performed a number of well
maintenance jobs to improve performance and remediate equipment problems.
As of November 2001, injection is 600 BWPM into three wells and production
is about 8 to10 BOPD from five wells. A second water source and an
additional injector are under consideration. According to Sonny Hope,
owner of Vintage, “Given that we
have had additional down time due to remediation work and lost a water
source, the jury is still out on the longer term performance of the Unit
under the waterflood, but we are hoping to do at least as well as the
forecast.”
Neural Networks and Geostatistical Mapping
Two techniques were employed to help
solve the problem of limited data. An artificial neural network was used
to correlate the older logs with the modern logs and geostatistical
mapping software was used to interpolate and map reservoir properties
across the Unit.
A neural network is a computational approach that simulates the way
humans deal with less-than-precise information to solve problems. A neural
network uses a known data set to train a computer to relate one set of
data with another, then, predicts outcomes based on that learned
“behavior” (see sidebar at end of article).
Geostatisical mapping differs from
conventional “nearest neighbor” mapping algorithms in that all the
available data is used to construct the contours, not just the offset
wells. A semivariogram is required to determine the data points to include
in the construction of a kriged map. In the case of MNA, there was little
difference in the appearance of nearest neighbor and kriged maps, but in
some cases the difference is significant and meaningful.
For more information on the
details of how this study was carried out or how the results might be
applied in another field, contact DOE's Jim Barnes, phone 918-699-2076,
email
Jim.Barnes@npto.doe.gov
or Walt North, phone 918-699-2026, email
Walt.North@npto.doe.gov.
Project
2. Reducing Risk When Deciding to Initiate a Waterflood
How can relatively sparse, publicly available data be used to quickly
determine the waterflood potential for selected fields? This was the
problem Coral Production Company (Coral) faced when considering the
installation of a waterflood in the Herboldsheimer Field in Western
Nebraska. The success of waterfloods in the area was mixed and failures
common. A method was needed to quickly and inexpensively evaluate the
potential for a waterflood, in order to justify spending the time and
money required to evaluate potential downhole mechanical problems and
provide some degree of confidence that solving such problems would be
worthwhile. In this example, fuzzy logic was used to prioritize the
relevance of available parameters on the outcome of the proposed
waterflood and a trained neural network was used to develop a correlation
to estimate future oil production.
Coral was considering a waterflood in the 5,300 foot J-Sand of the
Cliff Farms Unit (formerly the Herboldsheimer Field “J” Sand
Waterflood Unit), in the Western Nebraska Panhandle section of the Denver-Julesberg
Basin (D-J Basin), Cheyenne County, Nebraska. Waterflood failures in the
area are believed to be caused by depositional channels, high free-gas
saturations, wettability problems, and fractures.
As is often the case, the detailed information necessary to perform a
reservoir simulation was unavailable and acquiring the information cost
prohibitive. However, a significant amount of basic information is
available in the public domain through the agency responsible for
regulation of the state's oil and gas resources (e.g., field location,
discovery date, producing zone, depth, water injection date, number of
wells, and oil, gas, and water production records). Additionally, oil
gravity, estimates of average porosity and permeability, bottomhole
pressures, net pay, connate water saturation, and formation volume factor
are sometimes available from waterflood unitization application files and
hearing records. The challenge was to develop a correlation among these
parameters that could be used to at least roughly predict waterflood
performance.
Data
Compilation
Correlations Company was contracted to review the state hearing records
for 140 waterflood unitization hearing cases for “J” Sand waterfloods
in the vicinity of the Cliff Farms Unit. The available relevant data along
with the cumulative oil produced at the time of the hearing (primary oil
production) as well as the last reported cumulative oil production
(secondary oil production) of each waterflood unit was compiled. Since D-J
Basin sandstone waterfloods produce the bulk of their secondary oil during
the first five years of operation, and since most of the waterflood units
have produced secondary oil for at least twenty years, no effort was made
to estimate remaining incremental production, even if the floods were
still in operation. Similarly, primary oil recovery was assumed to have
been complete at the time the floods were initiated. A total of 30
descriptive parameters that could influence success or failure were sought
for each of the floods. Not all parameters were available in every case.
The ratio of secondary barrels produced to barrels produced during
primary production (S/P ratio) was chosen as a quantitative ranking factor
for secondary recovery success. The compiled parameter values were used to
construct correlations between the dependent variables (those parameters
which might affect waterflood success) and the S/P ratio for each
waterflood unit. As might be expected, initial analysis using conventional
cross-plots indicated little or no correlation between single
dependant variables and S/P ratio. The assumption, however, was that some
multivariant relationship among the variables might exist.
Fuzzy ranking was used to identify numerical and visual correlations
between S/P and individual parameters (Figure 1). An artificial
neural network was applied to develop a representative correlation between
the dependent variables likely to influence waterflood recovery and S/P
ratio. A two-layer, 6-3-1 neural network (see sidebar) was trained to
develop this correlation (Figure 2).
Assessment
of Waterflood Potential
The correlation was applied to predict the S/P ratio and waterflood
recovery (primary recovery times the S/P ratio) for the Cliff Farms Unit (Figure
3). The correlation predicted a Cliff Farms Unit S/P ratio of 0.315,
which, based on 326,000 barrels of primary oil, translated into an
incremental waterflood recovery in excess of 100,000 barrels of secondary
oil. Bottom hole pressure data obtained from the previous lease holder was
used to identify a prospective waterflood area and a waterflood pilot
project was initiated in July 2000, with water injection into one well
with two offsetting producers on 40-acre spacing. Not enough water has
been injected to fully evaluate the pilot at this date, but injection
continues. Parameters similar to those used to predict the Cliff Farms
Unit S/P ratio, could be used to evaluate waterflood potential for any
group of D-J Basin wells in the Nebraska panhandle.
For more information on the
details of how this study was carried out or how the results might be
applied in another field, contact DOE's Jim Barnes, phone 918-699-2076,
email
Jim.Barnes@npto.doe.gov
or Walt North, phone 918-699-2026, email
Walt.North@npto.doe.gov.
Project
3. Reducing the Cost of Identifying Candidates for Remedial Work
One problem many small operators face is the “can’t-afford-to but
can’t-afford-not-to” dilemma. They have a number of marginal wells,
some of which might profitably respond to remedial work, but the effort to
gather data and evaluate each and every well to determine which ones are
the best candidates can be daunting. This can be particularly true for
stripper gas wells where the potential benefits are less. What is needed
is a cost effective method for determining which of these wells have the
highest likelihood of responding to performance enhancing investments, so
they can be targeted for more serious evaluation. A procedure that could
also identify the likely source of the wells performance problem would be
even better.
Experience indicates that many stripper gas well problems are
manifested as abnormal production declines. Previous DOE research with
BDM-Oklahoma resulted in a software application called Priority
designed to help operators quickly locate problem wells by identifying
such abnormal declines. But even when operators identified decline
abnormalities, the underlying production problems often went undetected
and even when detected, the process of returning the wells to production
was slow and inconsistent. This resulted in substantial downtime and loss
of revenue. The proposed solution was the development of a procedure
employing data collection forms and decision trees to systematically and
economically identify the source of abnormal production declines and
suggest corrective action.
Data
Compiled for Methodology Development
James Engineering, Inc. of Marietta, Ohio was selected to study the
causes and effects of problems associated with abnormal production
declines in stripper gas wells. A study group of wells producing primarily
from the Clinton Sand was located in Ohio.
Each well in a sub-group of 376 wells was reviewed for abnormal
production decline by comparing a Clinton Sand production decline type
curve to the well’s historical production rate vs. time semi-log plot. A
total of 270 wells in the group (72%) exhibited abnormal production
decline during the most recent five year period.
The source(s) of abnormal production decline behavior can occur at any
point from the producing formation to the custody transfer point, and each
problem will be manifested in a different way. Typical examples include:
reservoir damage, reservoir depletion, fluid accumulation, precipitate
plugging, mechanical failure (of casing, tubing, plungers, rods or pumps),
gathering system restrictions, and metering inaccuracies.
Through a careful assessment, the circumstances and source of each
well’s problem(s) were determined for the entire sub-group. The results
showed that more than 46% of the abnormal production declines were caused
by fluid accumulation, 24% by gas gathering restrictions, 23% by
mechanical failures, 4% by reservoir depletion, 2% by metering
inaccuracies, and less than 1% by reservoir damage. In only 2% of the
cases was the cause of the decline undetermined.
Based on the information generated by the 370-well analysis, three Data
Collection Forms were developed to assist in analyzing problem wells.
Individual data forms were developed for the most common methods of
production: tubing plunger, casing plunger, beam pump, and swab/flowing
wells. These forms provided a rigorous, disciplined framework through
which information could be consistently gathered and production
characteristics that would typically result in abnormal behavior or
changes in flowing bottomhole pressure identified.
In addition to the three data forms, a Decision Tree Triage Form was
developed. This form is a step-by-step record of a simple three-phase
approach: identify the problem, measure the problem, and solve the
problem. Following the decision tree insures that the most common problems
are diagnosed and solved by the simplest analysis, only expanding the
analysis as the problem requires. The Decision Tree Triage Form
incorporates information from the Data Collection Forms.
Twenty-four Clinton Sand wells with abnormal production declines were
identified based upon decline curve analysis and the Priority
monthly production monitoring reports. Data Collection Forms were
developed for each of the wells and the Decision Tree Triage was followed
for each well. Eight of the wells suffered from mechanical failure of some
sort, fifteen had fluid accumulation, and one required compression. In
each case a review was carried out to determine the estimated cost to
remediate and three economic indicators: thousands of dollars per Mcf
equivalent per day rate increase (M$/Mcfeqd), months to payout, and net
present value. For the 24 wells evaluated, the remediation costs varied
from $1000 to simply swab and return to production, to $18,000 to run
tubing and put on pump. Payouts ranged from 2 months to five years.
Field Test
of Methodology
Two of these wells were selected for a field demonstration of the
methodology. Both of these wells were diagnosed with the most common
problem, fluid accumulation, and the two had the greatest potential for
production increase of the 24 wells reviewed. The C. Williams #1 was being
produced with a tubing plunger. A pumping unit was installed at a cost of
$11,000 and the well pumped four times a week at four hours per cycle. The
well was also put on a separate meter. It is producing an average of 38
Mcfd and 2 Bwpd, resulting in a payout of 4 months based on an incremental
30 Mcfd and $3/Mcf.
The second well, the L. Richey #1, was not being pumped at all, only
routinely swabbed. Tubing was run and it was put on pump at a cost of
$17,600. The well is pumped twice a week at four hours per cycle and
produces an average of 26 Mcfd and 0.75 Bwpd. Based on an incremental
production of 18 Mcfd the payout is 11 months.
This effort produced a number of insights. Perhaps more than was
generally recognized, most sub-optimal production in Ohio Clinton wells
appears to result from fluid accumulation rather than the need to clean
out wellbore damage or refracture.
More importantly, the methodology developed was shown to be effective
as a way to categorize and prioritize well problems. According to Tim
Knobloch, a consulting engineer with James Engineering, Inc., “One of the most important outgrowths of this work was the recognition
that the production manager and the well tender working together to
systematically gather and record information in a consistent format is the
most important step towards identifying well problems. This seems like
common sense, but enforcing the practice with a framework like the Well
Data Collection Forms and following the same step-by-step method to
diagnose each well provides the discipline needed to streamline the
management of stripper gas wells.” The preliminary results of the
field tests indicate that the tools developed are practical and suitable
for most stripper well operators using data commonly available.
The decision trees, data collection sheets, and economic analyses were
refined and incorporated into a procedural guide. Copies of this guide and
Priority software are available from NETL. An SPE Paper written by Jerry
James, Gene Huck, and Tim Knobloch entitled Low Cost Methodologies to
Analyze and Correct Abnormal Production Decline in Stripper Gas Wells
(SPE 72359) has also been published on this topic.
For more information on the
details of the "Priority" software project, contact DOE's Jim
Barnes, phone 918-699-2076, email
Jim.Barnes@npto.doe.gov
or Walt North, phone 918-699-2026, eamil
Walt.North@npto.doe.gov,
and for information on the "Procedure Guide", contact DOE's Gary
Covatch, phone 304-285-4589, email
gary.covatch@netl.doe.gov.
The Priority software is
available free from James Engineering, Inc., email
TimKnobloch@jeitsk@ee.net/.
Project 4. Reducing the Cost of Identifying
Candidates for Remedial Work
The stripper well business model – hundreds of wells covering
thousands of acres, each producing small volumes of gas – challenges the
available manpower and financial resources of smaller operators. These
operators need an easy way to screen their stripper gas wells to identify
wells that require remediation. While the Priority software described
above helps operators prioritize poor performers by encouraging a
systematic data collection and evaluation protocol and comparing each
well’s performance to a regional standard (type curve), another
solution, SWARM –
Stripper Well Remediation Methodology –
looks at changes in a well’s production compared to its neighbors and
identifies wells that have potential for remediation.
Developed for DOE by Schlumberger Holditch – Reservoir
Technologies and the result of 18 months of software design, SWARM
utilizes the concept that general trends identify changes in gas
production across a field. SWARM compares the cumulative production
of the subject well with all surrounding wells within a fixed distance,
taking into account depletion due to the variable date of first production
for each well. The operator pre-selects the cumulative production period
and the radial distance to surrounding wells. The program quickly performs
this evaluation for all wells in a field, and identifies those wells where
the performance falls outside of the general production trends.
Software Tested in Appalachian Basin
The software was beta-tested
in a stripper gas well field producing from the Whirlpool/Medina formation
in northwestern Pennsylvania using production history, location and well
data provided by Great Lakes Energy Company and Belden & Blake. The
software calculated appropriate performance indicators representative of
each well’s production history over a chosen interval (e.g., 4-year,
5-year, 7-year cum, etc., and average rate for last year of period).
Quickly screening over 700 wells, SWARM identified wells that stood out when compared to the
performance indicators of their offsets or other wells within their
“domain.” For example, Figure 4 shows the subject well’s
8-year cumulative production compared to that of all surrounding wells
within 4,000 feet. Although depletion is observed
(the negative slope of the trend line), the subject well’s
performance indicates it is a candidate for more detailed evaluation and
possible remediation.
Final work is underway to complete the user’s manual and the data
import interface for this software product. Databases for use as examples
are being prepared for the product and the tool should be available early
in 2002. Required input data is primarily the well’s location (x:y) and
production history. Converting mapped well locations to x:y coordinates
can be tedious, but a number of publicly available software products are
now available to simplify this effort. The only other requirement for
using SWARM is a basic understanding of Microsoft Access database
software.
For more information on the
availability of SWARM software contact Gary Covatch, Project Manager with
NETL-SCNG, at 304-285-4589 or at
gary.covatch@netl.doe.gov/.
Project 5. Reducing Operating Costs Through
a Low-Cost Well Surveillance System
Another example of cost reduction involves the design of a low-cost
method for extending the efficiency-improving benefits of a well
surveillance system to marginal oil wells. Such wells are often produced
with beam pumping units on time clocks and are often operated
“pumped-off” because equipment wear and power losses are slight in
shallow wells and fluid drawdown is maximized. The ability to monitor a
large number of marginal oil wells in order to identify and remediate
production problems in individual wells is essential to maintaining
maximum production rates. However, frequent manual inspection of shallow
wells is uneconomic and rarely maintained. Opportunities to correct well
problems promptly and maximize production efficiency are lost. DOE
recognized that the development of an affordable remote monitoring system
could allow significant labor savings, lowering operating costs and
extending the economic producing life of marginal wells.
Working with DOE the Hunter Living Trust (now Vaquero Energy Inc.), a
Santa Barbara CA producer, estimated that successful development and
operation of such a system in the 225-well Edison Field of Kern County, CA
could result in improvements of up to 1,000 BOPM over the 30,000 BOPM
average. Electricity costs would be reduced as worn pumps could be
identified and repaired promptly, which would increase overall average
pump efficiency. The estimated cost savings from using such a system were
$4,000 per month in personnel time and $2,000 per month in electricity
costs, due to greater pump efficiency. Additional environmental benefits
would include lowered greenhouse gas emissions due to reduced power
consumption by the more efficient pumping units and solar powered
instruments, and a reduction in the amount of automotive traffic needed
for visual inspection. The result has been the design and testing of a
Marginal Expense Oil Well Wireless Surveillance (MEOWWS) system (Nelson,
2000).
Non-Traditional
Sources of System Components Investigated
Even low production rate pumping wells generate a variety of mechanical
and electrical signals (e.g., motor current, wellbore acoustics, flowline
temperature and pressure, carrier bar acceleration, polish rod vibration,
etc.). Some of these signals, when properly monitored, provide data that
can be used to determine when a well's performance begins to deviate from
the optimum operating conditions. Existing, commercially available systems
do this well, but at a cost of thousands of dollars per well and the need
for expensive electronic data transmission systems. The challenge was to
design, evaluate, install and test inexpensive, commercially available,
“off the shelf, out of the box” wireless surveillance devices and
sensor technology from other non-oilfield industries which might be
applicable or adaptable to measure and transmit well-generated signals
inexpensively.
Preliminary inquiries were
extended to potential providers to identify possible equipment, software,
and wireless telemetry devices. Discussions regarding non-oil field
technology development and computational micro devices were conducted with
Dr. Massoud Mehdizadeh, a petroleum engineering consultant and faculty
member at California Polytechical State University, San Loius Obispo (Calpoly).
Other non-traditional oil field applications were investigated, including
attendance at trade shows (which were very effective) to locate suppliers
of low-cost, weatherproof sensors, wireless telemetry transmitters and
receivers (FM, microwave, and cellular), systems control computers,
technical support, systems engineering, and other related services. As a
result of the technology search, a vendor was located that offers wireless
vibration surveillance devices designed to track performance trends of
rotating machinery.
Dynamics
of Vibration Sensors
Low-cost, easy to install, self-contained vibration sensors (velocity
and acceleration sensors) with radio telemetry units (RTUs) were obtained
for evaluation. Each sensor unit contains a velocity/acceleration
vibration sensing transducer element, a pre-programmed microchip sensor
and circuit boards, a low power (< 1/4 Watt) spread-spectrum radio
transmitter, radio antenna, and serviceable 3.2 volt lithium battery. The
vibration sensor is a piezovelocity transducer employing a piezoceramic
shear-sensing element and dense seismic mass to produce a charge output
proportional to acceleration. Recent advances in miniaturization of hybrid
circuits allow the sensor to simultaneously provide acceleration and
velocity outputs. The radio transmitter operates in the 900 MHz spread
spectrum range (no special license required) with a transmission range of
up to 3/4 mile. For locations beyond the 3/4-mile range, radio repeaters
may be employed. The first sensor unit selected for this study is
self-contained in a waterproof cylinder, approximately 2 inches in
diameter by 6 inches in length (Nelson, 2000).
The sensor units were initially mounted in pairs for comparison (one
vertically and one horizontally) on the polish rod and on the walking beam
near the horse-head end (Figure 5). Due to the self-contained
nature of the sensors, no special wiring or complicated installation was
required. A nearby radio signal repeater, a base radio signal receiver,
and a base station PC (which interfaces with the base radio signal
receiver to store and process the vibration data) were acquired and
installed. Signals related to the vibration behavior of the pumping unit
were transmitted to the base station computer which in turn will activate
an alarm, alerting operating personnel when the well deviated from its
normal, pumped-off condition.
The sensors were installed on four wells in the field, two of which
were pumped off and two of which had fluid over the pump. Dynamometer
surveys were run in conjunction with the recorded vibration data to
establish the production conditions. The sensors were tested in both the
acceleration and velocity modes. A statistical analysis of both the
velocity and vibration data was made, including median, mean, and standard
deviation. While the velocity data did not prove useful in determining
when a well was pumped-off, the vibration data demonstrated that the
acceleration sensors were able to detect fluid pound when the well
pumped-off. Median values for the acceleration data were less than the
mean for pumped off wells, but similar for wells with fluid over the pump.
Results indicated that wells which are not pumped-off have vibration
behavior which fits a normal distribution, while pumped-off wells have
variations in vibration acceleration which consistently exceed 2 and 3
standard deviations (Nelson, 2000).
Potential
for Further Testing
Future work is being considered to test additional sensors, include
more wells in the project, the best sensor location (beam, polish rod,
etc.), the best sensor orientation (vertical or horizontal), minimum (400
Hertz or lower) sensor frequency, and the minimum sample rate (to extend
battery life). The operator would also like to evaluate the use of solar
cells to eliminate battery service, and the use of radio transmitter
repeaters to extend the transmission range. Future developments could also
include placing a radio receiver controlled computer in the pumping unit
power box to provide automatic pump off control and alternative sensors
such as non-invasive ultrasonic flow meters, temperature probes, and/or
pressure transducers.
According to Don Nelson, Vice President and Production Manager with
Vaquero Energy, Inc., “The change in oil price has directed our
attention away from this project to other priorities, but there is a good
chance that with a little bit more work one could develop a commercially
viable system.” Dr. Mehdizadeh at Calpoly, who has provided some
technical assistance to Vaquero during this project, believes that if the
per well cost for such a system can be kept under $500, there is promise
for commercial success. “There is still a lot we need to learn about the
nature of vibration in rod pumping units pounding on fluid downhole,”
says Mehdizadeh. “But the prospect of a low cost, wireless system that
can pinpoint problems for independent producers and help them avoid
unnecessary trips to remote well sites that is an exciting idea. It is
precisely the sort of technology that can be nudged along by programs like
the one DOE has in place.”
The initial system has demonstrated the feasibility of low-cost
wireless surveillance devices and remote sensor technology. Although more
work will be required to improve and develop a robust system, it appears
that successful MEOWWS monitoring equipment could be developed using the
technology demonstrated.
For more information on the
current status of the MEOWS system, contact DOE's Jim Barnes, phone
918-699-2076, email
Jim.Barnes@npto.doe.gov
or Walt North, phone 918-699-2026, email
Walt.North@npto.doe.gov.
Identifying
Ways to Reduce Risks and Costs Reduces Loss of Reserves
These five projects are some examples of new and innovative ways
independent producers are working to maintain production from the
country’s mature oil and gas fields. During the third phase of the Technology
Development with Independents Program now underway, DOE continues to
add leverage to projects that have industry support and promise to
significantly improve marginal well economics. All of these projects are
focused on the goal of keeping domestic wells profitably and safely on
stream for as long a period as possible. More information on these
projects is available at
www.pttc.org/
and at
www.netl.doe.gov/.
Figure
1: Relationship Between WOR and
S/P RatioBefore and After Fuzzy Ranking
Figure
2: Diagram of 6-3-1 Neural Network
Figure
3: Results of Correlation Developed Using Neural Network
Figure
4: Plot for Six Well Domain Highlights Poor Performer
Figure 5:
Placement of Sensors for Marginal Cost Surveillance System
Sidebar
on Neural Networks
Neural networks are non-linear computer models that attempt to mimic
the design and function of the biologic neurons in the human brain. Their
value lies in their ability to recognize patterns in data, establish
relationships, and learn from a collection of historical data. Although in
the beginning the neural net is “ignorant,” through a “learning”
process, it can become a model of the dependencies between the examples
and the result to be explained.
For example, if we had a collection of well log and core data for a
particular formation, from a number of wells in a particular area, we
could use traditional methods to develop empirical models or multiple
regression algorithms to predict core permeability from log data. The same
sort of traditional approach has been followed for predicting PVT
properties like bubble point and oil formation volume factor from solution
GOR, gas gravity, oil gravity and temperature (e.g., Standing’s
Correlations, etc.). We quantify the ability of the model or the algorithm
to predict the output value from the input variables with a correlation
coefficient, the closer the coefficient to 1.0, the better the confidence
in the prediction.
Alternatively, we can address these same problems by building an
artificial neural network that would attempt to relate the known input
data to the known output, learn from its successes, and adjust its
relating capability through iterative attempts. Such a neural network is a
network of processing units, each essentially an equation that takes
weighted signals from other units, combines them, transforms them and
outputs a numeric result. These linkages of processing units are somewhat
analogous to the network of linked neurons in the human body, hence the
name.
Many neural networks have their processing unit neurons structured in
layers that have similar characteristics and execute their transfer
functions at the same time. Basic neural network architecture is
illustrated here (see figure). The circles are the processing units
arranged in layers. The top row is the input layer, the middle row is the
hidden layer, and the bottom row is the output layer. The lines represent
weighted connections between processing units. The behavior of neural
networks is influenced primarily by the transfer functions, how they are
interconnected and the weights of those interconnections.
Weights are initially assigned randomly. The network calculates forward
and error is measured for all calculated output data by comparing it to
the “actual” data. The error is then “back-propagated” by
adjusting the weights. This process is repeated until either a maximum
number of iterations are reached or a sum of errors goal is reached. The
final weights constitute a “trained” network. Once the network has
been trained using the original data set, it can be used to predict
outputs from new input data simply by multiplying the input data through
the network once.
The number of neural network applications to reservoir engineering
problems is increasing. For example, a neural network designed to predict
permeability from log data in the Appalachian Basin was taught to predict
values with a correlation coefficient of 0.91, as compared to a
conventional multiple regression model’s coefficient of 0.61 (Mohaghegh,
et.al., 1997). Similarly, a neural network was taught to predict PVT
properties for Middle Eastern crude oils with a coefficient of 0.982
versus the conventional correlations’ coefficients that ranged from
0.812 to 0.845 (Gharbi, et.al, 1999).
A neural network was trained to a 0.94 coefficient and used to
predict bulk volume oil in 19 other wells (Weiss, et.al. 2001). A similar technique was used to evaluate the old well logs in
the MNA project.
The power of neural computation comes from the massive interconnection
among the processing units which share the load of the overall processing
task, and from the adaptive nature of the weights that interconnect the
processing units. Although neural networks have been around since the
1950s, one reason why neural networks have only become popular in the last
decade is that they need considerable processing power. A learning process
that takes 30 minutes on a modern PC would have required 320 hours’
worth of an IBM-XT's processing power in 1982 (Smith, 1996). Another
reason is that the widespread computerization of society gives us more and
more data records of all kinds to work with. A new problem now tends to be
how to sift through too much data (“data mining”). The Coral project
is an oilfield example of data mining. More information on this issue can
be found in “Data Mining at a Regulatory Agency to Forecast Waterflood
Recovery” (Weiss, et.al. 2000).
Neural networks are particularly useful with sensory data, or with data
from a complex (e.g. chemical, manufacturing, or commercial) process, or
where there may be an algorithm that is not known or has too many
variables. In such cases it may be easier to let the network learn from
examples.
Computers generally follow a logical path and use precise quantitative
information to solve problems. People, on the other hand, often solve
problems that are too complex to be understood quantitatively by using
knowledge that is imprecise or “fuzzy.” Much like human reasoning
processes, fuzzy set theory or fuzzy
logic relies on approximate or uncertain information to generate
decisions. (Note: Fuzzy set theory
encompasses fuzzy logic, fuzzy arithmetic, and fuzzy data analysis,
although the term fuzzy logic is often used to describe all of these.)
Fuzzy logic is a way to mathematically represent uncertainty and provide
tools for dealing with the imprecision intrinsic to many problems.
Conceived in 1965 by Lotfi A. Zadeh, a professor at the University of
California in Berkeley, fuzzy logic permits the definition of intermediate
values rather than the conventional evaluation (that is: true or false,
black or white, cold or hot, etc). To accomplish this, fuzzy logic
classifies data with boundaries that are not “crisp” or sharply
defined. In fuzzy logic applications, general terms such as “large,”
“medium,” and “small” are each used to capture a range of
numerical values. Fuzzy logic allows these sets to overlap (e.g.,
a 165 pound man may be classified in both the “large” and “medium”
categories, with varying degrees of membership in each group).
Its ability to handle approximate information in a systematic way makes
fuzzy logic ideal for modeling complex systems where an inexact model
exists or where ambiguity is common. A typical fuzzy system consists of a
rule base, membership functions, and an inference procedure.
Today, neural networks and fuzzy logic are found in a variety of
control applications including chemical process control, camera aiming
programs for the telecast of sporting events, systems for early
recognition of earthquakes, speech and handwriting recognition software,
spell checking software, automobile fuel-consumption optimizers, vibration
compensators in camcorders, and expert systems for the assessment of stock
exchange activities. Because these techniques are used to simulate the
thinking patterns of the human brain, they are considered elements of the
branch of computer programming termed “artificial” or “virtual”
intelligence.
A good Internet resource for information on neural networks is
maintained by the Pacific Northwest National Laboratory (PNNL). This site
includes up-to-date links for articles and papers on the topic, commercial
software demos and free software downloads. (www.emsl.pnl.gov:2080/proj/neuron/neural/neural.homepage.html).
DOE
Projects Using AI Techniques and Methods
The Department of Energy’s National Energy Technology Laboratory
(DOE-NETL) currently has several projects using neural networks and fuzzy
logic to improve exploration and production processes.
For example, with funding from the DOE, Gas Technology Institute,
teaming with specialists from West Virginia University, Intelligent
Solutions Inc., and TechnoMatrix Inc., is developing computer-assisted
methods for identifying and optimizing preferred management practices in
oil production operations, based on such “soft” computing technology.
Benson-Montin-Greer (BMG) Drilling Corporation of Farmington, NM, is
using new log interpretation methods based on neural networks to evaluate
Mesa Verde recompletion opportunities in Gavilan and West Puerto Chiquito
Mancos fields in New Mexico's San Juan Basin.
Luff Exploration Company is developing an intelligent computing system
to apply to reservoir-production models for analysis of seismic and
geologic data. The Red River formation is a shallow shelf carbonate
widespread in the Williston Basin of North Dakota, South Dakota and
Montana. Successful drilling locations are difficult to identify without
seismic surveys, an approach that is too expensive for many smaller
independent operators. The goals of this project are to locate and produce
oil reserves in the Red River by identifying nonlinear relationships among
twenty-five years worth of 3-D seismic, production, geological and
petrophysical data collected on the Red River. The approach will employ
neural networks, fuzzy logic and probabilistic reasoning in conjunction
with conventional techniques such as geostatistical mapping and pattern
recognition.
Finally, the New Mexico Institute of Mining and Technology Petroleum
Recovery Research Center (PRRC) is developing a Fuzzy Expert Exploration
Tool (FEE Tool) that will realistically model the decision-making
mechanisms of oil explorationists who routinely deal with incomplete
information and produce predictions with different levels of certainty.
Use of the FEE Tool, to be available via the Internet, is expected to
result in a reduction in exploration time and expense. Both large and
small operators have exhibited a great degree of interest in the project.
Please see
http://baervan.nmt.edu/REACT/reacthomepage.htm
for more information on this project.
A software application developed by DOE’s National Petroleum
Technology Office known as Neuro3
is currently available as part of a Risk Analysis toolset and can be
downloaded from the NPTO site (www.npto.doe.gov).
Common oil and gas applications of this neural network-based software
include forecasting of reservoir properties from wireline log signatures,
extension of reservoir properties for simulation, and seismic
interpretation. A 32-bit MS Windows application, Neuro3
has a spreadsheet interface to allow import and export of external data
sets, an extensive help system, and a tutorial with background information
on neural networks.
Neuro3 has been used by a number of operators. Jim Wilson, Acquisition
Manager for Ward Petroleum Corporation in Enid, Oklahoma is considering
Neuro3 software to develop a tool for valuing potential properties for
acquisition. “The idea is to use a neural network to build a
relationship between electric log values and behind pipe reserves,” says
Wilson. “In particular, we are looking at the Red Fork Formation in the
Anadarko Basin and trying to find a cost-effective way to assess the value
of potential acquisitions without having to do full-blown reserves studies
on dozens of properties. This sort of problem looks like a perfect
application for neural networks, assuming we can gather enough data to
train the network.”
For more information about Neuro3, contact Chandra Nautiyal, DOE's
Project Manager, phone 918-699-2021 or email
Chandra.Nautiyal@npto.doe.gov.
Sidebar
Figure: Neural Network Structure and Training (after Keltch, 2001)
References
Gharbi, R. B. C. and Eisharkawy, A. M., 1999. Neural Network for
Estimating the PVT Properties of Middle East Crude Oil, SPE Reservoir Eval.
& Eng., June, p. 255-265.
Keltch, B., 2001. “Application of Neural Networks to Oil and Gas
Problems,” presentation to the Bartlesville, OK, SPE chapter, November.
Mohaghegh, S., Bogdan, B., and Ameri, S., 1997. “Permeability
Determination From Well Log Data,” SPE Formation Evaluation, September,
p. 170-174.
Nelson, D. G., 2000. “Marginal Expense Oil Well Wireless
Monitoring,” SPE paper 62865 presented at SPE/AAPG Western Regional
Meeting, Long Beach, CA, June 19-23.
Smith, Dr. Leslie, 1996. Notes posted at
www.cs.stir.ac.uk/,
Centre for Cognitive and Computational Neuroscience, University of
Stirling, UK.
Weiss, W.
W., Wo, S., Weiss, J.: “Data Mining at a Regulatory Agency to Forecast
Waterflood Recovery,” SPE 71057 Presented at the 1999 SPE Rocky Mountain
Technical Conference, Keystone CO, 21-23 May 2000.
Weiss, W.
W., Stubbs, B.A. Balch, R.S.: “Estimating Bulk Volume Oil in Thin-Bedded
Turbidites,” SPE 70041 Paper presented at the SPE Permian Basin Oil
& Gas Recovery Conference, Midland, Texas, 14-17, May 2001.
Author: Karl Lang is Director of Custom Publishing at Hart/IRI
Fuels Information Services. He edits GasTIPS, a technical journal produced by
Hart for GTI. He also writes for a number of Hart energy publications. E-mail:
klang@chemweek.com
| 
