Real-Time Diagnostics of Gas Lift Systems

Using Intelligent Agents: A Case Study

G. Stephenson, SPE, Occidental Petroleum Corporation; R. Molotkov, SPE, Weatherford;
N. De Guzman, SPE; and L. Lafferty, SPE, Intelligent Agents Corporation

Summary
This paper describes a new method to continuously monitor and
diagnose the condition of wells producing through continuous gas
lift. The paper describes the application of this system in a mature
onshore gas lift field in the western United States and the results
obtained. A central problem related to the operation of gas lift
wells is the ability to identify underperforming wells and to address
the underlying issues appropriately and in a timely manner. This
problem is compounded by the trend toward leaner operations and
relative scarcity of application-specific domain knowledge. The
purpose of this method is to address these issues by leveraging
real-time data, gas lift domain expertise, and proven steady-state
analysis techniques in a desktop software application.
This system performs four key functions: Monitoring the wells
condition by collecting data, assessing the meaning of these data,
recommending actions for correcting problems and responding to
threats, and explaining recommendations.
The performance of the system has met initial expectations and
has provided additional unforeseen benefits. This paper cites specific cases that compare agent predictions to expert diagnoses and
quantify the benefits of taking the recommended actions. What was
found was that while the correct diagnoses of well performance
issues was beneficial, the real benefit was in allowing production
engineers to analyze a greater number of wells in far less time. To
that end, the paper will discuss the role of this system as it relates
to the overall production management workflow.
The success of this project has demonstrated that intelligent
agents can effectively perform functions historically performed by
a handful of experts. The paper will discuss key system-design features that enable this level of functionality as well as other potential
areas in which the technology can be extended in the future.
Introduction
One of the current challenges facing the upstream E&P industry
is the growing scarcity of specialist domain expertise and trained
personnel needed to efficiently operate oil and gas assets. In cases
in which these resources are limited or unavailable, automation
technology has often been touted as a solution. While the introduction of such technology has delivered numerous improvements in
operational efficiency, it also has introduced new challenges. One
such challenge involves the introduction of vast quantities of data
that result in minimal actionable information (Brul et al. 2008;
Hite et al. 2007). Operators are faced not only with the information
technology task of managing these data, but also with the business
challenge of leveraging the data to improve their profitability. In
response to this new challenge, a growing number of projects are
being initiated to help close this gap between data and information.
This paper discusses one such effort.
In this project, new technology has been developed to assist
production engineers in the well-by-well optimization of gas lift
systems. Well-by-well optimization has long been recognized as
having value (Mochizuki et al. 2006), but has often proven impractical to carry out on a routine basis because of the labor-intensive
nature of the work and the limited number of individuals with the
Copyright 2010 Society of Petroleum Engineers
This paper (SPE 124926) was accepted for presentation at the SPE Annual Technical
Conference and Exhibition, New Orleans, 47 October 2009, and revised for publication.
Original manuscript received for review 29 July 2009. Revised manuscript received for
review 29 July 2009. Paper peer approved 08 October 2009.

February 2010 SPE Production & Operations

required level of expertise to perform it. This project sought to

solve this problem by developing a system of intelligent agents
that leverage both real-time data and gas lift domain knowledge to
assist engineers in these well-by-well optimization tasks.
An intelligent agent is defined as an autonomous entity that
observes and acts upon an environment and directs its activity
toward achieving goals (Russell and Norvig 2003). In the context
of this project, intelligent agent systems are software applications
designed for surveillance engineers and production managers to
diagnose problems, provide advice on the proper course of action,
explain how the agent arrived at its conclusions, and take action
on behalf of the operator, if desired.
Optimizing Gas Lift Wells
Value of Optimization. For many years, production engineers
have recognized the value of gas lift optimization as a means to
generate incremental production and reduce operating expenses
(Kanu et al. 1981). Although gas lift wells represent a relatively
small portion of the well population, they generate a significant
portion of the industrys oil and gas production. Industry experience has shown that gas lift wells seldom operate under optimal
conditions (Sclumberger Gas Lift Design and Technology 1999).
As a result, significant production gains can be achieved through
the optimization of gas lift systems.
Objectives. In general, gas lift optimization efforts have the common objective of producing the greatest amount of oil with the
least amount of gas injection. For this to occur, a variety of related
tasks must be accomplished. These include: (1) injecting as deeply
as possible; (2) lifting the well from a single point of injection;
(3) injecting under stable, steady-state conditions; (4) injecting
the optimal amount of gas for the given production on the basis
of individual well performance; (5) minimizing back pressure; and
(6) allocating injection gas among a population of wells in the most
resource-efficient manner possible.
Optimization Project Classes. In order to achieve these objectives, operators often will initiate some form of optimization
project. Such projects can be described using three main classes
or tiers.
Tier 1. The first, most common tier of gas lift optimization
project is well-by-well optimization. Well-by-well efforts include
such tasks as troubleshooting wells, individual well modeling
using systems analysis, replacing gas lift valves, backpressure
reduction, and other tasks that focus on improving the production
performance of individual wells.
Tier 2. The second tier of optimization project can be defined
as full-field optimization. Full-field optimization projects focus
on managing the interactions of a population of wells with each
other, their production network, the gas distribution network, and,
in certain cases, of wells with with the reservoir over time. The
aim of such projects is to maximize the net present value of the
asset while honoring system constraints. Such projects generally
require sophisticated software models and specialists to run them
and are dependent on prior steps carried out in Tier 1 (Nadar et al
2008; Nadar et al. 2006).
Tier 3. The third tier of optimization project can be defined
as real-time enabled full-field optimization. Tier 3 is actually
an extension of Tier 2, in which real-time data are used to continuously update network and well models, enabling operators to
111

Gas Lift Optimization

100

Value, %

80

Well-by-well

60
40
20
Full-field

0
0

20

Full-field (RT-enabled)
60
80

40

100

Effort, %
Fig. 1Relative effort and value of optimization project classes.

sustain the production enhancements achieved in Tier 2 over an

extended period of time.
Research indicates that, of the three tiers, Tier 1 is by far the
most commonly employed. This is because well-by-well optimization is often the least expensive of the three and generally produces
the greatest performance gains. In terms of the Pareto principle,
well-by-well optimization is the 20% of effort that yields 80% of
the results. This is depicted in Fig. 1.
While a number of gas lift automation projects have been carried
out to address Tier 3 (McKie et al. 2001; Litvak et al. 2002), this project is somewhat unusual in that it focuses on leveraging technology to
allow engineers to more easily carry out tasks related to Tier 1.
Well Performance Issues. A variety of issues can impact the
performance of gas lift wells. These issues are frequently classified as either inlet issues, outlet issues, or downhole issues (API
RP 11V5 1999).
Inlet Issues. Inlet issues include conditions that inhibit or
obstruct the injection of gas into the well. Such conditions include
(1) frozen or plugged injection-control valves, (2) inadequate supply pressure to kick off or unload the well, and (3) unstable or
irregular gas supply pressure.
Outlet Issues. Outlet issues include conditions downstream of
the wellhead that impair a wells ability to flow. Such items include
excessive backpressure because of production chokes, undersized
flowlines or manifolds, and high separator pressures.
Downhole Issues. Downhole issues include events occurring
below the wellhead that impair a wells production performance.
Such conditions include (1) multipoint injection, (2) valve cycling, (3)
tubing-to-casing communication, (4) instability because of injection
in subcritical flow across an orifice, (5) plugged operating valves, (6)
inadequate differential pressure at depth, (7) reopening upper valves,
(8) flow-cutting gas lift valves, (9) temperature-locking injection-pressure-operated valves, and (10) circulating gas above the active fluid
level in the tubing, as well as a variety of other known issues.
Tools and Techniques. A variety of tools and techniques have been
developed and used over time to aid in gas lift troubleshooting and
diagnostics. These include (1) flowing pressure and temperature
gradient surveys, (2) annular fluid levels, (3) CO2 tracer surveys
(Peringod et al. 2009), (4) systems analysis (Mach et al. 1979),
(5) calculation of valve state (Mayhill 1974), (6) evaluation of gas
passage, (7) dynamic simulation (Hall and Decker 1996; Mantecon
2007), (8) surveillance of key performance parameters, and numerous other such techniques.
Historical Approach. Historically, well-by-well trouble-shooting
and optimization has often been performed on an ad-hoc or asneeded basis. In most cases, an internal or external expert would be
called upon to assist in evaluating a well once it had been identified
as having a problem. This expert then might use some or all of the
tools described above to assess the wells condition, identify the root
cause of performance problems, and recommend corrective actions.
112

This expert may even oversee the implementation of these corrective measures and evaluate the performance of the well after performing these services. In certain cases, a more proactive approach
might be used, in which an expert is placed in-house with the task
of systematically identifying underperforming wells and addressing
well performance issues in a prioritized, sequential manner.
Challenges With Historical Approach
The approach described presents a variety of challenges to operators. One of the most fundamental challenges is that this approach
tends to be both reactive and episodic in nature, resulting in missed
opportunites for production enhancement. In addition, much of this
work requires individuals with specialist artificial-lift-domain expertise, which is increasingly scarce as the demographics of the industry
change over time. Even in those cases in which a resident expert is
present in an asset, ones ability to detect and address the numerous
opportunities in a field is limited by the labor-intensive nature of the
work and the sheer volume of competing priorities. Finally, it is common for problems in gas lifted wells to go undetected for months or
even years because gas lift is such a forgiving artificial lift method.
Even those gas lifted wells that have a serious performance problem
and are not producing optimally will often continue to produce fluids.
With other forms of lift, failures tend to be catastrophic in nature and
are identified and addressed much more quickly.
Addressing the Need
Role of Intelligent Agents. In order to address these challenges,
an intelligent-agent system has been developed to provide realtime diagnostics of continuous gas lift wells. The role of the
intelligent-agent system is to enable every surveillance engineer,
regardless of experience or skill level, to make decisions that lead
to optimization of wells with the knowledge of a world-class gas
lift analyst.
Intelligent agents do this by providing engineers with the status
of all gas lifted wells under their control. Agents monitor the wells
situation by collecting and cleansing data, assessing the meaning
of these data, recommending actions for correcting problems and
responding to threats, and explaining their assessment results and
recommendations. Agents can detect the initial symptoms of a
problem and prompt for corrective action before the wells performance seriously degrades. The performance of these key functions
enables surveillance engineers to optimize a much larger number
of wells on a continuous basis.
Once armed with these tools, engineers no longer have to assess
the state of each well manually. The agents integrate continuous
data such as pressure readings with well-test data and predictions
from commercially available systems-analysis tools, using diagnostic principles stored in a knowledge base to determine each
wells condition and recommend corrective actions. The agent
reviews all gas lifted wells in the field and prioritizes recommended
actions in accordance with pre-established criteria that includes
increased production potential and possible cost efficiencies.
February 2010 SPE Production & Operations

Fig. 2Portion of gas lift knowledge base.

System Components. Knowledge Base. The central component

of the system is its knowledge base. The knowledge base is the
portion of the system that captures the mental model of the engineer. In this project, it was important to have a system that was
tolerant of missing or inaccurate data and could produce reliable
answers in less-than-ideal conditions. It was also desirable for the
system to be able to learn from experience and generate increasingly reliable answers over time. In order to accommodate these
needs, the knowledge base was assembled in a manner that was
not hierarchical, combined both deterministic and probabilistic
methods, and was overdefined.
Unlike typical troubleshooting methods (Ortiz 1990), this
knowledge base did not rely upon a top-down hierarchical workflow. Instead, knowledge was organized as a series of individual
cases (diagnoses), each with a unique set of attributes. This is
illustrated in Fig. 2. In all, more than 60 unique cases were
defined for wells operating through continuous gas lift with either
injection-pressure-operated (IPO) or production-pressure-operated
(PPO) gas lift valves.
By organizing the knowledge base in this manner, the system could accommodate a number of possible diagnoses for the
same condition, each with its own level of probability. Although
the individual attributes were each deterministically derived, the
arrangement of these into a set of discrete cases enabled the system
to provide a probability (or likelihood) for each diagnosis. This
arrangement also enabled the system to identify which individual
attributes contributed most strongly to a given diagnosis.
Another key aspect of the knowledge base is that it was overdefined. For a given condition, there could be multiple attributes
that would identify its occurence. As a result, even when certain
information was unavailable or inaccurate, the system still had the
ability to correctly identify a given condition. Having more than
one such attribute present would simply increase the certainty of
a given diagnosis.
Agent Attributes. Because of the real-time nature of this project,
the scope of work could include only those techniques that could be
performed automatically. Therefore, manual, episodic operations
such as flowing gradient survey acquisition and interpretation were
not considered. Instead, a series of attributes were collected that
were a function of analog inputs, the results of systems analysis,
calculation of valve state, calculation of gas passage, or a calculation based on a combination of these items. In all, more than 30
attributes could be collected for each analysis run.
February 2010 SPE Production & Operations

Technology. The agents underlying technology is an adaptation of methods that are used in applications such as document
classification and search on the Internet (Strategic Computing
1983; Broadwell 1998; Edwards and Geddes 1991; Geddes et al.
1996; Geddes et al. 1998; Geddes et al. 1990). Software programs
known as entity extractors can electronically read documents and
pick out the people, places, events, organizations, and other things
mentioned in the document. Other programs can determine which
entities typically occur together in a document of a particular type.
Given a query string such as US Airways Hudson River, search
programs can find documents and materials related to the 2009
crash landing of a US Airways flight in the Hudson River. Document searching is an application of pattern matching.
Relating the notion of pattern matching to gas lift analysis, we
know that in cases of valve cycling, the attributes casing heading:
yes and tubing heading: yes typically occur together. In wells
behaving normally, the attributes casing heading: no and tubing
heading: no typically occur. Gas lift experts understand the complex patterns of data that characterize various gas lift fault states.
The gas lift intelligent agent follows essentially the same process as a knowledgeable engineer. The engineer typically examines
well-test data, reviews trend data such as casing and tubing pressures, and refers to analytical models using systems-analysis tools.
In many cases, at least some of the data will be either missing or
suspect. Making sense out of this sometimes conflicting data is the
art practiced by an expert gas lift engineer.
The agent treats the art of gas lift diagnosis as a pattern-matching problem that involves overlaying the data that describe a wells
current state with information in the knowledge base that describes
problem states for gas lift systems. The set of attributes compared
is fairly large, on the order of more than 25 attributes. Like an
Internet search program, the agent returns an ordered list of the
best matches it finds. A site such as Google might return thousands
of matches, but the gas lift agent is configured to return only three
top candidate cases.
The agent has features that make it uniquely suitable for gas
lift. First, the pattern matcher can be tuned to handle special
circumstances. An attribute can be weighted (its importance can
be decreased or increased) if the attributes value is suspect or if
the attribute should be given special consideration during pattern
matching. Second, the pattern matcher can explain its results
by providing information about the attributes that most strongly
influenced its selection of a case. This feature provides a way for
113

Production
temperature

Flowline
pressure

Produced
fluids

Wellhead
pressure
Injection-gas
Injection-gas
temperature
control valve
Casinghead
Differential
Supply
pressure
pressure Static
pressure
pressure
Supply
gas

Gas-injection rate/
Gas flow meter

Gas lift valves

Annular fluid level

Bottomhole
pressure
Bottomhole
temperature

Fig. 3Gas lift system data elements.

users to see what the agent is thinking. Finally, over time, and
as needed, the agents knowledge base can be extended with additional cases that reflect conditions actually encountered in the field.
Thus, the agents performance can be refined over time.
Data Requirements. As depicted in Fig. 3, a variety of data
elements are required in order to fully define the flowing state of a
gas lift well. These data include real-time, well-test, and static data
elements. Real-time data include analog inputs such as flowline
pressure, supply pressure, casinghead pressure, flowline temperature, flowline pressure, and gas-injection rate. Well-test data include
information such as production rates (oil, water, and gas), gas-injection rates, tubing and casing pressures, and fluid gravities. Static
data include physical attributes such as perforation depths, well
deviation, completion geometry, inflow-performance data, and gas
lift valve specifications. Each of these data items is collected and
used to populate the intelligent-agent system and the well models
it uses in conjunction with systems-analysis operations.
System Architecture. Because oil production companies use a
wide variety of databases and analytical models, the agent has been
designed to be used in a variety of information technology (IT)
environments. Fig. 4 represents the systems architecture.
The system is implemented with a thin-client/server architecture. System components communicate with each other using a
Web services protocol.
Security Layer. The security layer controls user access to the
system using Windows login ID so that users do not have to log in
separately to the system. The security system is being enhanced to
provide basic role-based access control that distinguishes between
super users with administrative privileges and ordinary users.
Data Migration Layer. The data migration layer is responsible for fetching data from a production companys data stores,
including well-configuration data, well-test information, and trend
data. The agent currently is integrated with Weatherfords Life of
Well Information Software (LOWIS). However, data migration is
designed to enable integration with whatever data sources are used
by a producer, including real-time sources. The data mover pulls
data from their sources and pushes data to the middle tier, where
they are stored in the gas lift system database, which is a cache for
storing data used during analysis and the results from analysis.
Middle Tier. The middle tier includes the business logic for
analyzing a well. The application server module combines well-test
114

data and any available trend data. It invokes an analytical well model
to gather information about the wells configuration (e.g., types of
valves, their depths). The agent currently uses WellFlo as its analytical well model. However, consistent with the principle that it should
run in many IT environments, the agent will be integrated with
additional analytical models such as Prosper or WinGLUE. Using
data derived from the analytical model, the server performs calculations to determine whether valves are open, closed, or back-checked,
and it determines gas passage for each open valve. It combines this
information with all the other data gathered into a single record.
This record is an input context, which is a vector of attributes that
describes the current state of the well. The pattern matcher compares
the wells state with cases in the knowledge base and returns the top
three similar cases to the application server. The post-processor performs additional calculations primarily related to economic benefits
that might be realized by optimizing the well.
Client. The client is the agents primary user interface and is
discussed in the next section.
Client Graphical User Interface. The graphical user interface
(GUI) is a thin client that communicates to the middle tier using
Web services. It is not a Web application, but a browser-based
interface could certainly be implemented. The user interface is just
an interface, meaning that its function is simply to display data
rather than to perform any analysis functions. The GUIs purpose
is to present information to production engineers in a manner
consistent with the mental model for gas lift analysis. The mental
model describes how users typically think about gas lift analysis,
including the kinds of data that are useful when assessing a well.
A simple example of the mental model is that some users measure
depth in feet and others in meters. Accordingly, the GUI includes a
unit editor so that users can tailor unit values to their preferences.
The GUI is illustrated in Figs. 5 and 6. Fig. 5 is the system
dashboard.
The display provides a number of features for filtering and sorting data. The dashboard is essentially a tool for doing management
by exception. Users can select a well from the dashboard and drill
down to a detailed display.
Fig. 6 is the details display; it displays comprehensive information about the wells current diagnosis and supporting information
for this diagnosis as gathered from the latest well test, from trend
data, from the analysis model, and from the knowledge base.
February 2010 SPE Production & Operations

Security
Security
Service

Security
DB

Client

Security GUI

Middle Tier
Gas Lift
System DB

Anaylical
Well Apps

Application Server
Dashboard
Detailed Results
Well Mgmt

Analysis Controller
Valve Calculations
Gas Passage

Pattern
Matcher

Well
Models
Knowledge
Base

Post
Processor

Data Migration
Data Mover
Business
Logic

Configuation
DB

Configuration GUI
Data Sources

Fig. 4System architecture.

Fig. 5System dashboard.

February 2010 SPE Production & Operations

115

Fig. 6Details display.

The analysis history section (left side of the display) lists all
the analysis results performed for the well. Users can select any
of the items in the history list for review.
The well header pane provides basic well-configuration data.
The analysis results section (Fig. 7) lists three possible diagnoses
for the well: a primary, secondary, and tertiary condition. The likelihood value provided for each diagnosis is a rough measure of similarity
between the wells input context and a diagnosis case, normalized to
one across the three cases. The likelihood is not a probability.
The analysis trend data region (Fig. 8) provides graphs of
well-trend data such as casing and tubing pressures, gas-injection
rate, and supply pressure.

The diagnosis inputs section (Fig. 9) is the input context for

this analysis run. All of the attributes used by the pattern matcher
are included in this panel.
The valve mechanics section (Fig. 10) provides comprehensive information about the valves installed in the well, including
their mechanical status (open/closed/backchecked), calculated gas
passage, etc.
The well-test panel (Fig. 11) provides data from the wells
last well test and additional attributes calculated using the analytical well model.
The analysis model graphs region (Fig. 12) provides graphs generated by the analytical well model such as pressure drop curves.

Fig. 7Analysis results display.

116

February 2010 SPE Production & Operations

Fig. 8Analysis trend data.

Fig. 9Diagnosis inputs.

The impact model output section (Fig. 13) provides estimates

of revenue and production rates that may be possible if the wells
performance issues are corrected.
Testing. The system prototype was thoroughly tested in a laboratory environment prior to the field trial. The objectives of the
laboratory testing were to evaluate three major criteria: First, does
the systems data migration layer correctly import data into the
system, including both analog and well-test data? Second, does
the systems workflow layer follow the correct steps with respect
to gathering data and invoking the correct analytical calculations?

Finally, does the systems knowledge-base component generate

correct results?
The following discrete tests were performed:
Data Retrieval Testing. The purpose of this test was to ensure
that the system pulls the correct real-time and well-test data and
populates this information in its database and well models.
Testing of Analysis Modes. The system was required to support
three modes of analysis: (1) periodicanalysis is triggered each
time a new well-test file is shipped into the system, (2) continuousanalysis performed once a day on the basis of average values
of the analog signals over the previous 24 hours using the most

Fig. 10Valve mechanics panel.

Fig. 11Well-test panel.

February 2010 SPE Production & Operations

117

Gas-Injection Rate

Fig. 12Analysis model graphs.

recent test rates, and (3) manualanalysis process is initiated by

a user at any time.
Testing of the Agents Output Results and Recommendations
on Corrective Actions. The software was fed a set of well-test
data representing different well-performance conditions (e.g., normal operation by injecting at deepest point, multipoint injection,
circulation of gas above the top valve, overinjecting lift gas) and
allowed to perform the analysis and provide diagnostic results and
recommendations. At the same time, experienced gas lift experts
independently analyzed the same set of data using standalone
system-analysis software, and recorded their observations about
the performance of analyzed wells. In most cases, the experts
also were able to collect and evaluate current flowing pressure
and temperature gradient survey data to aid in this evaluation.
The results and recommendations of corrective actions generated
by the software were then compared to the ones provided by the
experts. The software was considered to pass the test when its
conclusions coincided with the conclusions of the independent
manual analysis.
Data Validation Testing (Handling Missing Data). This test
concerned the ability of the system to handle missing or invalid
data. If a well-test file is missing the following parameters, the
system will substitute values from corresponding analog data that
is concurrent with the time of the well test. These attributes include
flowing wellhead pressure, supply pressure, gas-injection rate, and
flowline temperature.
The ability of the system to perform boundary checking (determining if data is out of range) was also tested. In addition, the ability

of the software to detect no-flow conditions based on the slope of

change of flowline temperature and pressure was tested.
Analysis Workflow Testing. A specific analysis workflow is
incorporated in the agent. The workflow consists of a number of
consecutive computations and comparison of computation results
against the knowledge base. This test concerned the systems ability to successfully execute this workflow.
Training of the Knowledge Base and Agent Memory (Prediction). Two unit tests and one system-level test were performed on
the agent memory.
First, the memory was tested with input contexts fully defined
for each of the cases in the knowledge base. This test ensures
that the memory is trained correctly and that the knowledge base
performs as expected when queried using a fully defined context.
This is a unit-level test.
Second, the memory was tested using input contexts that were
partially defined for each of the cases in the knowledge base. This
case is consistent with the conditions expected to be found in the
field. The missing input contexts included all attributes that require
a measured fluid level as well as all attributes that require a separator pressure or backpressure. This is a unit-level test.
Finally, the memory was tested using real-world data. This
test is a system-level evaluation.
The software successfully passed laboratory testing. It gave
reliable diagnostic results and recommended corrective actions
which, in most of the cases, were in line with the results obtained
by experienced gas lift engineers by analyzing the same set of
data independently using standalone software. Those few cases in

Fig. 13Impact model output.

118

February 2010 SPE Production & Operations

Injection Rate, Casinghead Pressure,

Flowing Wellhead Pressure,
Flowline Temperature

1,700
1,600
1,500
1,400
1,300
1,200
1,100
1,000
900
800
700
600
500
400
300
200
100
0

180
170
160
150
140
130
120
110
100
90
80
70
60
50
40
30
02/15/2009 2/28/2009 03/15/2009

03/31/2009

04/15/2009 04/30/2009

Time
Gas-injection rate
Flowline temperature

Casinghead pressure
04/13/2009

Flowing wellhead pressure

Fig. 14Well 1 historical data trends.

which the ternary diagnostic results were not in agreement with

those of the specialists were investigated. The software knowledge
base was expanded or refined to accommodate those cases.
Field Trial. Following the laboratory testing, a field trial was conducted in a large, mature waterflood located in the western United
States. This onshore field has several thousand wells operating on
a variety of forms of lift, including natural flow, reciprocating rod
lift, electric submersible pumping, progressing cavity pumping,
and continuous gas lift. At the time of the field trial, 84 of these
wells were actively producing through gas lift. Each of these wells
is fully instrumented and controlled using a remote terminal unit
(RTU), which, in turn, communicates through radio to a host system. Typical instrumentation includes the following:
Gas lift supply pressure, psig
Casinghead pressure, psig
Flowing wellhead pressure, psig
Flowline temperature, F
Gas-injection rate, Mscf/D
As with the laboratory testing, each of these analog inputs,
along with relevant well-test parameters, were imported into the
intelligent-agent system for use in its analysis.
The goals of the field trial were similar to those of the laboratory
testing. First, the system had to successfully import well-test and analog input data from the real-time host system. It then had to follow the
correct steps with regard to importing data and invoking its analytical
processes. Finally, it had to generate results that were consistent with
those that were independently obtained from the gas lift experts.
Deployment. In preparation for this field trial, the system was
deployed to a field-trial server isolated from the local operational
server but available on the corporate wide area network (WAN).
The field trial machine was a midlevel server running Windows
XP. LOWIS, WellFlo, SQL Server, and the agent were installed on
the machine. To ensure data integrity on the operational system,
local IT staff developed a manual procedure for copying data from
the live LOWIS instance to the LOWIS instance on the field trial
server. The client (i.e., the user interface) was installed on several
user desktop computers both at the field and at the corporate offices
in Houston. Thus, the agent functioned as if it were operationally
deployed, except that it used a copy of the local data.
Results
Consistent with the performance seen in the laboratory test environment, the system produced impressive results once deployed
February 2010 SPE Production & Operations

in the field. In general, the system provided diagnoses and recommendations that were consistent with the manual diagnoses
of the experts, often with a startling degree of accuracy. The
following case studies illustrate three such examples of system
performance.
Case Studies. Well 1. This well had been producing stably for the
previous 15 months with an average production rate of 1,323 BLPD,
a 98% water cut, 895 Mscf/D of produced gas, and 881 Mscf/D of
injected gas. As can be seen in Fig. 14, the flowing wellhead pressures, casinghead pressures, gas-injection rates, and temperatures
were stable over the previous several months, with only minor fluctuations in flowline temperatures because of day/night patterns.
After analyzing the conditions for this well, the agent made the
following primary diagnosis:
Scenario: Steady single-point injection at deepest valve in well.
Condition: No apparent problems at this time.
Likelihood: 42.68%.
Action recommended: Gas lift design is appropriate for
conditions.
This diagnosis was influenced by a variety of attributes, but
the key contributors to this particular diagnosis were that (1) only
one active gas lift valve (the deepest) was determined to be open,
(2) the gas-injection rate was not fluctuating, and (3) the deepest
active gas lift valve was not closed. This diagnosis along with
the secondary and tertiary diagnoses can be seen in Fig. 15. As
Fig. 15 illustrates, other (less likely) diagnoses included Steady
single point of injection not at deepest point in well and surging single point of injection through deepest valve in well, low
casing pressure.
Independent of this diagnosis, two gas lift engineers evaluated
the same information as well as a recent flowing pressure and
temperature gradient survey for the well (Fig. 16).
After modeling the well and evaluating the flowing gradient
survey, both engineers independently concluded that the well was
indeed operating through single-point injection at the deepest station in the well and that performance was close to optimal. This
was consistent with the diagnosis provided by the agent.
Well 2. This well had been producing at an average production rate of 272 BLPD, a 95% water cut, produced-gas rate of 18
Mscf/D, and a gas-injection rate of 542 Mscf/D. As illustrated in
Fig. 17, the well was unstable and experienced both tubing and
casing pressure heading with an extremely low casinghead pressure
(approximately 200 psi). After analyzing the conditions for this
well, the agent made the following primary diagnosis:
119

Fig. 15Well 1 diagnoses and recommendations.

Scenario: Gas is injected/valve mechanics indicate that valves

should be closed.
Condition: Possible communication at or above 4355.12 ft.
Likelihood: 50.98%.
Action recommended: It may be possible to confirm communication as follows: (1) first, shut in tubing, (2) continue to inject
gas into annulus, (3) once the pressures equalize, bleed down the
annulus as quickly as possible, then (4) observe tubing pressure.
If tubing pressure drops with casing pressure, this confirms that
communication exists. It is also useful to shoot the fluid level in
the casing before and after bleeding down the casing. If fluid level
rises, this also indicates communication. Alternatively, run a flowing gradient survey to identify point(s) of injection.

The primary contributors to this diagnosis were that (1) the

casing pressure at depth was less than the opening pressure for
all valves, (2) the gas-injection rate into the well was greater than
the throughput of all active valves, and (3) on the basis of valve
mechanics alone, no valves were determined to be open.
Following this analysis, troubleshooting operations were conducted in the field using slickline-based methods. Operations personnel confirmed that a hole was present in the tubing at a depth of 4,360
ft, just 5 ft below the communication point predicted by the agent.
A decision was then made to replace the tubing string. The well
was subsequently recompleted to restore tubing integrity. Following
the recompletion, the well returned to operation with a stabilized
production rate of 330 BLPD, a 21% increase in production.

Fig. 16Well 1 survey matching.

120

February 2010 SPE Production & Operations

InjectionRate, Casinghead Pressure,

Flowing Wellhead Pressure,
Flowline Termperature

CHP, FWP, FLT, QCI

1,500
1,400
1,300
1,200
1,100
1,000
900
800
700
600
500
400
300
200
100
0

Post-workover
Preworkover

150
140
130
120
110
100
90
80
70
60
50
40

30
05/24/2009 05/31/2009 06/07/2009 06/14/2009
Time
Flowing wellhead pressure
Casinghead pressure
05/31/2009

05/10/2009 05/17/2009
Gas-injection rate
Flowline temperature

Fig. 17Well 2 historical data trends.

InjectionRate, Casing Head Pressure,

Flowing Wellhead Pressure,
Flow Line Termperature

Well 3. This well had been producing at an average production

rate of 1,197 BLPD, a 95% water cut, and a produced-gas rate of
128 Mscf/D, with a gas-injection rate of 631 Mscf/D. As can be
seen in Fig. 18, the well was unstable and experienced severe tubing and casing heading, along with fluctuations in gas injection.
After analyzing the conditions for this well, the agent made the
following primary diagnosis:
Scenario: Multipoint injection through valves 1, 2.
Condition: Valves failing to close properly. This could be due
to flow cutting, bellows failure, or mechanical obstruction.
Likelihood: 37.86%.
Action recommended: Consider running flowing gradient survey to confirm multipoint injection. Replace valves as appropriate.
The primary attributes contributing to this diagnosis were that
(1) the gas injection rate was greater than the combined capacity
of all active valves, (2) the casing pressure at depth was less than
the opening pressure of all valves, and (3) the gas-injection rate
was greater than the capacity of the deepest open valve.
After independently evaluating these data along with a recent
flowing gradient survey obtained for the well, each of the gas lift
800
750
700
650
600
550
500
450
400
350
300
250
200
150
100
50
0
03/15/2009

experts determined that the well was indeed injecting through the
top two gas lift valves. This well is currently being evaluated for
possible corrective action.
Benefits Realized. Based on its initial deployment and use in the
field, it is evident that the sytem provides a number of clear benefits
to the operator. Among these benefits are the following.
1. Well performance issues and optimization opportunities can
be identified much faster than is possible through conventional
means.
2. Operations personnel are able to analyze more wells in less
time.
3. The system provides a means to prioritize work opportunities
and direct operations and well servicing resources toward those
efforts that provide the greatest value.
In addition to the benefits described, the system plays a key
role in the production management workflow.
Role in Production Management Workflow. A variety of
production-management activities are performed in any oil- and
gas-producing asset in order to optimize well performance and
173
172
171
170
169
168
167
166

03/31/2009

04/15/2009

Gas-injection rate
Casing head pressure

04/30/2009
Time

05/15/2009

Flowing wellhead pressure

Flowline temperature

05/31/2009

165
06/15/2009

06/14/2009

Fig. 18Well 3 historical data trends.

February 2010 SPE Production & Operations

121

Step 1:
Get Data

Step 2:
ID Opportunity

Step 3:
Analyze

Step 4:
Execute

Step 5:
Look Back

Fig. 19Production-management workflow.

maximize uptime. These activities can be described by a common

workflow such as the one depicted by Fig. 19. The first step in this
process is to get data. This includes all of the manual or automatic
collection, sorting, and aggregating of data needed to perform basic
surveillance and analysis of the wells. The second step is to use
these data to identify opportunities for improvement. Next, these
opportunities are analyzed to validate the opportunity, identify
corrective actions, identify the cost and benefit of such work, and
rank the opportunities as candidates for well work. Next, the work
is executed in the field. Once the work is completed, scorecarding
is then performed to evaluate the effectiveness of this work and
provide lessons learned that can be used to aid future decisions.
Upon completion of these steps, the process is repeated.
The intelligent-agent system plays a key role in this process by
automating most, if not all, of the first three steps. By collecting and
aggregating all of the pertinent data, identifying opportuntities for
improvement, and providing detailed analysis of these opportunities,
the system frees personnel to focus a majority of their energy on the
actual execution of work and evaluation of results. In this way, the
system shifts the focus of the production management process from
purely reactive activities to proactive high-value activities.
Areas for Improvement and Future Development. Intelligent
gas lift agents are designed to learn more and, therefore, improve
over time. This ongoing process is vital for the continued value and
use of the agents. As the system is exposed to a greater number of
real-world cases and conditions, the knowledge base will need to
be expanded and the agents trained to better match reality.
Training of Agents. The agent has been trained to date using the
knowledge base without including many real-world observations.
However, further training is possible and desirable. It is expected
that the fully trained agent will have on the order of a few hundred
cases in addition to the base cases in the knowledge base. An
important point is that the agent is not configured to learn on its
own. Rather, the agent is configured for directed learning, meaning
that the agent learns what we teach it. The question, then, is: What
is the protocol for training the agent using real-world data?
An old adage related to computers is garbage in/garbage out,
meaning that the results produced by a system are only as good as
the data fed to the system. The same adage applies to an intelligent
agent. In other words, poorly trained agents perform poorly. Agent
training has to be done in a carefully controlled manner.
The agents knowledge base is well conceived and implemented.
However, the real world can be unexpectedly complicated, and the agent
may not always perform as users expect. While field users should be
able to provide feedback related to the agents performance, they should
not be able to train the agent themselves. After all, different usersand,
indeed, different expertshave different opinions. Accordingly, the following protocol has been adopted for agent training:
1. A feedback button has been incorporated on the agents
well details display. Users who disagree (or agree) with the agents
diagnosis can state their position and electronically forward a
packet of well data back to a customer support desk.
2. A gas lift expert at the support center assesses the merits of
feedback and determines whether the case should be added to the
knowledge base. This assessment process may involve some backand-forth communication between experts at the support center
and users in the field.
3. The engineering support staff adds cases to the knowledge
base as directed by the gas lift expert.
4. Periodically, a new release of the agent is distributed that
incorporates new cases.
Stated succinctly, the knowledge base is treated as a corporate
asset, and procedures have been adopted to ensure that the knowledge bases integrity is preserved.
122

Future Development. Future developments in intelligent-agent

technology in the production domain will include support of new work
processes and development and deployment of agents for other activities in artificial lift, flow assurance, and secondary recovery. Agents
also will be built in a manner enabling collaboration with other agents.
This will help manage the tradeoff between maximizing gains in one
area while hindering the optimization of other areas.
Conclusions
The field trial was successful in demonstrating the benefits that
can be derived from using agents to enable production engineers
to manage and optimize many more gas lifted wells than has been
possible in the past. A variety of insights were gained from this
project. These include the following:
1. Agents are able to process the massive amounts of data that
needed to be monitored from all of the gas lifted wells.
2. The agents used the knowledge base effectively to reach conclusions about well conditions, explain the diagnoses given,
and recommended the right remedial actions.
3. The knowledge base used by the agents system was sufficiently
comprehensive so that the system functioned efficiently from
the beginning. This requires comprehensive domain expertise
but makes performance less dependent upon training from
representative data sets that are hard to acquire. Continuous
training is then carried out in the future.
4. The agents were required to increase the productivity of both the
experienced and inexperienced operators. The agents overcame
the lack of experience of individual surveillance personnel. They
enabled nonengineers to use the results from sophisticated tools that
would otherwise not be available to them. The criteria for success
from this type of operator was met: the agent had to be fast, easy
to use, and accurate. This will be critical in the future because of
the much-publicized big crew change occurring in the industry.
Such a tool will allow experienced operators to analyze many more
wells than they could in the past in a given period of time.
5. The knowledge incorporated by the agents was sufficient from
the start so that the agents correctly diagnosed the conditions
in the vast majority of cases. The agents received further training in cases that were discovered during the field trial. These
were unusual cases and required modification of the knowledge
base. These modifications were easily achieved.
6. The agents were able to help optimize the performance of many
wells that were not even looked at in the past because there
were not enough experienced people assigned to perform the
work. This resulted in increased production and income.
7. The gas lift agents overcame a fundamental problem that has
plagued the industry in the use of certain types of artificial
intelligence systems: One does not know why the black box
has reached specific conclusions. The fact that agents emulate
normal decision making through the display of the scenario, condition, likelihoods, attributes considered, and the weight given
to each attribute, models used, well-test data, and actions to be
taken means that the agent can be trusted by the operator.
8. The agents were able to identify systems-analysis models that
were not reflective of current conditions and needed to be
investigated and tuned. This is an extremely helpful feature for
surveillance engineers because it ensures accuracy of the information used to make decisions. Often, models are not maintained because it is impractical to do so. Agents identify which
models are no longer accurate as conditions change, enabling
engineers to quickly update those models in need of attention.
This provides additional value in cases in which the models are
being used in conjunction with a full-field network model.
9. The incorporation of production costs provided the basis for the
selection of which wells should receive remediation by virtue
February 2010 SPE Production & Operations

of the greatest ROI or other applied criteria. This was important

in the demonstration of the specific value that can be created
on a well-by-well basis.
10. The use of gas lift agents appears to overcome some of the
difficulties encountered in processing and cleansing large data
sets, fusing attributes from multiple sources, and enabling
operators to understand what is going on by capturing their
mental model. This provides the possibility of using agents in
other production applications.
Acknowledgments
The authors would like to thank their respective employers, Occidental Petroleum Corporation, Intelligent Agents Corporation,
and Weatherford International and its affiliates, for permission to
publish this paper. The authors would also like to thank their many
coworkers for their contributions to the project development.
References
API RP 11V5, Operation, Maintenance, Surveillance, and Troubleshooting
of Gas-Lift Installations, second edition. 1999. Washington, DC: API.
Broadwell, M. 1998. Pilots Associate: The Intelligent Edge. Presented at
LokTec X, The Annual Lockheed Technical Symposium, Washington,
DC, November.
Brul, M., Charalambous, Y., Crawford, M.L., and Crawley, C. 2008.
Bridging the Gap Between Real-Time Optimization and InformationBased Technologies. Paper SPE 116758 presented at the SPE Annual
Technical Conference and Exhibition, Denver, 2424 September. doi:
10.2118/116758-MS.
Edwards, G.R. and Geddes, N.D. 1991. Deriving a Domain-Independent
Architecture for Associate Systems From Essential Elements of Associate
Behavior. In Associate Technology: Opportunity and Challenges, ed. P.E.
Lehner, 1729. Fairfax County, Virginia: George Mason University.
Gas Lift Design and Technology course manual, Chap. 6, 23. 1999.
Houston, Texas: Schlumberger Oilfield Services.
Geddes, N.D. and Sewell, D.R. 1990. A Plan and Goal Based Method for
Computer-Human Design. In Human-Computer InteractionINTERACT 90: Proceedings of the IFIP TC13 Third International Conference
on Human-Computer Interaction, Cambridge, UK, 2731 August, 1990,
ed. D. Diaper, D.J. Gilmore, G. Cockton, and B. Shackel, 283288.
Amsterdam, The Netherlands: North-Holland Publishing Company.
Geddes, N.D., Brown, J.L., Beamon, W.S. III, and Wise, J.A. 1996. A
shared model of intentions for Free Flight. Proc., 15th AIAA/IEEE
Digital Avionics Systems Conference, Atlanta, Georgia, USA, 2731
October, 179184. doi: 10.1109/DASC.1996.559154.
Geddes, N.D., Smith, D.M., and Lizza, C.S. 1998. Fostering collaboration in systems of systems. Proc., 1998 IEEE International Conference on Systems, Man,
and Cybernetics, Vol. 1., 950954. doi: 10.1109/ICSMC.1998.725538.
Hall, J.W. and Decker, K.L. 1996. Gas-Lift Unloading and Operating
Simulation as Applied to Mandrel Spacing and Valve Design. Paper
SPE 29450 presented at the SPE Production Operations Symposium,
Oklahoma City, Oklahoma, USA, 24 April. doi: 10.2118/29450-MS.
Hite, J.R., Crawley, C., Deaton, D.F., Farid, K., and Sternevsky, M. 2007. Barriors to Implementation of Real-Time Operations Technology. Paper SPE
110236 presented at the SPE Annual Technical Conference and Exhibition,
Anaheim, California, USA, 1114 November. doi: 10.2118/110236-MS.
Kanu, E.P., Mach, J., Brown, K.E. 1981. Economic Approach to Oil
Poduction and Allocation in Continuous Gas Lift. J. Pet Tech 33 (10):
18871892. SPE-9084-PA. doi: 10.2118/9084-PA.
Litvak, M.L., Hutchins, L.A., Skinner, R.C., Darlow, B.L., Wood, R.C., and
Kuest, L.J. 2002. Prudhoe Bay E-Field Production Optimization System
Based on Integrated Reservoir and Facility Simulation. Paper SPE 77643
presented at the SPE Annual Technical Conference and Exhibition, San
Antonio, Texas, USA, 29 September2 October. doi: 10.2118/77643-MS.
Mach, J., Proano, E., and Brown, K.E. 1979. A Nodal Approach for Applying Systems Analysis to the Flowing and Artificial Lift Oil or Gas Well.
Paper SPE 8025 available from SPE, Richardson, Texas.
Mantecon, J.C. 2007. The Virtual Well: Guidelines for the Application of
Dynamic Simulation to Optimize Well Operations, Life Cycle Design,
and Production. Paper SPE 109829 presented at the SPE Annual Technical Conference and Exhibition, Anaheim, California, USA, 1114
November. doi: 10.2118/109829-MS.
February 2010 SPE Production & Operations

Mayhill, T.D. 1974. Simplified Method for Gas-Lift Well Problem Identification and Diagnosis. Paper SPE 5151 presented at the Fall Meeting
of the Society of Petroleum Engineers of AIME, Houston, 69 October.
doi: 10.2118/5151-MS.
McKie, C.J.N., Rojas, E.A., Quintero, N.M., Chacn Fonseca, J.R., and
Perozo, N.J. 2001. Economic Benefits From Automated Optimization
of High Pressure Gas Usage in an Oil Production System. Paper SPE
67187 presented at the SPE Production Operations Symposium, Oklahoma City, Oklahoma, USA, 2427 March. doi: 10.2118/67187-MS.
Mochizuki, S., Saputelli, L.A., Kabir, C.S., Cramer, R., Lochmann, M.J.,
Reese, R.D., Harms, L.K., Sisk, C.D., Hite, J.R., and Escorcia, A. 2006.
Real-Time Optimization: Classification and Assessment. SPE Prod &
Oper 21 (4): 455466. SPE-90213-PA. doi: 10.2118/90213-PA.
Nadar, M.S., Kulkarni, R., Lemanczyk, Z.R., and Aggarwal, P. 2006.
Development of an Integrated Production Network Model for the Heera
Field. Paper SPE 101089 presented at the SPE Asia Pacific Oil and Gas
Conference and Exhibition, Adelaide, Australia, 1113 September. doi:
10.2118/101089-MS.
Nadar, M.S., Schneider, T.S., Jackson, K.L., McKie, C.J.N., and Hamid,
J. 2008. Implementation of a Total-System Production-Optimization
Model in a Complex Gas-Lifted Offshore Operation. SPE Prod & Oper
23 (1): 513. SPE-103670-PA. doi: 10.2118/103670-PA.
Ortiz, J.L. 1990. Gas-Lift Troubleshooting Engineering: An Improved
Approach. Paper SPE 20674 presented at the SPE Annual Technical Conference and Exhibition, New Orleans, 2326 September. doi:
10.2118/20674-MS.
Peringod, C., Clark, J., Barboussat, D., Peacock, L., and Holloway, J.
2009. Gaslift diagnosis using WellTracer surveysuccessful trials in
20 wells of Petroleum Development Oman. Paper presented at 32nd
Annual ASME/API/ISO Gas-Lift Workshop, The Hague, 26 February.
doi: 10.2118/124926-MS.
Russell, S. and Norvig, P. 2003. Artificial Intelligence: A Modern Approach,
second edition, Chap. 2, http://aima.cs.berkeley.edu/. Upper Saddle
River, New Jersey: Series in Artificial Intelligence, Prentice Hall.
Strategic Computing. New-Generation Computer Technology: A Strategic Plan for Its Development and Application to Critical Problems
in Defense. 1983. Technical Report, Accession No. AD-A141 982,
Defense Advanced Research Projects Agency, Arlington, Virginia (28
October 1983).
Neil de Guzman is the president and CEO of Intelligent Agent
Corporation. He founded Intelligent Inspections Corporation
to develop autonomous robots and artificial intelligent systems under strategic alliances with BP, Statoil, Marathon,
Baker Hughes, and Halliburton. He founded Intelligent Agent
Corporation in 2004 to develop intelligent agents. He holds a
number of patents related to the application of intelligent
agents in producing oil and gas wells and reservoirs. He holds a
BA degree in economics and an MBA from Columbia University.
Larry Lafferty is the CTO of Intelligent Agent Corporation. He has
more than 20 years experience in research and development,
software engineering and program management, focused primarily on the development of intelligent systems. Lafferty has
worked at Harris Corporation, Ensco, and ISX Corporation where
he served as software technical lead for a number of intelligent
systems development efforts for the military and intelligence
communities. During the past five years, he has been adapting
these advanced software technologies to solve problems in the
oil and gas industry. Greg Stephenson is the production surveillance lead for Occidental Oil and Gas Corporation. Throughout
his career, he has served in a variety of roles focusing on production optimization, artificial lift, and production automation. In
addition to these duties, Stephenson has taught industry courses
throughout the world related to the selection, design, and application of artificial lift systems. He actively serves on a variety of
industry committees. He is the author of numerous technical
publications and holds multiple patents related to artificial lift
and completion technology. He holds a BS degree in petroleum
engineering from Texas Tech University. Roman Molotkov is the
product line manager for artificial lift optimization at Weatherford
International. He holds an MS degree in petroleum engineering from Russian State University of Oil and Gas. Throughout his
career, Molotkov has served in a variety of positions in operations, strategic marketing, and new product development in the
oilfield services industry.
123