You are on page 1of 3

ARTIFICIAL LIFT

Integrating Data Mining and Expert Knowledge


for an Artificial-Lift Advisory System

The full-length paper describes a new tion with the asset team to ensure the Current Situation. Under the current
workflow to accelerate and improve validity and consistency of the system working environment, two main situa-
decisions regarding where and when recommendations. While the iterative tions challenge the asset team:
to apply an artificial-lift system in fields validation is necessary for the BBN 1. Analysis of the huge amount of
with a considerable number of active setup, it also enables the asset team data involves a significant time delay,
wells. The workflow deploys a hybrid to review and visualize their complex which leads to lost production because
combination of a user-driven expert decision-making process thoroughly of delayed well intervention (e.g.,
system and a data-driven knowledge- and helps to define more-efficient installing the right artificial-lift system
capturing system calibrated with his- decision logic. in time).
torical data. The two systems interact In parallel, a data-mining analysis 2. There is insufficient processed
to determine the right point in time on historical data is performed to find data to make an informed decision,
to support a particular well with an the decision logic as it was applied which means that the asset team initi-
artificial-lift system. in the field. The data-mining study ates well interventions too early and
incorporates operational data (e.g., tub- hence blocks the intervention resourc-
Introduction inghead pressures, line pressures, and es for other wells that might require the
Typically, asset expertise is stored in calculated gas rates), well-test data, intervention investment sooner.
spreadsheets that define the limits for and information from well-intervention Currently, AIB has a contractual
condition parameters within which reports. The objectives of the data- scheme in which several service com-
a specific artificial-lift system needs mining analysis are: panies participate by measuring well
to be installed. As a first step toward • To identify patterns in operational parameters in the field and creating
decision automation, these limits are and well-test data that lead to the deci- engineering analysis of candidates to
implemented in a rule-based reason- sion of a new artificial-lift system. install the most appropriate artificial-
ing system deploying if-then-else logic • To identify potential correlations lift system. Because of the large number
to automate the knowledge. After suc- of well-intervention activity and incre- of wells in the asset, the associated
cessful validation of the rule-based mental production after intervention capture of required measurements, and
approach, the information about the to quantify the success of a specific the distance to the location of interest,
limits and the combination and inter- artificial-lift installation under specific the existing analysis process was per-
action of the parameters is fed into a conditions. formed at an average rate of only two
Bayesian belief network (BBN) capable wells per week.
of performing a more robust reasoning Case Study The current strategy focuses on
process under uncertainty or miss- Activo Integral Burgos (AIB) is a typi- accelerating the quantity and improv-
ing parameters. The setup process of cal example of a large gas field where ing the quality of the recommenda-
the BBN requires frequent interac- production declined as a result of gas- tions at the well level that help increase
loading backpressure and reduced or at least maintain the production of
This article, written by Assistant Tech- permeability in the target formation. the current asset. The asset team wants
nology Editor Karen Bybee, contains The fast decline of the gas wells to take advantage of its investment in
highlights of paper SPE 128636, “Inte- during their first year of production the automation of current processes
grating Data Mining and Expert Knowl- drove a change from a reactive to a related to operational data storage and
edge for an Artificial Lift Advisory proactive management approach to engineering workflows, such as gas-
System,” by E. De la Vega, G. monitor the field and select candi- rate estimation and event detection
Sandoval, and M. Garcia, SPE, Pemex, dates for workovers and artificial-lift and notifications.
and G. Nunez, SPE, A. Al-Kinani, SPE, systems. However, the large number An investigation of current condi-
R.W. Holy, SPE, H. Escalona, SPE, and of wells in AIB (approximately 3,500 tions also shows that while all deci-
M. Mota, Schlumberger, originally pre- active wells) and the fact that 95% sions are made on the basis of field
pared for the 2010 SPE Intelligent Energy of the data is obtained by manu- data, analysis relies heavily on an
Conference and Exhibition, Utrecht, The al surveillance suggest the need for engineer’s experience and judgment.
Netherlands, 23–25 March. The paper some level of automation to aid deci- Engineers typically move from one job
has not been peer reviewed. sion making. position to the next, and the arrival

For a limited time, the full-length paper is available free to SPE members at www.spe.org/jpt.

JPT • JULY 2010 57


Gas Rate After (MMscf/D) Gas Rate Before (MMscf/D) Line Pressure Before (psi) Type Numeric ( )

0.1 0.3 0.6 0.8 1.1 1.3 1.6 1.8 2.1 2.3 0.1 0.4 0.7 1.0 1.3 1.6 1.9 2.2 2.5 2.8 122 294 465 636 807 978 1,149 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
Wellhead Pressure After (psi) Wellhead Pressure Before (psi) Liquid Level Before (ft) Last Well-Test Gas Rate (MMscf/D)

190 487 783 1,080 1,377 1,874 1,971 178 478 778 1,078 1,378 1,678 1,978 89 503 917 1,330 1,744 2,157 2,571 0.1 0.5 0.8 1.1 1.5 1.8 2.1 2.5 2.8 3.1

Last Well-Test Water Rate (STB/D)

Type Numeric
Red: AL System A
Green: AL System B
Blue: AL System C
1 19 38 56 74 92 111 129 147

Fig. 1—SOM to investigate historical decision making regarding artificial-lift installations.

of new engineers makes knowledge expert-knowledge application and criteria. Additionally, the user-driven
transfer to those in charge of the reasoning analysis, and finally report- expert system can be revised easily and
field additionally difficult. Therefore, ing is implemented to run once a day. changes can be undertaken should the
the following main questions need to Results are presented to the produc- need arise (e.g., technical or economi-
be asked: tion engineer in the form of a list in cal limitations).
• How to capture the knowledge of which candidate wells are ranked by During the implementation of the
experienced personnel and make it priority of intervention according to user-driven expert system, two differ-
available to the whole organization? their critical condition. The engineer ent methods have been applied: (1) a
• How to apply this knowledge then is empowered to initiate the nec- rule-based approach that can be con-
in a general and unbiased manner essary engineering workflows (e.g., sidered deterministic because the out-
throughout the assets of the organiza- nodal analysis) more effectively and come is either “artificial lift=true” in
tion? make the final decision on how to case the input conditions are favorable
• How to keep this knowledge distribute resources. or “artificial lift=false” in case they
updated by use of new experience? are not and (2) a stochastic approach
The solution presented in the full- User-Driven Expert System implemented through a BBN in which
length paper is based on an expert The user-driven expert system is an the outcome is not just true or false
system and automated reasoning tools automated workflow that fully relies but a probability distribution denot-
that are able to capture and apply on the knowledge and experience of ing the “degree of applicability of an
knowledge and make it available to the engineer. The asset team has the artificial-lift system” that takes into
the whole organization. In general, ability to incorporate its complex busi- account the uncertainty of a certain
the combination of these two tech- ness reasoning into surveillance rules decision or choice.
nology-based approaches enables the to detect potential artificial-lift candi-
asset team to process the field data dates. The advantage of this system Rule-Based Approach. The rule-based
and apply the knowledge system- is the trustworthiness and, most sig- approach consists of implementing the
atically at the field level, obtaining nificantly, the traceability of decisions engineer’s reasoning into an automated
results faster and on time for all wells leading to the artificial-lift candidates routine comprising a sequence of if-
in the field and independently of detected by the system, because the then-else loops. In this way, complex
asset, well, or engineer. An automat- inference rules have been created in knowledge can be captured and used
ed workflow consisting of data cap- close cooperation with the asset engi- in real-time surveillance and detection
ture, data validation and processing, neers and fully reflect their provided workflows. These workflows can be

58 JPT • JULY 2010


SPE Bookstore
scheduled to run at any time and as It can be seen in Fig. 1 that there www.spe.org/store
often as desired, enabling immediate is a distinction between when to use
detection when the conditions of a well Method A (red) and when not to
indicate need of an artificial-lift system (green or blue). Artificial Lift Method
and enabling early notification to the A typically is applied for low- to very-
asset team and engineers. low-gas-rate wells with low pressures
but with a moderately high static-
Data-Driven Expert System. Data- liquid level in the well. However, it was
driven approaches are used widely considerably more difficult to clearly
to reveal hidden information in large identify the conditions under which NEW BOOK
data sets. Typically, in a brownfield Method C was applied in the well to
with a large number of wells and decrease the likelihood of liquid load-
a production history spanning many ing in the well. The operating condi-
years or decades, the amount of data tions that lead to the choice of Method
available overwhelms the convention- C are very much like those of Method
al approach to data screening (e.g., B (higher gas rates, higher pressures).
spreadsheets or production plots). A The data that were used to calibrate
data-driven approach allows the engi- the SOM then were clustered to find
neer to visualize data in different con- groups of similarly behaving wells.
texts to facilitate the analysis of cor- The wells per cluster are selected by
relations and the process of knowledge the clustering algorithm to minimize
discovery. The presented workflow the variation of the individual items
applied an unsupervised clustering (wells) in each cluster. Using the clus-
algorithm and a special visualization tering information, the problem space
technique using Kohonen’s self-orga- could be reduced from several hun-
nizing maps (SOMs) dred individual wells to five clusters Transient Well Testing
of similarly behaving wells. This gen-
SOMs. SOMs display multidimension- eralization helps the asset team (and Monograph Series, Vol. 23
al data sets in a 2D map. The SOM uses the surveillance engineer) to classify Medhat M. Kamal
the Euclidian distance between two a well immediately according to its
points in the multidimensional space production features (which are driving
as a measure of their similarity. The the cluster number) rather than having Transient well testing is one of the
more similar the measurement, the to investigate all available production most important diagnostic tools
closer the points are placed together in information of a well before knowing used by petroleum engineers to
the 2D projection—the “map.” how to classify it. characterize hydrocarbon assets
In this work, SOMs have been applied and predict future performance.
to identify a pattern in the historical Conclusions This new monograph has the latest
decision making of the asset team. After analyzing the implemented work- reference information on designing,
The objective is to discover the pattern flow and the results obtained, the fol- running, and analyzing different
under which a decision for a specific lowing conclusions were reached: types of transient tests in oil and
artificial-lift installation was typically 1. The main benefit from applying gas reservoirs.
made. Ideally, this pattern corresponds this methodology is the reduction
to the rule-based expert system as of lost production by the ability to
discussed earlier; however, sometimes install an artificial-lift system sooner. Contents: #BTJDDPODFQUTPGUSBOTJFOUUFTUJOHt
it does not as a result of exceptional 2. It is possible to capture the 8FMMUFTUJOHNFBTVSFNFOUTt7BMVFPGJOGPSNBUJPO
issues (e.g., no infrastructure available knowledge and field experience and t#BTJDJOUFSQSFUBUJPOPGIPNPHFOFPVTSFTFSWPJST
t8FMMCPSFFGGFDUTt0VUFSCPVOEBSJFTt
to support a certain installation). implement them in an automated $PNQVUFSBTTJTUFEJOUFSQSFUBUJPOt1SFTTVSF
Fig. 1 presents the depiction of workflow, which can be used in a USBOTJFOUBOBMZTJTTPGUXBSFt/BUVSBMMZGSBDUVSFE
several realizations of the same SOM. systematic way, acting as a fast field SFTFSWPJSTt8FMMUFTUBOBMZTJTPGIZESBVMJDBMMZ
Each realization corresponds to a screening tool. The asset team then GSBDUVSFEXFMMTt%FMJWFSBCJMJUZUFTUJOHBOE
parameter that has been used to cali- can focus its attention on the candi- VOEFSHSPVOEHBTTUPSBHFt4MBOUFEBOEQBSUJBMMZ
brate the SOM (gas rate before and dates identified, saving valuable time. QFOFUSBUJOHXFMMTt)PSJ[POUBMXFMMTt5FTUJOH
VOEFSNVMUJQIBTFnPXDPOEJUJPOTt*OUFSQSFUBUJPO
after artificial lift installation, line 3. As knowledge is captured—
PGTJNVMUBOFPVTMZNFBTVSFEUSBOTJFOUnPXSBUF
pressure before artificial-lift installa- including multiparameter criteria— BOEQSFTTVSFEBUBt1FSNBOFOUHBVHFTBOE
tion, wellhead pressures before and and applied systematically, the meth- QSPEVDUJPOBOBMZTJTt.VMUJMBZFSSFTFSWPJSTt
after artificial-lift installation, liquid odology can be applied in an unbi- *OKFDUJPOXFMMUFTUJOHt'PSNBUJPOUFTUJOHt
level in the tubing before installa- ased way to the whole field, allowing %SJMMTUFNUFTUJOH
tion). The output can be seen in the identification of new opportunities
top-right realization of the parameter for artificial lift systems that could
(i.e., “type numeric”, representing the not have been identified simply by
artificial-lift system that was chosen selecting one parameter as the rank-
for the particular well). ing criterion. JPT

Visit our online bookstore at


www.spe.org/store
JPT • JULY 2010

You might also like