IPTC-19423-MS

Artificial Lift Selection Using Machine Learning

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/19IPTC/2-19IPTC/D021S042R003/1127193/iptc-19423-ms.pdf/1 by Bandung Inst. of Tech. user on 17 July 2021

Thanawit Ounsakul, Thum Sirirattanachatchawan, Wiwat Pattarachupong, Yaovanart Yokrat, and Peerapong
Ekkawong, PTTEP

Copyright 2019, International Petroleum Technology Conference

This paper was prepared for presentation at the International Petroleum Technology Conference held in Beijing, China, 26 – 28 March 2019.

This paper was selected for presentation by an IPTC Programme Committee following review of information contained in an abstract submitted by the author(s).
Contents of the paper, as presented, have not been reviewed by the International Petroleum Technology Conference and are subject to correction by the author(s). The
material, as presented, does not necessarily reflect any position of the International Petroleum Technology Conference, its officers, or members. Papers presented at
IPTC are subject to publication review by Sponsor Society Committees of IPTC. Electronic reproduction, distribution, or storage of any part of this paper for commercial
purposes without the written consent of the International Petroleum Technology Conference is prohibited. Permission to reproduce in print is restricted to an abstract of
not more than 300 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgment of where and by whom the paper was presented.
Write Librarian, IPTC, P.O. Box 833836, Richardson, TX 75083-3836, U.S.A., fax +1-972-952-9435.

Abstract
The artificial lift selection process performed by human involves iterating of several design parameters.
Moreover, the human's curated selection required the decision making with unbiased, repeatable and
reliable. Capturing the lesson learned from the previous mistake into the new design and lack of look back
in the past performances are the limits of human. The supervised machine learning method can apply to
improve selection process. This approach can minimize the life-cycle cost of artificial lift wells by using
machine learning which incorporate the past performances and lesson learnt from installations. The data is
prepared into a structured dataset. The dataset is pre-processed to determine the "Good" and "Bad" wells
based on their life-cycle cost, then used for training and validating the classification models. The most simple
and accurate model is adopted for future artificial lift selection and current wells’ performance assessment.
Finally, the performance of new wells is continuously added for further model's training.
The artificial lift suggested by the machine learning expects reducing life-cycle cost in the ongoing trial
in the fields. In term of assessing tool, the selection model reveals some discrepancy in the current installed
artificial lift. This alerts the operator to look inside the potential problems. However, the subject matter
experts still need to give an adequate interaction in case of false alarm. Therefore, the discovered pattern
for good artificial lift selection will help improve the fields’ production. In addition, the endless learning
capability of machine learning allows the new data feeds into the existing dataset and further incorporates
the model in order to response to the dynamic change of the fields’ conditions. In conclusion, machine
learning process is more comprehensive comparing to the selection made by conventional process where
only few tables used for the artificial lift selection and overlook the value of data captured.
The Artificial Intelligence is one of the emerging technologies which provides the breakthrough results.
This paper presents the artificial intelligence trend in oil and gas industry. It is a promising tool which help
solving human's complex problems. Ultimately, adding the durable competitive advantage to the oil and
gas industry.
2 IPTC-19423-MS

Introduction
The artificial lift selection is important in order to achieve ultimate recovery and profitability of oil reservoir.
To select the best lifting method for a well, there are many considerations, for example, well construction,
reservoir parameters, infrastructure, etc. Table 1 illustrates the broad aspects of factors that be considered
during artificial lift selection. These parameters change over the time resulted in change in proper lifting
method from one to another. It is such a responsibility of well enginner to monitor it and redesign/ reselect
the optimum lifting method.

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/19IPTC/2-19IPTC/D021S042R003/1127193/iptc-19423-ms.pdf/1 by Bandung Inst. of Tech. user on 17 July 2021

The documentation for the success of oil production and past failure are also the key for the future
selection. It could adapted for dynamic conditions of mature field with unbiased, repeatable and reliable.
However, the current lifting selection still based on strong human's opinions. It tends to complete by their
most familiar lifting methods [1].

Table 1—Artificial Lift Design Considerations

Another pain point of any field where has the vast amount of wells and variety kind of artificial lift
method is surveillance. Regading dynamic changes, the right timing of artificial lift conversion could be
maximize the profitability. However, it is hard for human to capture and act in the proper time. The modern
surveillance system and "Internet of Thing" or IoT helps to ease the operators’ tasks on making a systematic
decision and automation. The application for IoT devices is on the rise which can be used in conjunction to
the "rule-based" [2] and could be "Artificial Intelligence-based" decision making in the near future.

Selection Using Machine Learning

Artificial Intelligence (A.I.) is an emerging technology that recently provided breakthrough results for big
data analysis. Machine Learning (ML) is a field of A.I. which its ability improves by learning from more
data over time [3]. Machine learning is a promising tool which help human's understanding of complex
problems, and, in this case, could furnish artificial lift selection. The A.L. selection model is constructed
from the decade of mature fields data and algorithms which can be written as:

or

The data is extracted from several databases, transformed and loaded into a data warehouse for further
analytics. The raw data comprises many attributes as listed in Table 1. This data is nowadays used
predominantly by the engineer to perform the A.L. selection. It will be more systematic approach if it
IPTC-19423-MS 3

transforms to a computer task likes "supervised machine learning". This learning process requires human
interaction to train the model. The human provides the training dataset which comprises attributes (i.e. input
- X) and labels (i.e. output - Y). Given the input and output, the algorithm needs to map or connect it together
[4]. To find the parameters that contribute the good production performance, it is essential that the training
dataset comes from well with low cost per barrel. Hence, the dataset will be manipulated, cleaned, arranged
to consider only low cost per barrel oil production.
The available dataset for training the model contains 30,000 samples which can be distinguished into

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/19IPTC/2-19IPTC/D021S042R003/1127193/iptc-19423-ms.pdf/1 by Bandung Inst. of Tech. user on 17 July 2021

natural flow and 4 different A.L. types i.e. beam pump (BP), gas lift (GL), electrical submersible pump
(ESP), and progressive cavity pump (PCP). Figure 1 shows the size of each lift method prior pre-processing.
It can be seen that the number of example in each A.L. is imbalance. The imbalance dataset could cause
the "Accuracy paradox" or the misleading of its accuracy [5]. So the upsampling technique is required for
ESP and PCP's example datasets to increase the number of example while the downsampling one applied to
the BP and GL's datasets. Ultimately, the datasets for training are balanced around 7,000 examples across
all classes (i.e. lift method).

Figure 1—Example wells for training the selection model

Since there are 50 attributes in a dataset, it is important to select the right amount of it. Too many attributes
for training resulted in time consuming and overfitting. The attribute forward selection (AFS) is selected
in this study due to the efficiency is over the backward method on a large number of attributes. The AFS
process starts with no attribute and adds in turn, one at a time until no significant improvement (i.e. no
information gain) [6]. By the result, the number of attributes reduces to 17. These attributes correlation
are shown in the matrix (Figure 2). It can be seen from this data exploration that some parameters have
a strong correlation to the specific lifting method (i.e. positive and negative correlation). An example, the
well with high sand production (i.e. SAND_PPTB) is best handled by PCP as shown in the plot with "Blue".
In contrast, the ESP will not a solution for the sandy wells, shown in the plot with "Orange". Although this
is a simple knowledge, engineer's opinion is in "Qualitative" rather than "Quantitative.
4 IPTC-19423-MS

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/19IPTC/2-19IPTC/D021S042R003/1127193/iptc-19423-ms.pdf/1 by Bandung Inst. of Tech. user on 17 July 2021

Figure 2—Correlation plot of attributes in training dataset and lift methods

Figure 3—Artificial Lift Selection model training, validation and deployment

The selected attributes from AFS are then used for training the A.L. selection model. Since the model is
the function of multivariate, datamining tools are required. The reconstruction of the selection model from
the dataset based on available algorithms for classification problem, for instant, decision tree, Naive Bayes,
neural network, etc. The good performance examples in dataset take the turn on training and testing the
models or calls "cross-validation technique". On each iteration, the dataset is divided into an equal size with
K subsets. The K-1 subsets are used for training and 1 subset for testing. The cross-validation process with
5 subsets is shown in Figure 4.
IPTC-19423-MS 5

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/19IPTC/2-19IPTC/D021S042R003/1127193/iptc-19423-ms.pdf/1 by Bandung Inst. of Tech. user on 17 July 2021

Figure 4—An example of cross validation process with 5 subsets

The selection of an algorithm involves 2 factors which are speed and accuracy. The best algorithm
could provide a fair speed with compromised accuracy. However, when an engineer introduces the machine
learning, it was perceived as a black box filling with mathematic equations. It should not be the case for
artificial lift selection which is explainable. At the end of the day, the engineer can verify the reasonableness
of the model against their technical knowledge.

Results and Discussions

The evaluation of machine learning algorithm is an essential process to ensure the quality of the model. There
are 3 algorithms used in order to compare. The aim is to guess of which lifting method was installed based
on other attributes. Since the training dataset comprises only good performance well, this reconstruction of
lifting method will deliver the proper A.L. Among 3 algorithms, Naive Bayes provides the shortest training
time with moderate accuracy and F1 score [7]. Neural network by far the slowest algorithm, yet provides
better performance than the Naive Bayes but poorer performance when comparing to the decision tree.
The decision tree model yields the satisfactory performance in both accuracy and F1 score with moderate
training time. Hence, the decision tree algorithm is adopted for A.L. selection model. Decision tree algorithm
provides not only a good prediction performance and training speed but also intuitive understanding. This
algorithm has a process similar to human's decision making. Therefore, the model will easily explainable
by the engineer.

Table 2—Comparison of classification performance from 3 algorithms

The decision tree model is tested with the new data for selecting the artificial lift for 9 wells. For
comparison purpose, the well life-cycle cost and production are normalized and cross-plotted in Figure 5.
6 IPTC-19423-MS

The black dots represent human selection. It can be seen from this visualization that there 2 obvious groups
of well's performance. One with the low cost per barrel and another with the high cost. The Artificial Lift
selected by machine learning was performed of 9 wells resulted in lowering the normalized cost as seen
in yellow dots.

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/19IPTC/2-19IPTC/D021S042R003/1127193/iptc-19423-ms.pdf/1 by Bandung Inst. of Tech. user on 17 July 2021

Figure 5—Cross-plot between Normalized Cost and Normalized Production

Despite the low life-cycle cost incurred during the trial period, there still more room for improvement
especially on the quantity and quality of attributes in the dataset. In general, the number of attributes now
tend to overfit the model. Also, some indirect attributes should be defined more proper. For example, the
distance related attribute like the distance from gas lift network is only approximate (i.e. Euclidean distance).
In term of well surveillance, the lift selection model can use in conjunction with the well test data. The
online data feeding to model provides the real-time advisory on possible well conversion. The advisory is
ready upon the completion of well testing. Again, this is a typical workflow which an engineer may interact
with the change of well's performance. The maximum profitability from the well relies on the engineer to
evaluate the test data and provides the recommendations in a timely manner to convert the well or not. Unlike
the conventional process, the model created by Machine Learning can help automatic screen, pinpoint the
candidate and recommend the proper A.L. for well conversion. Therefore, the A.L. conversion could convert
in the right timing.

Conclusions
• The decades of production from a mature oil field not only mark the higher high cumulative oil
production but also higher high value in the data. These datasets can use in many ways e.g. training
the machine.
• According to the Attribute Forward Selection, there are several parameters affect the artificial lift
selection. However, the top 5 parameters accounted for 80% weight to the A.L. selection.
IPTC-19423-MS 7

• Three classification algorithms used to generate the A.L. selection model. However, the decision
tree algorithm provides the best accuracy at the moderate training time.
• Based on 9 wells trial, the A.L. selection using machine learning tend to reduce the life-cycle cost
of the well. It could be a promising technique in the upcoming future.
• The machine learning process provides a more comprehensive approach than man's curated
selection process.

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/19IPTC/2-19IPTC/D021S042R003/1127193/iptc-19423-ms.pdf/1 by Bandung Inst. of Tech. user on 17 July 2021

References
1. J. Clegg, S. Bucaram and N. Hein, "Reccommendations and Comparisons for Selecting Artificial-
Lift Methods," Journal of Petroleum Technology, vol. December, pp. 1128–1167, 1993.
2. W. G. Elmer, "Artificial Lift Applications for the Internet of Things," in SPE Annual Technical
Conference and Exhibition, San Antonio, Texas, 2017.
3. M. COPELAND, "The official NVIDIA blog," 29 July 2016. [Online]. Available: https://
blogs.nvidia.com/blog/2016/07/29/whats-difference-artificial-intelligence-machine-learning-
deep-learning-ai/.
4. R. v. Loon, "bigdata-madesimple," 05 Feb 2018. [Online]. Available: http://bigdata-
madesimple.com/machine-learning-explained-understanding-supervised-unsupervised-and-
reinforcement-learning/. [Accessed 06 Aug 2018].
5. J. S. Akosa, "Predictive Accuracy: A Misleading Performance Measure for Highly Imbalanced
Data," in SAS Proceeding, Oklahoma, 2017.
6. P. Kalyani and M. Karnan, "Attribute Reduction using Forward Selection and Relative Reduct
Algorithm," International Journal of Computer Applications, vol. 11, no. No.3, pp. 8–12, 2010.
7. N. Lavesson, in Evaluation And Analysis Of Supervised Learning Algorithms And Classifiers,
SWEDEN, Department of Systems and Software Engineering, Blekinge Institute of Technology,
2006, p. 13.