University of North Dakota

UND Scholarly Commons

Petroleum Engineering Student Publications Department of Petroleum Engineering

2-13-2023

Evaluation of Liquid Loading in Gas Wells Using Machine Learning

Abderraouf Chemmakh
University of North Dakota, [email protected]

Olusegun Stanley Tomomewo

Kegang Ling
University of North Dakota, [email protected]

Ahmed Shammari

How does access to this work benefit you? Let us know!

Follow this and additional works at: https://commons.und.edu/pe-stu

Part of the Petroleum Engineering Commons

Recommended Citation
Abderraouf Chemmakh, Olusegun Stanley Tomomewo, Kegang Ling, et al.. "Evaluation of Liquid Loading
in Gas Wells Using Machine Learning" (2023). Petroleum Engineering Student Publications. 3.
https://commons.und.edu/pe-stu/3

This Article is brought to you for free and open access by the Department of Petroleum Engineering at UND
Scholarly Commons. It has been accepted for inclusion in Petroleum Engineering Student Publications by an
authorized administrator of UND Scholarly Commons. For more information, please contact
[email protected].
 Petroleum & Petrochemical Engineering Journal
ISSN: 2578-4846
MEDWIN PUBLISHERS
Committed to Create Value for Researchers

Evaluation of Liquid Loading in Gas Wells Using Machine

Learning

Chemmakh A1*, Tomomewo O1, Ling K1 and Shammari A2

Research Article
1
Department of Petroleum Engineering, University of North Dakota, USA
Volume 7 Issue 1
2
SLB, Norway
Received Date: January 16, 2023
Published Date: February 13, 2023
*Corresponding author: Abderraouf Chemmakh, Department of Petroleum Engineering,
DOI: 10.23880/ppej-16000333
University of North Dakota, Grand Forks, North Dakota, USA, Tel: +17012159164; Email:
[email protected]

Abstract
The inevitable result that gas wells witness during their life production is the liquid loading problem. The liquids that come
with gas block the production tubing if the gas velocity supplied by the reservoir pressure is not enough to carry them to
surface. Researchers used different theories to solve the problem naming, droplet fallback theory, liquid film reversal theory,
characteristic velocity, transient simulations, and others. While there is no definitive answer on what theory is the most valid
or the one that performs the best in all cases. This paper comes to involve a different approach, a combination between
physics-based modeling and statistical analysis of what is known as Machine Learning (ML). The authors used a refined ML
algorithm named XGBoost (extreme gradient boosting) to develop a novel full procedure on how to diagnose the well with
liquid loading issues and predict the critical gas velocity at which it starts to load if not loaded already. The novel procedure
includes a combination of a classification problem where a well will be evaluated based on some completion and fluid
properties (diameter, liquid density, gas density, liquid viscosity, gas viscosity, angle of inclination from horizontal (alpha),
superficial liquid velocity, and the interfacial tension) as a “Liquid Loaded” or “Unloaded”. The second practice is to determine
the critical gas velocity, and this is done by a regression method using the same inputs. Since the procedure is a data-driven
approach, a considerable amount of data (247 well and lab measurements) collected from literatures has been used. Convenient
ML technics have been applied from dividing the data to scaling, modeling and assessment. The results showed that a well-
constructed XGBoost model with an optimized hyperparameters is efficient in diagnosing the wells with the correct status and
in predicting the onset of liquid loading by estimating the critical gas velocity. The assessment of the model was done relatively
to existing correlations in literature. In the classification problem, the model showed a better performance with an F-1 score
of 0.947 (correctly classified 46 cases from 50 used for testing). In contrast, the next best model was the one by Barnea with
an F-1 score of 0.81 (correctly classified 37 from 50 cases). In the regression problem, the model showed an R2 of 0.959. In
contrast, the second best model was the one by Shekhar with an R2 of 0.84. The results shown here prove that the model and
the procedure developed give better results in diagnosing the well correctly if properly used by engineers.

Keywords: Gas Wells; Machine Learning; XGBoost

Introduction is when the co-produced liquids with gas (water and

condensate) can accumulate in the wellbore if the reservoir
The problem of liquid loading in gas wells has been energy depletes. The liquid loading happens in all types of gas
discussed widely in literature. A definition of the problem wells; vertical, inclined and horizontal which causes severe

Evaluation of Liquid Loading in Gas Wells Using Machine Learning Pet Petro Chem Eng J
2
Petroleum & Petrochemical Engineering Journal

slugging [1]. This phenomenon could be studied in many gas simulations to predict the onset of liquid loading by tracking
fields around the world conventional and unconventional the liquid holdup [33,34,35]. Nagoo, et al. [36] considered
[2] such as Marcellus gas field in the United states [3], that the flow in gas wells is extremely complicated and
Turkmenistan gas fields [4], China gas fields [5], the huge gas cannot be described by only one theory of either droplet
field of Hassi R’Mel in Algeria [6], Reggane and Ahnet fields fallback or the liquid film reversal, rather they developed a
in Algeria [7,8] and other fields around the world such as new analytical equation based on the axial buoyancy vector,
Qatar, Oman and Nigeria gas fields. In order to intervene in the convective inertial and the interfacial tension forces to
the wells and reduce the severity of the liquid loading, one of elaborate an easy-to-use equation reliable enough to give the
the deliquefaction technics that follow should be used: onset of liquid loading as they state in their paper. On the
• Mechanical intervention: velocity string, plunger lift, other hand, only few researchers tried to solve the problem
scheduled opens and shuts of the wells manually or by learning from previous lessons and analyzing the data
automated, and downhole pumps. from a statistical point of view. To the best of our knowledge,
• Chemical intervention: Foaming agents, or surfactants, Ansari, et al. [3] were the only ones who used a statistical
and gas lift. approach and involved Machine Learning technics to solve
the problem of the prediction of the onset of liquid loading.
These interventions proved to be efficient if used They used data from 160 gas wells in the Marcellus shale
properly at the right time of the well’s life. The term “right and developed an Artificial Neural Network model that was
time” refers to the time of the start of liquid loading or the able to outperform all the correlations compared with. The
onset of liquid loading. Building on this, knowing when the novelty of this paper is to develop a full procedure using
well will start to load is as critical as the intervention itself. Machine Learning (XGBoost specifically) on how to diagnose
For this reason, several researchers investigated theoretically a well that is suspect of liquid loading and predict the liquid
and in practice the mechanism that lead to liquid loading. loading state from a classification perspective, a direct value
The theories vary while describing the problem and how to prediction of the critical gas velocity and how to combine the
approach it. The early hypothesis was proposed by Turner, et knowledge of both to decide when to intervene in the wells
al. [9] of what is known later as the Droplet Fallback model. to minimize the severity of liquid loading. The following
This later suggests that liquids are transported as entrained sections will discuss in great details on a step-by-step basis
droplets inside the gas core. This conclusion was elaborated how to develop the model and how to use it.
after analyzing their field data with the two methods: droplet
fallback and liquid film reversal. Researchers afterwards Data Collection and Processing
worked on improving the model by suggesting modifications
and including new parameters based on lab experiments or In order to model the onset of liquid loading properly,
new field data naming Belfroid and Zhou [10-16]. Another meaningful data needed to be collected from the literature.
theory is the liquid film reversal model adopted by Barnea There are a good number of published articles on liquid
[17] who derived the analytical model starting from the loading, but only a few published data that can be used in
momentum balance equations. Researchers such as (Alsaadi such a modeling approach. Data were collected from field
[19]; Chen, et al. [20]; Fan, et al. [21]; Liu, et al. [22]; van ’t studies Luo, et al. [27]; Turner, et al. [9]; Veeken, et al. [14]
Westende, et al. [23]; C Vieira, et al. [24]; Waltrich, et al. [25]) and labs [19,37]. Choosing these sources is due to the fact
and other investigated the theory by doing lab experiments that the published data has all parameters which could
and observed the same thing, which is that the onset of liquid influence the onset of liquid loading, naming; (Superficial
loading coincides with the start of the film reversal. Other liquid velocity, liquid density, gas density, liquid viscosity,
researchers based their studies on this theory and tried to gas viscosity, interfacial tension, the inclination (degrees)
improve the model to work better in all cases [26-29] or of the well or the experimental setup, diameter, superficial
to be more realistc [30]. Other researchers used different gas velocity and the status of the well or the experimental
theories such as Lea, et al. [31] who suggested that the setup whether it is loaded or not). Since the data comes from
onset of liquid loading occurs at the minimum of the outflow different systems, it has to be brought to a united system, the
performance curve in the nodal analysis. Adesina, et al. [18]; SI units, and the figures below show the distribution of the
Guo, et al. [12] used the minimum kinetic energy criterion parameters.
to the 4-phase (gas, oil, water, and solid particles) mist-flow
model and elaborated a closed-form analytical equation to As can be seen in the plot, superficial gas velocity
predict the minimum gas flow rate. Gaol, et al. [32] did not covers a wide range of values which commonly found
consider the onset of liquid loading as a single event, and in field observations, and liquid superficial velocity also
developed a model based on the characteristic velocity as encompasses a wide range of values that represent wells that
a scaling variable to track the overall liquid content inside produce a minimal amount of water or condensate (liquid)
the wellbore. Some other researcher relied on transient flow or wells which have a considerable amount of liquid. The

Chemmakh A. Evaluation of Liquid Loading in Gas Wells Using Machine Learning. Pet Petro Copyright© Chemmakh A.
Chem Eng J 2023, 7(1): 000333.
3
Petroleum & Petrochemical Engineering Journal

wells and experimental setups which data are coming from where field engineers measure their fluid properties and rely
included mostly vertical cases, with some cases from nearly on to determine the status of the well whether is loaded or
vertical to nearly horizontal and with a small diameter or not. This explains the high values shown in the plot which
relatively large diameter up to 7” (0.1772 m). The most also refers to a high-pressure producing well. The interfacial
critical observation could be the status of the well, as we tension plot also indicates that both water and condensate
notice here most of the wells are loaded with a ratio of 3:1 are coproduced with gas.
between loaded and unloaded.
The total measurements obtained from the data
The fluids properties plotted in Figure 1 showing liquid gathered resulted in 246 measurements for both field and
density and viscosity reveal that both water and condensate lab experiments (196 samples from field studies and 50
are produced from the wells in which the data are collected, samples from lab experiments).
Gas density is measured at the wellhead conditions, the point

Figure 1: Data distribution frequency histograms.

Chemmakh A. Evaluation of Liquid Loading in Gas Wells Using Machine Learning. Pet Petro Copyright© Chemmakh A.
Chem Eng J 2023, 7(1): 000333.
4
Petroleum & Petrochemical Engineering Journal

Modelling Method predicted using K additive functions as follows:

k

Chen, et al. [20] at the University of Washington,
developed XGBoost for the first time and described it as a
= ( xi )
yi φ= ∑ f ( x ),
k =1
k i f k ∈

scalable Machine Learning system for tree boosting. The Here 𝔽 is the space of regression trees, and k is the
basis of XGBoost is set on decision trees as the classifier number of trees in the model. The solution of the model is
or regressor, including some modifications of the objective obtained by minimizing the regularization loss function and
function by adding some regularization to limit the over- by finding the best set of functions to embed in the model:
fitting. An important novelty in the XGBoost is the boosting
mechanism introduced, which not only boosts the decision
trees but also makes the computation faster than any other
 (φ )
= ∑l ( y , y ) + ∑Ω ( f )
i
i i
k
k

ML algorithm. What makes the XGBoost different than where,

Random-Forest and other tree-based algorithms is the
1
architecture that keeps building on previous trees and keeps Ω ( f ) =γ T + λω 2
minimizing the error until it becomes infinitely small (Figure 2
2). Here l represents the loss function, or the difference between
Chen, et al. [20] stated on their paper the mechanism on the predicted output  yi and the actual output yi . Ω
how the algorithm works by the following example: measures the complexity of the model and controls the over-
For a dataset of m features and n number of examples where fitting of the model. T is the number of leaves of the tree, w is
D= { ( xi , yi ) ( D =
n, xi ∈  m , yi ∈ ) , the output yi is the weight of each leaf.

Figure 2: XGBoost Model Schematic.

The term Boosting refers to the process of adding new where,

function f as the model keeps training while working on (
gi = ∂ yˆ (t −1) l yi , yˆ (t −1) )
minimizing the objective function in decision trees. Trees or
l ( y , yˆ ( ) )
t −1
functions are added as follows: hi = ∂ 2yˆ (t −1) i
n
(t ) ∑l ( y , yˆ ( )
+ f i ( xi ) + Ω ( f i )
t −1)
= i i Here gi , hi are first and second order gradient statistics
i =1 on the loss function.


( ) ( ) (
2 2 2 
1
split= 
∑ i∈I L
g i
+
∑ i∈I R
g i
−
∑ g 
i∈l i
+γ
) The complete model equations and details can be found
in the article by Chen, et al. [20].
2  ∑ i∈I hi + λ ∑ i∈I hi + λ ∑ i∈lhi + λ 
 
L R

Chemmakh A. Evaluation of Liquid Loading in Gas Wells Using Machine Learning. Pet Petro Copyright© Chemmakh A.
Chem Eng J 2023, 7(1): 000333.
5
Petroleum & Petrochemical Engineering Journal

As can be seen, the model has multiple parameters, often interpretable numeric value; for this purpose, the following
referred to as hyper parameters that construct a full set of coding was considered (Table 1):
equations to be solved to obtain a final prediction of the
critical gas velocity. The model equations are translated to Well status Status Code
a coded algorithm that trains on the data supplied and then
tested to get the best performance. It is worth mentioning Loaded 1
that all the coding related to this research was done in Python. Unloaded 0
This open-source programming language was used to model Table 1: Status encoding.
and tune the model’s hyperparameters to produce the final
predictive model. The process of tuning the parameters will After encoding the well’s status, the first practice is to
be discussed in the next section. choose the features to be asserted as inputs for the model.
Since the problem being solved is not a purely statistical
Modelling Approach problem, so it will not be treated as conventional statistical
problems where features are selected using technics like
The standard field practice is to calculate the critical gas
recursive features elimination or principal component
velocity at which liquid loading occurs, using correlations
analysis; rather, the features will be chosen based on previous
such as Barnea [17]; Luo, et al. [27]; Turner, et al. [9] or
observations from analytical models such as the liquid film
transient multiphase flow simulations. This process is done
model by Barnea or correlations such as Turner correlation.
before and after the onset of liquid loading and results in
The reason behind this comes from the engineering
classifying the well as ‘Loaded’ or ‘Unloaded’.
perspective of the problem, where each parameter presents
Since we are employing Machine Learning to model this
specific influence on the problem.
problem, the solution provided will contain two parts:
• The first part is a classification problem where a
Considering the analytical film model by Barnea
classification model of XGBoost will be developed and
(Appendix) the following parameters coming from the
compared with the correlations commonly used in
momentum balance equation are chosen to be inputs for the
literature to see which method better classifies the wells
model (diameter, liquid density, gas density, liquid viscosity,
to their correct state.
gas viscosity, angle of inclination from horizontal (alpha) and
• The second part is a regression problem where a
superficial liquid velocity). Considering the model of Turner
regression model of XGBoost will be developed to predict
(Appendix), the interfacial tension is added to the inputs list.
a value of a critical gas velocity at which liquid loading
occurs, then compare the results of the prediction with
Building an effective classification XGBoost model
the same correlations.
requires tuning the hyperparameters used inside the model
to generate a prediction. To tune the parameters, multiple
The reason behind this modeling approach is that a
values should be tested; for that, the Randomized Search
correlation or a model may be conservative and predict
technic from Scikit-Learn was used.
critical gas velocity values higher than what was observed
in the field or the lab leading to wrong interventions, which
RandomizedSearchCV implements a “fit” and a
may cause extra expenses that could have been avoided.
“score” method; it uses a random combination of the
In the reverse case, a model could be too optimistic and
hyperparameters from the supplied lists of the values to train
consistently predict values less than what was observed
and test the model by applying specified cross-validation
in the field, which may lead to more damages that could
over the data. Randomized Search chooses the combination
have been avoided. Contemplating this problem, this paper
of the hyperparameters, which gave a better performance on
proposes a two-step modeling approach to generate a cost-
the testing data.
effective tool that will not only classify the wells as loaded
The list of the parameters considered to be tuned is below:
or not but also provide a close approximation of what was
• Number of estimators: this parameter controls the
observed in the field at the onset of liquid loading.
number of decision trees used by the XGBoost. Values
list [20,50,100,200,300,500,1000].
Since the two steps fall under the Machine Learning
• Lambda: this is used to handle the regularization part of
modeling, data will be divided randomly with a ratio of 80%
XGBoost and avoid overfitting. Values list [0.01,0.1,1].
training to 20% testing.
• Max-depth: the maximum depth of a tree is used to
control overfitting as well. More trees depth means
Classification problem: This section aims to develop a
that the model is learning relations very specific to a
model that classifies the well to a correct status, whether
particular sample. Values list [3,4,5,7,8,9,10,12,14,20].
loaded or not. The first step is to encode the status to an

Chemmakh A. Evaluation of Liquid Loading in Gas Wells Using Machine Learning. Pet Petro Copyright© Chemmakh A.
Chem Eng J 2023, 7(1): 000333.
6
Petroleum & Petrochemical Engineering Journal

• Gamma: it specifies the minimum loss reduction required • If the critical rate predicted by the models is smaller than
to make a split. Values list [0,0.1,0.25,0.5,1]. the field rate and the actual status of the well is loaded,
• The learning rate: it defines the step size shrinkage used the case is labeled “False Unloaded FU”.
in update to prevent overfitting. Values list [0.01,0.05,0. • If the critical rate predicted by the models is bigger than
1,0.2,0.3,0.5,0.6,1]. the field rate and the actual status of the well is unloaded,
the case is labeled “False Loaded FL”.
The final optimized hyperparameters are presented below • If the critical rate predicted by the models is smaller
(Table 2): than the field rate and the actual status of the well is
unloaded, the case is labeled “True Unloaded TU”.
Parameter Value selected
The confusion matrix presented in Figure 3 displays the
Number-of-estimators 300 obtained results for the model’s predictions on the testing
Lambda 1 dataset (50 samples). The confusion matrix is an effective
Max-depth 20 tool to compare the field observations to the predicted ones.
Turner’s model, for example, could predict the correct status
Gamma 0.1 of the well only in 15 measurements (3 True Loaded well and
Learning rate 0.2 12 True Unloaded wells) and misclassified 35 measurements
Table 2: Hyperparameters selected for the XGBoost classifier. (34 False Unloaded and 1 False Loaded) which could be
interpreted as an underprediction of the critical gas velocity
Since the goal is to compare the performance of the model leading to very late alert by classifying a well as flowing while
to other correlations in the literature and looking at the fact it is already loaded or started loading. Similar observation
that the correlations generate a solid value of the critical gas for the model of Belfroid where it performed similarly to
velocity, the following procedure was followed to bring the Turner’s model. These two models represent the droplet
results to the same page, as was done by Chemmakh, et al. fallback theory models, and as can be seen here, the models
[30] to compare the predicted values by different correlations did not perform well as aligned with other research article
and the field observation rates as follows: findings naming Luo, et al. [27]; Shekhar, et al. [29]; Veeken,
• If the critical rate predicted by the models is bigger than et al. [14]; Cleide Vieira, et al. [38] that stated the Turner’s
the field rate and the actual status of the well is loaded, model or the droplet fallback models tend to underpredict the
the case is labeled “True Loaded TL”. critical gas velocity and giving late alerts to field engineers to
intervene in the wells.

Figure 3: Confusion matrix for model’s predictions.

Chemmakh A. Evaluation of Liquid Loading in Gas Wells Using Machine Learning. Pet Petro Copyright© Chemmakh A.
Chem Eng J 2023, 7(1): 000333.
7
Petroleum & Petrochemical Engineering Journal

The model of Barnea, on the other hand, performs to be very sure of the prediction. Also, it gives an insight
slightly better by correctly classifying 37 wells/lab into how many predicted loaded wells are in fact loaded.
measurements as True Loaded (28) or True Unloaded
TL
(9); the model of Shekhar also presented better results by • Recall: is used when predicting loaded wells is
correctly classifying 32 wells/lab measurements which is TL + FU
similar to the results of the liquid film model which takes the a priority as it gives the portion of correctly classified
entrainment into consideration, developed by Chemmakh, et wells among the loaded wells in the field.
al [30]. The three models representing the liquid film model
TL + TU
theory with various modifications, performed better than • Accuracy: is the ratio of the correct
the droplet entrainment but still underpredicting the results TL + FL + TU + FU
by misclassifying at least 9 wells/lab measurements as False predictions to the total number of wells. The accuracy
Unloaded which recommends late alert to intervene in the may not be the best metric to use when the data is
well to reduce liquid loading by considering it as flowing imbalanced (the number of loaded vs. unloaded wells is
normally while the liquid loading has already started. not even), which is the case for the current data set (37
loaded and 17 unloaded cases). An example would be if a
Contrary, the results obtained from the XGBoost model model predicts that all wells are loaded, then it would
were reasonably satisfying. The model correctly classified 46 have an accuracy of 0.74, which is not helpful in the field
wells/lab measurements from a total of 50 measurements. even though it seems high.
36 True loaded and 10 True Unloaded wells show how • To overcome the drawbacks of accuracy and to get
good the model predictions are by correctly predicting the a balance between precision and recall F1-score is
status of the wells from the supplied inputs. 3 False Loaded introduced, which takes the harmonic mean of precision
cases show that the model overpredicted the results only and recall.
3 times. Only 1 False Unloaded case means that the model Precision × Recall
F1score= 2 ×
underpredicted the status of the well only 1 time which Precision + Recall
reduces the cost of the model predictions by avoiding wrong
alerts and missing wells that are loaded and delaying the TL
necessary interventions, making the decisions recommended or F1score = TL + 0.5 ( FL + FU )
by XGBoost classifier right on time in most cases in the
( )
collected dataset (Table 3). Since the F1-score takes both precision and recall into
consideration, then any model with a low value of each of
To further compare the results, the following table them would have as well a low value of F1-score keeping the
presents the metrics often used for classification problems highest value of F1-score as a reference of which model is the
[39,40]: best.
TL
• Precision: which is used when the main goal is
TL + FL

False False True True

Names Precision Recall Accuracy F1_Score
Unloaded Loaded Unloaded Loaded
Turner 34 1 12 3 0.75 0.081081 0.3 0.146341
Belfroid 35 0 13 2 1 0.054054 0.3 0.102564
Barnea 9 4 9 28 0.875 0.756757 0.74 0.811594
Shekhar 16 2 11 21 0.913043 0.567568 0.64 0.7
Olimans 15 3 10 22 0.88 0.594595 0.64 0.709677
XGB-
1 3 36 10 0.923 0.972 0.92 0.947
Classifier
Table 3: Model’s performance metrics.

To assess the results, comparing each metric individually only 2 loaded wells from a total of 37. On the other hand,
may lead to erroneous results. For instance, looking at the the XGBoost model has a precision of 0.91 but correctly
precision alone would choose the model of Belfroid as the predicted 36 loaded wells from 37. Looking at the recall, the
best model having a precision of 1, but in fact it could predict XGBoost has the highest score of 0.972, outperforming the

Chemmakh A. Evaluation of Liquid Loading in Gas Wells Using Machine Learning. Pet Petro Copyright© Chemmakh A.
Chem Eng J 2023, 7(1): 000333.
8
Petroleum & Petrochemical Engineering Journal

other models clearly. As mentioned before, Recall is used selected based on the observation from previous analytical
when capturing the loaded wells is a priority. However, it models and correlations.
is still not very useful when used alone since a model that
predicts all wells are loaded would have a recall of 1 and still Since the model being developed is an XGBoost model,
not functional. The accuracy metric indicates that XGBoost a similar process of optimizing the hyperparameters
is the best model, but the result may not be very insightful is integrated into the model. After trying all different
since the dataset is not symmetric. When using the F1- combinations, the following parameters were selected and
score, the metric that takes both Precision and Recall into presented the highest performance (Table 4):
consideration, XGBoost model outperforms all the models
with a score of 0.947, followed by the Barnea’s model with Parameter Value selected
a score of 0.81 and the two models derived from the liquid
film reversal theory, then comes the models of Turner and Number-of-estimators 130
Belfroid presenting the droplet fallback theory which again Lambda 1
aligns with the results found in Luo, et al. [14]; Shekhar, et Max-depth 50
al. [29] with only one difference is that the Machine Learning
algorithm performed better than all models. Gamma 1
Learning rate 0.5
Regression Problem: Predicting the status of the well is Table 4: Hyperparameters selected for the XGBoost
crucial but predicting when the well will start loading is regressor.
equally vital, if not more. For this reason, a model of XGBoost
was developed to predict the critical gas velocity. The first The model with the best performance was chosen and
step in modeling a regression problem is similar to the compared with the different correlations.
classification problem, where the inputs need to be specified.
As was done in the previous section, the same inputs were

Figure 4: models’ predictions VS real field velocity measurements.

Figure 4 presents the critical gas velocity predicted by velocity both in field and lab experiments. The results are
the different models compared to the observed critical gas best when the predictions are on 45 degrees line with the

Chemmakh A. Evaluation of Liquid Loading in Gas Wells Using Machine Learning. Pet Petro Copyright© Chemmakh A.
Chem Eng J 2023, 7(1): 000333.
9
Petroleum & Petrochemical Engineering Journal

observed critical velocities. The graph showing the models entrainment in the Barnea’s liquid film model may not be the
of Turner and Belfroid confirms the previous section’s best practice.
observation that the models tend to underpredict the critical
gas velocity by placing the wells in the unloaded region Results and Discussion
while they are already loaded. The graph showing the results
of XGBoost model predictions demonstrates how close In this research, a Machine Learning method has been
the prediction could be compared to the observed results utilized to model the problem of the onset of liquid loading in
since most points are either on the 45 degrees line or very gas wells. The process was divided on two parts:
close to it. Further comparison of the results requires the • A classification problem: here the main goal was to predict
introduction of the metrics suitable for regression problems the status of the well weather ‘Loaded’ or ‘Unloaded’.
naming: In this section the discussion above highlights how the
Coefficient of determination R2: XGBoost could be used as a tool to better predict the
status of the wells learning from previous experience and
∑ (y
n
− ytrue ) 2 data. The results revealed that XGBoost outperformed all
i =1 predicted
R 2
= other models and correlations compared with.
∑ (y
n
i =1 predicted − yaverage ) 2 • A regression problem: here the goal was to predict when
the well will start to load by predicting the critical gas
Normalized Root Mean Squared Error NRMSE: velocity at which the liquids will start to accumulate in
the well. The XGBoost model performed better than all
other models here as well by predicting the results as
∑
n
( y predicted − ytrue ) 2 1 close as possible to the field observations with an R2
NRMSE = i =1

n ytrue ( max ) − ytrue ( min ) score of 0.96.

• The only shortcoming of this model is the fact that it is an
Normalized Mean Absolute Error MAE: implicit model that is not quite clear for engineers how
to use it. However, with the advancement in the coding
∑
n
i =1
y predicted − ytrue 1 skills among engineers, the huge amount of open access
NMAE =
n ytrue ( max ) − ytrue ( min ) resources, and the procedure descried in this paper will
give a full roadmap on how to use such an algorithm and
predict the onset of liquid loading in gas wells and plan
The following table sums all the results obtained (Table 5): early for the necessary interventions.

Names R2 NMAE NRMSE References

XGB 0.959851 1.203637 1.617146 1. Khetib Y, Rasouli V, Rabiei M, Chellal HAK, Abes A, et al.
Turner 0.603237 3.766203 5.083678 (2022) Modelling Slugging Induced Flow Instabilities
Belfroid 0.629135 3.744506 4.914962 and its Effect on Hydraulic Fractures Integrity in
Long Horizontal Wells. 56th U.S. Rock Mechanics/
Olimans 0.423974 4.056983 6.125383 Geomechanics Symposium, USA.
Barnea 0.733511 3.201529 4.166311
2. Ouadi H, Mishani S, Rasouli V (2023) Applications of
Shekhar 0.845328 2.433582 3.174083
Underbalanced Fishbone Drilling for Improved Recovery
Table 5: Metrics of model’s performance. and Reduced Carbon Footprint in Unconventional Plays.
Petrol Petrochem Eng J 7(1): 1-22.
Comparing R2, XGBoost has the highest score with 0.959,
followed by Barnea’s model. Comparing NMAE XGBoost has 3. Ansari A, Fathi E, Belyadi F, Takbiri-Borujeni A, Belyadi
the lowest value with 0.037, followed by Shekhar’s model H (2018) Data-Based Smart Model for Real Time Liquid
with 0.076. Comparing the RMSE, XGBoost has the lowest Loading Diagnostics in Marcellus Shale Via Machine
score with 0.050 confirming the outperformance of the Learning. SPE Canada Unconventional Resources
model over all models compared with. The results also prove Conference, Canada.
the previous findings that the liquid film models outperform
the droplet fallback models. The model of liquid film model 4. Zhang P, Cheng X, Liu R, Yang J, Zheng K (2015) Abnormal
including the entrainment fraction performed the worse Liquid Loading in Gas Wells of the Samandepe Gasfield
and this is also aligning with the conclusions of the work in Turkmenistan and Countermeasures. Natural Gas
by Chemmakh, et al. [30] that recommended including the Industry B 2(4): 341-346.

Chemmakh A. Evaluation of Liquid Loading in Gas Wells Using Machine Learning. Pet Petro Copyright© Chemmakh A.
Chem Eng J 2023, 7(1): 000333.
10
Petroleum & Petrochemical Engineering Journal

5. Ming R, He H (2017) A New Approach for Accurate Whole Range of Pipe Inclinations. Int J Multiph Flow
Prediction of Liquid Loading of Directional Gas Wells in 12(5): 733-744.
Transition Flow or Turbulent Flow. J Chem 4969765.
18. Adesina F, Damilola FO, Olugbenga F (2013) An Improved
6. Djezzar S, Boualam A (2021) Impact of Natural Fractures Predictive Tool for Liquid Loading in a Gas Well. SPE
on the Productivity of Lower Devonian Reservoirs Nigeria Annual International Conference and Exhibition,
in Reggane Basin, Algeria. 55th U.S. Rock Mechanics/ Nigeria.
Geomechanics Symposium, USA.
19. Alsaadi Y (2013) Liquid Loading in Highly deviated Gas
7. Djezzar S, Boualam A (2020a) Analysis and Modeling Wells. Master Thesis, The University of Tulsa, Oklahoma,
of Tight Oil Fractured Reservoir. SEG International USA.
Exposition and Annual Meeting.
20. Chen D, Yao Y, Fu G, Meng H, Xie S (2016) A New Model
8. Djezzar S, Boualam A (2020b) Fractures Characterization for Predicting Liquid Loading in Deviated Gas Wells. J
and Their Impact on the Productivity of Hamra Quartzite Nat Gas Sci Eng 34: 178-184.
Reservoir (Haoud-Berkaoui field, Oued-Mya basin,
Algeria). 54th U.S. Rock Mechanics/Geomechanics 21. Fan Y, Pereyra E, Sarica C (2018) Onset of Liquid-Film
Symposium, USA. Reversal in Upward-Inclined Pipes. SPE Journal 23(5):
1630-1647.
9. Turner RG, Hubbard MG, Dukler AE (1969) Analysis and
Prediction of Minimum Flow Rate for the Continuous 22. Liu Y, Luo C, Zhang L, Liu Z, Xie C, et al. (2018)
Removal of Liquids from Gas Wells. J Pet Technol 21(11): Experimental and modeling studies on the prediction of
1475-1482. liquid loading onset in gas wells. J Nat Gas Sci Eng 57:
349-358.
10. Belfroid SPC, Schiferli W, Alberts GJN, Veeken CAM,
Biezen E (2008) Prediction Onset and Dynamic 23. van’t Westende JMC, Kemp HK, Belt RJ, Portela LM,
Behaviour of Liquid Loading Gas Wells. SPE Annual Mudde RF, et al. (2007) On the Role of Droplets in
Technical Conference and Exhibition, USA. Cocurrent Annular and Churn-Annular Pipe Flow. Int J
Multiph Flow 33(6): 595-615.
11. Coleman SB, Clay HB, McCurdy DG, Norris HL (1991)
New Look at Predicting Gas-Well Load-Up. J Pet Technol 24. Vieira C, Kallager M, Vassmyr M, Forgia NLa, Yang Z (2018)
43(3): 329-333. Experimental Investigation of Two-Phase Flow Regime
in an Inclined Pipe. 11th North American Conference on
12. Guo B, Ghalambor A, Xu C (2006) A Systematic Approach Multiphase Production Technology, Canada.
to Predicting Liquid Loading in Gas Wells. SPE Production
& Operations 21(1): 81-88. 25. Waltrich PJ, Posada C, Martinez J, Falcone G, Barbosa JR
(2015) Experimental Investigation on the Prediction
13. Nosseir MA, Darwich TA, Sayyouh MH, El Sallaly M (2000) of Liquid Loading Initiation in Gas Wells Using a Long
A New Approach for Accurate Prediction of Loading Vertical Tube. J Nat Gas Sci Eng 26: 1515-1529.
in Gas Wells Under Different Flowing Conditions. SPE
Production & Operations 15(4): 241-246. 26. Pagou AL, Wu X, Zhu Z, Peng L (2020) Liquid Film Model
for Prediction and Identification of Liquid Loading
14. Veeken K, Hu B, Group SPT, Schiferli W (2010) Gas-Well in Vertical and Inclined Gas Wells. J Pet Sci Engin 188:
Liquid-Loading-Field-Data Analysis and Multiphase- 106896.
Flow Modeling. SPE Production & Operations 25(3):
275-284. 27. Luo S, Kelkar M, Pereyra E, Sarica C (2014) A New
Comprehensive Model for Predicting Liquid Loading in
15. Wang Z, Guo L, Zhu S, Nydal OJ (2018) Prediction of the Gas Wells. SPE Production and Operations 29(4): 337-
Critical Gas Velocity of Liquid Unloading in a Horizontal 349.
Gas Well. SPE Journal 23(2): 328-345.
28. Mamudu EEO, Cynthia IB, Samuel ES, Muhammad GBA,
16. Zhou D, Yuan H (2010) A New Model for Predicting Gas- Angela OM (2019) Application of Non-Uniform Film
Well Liquid Loading. SPE Production and Operations Thickness Concept in Predicting Deviated Gas Wells
25(2): 172-181. Liquid Loading. MethodsX 6: 2443-2454.

17. Barnea D (1986a) Transition from Annular Flow and 29. Shekhar S, Kelkar M, Hearn WJ, Hain LL (2017) Improved
from Dispersed Bubble Flow-unified Models for the Prediction of Liquid Loading in Gas Wells. SPE Production

Chemmakh A. Evaluation of Liquid Loading in Gas Wells Using Machine Learning. Pet Petro Copyright© Chemmakh A.
Chem Eng J 2023, 7(1): 000333.
11
Petroleum & Petrochemical Engineering Journal

and Operations 32(4): 539-550. Horizontal Well Liquid Loading Prediction: Theory and
Extensive Field Validation. SPE Artificial Lift Conference
30. Abderraouf C, Ling K, Olusegun T, Ahmad S (2023) and Exhibition – Americas, USA.
Evaluation of the Onset of Liquid Loading by a Proper
Inclusion of Droplet Entrainment. J Energy Resour 37. Vieira C (2020) Modelling and Experimental Study on
Technol. the Production of Gas Wells with Associated Liquid.
NTNU, Norway.
31. Lea J, Nickens H, Wells M (2003) Other Methods to Attack
Liquid-Loading Problems. In: Lea J, et al. (Eds.), Gas Well 38. Vieira C, Stanko M (2020) Effect of Droplet Entrainment
Deliquification, Solution to Gas Well Liquid Loading in Liquid Loading Prediction. BHR 19th International
Problems. Gulf Professional Publishing, pp: 271-282. Conference on Multiphase Production Technology,
France.
32. Gaol AH, Valkó PP (2016) Modeling Wellbore Liquid-
Content in Liquid Loading Gas Wells Using the Concept 39. Laoufi H, Megherbi Z, Zeraibi N, Merzoug A, Ladmia A
of Characteristic Velocity. SPE Low Perm Symposium, (2022) Selection of Sand Control Completion Techniques
USA. Using Machine Learning. International Geomechanics
Symposium, UAE.
33. Hu B, Veeken K, Yusuf R (2010) Use of Wellbore-
Reservoir Coupled Dynamic Simulation to Evaluate 40. Mouedden N, Laalam A, Rabiei M, Merzoug A, Ouadi H, et
the Cycling Capability of Liquid-Loaded Gas Wells. SPE al. (2022) A Screening Methodology Using Fuzzy Logic to
Annual Technical Conference and Exhibition, Italy. Improve the Well Stimulation Candidate Selection. 56th
U.S. Rock Mechanics/Geomechanics Symposium, USA.
34. Yusuf R, Veeken K, Hu B (2010) Investigation of Gas Well
Liquid Loading with a Transient Multiphase Flow Model. 41. Wallis GB (1969) One-Dimensional Two-Phase Flow.
SPE Oil and Gas India Conference and Exhibition, India. McGraw-Hill.

35. Zhang H, Falcone G, Valko P, Teodoriu C, Texas A (2009) 42. Barnea D (1986b) Transition from Annular Flow and
Numerical Modeling of Fully-Transient Flow in the Near- from Dispersed Bubble Flow-unified Models for the
Wellbore Region During Liquid Loading in Gas Wells. Whole Range of Pipe Inclinations. Int J Multiph Flow
Latin American and Caribbean Petroleum Engineering 12(5): 733-744.
Conference, Colombia.
43. Taitel Y, Dukler A (1976) A Model for Predicting Flow
36. Nagoo AS, Kulkarni PM, Arnold C, Dunham M, Sosa J, Regime Transitions in Horizontal and Near Horizontal
et al. (2018) A Simple Critical Gas Velocity Equation Gas-Liquid Flow. AIChE Journal 22(1): 47-55.
as Direct Functions of Diameter and Inclination for

Chemmakh A. Evaluation of Liquid Loading in Gas Wells Using Machine Learning. Pet Petro Copyright© Chemmakh A.
Chem Eng J 2023, 7(1): 000333.