Professional Documents
Culture Documents
Predicción de La Rop Con Ia - Ingles
Predicción de La Rop Con Ia - Ingles
d
College of Engineering, University of Al-Qadisiyah, Al-Qadisiyah 58002, Iraq.
te
Email: salih.rushdi@qu.edu.iq
Tel: + 964 7901230881
di
Mortadha Alsaba
Australian College of Kuwait, Safat 13015, Kuwait.
e
Email: m.alsaba@ack.edu.kw
py
Tel: + 965 98059212
Mohammed F. Al Dushaishi
Co
School of Engineering, Texas A&M International University, Laredo, Texas 78041, United States.
Email: mohammed.aldushaishi@tamiu.edu
Tel: +1 5733561764
ot
ABSTRACT
Predicting the rate of penetration (ROP) is a significant factor in drilling optimization and minimizing expensive drilling
tN
costs. However, due to the geological uncertainty and many uncontrolled operational parameters influencing the ROP, its
prediction is still a complex problem for the oil and gas industries. In the presented study, a reliable computational approach
for prediction of ROP is proposed. Firstly, fscaret package in R environment was implemented to find out the importance and
rip
ranking of the inputs parameters. According to the feature ranking process, out of the 25 variables studied, 19 variables had
the highest impact on ROP based on their ranges within this dataset. Secondly, a new model that is able to predict the ROP
using real field data, which is based on artificial neural networks (ANNs), was developed. In order to gain a deeper
sc
understanding of the relationships between input parameters and ROP, this model was used to check the effect of the weight
on bit (WOB), revolutions per minute (RPM), and flow rate (FR). Finally, the simulation results of three deviated wells
showed an acceptable representation of the physical process, with reasonable predicted ROP values. The main contribution of
nu
this research as compared to previous studies is that it investigates the influence of well trajectory’s (azimuth and inclination)
and mechanical earth modeling parameters on the ROP for high-angled wells. The major advantage of the present study is
Ma
optimizing the drilling parameters, predicting the proper penetration rate, estimating the drilling time of the deviated wells
and eventually reducing the drilling cost for future wells.
Keywords: drilling; rate of penetration; high-angled wells; feature ranking; artificial neural networks
ed
1. Introduction
pt
In the past 10 years, the petroleum industry has faced significantly higher operating costs associated with drilling. These
expensive operational costs prompted a renewed emphasis on optimizing drilling performance [1]. The rate of penetration
ce
(ROP) is one of the main factors that can be varied to improve drilling performance [2]. Penetration rate or drilling speed is
defined as the depth of penetration achieved per unit of time [3]. In most cases, considerable drilling cost reduction is
Ac
accomplished by maximizing the ROP. This is chiefly accomplished by finding the optimum drilling parameters [4].
However, applying very fast ROP results in improper hole cleaning. In fact, one of the predictable problems may occur if the
drilling cuttings are not removed from the bottom hole properly [5]. Therefore, it is essential to optimize a relationship
between the ROP and drilling parameters to obtain superior drilling performance with less cost, while maintaining safety.
The ROP depends on so many variables that one cannot be changed without affecting the others [6]. Generally, these
variables can be classified into three categories: rig- and bit-related parameters, one-dimension mechanical earth model (1D
MEM) parameters, and drilling fluid properties [7]. These categories are divided into two major subgroups: controllable
parameters and uncontrollable or environmental parameters. The controllable variables, known as the operational parameters,
refer to those parameters that can be modified quickly to adjust the ROP without affecting the operations economics, such as
the weight on bit (WOB), revolutions per minute (RPM), flow rate (FR), total flow area (TFA), and bit type (IADC).
However, it is difficult to change the uncontrollable parameters because of economical or geological reasons, such as mud
d
(ML) seems to be an attractive alternative to model this complicated physical process [16]. A key characteristic regarding
te
ML is its ability to recognize complex patterns with good predictive accuracy through a typical learning process [17]. Thus,
ML can provide intuitive solutions to complicated problems with no need for the formal description of the underlying physics
di
[18, 19]. In addition, the ROP predictions by these methods are not bound to bit specifications, borehole conditions, or
bottom hole assembly (BHA). Among the ML methods, artificial neural networks (ANNs) can most powerfully and
e
efficiently produce accurate ROP predictions [20-22]. Usually, ANNs are utilized to identify complex relationships when
sufficient data exist. The increased availability of drilling parameter measuring tools has resulted in datasets with many
py
recorded variables related to the drilling process. A number of studies have succeeded in modeling many of the oil and gas
industry processes using ANNs [23-26]. In general, their conclusions find ANNs to be a good modeling method due to their
Co
learning, generalization, and nonlinear behavior properties.
This research seeks to assess the potential of using ANNs to predict the drilling performance in terms of the ROP using
datasets from previously drilled wells.
ot
2. Methodology tN
2.1. Data acquisition
Database preparation is a very challenging and time-consuming process. The ANNs method can learn the trend or
behavior of previous experiences with certain experiments or data-set. Thus, the quality and authenticity of the datasets
rip
(input) play a significant role in the reliability of the prediction process and the decisions made by the ANNs technique. For
these reasons, the learning datasets should be large enough and also of high enough quality to cause the model to refine itself.
Otherwise, lots of noise might be created, which negatively affects the model [27].
sc
Due to drilling requirements and the similarity of wells located close to one another, collecting past data and utilizing the
data in a useful manner have an important impact on drilling cost reduction. As a matter of fact, there are various factors that
can affect ROP. Previous studies showed that the ROP is largely dependent on the rig- and bit-related parameters such as
Ma
measured depth (MD), rotation per minute (RPM), torque (TQ), weight on bit (WOB), flow rate (FR), circulating pressure
(CP), bit type (IADC), bit working-hours (BW), total flow area (TFA), and wellbore trajectory (azimuth [AZI] and
inclination [INC]).
In this paper, a total of 13,125 cases (datasets) were collected from MEM data, geophysical logs, daily drilling reports
ed
(DDRs), daily mud reports (DMRs), final well reports (FWRs), and mud logs (master logs) of 14 deviated wells drilled in the
Zubair Formation located in Southern Iraq. All of the studied wells were highly deviated (i.e., inclinations higher than 45°),
where water-based mud and conventional bottom hole assembly (BHA) were used to drill these wells. Moreover, the size of
pt
the hole section was 8 1/2”, which implies that this was a production section. Thus, the mud type and hole diameter were not
considered as input parameters. It is important to note that in this study, a drag bit with a polycrystalline diamond compact
ce
d
were complete and of good quality. Moreover, the necessary data, including gamma-ray logs, density logs, porosity logs,
te
sonic logs (compression and shear wave velocities), resistivity logs, formation micro-imager (FMI) logs, caliper logs, mud
logs (master logs), mini-frac tests, and laboratory measurements, were available for most of the selected wells. Subsequently,
di
the data were used to build a basic 1D MEM [32].
e
2.1.3.1. Rock mechanical parameters
py
Rock mechanical properties consist mainly of elastic parameters and strength parameters. These properties are considered
primarily to determine rock strength and formation drillability. The continuous profile of rock mechanical properties provides
a good indication of the natural variation in the formation hardness in different layers within the interval of interest.
Co
Laboratory tests are considered the most direct and reliable way of determining rock mechanical properties. Therefore, both
triaxial and triaxial-multistage testing techniques were carried out on retrieved core samples from the Zubair Formation to
measure the rock mechanical properties. Valid local empirical correlations were established between laboratory-derived rock
mechanical properties and geophysical well logs, such as porosity (∅) and compressional acoustic wave velocity (vp). Further
ot
details of estimating the rock mechanical properties of Zubair sandstone and shale formations using wireline measurements
were comprehensively discussed in the studies conducted by Abbas et al. [33, 34].
tN
Rock strength parameters such as unconfined compressive strength (UCS), and internal friction angle (φ) indicate to the
ability of the rock formation to withstand the in-situ stress environment around the wellbore. The UCS and φ were
determined using consolidated drained (CD) multistage triaxial tests for sandstone and consolidated undrained (CU) triaxial
rip
tests for shale. In this study, the UCS and φ were determined using Eqs. 1, and 2 for the sandstone formation as a function of
porosity (∅), while Eqs. 3 and 4 were used to estimate these parameters for the shale formation as a function of the
compressional wave velocity (vp) [33, 34].
sc
ϕ shale = 17.134e
0.239 v p
Ma
(4)
The elastic parameters (such as Young's modulus (E) and Poisson's ratio (v)) demonstrate the deformation behavior for
isotropic elastic materials. The static values of these parameters were calculated using porosity (∅) for the sandstone
formation and compressional acoustic wave velocity (vp) for the shale formation, as follows [33, 34]:
ed
Eshale = 0.2966e
0.6984 v p
(7)
ce
−0.353v p
vshale = 0.7621e (8)
The mechanical responses, as well as the properties of both the sandstone and shale formations, were significantly
Ac
different. Therefore, by classifying rocks according to their mechanical stratigraphy, it became possible to use different
correlations for different formations to best estimate their rock mechanical properties. The differentiation of non-shale from
shale was realized by applying a threshold of 75 gAPI to the gamma-ray logs in the studied wells [35].
d
te
2.1.3.3. Vertical stress
The overburden weight per unit area is called overburden stress. The overburden or vertical stress is the most basic input
di
parameter in the 1D MEM analysis. In areas with low tectonic activity, vertical stress represents the weight of overlying
formations [38]. The vertical stress in the Zubair Formation was calculated by integrating the bulk density log over the
e
vertical depth, using Eq. 10.
py
z
σ v = ∫ ρ ( z ) gdz (10)
0
Co
where g represents the acceleration constant due to gravity (m/s2), z is the vertical depth (m), and ρ is the rock bulk density
(g/cm3).
ot
The determination of minimum and maximum horizontal stress magnitude involves technical challenges in
geomechanical modeling. Since 1930 s, many researchers, geologists, geophysicists and engineers attempted to determine the
tN
magnitude of horizontal in-situ stresses. Hence, various laboratory and field methods are introduced for the estimation of the
horizontal in-situ stress magnitude as hydraulic, relief, jacking, strain recovery, and focal mechanism [39]. The poro-elastic
relations are the most commonly used method for estimation of horizontal in-situ stresses along the well length [40].
rip
Assuming flat-layered, poro-elastici deformation in the formation rock, a pair of particular constant strains, ɛy and ɛx, is
applied to the formation in the direction of the maximum and minimum stress, respectively. For a fluid-saturated porous
material that is assumed to be linear, elastic, and isotropic, considering anisotropic tectonic strain, the horizontal stresses
sc
(minimum and maximum) are expressed in Eqs. 11 and 12, respectively [41]. This approach estimates the magnitudes of the
horizontal principal stresses along the well length using the static Young's modulus, static Poisson's ratio, regional pore
nu
ν 1 − 2ν E νE
σH = σν + α Pp + ε + εx (12)
1 −ν 1 −ν 1 −ν 2 y
1 −ν 2
ed
In the above equations, α is Biot’s coefficient, which is maintained at unity to account for the brittle failure of rocks
(conventionally α = 1), E is the static Young’s modulus (GPa), ɛx is the strain in the minimum horizontal stress direction, and
pt
ɛy is the strain in the maximum horizontal stress direction. The two horizontal strains (ɛy and ɛx) can be measured by Eqs. 13
and 14, respectively [42].
ce
σ vv
v2
εy
= 1 − (13)
E 1− v
Ac
σ vv 1
=εx − 1 (14)
E 1− v
The magnitude of the minimum horizontal stress obtained from the above formulae can be calibrated against the direct
measurement procedures based on the field-tests (i.e. extended leak off test (ELOT), a standard leak off test (LOT), or a mini-
frac test) [43].
d
method are to simplifying training and perform variety of predictive modeling technics. The main advantage of this extension
te
is that it requires minimum user involvement. Also, the variety of the methods used in combination with the scaling
according to RMSE or MSE obtained from models profit the user. The idea is based on the assumption that the variety of
di
models will balance the roughness of calculations (default model settings were applied). This method is a simple package
providing fast and automated feature ranking based on the caret package [47]. It uses a large number of 103 various models
e
in order to determine the importance of features and rank them. In the end, the importance of variables is averaged through all
models to give reliable results and is scaled from 0 to 100.
py
In this work, the following parameters were collected as input: drilling operation parameters, formation type and lithology
of the rock, and fluid mud characteristics. All these input parameters are considered the independent variables to estimate the
Co
output i.e. ROP. Out of the 25 variables studied, six constant or nearly constant variables had to be excluded, and the
remaining 19 variables were considered to be the input parameters to predict the ROP, as shown in Table 1. The ranges for
these variables are very significant factors, and if new input data that fall outside of these ranges were applied to machine
learning models, the results of the ROP predictions would be not reliable [48].
ot
Table 1—Ranges of input parameters
tN
Parameter Minimum Maximum
Measure drilling depth (MD) (m) 2213 3834
Weight on bit (ton) 3.5 20
rip
ANNs are information processing systems that are a rough approximation and simplified simulation of a biological
ce
learning process and have performance characteristics similar to those of biological neural networks. These are adaptive,
parallel information processing systems, which are able to develop associations, transformations or mappings between
objects or data and have proven to have potential in solving problems that require pattern recognition [49]. The processing
Ac
elements of ANNs are artificial neurons. These neurons consist of four basic components that include input data, connection
strengths (weights), a transfer function (activation function), and output values. The inputs to each neuron are multiplied by
the adjusted weight elements (w1j ... wnj), and these modified inputs are summed up with a value called “bias.” Then, the
weighted and summed output is fed through a nonlinear function (i.e., a transfer function) to produce the results. Usually, the
neurons are interconnected in a multilayer network topology that consists of three layers: (1) an input layer, (2) one or more
hidden layers, and (3) an output layer (Fig. 1). The hidden layer(s) are the coefficients that provide the relationship between
the input and output layers.
d
te
The most common types of ANNs are feed-forward networks, which are the most efficient ones [50]. In such ANNs
architecture, the information will propagate in one direction from input neurons through the transfer function of the hidden
di
neurons to the outputs. For supervised learning, a back-propagation algorithm with a Levenberg-Marquardt training function
is probably the best-known learning algorithm for neural networks [51]. Back-propagation refers to the mechanism adjusting
and modifying the connection weights and biases of the network. In the beginning, the training algorithm is created by
e
assigning random weights and biases. In the first stage of this cycle, the input data are transformed through each layer until
py
an output is calculated. An error signal is computed for each output unit by comparing the generated output with the desired
output (target). These error signals are propagated back from the output layer to each neuron in the hidden layer, which, in
general, contributes directly to the output. Based on the relative contribution of each unit in the hidden layer to create the
Co
original output, these units will receive only a portion of the total error signal. The neurons use transfer functions to generate
their output from the net input. The most commonly used transfer functions for back-propagation are PURELIN, TANSIG,
and LOGSIG. The number of neurons in the input layer corresponds to the number of parameters being fed to the network as
ot
input. The same is true for the target layer. Therefore, the model was built using ANNs with 19 neurons in the input layer and
only one neuron in the output layer. The optimum number of neurons and layers was selected based on trial and error. On the
tN
other hand, many hidden layers of various numbers of neurons have been tested with the TANSIG and LOGSIG transfer
functions, while the output layer was one neuron, using the PURELIN transfer function.
In the modeling process, the database is randomly divided into two parts: a training dataset is applied to develop and
adjust the weights in a network, and a testing dataset is applied to examine the final performance of the ANNs. Among
rip
13,125 datasets of acquired data, which contained 19 input parameters and one output parameter (ROP), the ratio of 4:1was
applied for training and testing the developed ANNs model, respectively.
sc
to speed up the learning process and ensures that every parameter receives equal attention by the network and improving the
overall network performance. In addition, the ANNs model cannot be trained with textual or symbolic data. In this case, the
bit type must be translated into a numeric form. There are many techniques to translate textual or nonnumeric information
Ma
into numeric data. Some of the common textual translation methods consist of numbering classes, unary encoding, and binary
encoding. Numbering classes are used in this study to translate the bit type into a numeric form. The three-digit numbers of
the International Association of Drilling Contractors (IADC) were set as an equivalent for each drilling bit type. Furthermore,
these input parameters fall within various ranges; for example, the measured depths are described in thousands of meters,
ed
while borehole inclinations are described in single-digit values or decimals. Therefore, the next step is to preprocess the
inputs and target data, such as normalization, into a form that can be used with the ANNs.
pt
The data are normalized depending on either the applied-transfer function in designing the ANNs [52]. A transfer
function used in training can be either unipolar, with values between 0 and 1, or bipolar, with values between −1 and +1.
Regarding the FFBP with TANSIG hidden layers, the normalization can be applied to fall input and target data in the range
ce
of −1 to +1 (Eq. 15). On the other hand, when LOGSIG transfer function is used, the input and target data are scaled in the
range of 0 to 1 (Eq. 16).
Ac
2( X − X Min )
=Xi −1 (15)
( X Max − X min )
( X i − X min )
Xi = (16)
( X max − X min )
where Xi is the normalized value, Xmin is the minimum of original values, Xmax is the maximum of original values, and X is the
original value.
6
1 n
=
RMSE ∑
n i =1
( f ( xi ) − yi ) 2 (17)
d
te
Absolute error expresses the difference (error) between each of the predicted output and its corresponding actual value.
Relative error is calculated as the ratio of the mean of the absolute error and mean value of the predicted output. Percent
di
relative error is the relative error measured in percentage. Hence, the formula is as expressed in Eq. 18.
e
1 n xi − yi
= ∑ ] ∗100
py
AAPE [ (18)
n i =1 xi
Co
𝑅𝑅2 measures the statistical correlation between the predicted and actual values. The value ranges between 0 and 1. The value
of 0 indicates no similarity and a bad performance while 1 represents a perfect correlation between the model output and
actual value to be predicted indicates perfect generalization accuracy. The formula for correlation coefficient depicted in (19).
ot
n
∑ ( f (x ) − y )
i i
2
R2 = 1 − i =1
tN
n n
(19)
i
=i 1 =i 1
∑ f (x ) − ∑ ( y )
2
i
2
/n
rip
In the above equations, yi is the actual data; f(xi) represents the predicted data; xi and xn are the input parameters; and n is
the total number of records. The best-developed model will have a lower RMSE and AAPE but a higher R2.
sc
The aforementioned empirical correlations were used to derive the rock mechanical properties from the neutron and sonic
logs. The suggested mechanical stratigraphy was used to apply correlations of shale to the shale formation and relationships
Ma
of sandstone to the sandstone formation. Then, each derived mechanical properties was combined to build a single log from
the top to the bottom of the Zubair Formation. The laboratory-measured rock mechanical properties data were added to the
plot to assess the level of agreement between the experimentally measured mechanical properties and the derived mechanical
properties logs [54, 55]. Figure 2 presents examples of the rock mechanical property logs for one of the cored wells in the
ed
Zubair Formation. As can be seen, there is a high degree of positive correlation between the calculated rock mechanical
property logs and those derived from the laboratory.
pt
ce
Ac
d
te
e di
py
Co
Fig. 2– Predicted rock mechanical properties logs and laboratory measurements.
In the current study, the pore pressure of the depleted sandstone formations was computed using constant pressure
gradients that were determined from repeated formation test (RFT) measurements. Due to the discontinuity of the RFT data
ot
for the shale formations, the pore pressure was estimated using the modified Eaton (1969) technique of Eq. 3. The resultant
formation pore pressure profile was calibrated against the actual pore pressure measurements.
tN
The vertical stress magnitude on the Zubair Formation was calculated using the density log, according to Eq. 10. For
shallow zones, where the density log was not recorded because it was not of geologic interest, a compaction relationship
based on the regional density data was used to create a synthetic density curve to fill the gaps where the direct measurement
rip
was unavailable. Moreover, the continuous estimation of the minimum and maximum horizontal stress magnitudes along the
well length was obtained from poro-elastic formulations using Eqs. 11 and 12, respectively. The minimum horizontal stress
was calibrated against the direct measurements of closure pressures that were available from the mini-frac test. The predicted
sc
minimum horizontal stress from the poro-elastic formulation showed reasonably good agreement with the closure pressures.
Fig. 3 illustrates the estimate of the vertical stress, maximum horizontal stress, minimum horizontal stress, and pore pressure
magnitudes.
nu
Ma
ed
pt
ce
Ac
Fig. 3– Estimation of the pore pressure and in situ principal stress magnitudes at a single well location.
8
d
σh may have a negligible impact on the ROP for other formation types.
te
e di
py
Co
ot
tN
Fig. 4– Ranking of variables.
3.3. Network architecture design
rip
The TANSIG and LOGSIG transfer functions were examined for one, two, and three hidden layers. In addition, the
optimal number of neurons in hidden layers was determined by increasing the numbers incrementally. For the TANSIG
transfer function, the RMSE reached its lowest value at 1.37 with 30 nodes and three hidden layers, while the RMSE attained
sc
its lowest point at 2.83 with 32 nodes and two hidden layers for the LOGSIG transfer function. Comparing the RMSE
performance indicated that the architecture of 30 neurons in three hidden layers with the TANSIG transfer function yielded
nu
ROP for the developed model for the training and testing datasets, respectively. The ANNs model achieved a high R2 of 0.92
for the training dataset, while R2 was found to be 0.97 for the testing dataset. In addition, error distribution plots were also
used to allow more statistical analysis of ANNs performance (Figs. 6a and 6b). Mean and standard deviation (SD) were in
turn equal to 0.26 and 1.03 for training dataset, while the mean and standard deviation (SD) were equal of -0.16 and 0.72 for
ed
the testing dataset. Small values of mean and standard deviation (SD) verify the robustness of ANNs model. The error
distribution indicates 93% of predicated ROP values have errors in the range of ±2 at the training phase, while 99% of
predicated ROP values have errors in the range of ±2 m/hr at the testing phase which is an acceptable error for ROP. An
pt
inspection of Figs. 5 and 6 clearly show that the performances of the ANNs model for the training and testing datasets varied
slightly. Furthermore, the developed ANNs model provided very good results for the two datasets.
ce
Ac
R 2 =0.92 R 2 =0.97
15 15
10 10
5 5
0 0
d
0 5 10 15 20 0 5 10 15 20
Predicted ROP (m/hr) Predicted ROP (m/hr)
te
Fig. 5– Model outputs vs. real data: (a) training dataset, (b) testing dataset.
di
(a) (b)
e
py
Co
ot
tN
rip
Fig. 6– Error distribution statistics for the developed ANNs model: (a) training dataset, (b) testing dataset.
sc
To demonstrate the robustness of the proposed ANNs model, the RMSE, AAPE, and R2 were calculated for the training
and testing datasets. As seen in Table 2, the predicted performance for the testing dataset is better than for the training
nu
dataset. The high performance of the ANNs model in the testing phase can be considered a good indication of the model’s
capabilities.
Ma
R2 0.92
RMSE 1.23
Testing AAPE 8.8%
R2 0.97
pt
The impact of the most effective controllable parameters (i.e., WOB, RPM, and FR) was investigated to determine their
effect on the ROP. Figs. 8a, 8b, and 8c show the effect of these variables on the ROP; for each of these variables, the other
variables were kept constant at their midrange or recommended values (Table 3), while the parameters under study (i.e.,
Ac
WOB, RPM, and FR) varied within the range of the input parameters used in the ANNs training process.
10
d
*Parameter under study is varied.
te
Figure 7a demonstrates a semi-straight line response of the ROP to the WOB until a critical point (i.e., founder point).
The founder point is the point at which the ROP is no longer a linear response to the applied WOB. This point depends on the
di
bit design and the structure of the bit teeth or cutters [57]. Increasing the WOB will push the bit cutters further down into a
formation and disintegrate more rocks. However, if too much WOB is applied (i.e., the founder point is reached), rapid bit-
e
wearing will occur, resulting in a lower ROP. In Fig. 7a, the WOB was increased in 0.5-ton steps, and the ROP responded by
py
increasing 0.47 m/hr with each step, up to 13.5 tons. The bit founder point was between 12 and 13.5 tons, and the next
increase in the ROP was at less than 0.45 m/hr (0.12 m/hr). To achieve a high ROP with a long bit life, the drilling team must
limit the WOB so that it remains at or below the founder point (13.5 tons).
Co
The relationship between the ROP and RPM is presented in Fig. 7b. An increase in the RPM generally enhanced the
ROP, but slight growth in the ROP was seen at a high RPM of more than 130 rev/min. When starting at a low RPM, an
increase in the RPM could raise the shearing force of the bit cutters to disintegrate more rocks and therefore increase the
ROP. This improvement continued until hole-cleaning problems occurred and reversed the result. When the RPM range was
ot
limited, a later reverse of the performance was not observed (Fig. 7b).
Figure 7c shows that the ROP is directly proportional to the FR, where increasing the FR improves the lifting capacity of
tN
the drilling fluids to transport the cuttings, which consequently raises the ROP. However, an excess amount of FR may cause
a positive effect on the hydraulic impact force by increasing the equivalent circulating density (ECD), which decreases the
ROP. This behavior was not observed because the FR range was limited to 2200 l/min (Fig. 7c).
rip
(a) (b)
sc
nu
Ma
ed
(c)
pt
ce
Ac
Fig. 7– Effect of operating parameters on the ROP: (a) effect of the WOB on the ROP, (b) effect of the RPM on the ROP, (c) effect
of the FR on the ROP.
11
d
advantageous for prediction because the ANNs model underwent many scenarios to better learn about the ROP behavior in
the Zubair Formation.
te
Thus, the developed model was successful at indicating changes in the ROP due to a change in input parameters. For
example, after a sudden decrease in the WOB and RPM occurred, the actual ROP experienced a sudden significant decrease,
di
as shown in well 1 in the short interval between 3012 m and 3026 m (Fig. 8). This was detected and reflected on the value of
the predicted ROP at the mentioned depth. In addition, the sudden increase in the ROP that was associated with an increase in
e
the WOB and RPM was identified by the predicted ROP over several short intervals in well 2 from 3460 m to 3466 m and
py
also from 3585 m to 3590 m and for well 3, between 3224 m and 3229 m (Fig. 9 and Fig. 10, respectively). However, no
unjustified changes in the actual ROP without any clear known changes in the input parameters was detected by the
developed model; therefore, none was reflected in the predicted ROP. Changes in the predicted ROP were smooth, an
Co
illustrative example of which can be found in well 3, at a depth 3278 m to 3300 m and 3453 m to 3500 m, where a sudden
change in the actual ROP was indicated, although there were no changes in the input parameters (Fig. 10).
Measured ROP
ot
(a) Predicted ROP
(b) Residual Errors
3000 3000
tN
3050 3050
rip
sc
3100 3100
nu
Depth(M)
Depth(M)
3150 3150
Ma
3200 3200
ed
3250 3250
pt
ce
3300 3300
0 5 10 15 20 -10 -5 0 5 10
ROP (m/hr) Residual Error of ROP
Ac
Fig. 8– ROP prediction for well 1: (a) predicted and measured ROP along the depth, (b) residual errors of the predicted ROP.
12
3400 3400
3450 3450
d
Depth(M)
Depth(M)
te
3500 3500
di
3550 3550
e
py
3600 3600
Co
3650 3650
0 5 10 15 20 -10 -5 0 5 10
ot
ROP (m/hr) Residual Error of ROP
tN
Fig. 9– ROP prediction for well 2: (a) the predicted and measured ROP along the depth, (b) residual errors of the predicted ROP.
Measured ROP
(a) Predicted ROP
(b) Residual Errors
rip
3200 3200
sc
3250 3250
nu
3300 3300
Ma
Depth(M)
Depth(M)
3350 3350
ed
3400 3400
pt
ce
3450 3450
Ac
3500 3500
0 5 10 15 20 -10 -5 0 5 10
ROP (m/hr) Residual Error of ROP
Fig. 10– ROP prediction for well 3: (a) the predicted and measured ROP along the depth, (b) residual errors of the predicted
ROP.
13
4. Conclusions
Introducing an artificial neural network to petroleum industry eases the process of recognizing complex patterns and
relations between different parameters without any assumptions about the nature and distribution of the data. Presented work
is an example of the implementation of the artificial neural network to model the rate of penetration using different
parameters such as rig- and bit-related parameters, 1D MEM parameters, and drilling fluid properties. Moreover, following
this technique can help to unfold knowledge from the dataset.
Applying fscaret package in R environment enabled to find the importance and ranking of all 19 variables. The results of
feature ranking demonstrate that the WOB, RPM, and FR had the most impact on the ROP. According to the performance
indicators, the structure of the FFBP, with a five-layer network, with an architecture of {19 30 30 30 1} neurons provided the
best performance for ROP prediction. Among several tested transfer functions, TANSIG for hidden layers and PURELIN for
d
the output layer yielded better efficiency in predicting the ROP.
te
Finally, the newly developed model can only be used within the range of applied datasets, and caution should be taken
when selecting the input parameters. However, for a new system with variables beyond the range of the input parameters
di
used in the training operations of this study or for a new bit type/size (e.g., roller cone bits), it is necessary to develop new
ANNs models. Also, this model is valid for other datasets that fall within the range of the dataset employed in the training
e
process. Therefore, the use of a wider range of input parameters for the training process can enable the ANNs to have wider
applications, in select cases.
py
Acknowledgments
Co
The authors would like to gratefully acknowledge Basrah Oil Company (BOC) and Iraqi Drilling Company (IDC) in Iraq
for providing technical data and their permission to publish the results.
ot
Nomenclature
fscaret Automated Feature Selection from 'caret'
tN
ROP Rate of Penetration
1D MEM One-Dimension Mechanical Earth Model
MD Measured Depth
rip
FR Flow Rate
CP Circulation Pressure
nu
BW Bit Working-hours
TFA Total Flow Area
AZI Azimuth
Ma
INC Inclination
UCS Unconfined Compressive Strength
σv Vertical Stress
σh Minimum Horizontal Stress
ed
PV Plastic Viscosity
YP Yield Point
ce
References
Ac
[1] Hegde, C., and Gray, K., 2018, “Evaluation of coupled machine learning models for drilling optimization,” J. Nat. Gas
Sci. Eng. 56, pp.397–407. http://dx.doi.org/10.1016/j.jngse.2018.06.006.
[2] Yi, P., Kumar, A., and Samuel, R., 2014, “Realtime Rate of Penetration Optimization Using the Shuffled Frog Leaping
Algorithm,” ASME J. Energy Resour. Technol., 137(3), p. 032902. https://doi.org/10.1115/1.4028696.
[3] Al-AbdulJabbar, A., Elkatatny, S., Mahmoud, M., Abdelgawad, K., and Al-Majed, A., 2018, “A Robust Rate of
Penetration Model for Carbonate Formation,” ASME J. Energy Resour. Technol., 141(4), p. 042903.
https://doi.org/10.1115/1.4041840.
[4] Gidh, Y. K., Ibrahim, H., and Purwanto, A.,2011, “Real-time drilling parameter optimization system increases ROP by
predicting/managing bit wear,” SPE Digital Energy Conference and Exhibition, Woodlands, Texas, April 19–21, Paper
14
d
[9] Ataei, M., KaKaie, R., Ghavidel, M., and Saeidi, O., 2015, “Drilling rate prediction of an open pit mine using the rock
te
mass drillability index,” Int. J. Rock Mech. Min. Sc. 73, pp.130-138. http://dx.doi.org/10.1016/j.ijrmms.2014.08.006.
[10] Deng, Y., Chen, M., Jin, Y., Zhang, Y., Zou, D., and Lu, Y., 2016, “Theoretical and experimental study on the
di
penetration rate for roller cone bits based on the rock dynamic strength and drilling parameters,” J. Nat. Gas Sci. Eng. 36,
pp.117–123. http://dx.doi.org/10.1016/j.jngse.2016.10.019.
e
[11] Hegde, C., Daigle, H., Millwater, H., and Gray, K., 2017, “Analysis of rate of penetration (ROP) prediction in drilling
using physics-based and data-driven models,” J. Pet. Sci. Eng. 159, pp.295–306.
py
https://doi.org/10.1016/j.petrol.2017.09.020.
[12] Rashidi, B., Hareland, G., and Wu, Z., 2015, “Performance, simulation and field application modeling of rollercone
Co
bits,” J. Pet. Sci. Eng.133, pp.507–517. http://dx.doi.org/10.1016/j.petrol.2015.06.003.
[13] Kahraman, S., Balcı, C., Yazıcı, S., and Bilgin, N., 2000, “Prediction of the penetration rate of rotary blast hole drills
using a new drillability index” Int. J. Rock Mech. Min. Sci. 37(5), pp.729–743. http://dx.doi.org/10.1016/s1365-
1609(00)00007-1.
ot
[14] Eren, T., and Kok, M. V., 2018, “A new drilling performance benchmarking: ROP indexing methodology,” J. Pet. Sci.
Eng. 163, pp.387–398. http://dx.doi.org/10.1016/j.petrol.2018.01.002.
tN
[15] Soares, C., Daigle, H., and Gray, K., 2016, “Evaluation of PDC bit ROP models and the effect of rock strength on model
coefficients,” J. Nat. Gas Sci. Eng. 34: pp.1225–1236. https://doi.org/10.1016/j.jngse.2016.08.012.
[16] Elkatatny, S., 2018, “Application of Artificial Intelligence Techniques to Estimate the Static Poisson's Ratio Based on
rip
Wireline Log Data,” ASME J. Energy Resour. Technol., 140(7), p. 072905. https://doi.org/10.1115/1.4039613.
[17] Paul, A., Bhowmik, S., Panua, R., and Debroy, D., 2018, “Artificial Neural Network-Based Prediction of Performances-
Exhaust Emissions of Diesohol Piloted Dual Fuel Diesel Engine Under Varying Compressed Natural Gas Flowrates,”
sc
Drilling Technology Conference and Exhibition, Bangkok, Thailand, August 27-29, Paper No. SPE-191097-MS.
http://dx.doi.org/10.2118/191097-ms.
[19] Alkamil, E. H., Al-Dabooni, S., Abbas, A. K., Flori, R., and Wunsch, D. C., 2018, “Learning From Experience: An
Ma
Automatic pH Neutralization System Using Hybrid Fuzzy System and Neural Network,” Procedia Computer Science,
140, pp.206-215. http://dx.doi.org/10.1016/j.procs.2018.10.330.
[20] Mantha, B., and Samuel, R., 2016, “ROP optimization using artificial intelligence techniques with statistical regression
ed
coupling,” SPE Annual Technical Conference and Exhibition, Dubai, UAE, September 26–28, Paper No. SPE-181382-
MS. http://dx.doi.org/10.2118/181382-ms.
[21] Abbas, A. K., Alameedy, U., Alsaba, M., and Rushdi, S., 2018, “Wellbore Trajectory Optimization Using Rate of
pt
Penetration and Wellbore Stability Analysis,” SPE International Heavy Oil Conference and Exhibition, Kuwait City,
Kuwait, December 10-12. SPE-193755-MS. http://dx.doi.org/10.2118/193755-ms.
ce
[22] Bhowmik, S., Panua, R., Debroy, D., and Paul, A., 2017, “Artificial Neural Network Prediction of Diesel Engine
Performance and Emission Fueled With Diesel–Kerosene–Ethanol Blends: A Fuzzy-Based Optimization,” ASME J.
Energy Resour. Technol., 139(4), p. 042201. http://dx.doi.org/10.1115/1.4035886.
Ac
[23] Rostami, H., and Khaksar Manshad, A., 2014, “A New Support Vector Machine and Artificial Neural Networks for
Prediction of Stuck Pipe in Drilling of Oil Fields,” ASME J. Energy Resour. Technol., 136(2), p. 024502.
http://dx.doi.org/10.1115/1.4026917.
[24] Wang, Y., and Salehi, S., 2015, “Application of Real-Time Field Data to Optimize Drilling Hydraulics Using Neural
Network Approach,” ASME J. Energy Resour. Technol., 137(6), p. 062903. http://dx.doi.org/10.1115/1.4030847.
[25] Manshad, A. K., Rostami, H., Rezaei, H., and Hosseini, S. M., 2015, “Application of Artificial Neural Network-Particle
Swarm Optimization Algorithm for Prediction of Asphaltene Precipitation during Gas Injection Process and Comparison
with Gaussian Process Algorithm,” ASME J. Energy Resour. Technol., 137(6), p. 062904.
http://dx.doi.org/10.1115/1.4031042.
15
d
[30] Shi, X., Liu, G., Gong, X., Zhang, J., Wang, J., and Zhang, H., 2016, “An efficient approach for real-time prediction of
te
rate of penetration in offshore drilling,” Math. Probl. Eng. Article ID 3575380. http://dx.doi.org/10.1155/2016/3575380.
[31] Mohammed, H. Q., Abbas, A. K., and Dahm, H. H., 2018, “Wellbore instability analysis for Nahr Umr Formation in
di
southern Iraq,” 52nd US Rock Mechanics/Geomechanics Symposium (ARMA), Seattle, Washington, June 17–20, Paper
No. ARMA 18–916.
e
[32] Gholami, R., Aadnoy, B., Foon, L. Y., and Elochukwu, H., 2017, “A methodology for wellbore stability analysis in
anisotropic formations: A case study from the Canning Basin, Western Australia,” J. Nal. Gas Sci. Eng. 37, pp.341–360.
py
http://dx.doi.org/10.1016/j.jngse.2016.11.055.
[33] Abbas, A. K., Flori, R. E., and Alsaba, M., 2018, “Estimating rock mechanical properties of the Zubair shale formation
Co
using a sonic wireline log and core analysis,” J. Nal. Gas Sci. Eng. 53, pp.359–369.
https://doi.org/10.1016/j.jngse.2018.03.018.
[34] Abbas, A. K., Flori, R. E., Alsaba, M., Dahm, H., and Alkamil, E. H., 2018, “Integrated approach using core analysis
and wireline measurement to estimate rock mechanical properties of the Zubair reservoir, Southern Iraq,” J. Pet. Sci. Eng.
ot
166, pp.406–419. https://doi.org/10.1016/j.petrol.2018.03.057.
[35] Ahmed, M., Al-Shehri, H. A., Haidary, S. A., and Povstyanova, M., 2016, “A comprehensive geomechanical study to
tN
understand drilling challenges in the Manifa field offshore, Saudi Arabia,” Kingdom of Saudi Arabia Annual Technical
Symposium and Exhibition, Dammam, Saudi Arabia, April 25–28, Paper No. SPE-182833-MS.
https://doi.org/10.2118/182833-ms.
rip
[36] Zhang, J., 2011, “Pore pressure prediction from well logs: Methods, modifications, and new approaches,” Earth Sci.
Rev. 108(1–2), pp.50–63. http://dx.doi.org/10.1016/j.earscirev.2011.06.001.
[37] Eaton, B.A., 1969, “Fracture gradient prediction and its application in oilfield operations,” J. Petrol. Tech. 21 (10),
sc
pp.1353–1360. http://dx.doi.org/10.2118/2163-pa.
[38] Jaeger, J. C, Cook, N. G, and Zimmerman, R. W., 2007, “Fundamentals of Rock Mechanics,” fourth ed. Wiley-
Blackwell.
nu
[39] Najibi, A. R., Ghafoori, M., Lashkaripour, G. R., and Asef, M. R., 2017, “Reservoir geomechanical modeling: In-situ
stress, pore pressure, and mud design,” J. Pet. Sci. Eng. 151 pp.31–39. http://dx.doi.org/10.1016/j.petrol.2017.01.045.
[40] Gholami, R., Rabiei, M., Aadnoy, B., and Rasouli, V., 2017, “A methodology for wellbore stability analysis of drilling
Ma
into presalt formations: A case study from Southern Iran,” J. Pet. Sci. Eng. 167, pp.249–261.
http://dx.doi.org/10.1016/j.petrol.2017.11.023.
[41] Thiercelin, M., and Plumb, R., 1994, 1994, “A core-based prediction of lithologic stress contrasts in East Texas
ed
http://dx.doi.org/10.1016/j.petrol.2016.01.003.
[43] Abbas, A. K., Flori, R. E., and Alsaba, M., 2019. “Stability analysis of highly deviated boreholes to minimize drilling
ce
risks and nonproductive time,” ASME J. Energy Resour. Technol., 141(8), p. 082905.
http://dx.doi.org/10.1115/1.4042733.
[44] Chandrashekar, G., and Sahin, F., 2014, “A survey on feature selection methods,” Comput. Electr. Eng.; 40 (1), pp.16–
Ac
28. https://doi.org/10.1016/j.compeleceng.2013.11.024.
[45] Eskandarian, S., Bahrami, P., and Kazemi, P., 2017, “A comprehensive data mining approach to estimate the rate of
penetration: Application of neural network, rule based models and feature ranking,” J. Pet. Sci. Eng. 156, pp.605–615.
https://doi.org/10.1016/j.petrol.2017.06.039.
[46] Szlek, J., and Mendyk, A., 2018, “Fscaret: Automated feature selection from caret,” https://cran.r
project.org/web/packages/fscaret/index.html (accessed 3 November 2018).
[47] Kazemi, P., Khalid, M. H., Szlek, J., Mirtič, A., Reynolds, G. K., Jachowicz, R., and Mendyk, A., 2016, “Computational
intelligence modeling of granule size distribution for oscillating milling,” Powder Technol. 301, pp.1252–1258.
https://doi.org/10.1016/j.powtec.2016.07.046.
16
d
pipe sticking by support vector machine compared with conventional artificial neural networks: An example of Iranian
te
offshore oil fields,” SPE Drill. & Compl. 27 (04), pp.586–595. http://dx.doi.org/10.2118/163062-pa.
[53] Akande, K. O., Owolabi, T. O., and Olatunji, S. O., 2015, “Investigating the effect of correlation-based feature selection
di
on the performance of support vector machines in reservoir characterization,” J. Nal. Gas Sci. Eng. 22, 515-522.
http://dx.doi.org/10.1016/j.jngse.2015.01.007.
e
[54] Abbas, A. K., Flori, R. E., and Alsaba, M., 2018, “Laboratory Geomechanical Characterization of the Zubair Shale
Formation,” 52nd US Rock Mechanics/Geomechanics Symposium (ARMA), Seattle, Washington, USA, June 17–20,
py
Paper No. ARMA 18– 78.
[55] Abbas, A. K., Dahm, H. H., Flori, R. E., and Alsaba, M., 2018, “Laboratory Measurements of Petrophysical and
Co
Geomechanical Properties for Zubair Sandstone Formation in Southern Iraq,” 52nd US Rock Mechanics/Geomechanics
Symposium (ARMA), Seattle, Washington, USA, June 17–20, Paper No. ARMA 18–243.
[56] Soares, C., and Gray, K., 2019, “Real-time predictive capabilities of analytical and machine learning rate of penetration
(ROP) models,” J. Pet. Sci. Eng. 172, pp.934–959. http://dx.doi.org/10.1016/j.petrol.2018.08.083.
ot
[57] Chen, X., Gao, D., Guo, B., Feng, Y., 2016, “Real-time optimization of drilling parameters based on mechanical specific
energy for rotating drilling with positive displacement motor in the hard formation,” J. Nat. Gas Sci. Eng. 35, pp.686–694.
tN
http://dx.doi.org/10.1016/j.jngse.2016.09.019.
rip
sc
nu
Ma
ed
pt
ce
Ac
17