Journal of Geophysics and Engineering

J. Geophys. Eng. 15 (2018) 1146–1159 (14pp) https://doi.org/10.1088/1742-2140/aaac5d

Drilling rate prediction from petrophysical

logs and mud logging data using an optimized
multilayer perceptron neural network

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024
Mohammad Anemangely1, Ahmad Ramezanzadeh1, Behzad Tokhmechi1,
Abdollah Molaghab2 and Aram Mohammadian2
1
School of Mining, Petroleum and Geophysics engineering, Shahrood University of Technology,
Shahrood, Iran
2
National Iranian South Oil Company, Ahvaz, Iran

E-mail: [email protected]

Received 19 August 2017, revised 29 January 2018

Accepted for publication 1 February 2018
Published 13 April 2018

Abstract
Rate of penetration (ROP) enhancement serves as a key factor in reducing drilling time and hence
drilling costs. ROP enhancement requires identification of the parameters affecting this rate.
However, the large number of effective parameters, which are further immersed in noise, makes it
difficult to present a highly accurate and comprehensive model. In the present research, in order to
predict the drilling ROP in one of the vertical wells drilled into the Karanj Oilfield, a hybrid model
composed of a multilayer perceptron (MLP) neural network together with either a particle swarm
optimization (PSO) algorithm or a cuckoo optimization algorithm (COA) was used. For this purpose,
first petrophysical logs and drilling data were denoised using the Savitzky–Golay filter. Then, the
‘plus-l-take-r’ method was used to select superior features. Feature selection results indicated that an
increase in the number of input parameters tends to reduce the error associated with the estimator
model; however, the error reduction rate was seen to be negligible for models with five or more input
parameters. Therefore, five parameters were considered as input parameters in MLP-COA and MLP-
PSO hybrid models: rotary speed, weight on bit, shear wave slowness, compressional wave
slowness, and flow rate. A comparison of errors and coefficients of determination in the training
phase of the two models indicated that MLP-COA model tended to converge way faster and more
accurately. The small difference in generated error using this model between training and testing
phases indicated the high reliability and generalizability of the model. Comparing the results of the
model trained with raw and denoised data against the same set of selected features clearly showed the
positive effect of the denoising step on the accuracy of the model. Validation of the proposed model
via the multilinear regression method was indicative of the superior performance of the MLP-COA
model, so that it could be confidently stipulated that this model can be used to estimate the ROP at
other vertical wells near the studied well. Further, provided the required information is available, this
method can predict the ROP to high accuracy in vertical oil and gas wells.

Keywords: rate of penetration, artificial intelligence, noise reduction, feature selection,

regression
(Some figures may appear in colour only in the online journal)

1. Introduction operation. The results of a survey (Li et al 2015) showed that

the drilling phase takes one third of the total time to construct a
Drilling rate or rate of penetration (ROP) refers to the length to well, with the fraction being one half in deep wells. Drilling
which a drilling tool penetrates into rock during drilling rate prediction serves as a key step toward determining optimal

1742-2132/18/0401146+14$33.00 1146 © 2018 Sinopec Geophysical Research Institute Printed in the UK

J. Geophys. Eng. 15 (2018) 1146 M Anemangely et al

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024
Figure 1. Some of the parameters known to affect the ROP.

values of drilling parameters, so as to reduce drilling costs. As earth model, from offset well data is a simple method that is
such, it has gained a great deal of attention from drilling widely used to predict petrophysical and mechanical rock
engineers over recent decades and is still an active field of properties at the proposed location for drilling a well.
research where researchers seek more accurate results. There- Therefore, these data can be used for drilling rate estimation.
fore, it is extremely important to compare results of different The methods used in previous research works can be
techniques to find ROP predictions of the highest accuracy categorized into two groups: mathematical methods and soft
(Amar and Ibrahim 2012). computation techniques. The Bourgoyne–Young model
Before developing a model to predict the ROP, it is (Bourgoyne and Young 1974) has been among the most suc-
necessary to identify the parameters affecting this rate. cessful mathematical models ever proposed to predict ROP and
Numerous factors tend to affect the ROP (figure 1). The is being widely used by researchers (Ansari et al 2016). The
existence of nonlinear and complex interactions between each results of this model depend highly on its constants. One of the
of these parameters and the ROP (Mendes et al 2007) and the main limitations faced by this model is the way these constants
uncertainties associated with data interpretation (due to sud- are determined from available drilling data in the field of
den changes in rock properties—Paiaman et al 2009) have interest. Improper results of multivariate regression in deter-
made it difficult to present a comprehensive relationship. mining these constants motivated other researchers to use other
Numerous attempts have already been made to model the optimization methods (Bahari and Baradaran Seyed 2009,
ROP, with each model accounting for a different set of Rahimzadeh et al 2011, Anemangely et al 2017a). Moraveji
parameters to predict the rate. A study of previous research and Naderi (2016) applied a response surface methodology to
indicates that, in most of the modeling performed so far, only model the ROP. They then used the model and intelligent bat
mechanical parameters and fluid properties have been used to optimization algorithm to determine optimal ranges of different
predict the ROP. That is, despite the undeniable role played parameters under investigation. Considering the good perfor-
by rock properties in the ROP, the properties are yet to be mance of artificial intelligent (AI) methods in dealing with
well addressed in ROP prediction problems. Failure to con- complicated problems, these methods have been further
sider all effective parameters has contributed to the low acc- applied to ROP modeling and optimization problems. Arabja-
uracy of these models (Amar and Ibrahim 2012) while maloei and Shadizadeh (2011) used a neural network to predict
making them slow (Sultan and Al-Kaabi 2002, Paiaman ROP in wells drilled into the Ahwaz Oilfield (Iran). They
et al 2009, Abtahi 2011, Hamrick 2011, Huang et al 2011). further used multivariate regression and the Bourgoyne–Young
Basically, wireline logs are run once the drilling operation is model to predict the ROP. A comparison of the results of the
actually performed, and this may be a reason for ignoring three methods indicated good accuracy of the neural network.
rock characteristics in the penetration rate estimation. In order to optimize drilling parameters using the particle
Nowadays, petrophysical logs can be provided during the swarm optimization (PSO) algorithm, Duan et al (2015)
drilling operation using logging while drilling (LWD) tools employed a back-propagation neural network for ROP mod-
(Ma et al 2016, Tian-Shou and Ping 2015). Also, these data eling. Ansari et al (2016) began by selecting the parameters
can be estimated by applying cross-well imaging techniques imposing the largest contributions on to the ROP using mul-
for a well to be drilled between two wells (Raji et al 2017). tiple regression error analysis. Then, they used the selected
Furthermore, building 3D earth models, such as a geophysical parameters along with different support vector regression

1147
J. Geophys. Eng. 15 (2018) 1146 M Anemangely et al

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024
Figure 2. Research steps.

methods to predict the ROP in the well under study. The results increase the processing time, although an enhancement in
of this study showed that higher accuracy can be achieved when accuracy would also be achieved. As such, one should select a
the support vector regression is used with RBF as the kernel smart combination of these parameters with which the pro-
function. Bezminabadi et al (2017) used a multilayer perceptron cessing rate can be enhanced while preserving as much acc-
(MLP) neural network and multiple nonlinear regression (MNR) uracy as possible. This can be done via feature selection
for estimating the ROP. For this purpose, they gathered well techniques. However, before such a technique can be imple-
logging and mud logging data from a well in the southwest of mented, actual measured data (which are contaminated with
Iran. Their results showed that the MLP neural network was noise) should be denoised to be able to present a model with
superior to MNR. Parameter selection was based on the value of better accuracy and performance. Figure 2 shows the research
the correlation coefficient of each parameter to the ROP. Seeking
steps in the present study. In order to model the ROP, an MLP
the minimum difference between estimated results and observed
neural network will be used, with its parameters optimized
data, parameter selection was essentially based on residual error,
using two metaheuristic algorithms: the cuckoo optimization
leaving the correlation coefficient-based selection as an improper
selection approach (Anemangely et al 2017b). Based on this algorithm (COA) and PSO. In the following, different stages
feature selection method, they ignored shear and compressional of this research are explained in brief.
sonic logs in the ROP estimation. These logs provide information
on the mechanical characteristics of underground layers, such as
uniaxial compressive strength (UCS) and the tensile strength of 2.1. Data from the well under study
rock. Some researchers have used UCS to estimate the drilling The data studied in the present paper are taken from a vertical
rate (Bahari and Baradaran Seyed 2007, Zhang et al 2012). well drilled into the Karanj Oilfield in southwestern Iran. The
Considering the limitations dealt with in previous studies data used to predict the ROP included mud logging data and
(i.e. failure to account for formation parameters in ROP models
well logs. Mud logging data encompassed depth, weight on
and thereby low accuracy of the models) the present research
bit (WOB), rotary speed, drilling fluid density, drilling fluid
takes into consideration formation parameters along with mud
flow rate, and pump pressure. Moreover, the used petrophy-
logging data to estimate the ROP. In order to enhance the
estimator model in terms of accuracy and processing time, sical logs included compressional and shear sonic wave logs,
noise reduction and feature selection techniques were used. density logs, and gamma ray and porosity logs. The sampling
Identification of appropriate input parameters not only accel- rate of the well logging and mud logging data in this study
erates data processing, but also keeps the model from over- was 0.1524 and 1 m, respectively. As an up-scaling method,
fitting (Anemangely et al 2017b). Since the AI methods have the well logs were averaged to have them on the same depth
given better results than those of mathematical and statistical scale, so as to be able to integrate well logging and mud
approaches, a neural network was used for ROP modeling. logging data into one database. The area under study is
located within the reservoir and represented by 625 data
points in ASMARI (2404–2808 m) and PABDEH
2. Methodology (2809–3028 m) formations. Table 1 presents the effect of each
of the investigated parameters on the ROP. The ranges of the
As was previously mentioned, numerous parameters con- investigated parameters in the well under study are further
tribute to the ROP. However, constructing a model wherein reported in table 2. This range of depth was drilled using three
all of these parameters are accounted for would impractically PDC bits of 8.5 inch in diameter.

1148
J. Geophys. Eng. 15 (2018) 1146 M Anemangely et al

Table 1. Reasons for selecting the investigated parameters in this study.

Parameters A measure of Effect on ROP

Depth Vertical stress Elastic properties of many formations are influenced by the state of
stress. Many times, shales tend to raise problems for drilling opera-
tion in depth, mainly due to the effect of vertical stress.
Weight on bit (WOB) Axial thrust for rock crushing As the WOB increases to a critical value (dominating rock strength), the
ROP increases. If it exceeds this critical value, the ROP will reduce as
bottom of the borehole may no longer be efficiently washed.

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024
Rotary speed Shear cutting force for rock With increasing rotary speed to a critical value, the ROP increases.
fracturing However, beyond this critical value, reduction in well stability and
washout decrease the rate at which the ROP increases with rotary
speed.
Drilling fluid density Drilling fluid column pressure A reduction in drilling fluid weight results in the fragmentation of the
in the well bottom-hole rock under the effect of formation fluid pressure, thereby
increasing the ROP.
Flow rate — An increase in flow rate contributes to better removal of cuttings,
thereby enhancing the ROP.
Pump pressure — An increase in pump pressure results in enhanced bit nozzle force, and
hence better removal of cuttings, thereby enhancing the ROP.
Compressional and shear wave Mechanical properties of rock Volume of generated cuttings under a single cone is inversely propor-
slowness tional to compressive or shear strength of the rock.
Porosity log Rock compaction The higher the porosity, the easier the rock can be penetrated.
Density log Mineralogy and rock Tendency of formation to become muddy is influenced by its miner-
compaction alogy. Sedimentary rocks contain chert and quartz, making them
difficult to drill.
Gamma log Shale and clay contents Being capable of being hydrated, clays (particularly bentonite) may
surround the bit entirely when they come into contact with water,
thereby reducing the ROP.

Table 2. Ranges of the parameters investigated in the well under study.

Factor Coded factor Minimum value Mean value Maximum value

Depth (m) X1 2404 2716 3028
Gamma ray (GAPI) X2 7.859 49.772 128.104
Density (G/C3) X3 2.310 2.622 2.832
Porosity (V/V) X4 0.003 0.116 0.360
Compressional wave travel time (ms ft−1) X5 49.106 64.617 93.464
Shear wave travel time (ms ft−1) X6 96.508 121.165 179.363
Weight on bit (klbf) X7 0.161 6.928 18.996
Rotation speed (rpm) X8 0 135.099 222.750
Pump pressure (kPa) X9 1597.860 3926.296 5465.620
Mud density (pcf) X10 46.053 55.980 56.123
Flow rate (l/s) X11 12.732 20.982 25.789
Penetration rate (m h−1) ROP 0.252 3.677 8.865

2.2. Noise reduction Carvalho 2004, Garćia et al 2013). Geophysicists have fre-
quently argued that any environmental property measurement
Real data (i.e. inputs into processing algorithms) are influ-
is contaminated with noise (Cardiff and Kitanidis 2010). The
enced by numerous factors. Among others, noise is a key presence of noise in seismic data, particularly illumination
factor affecting the data (Wang et al 1995). Various defini- noises such as multiple reflections and artefacts resulting
tions have been proposed for noise in statistical and machine from inadequate seismic illumination, will end up with
learning literature. Most of the definitions confirm that noisy inappropriate results when low signal-to-noise ratio data is
data can affect the learning process (Quinlan 1986) and integrated (Zhou 2014).
increase learning time. In addition, the presence of noise The most important step to be taken for noise reduction is
within data can challenge the extraction of rules from data in to identify the source of the noise. Such causes as shift change
machine learning models, deteriorating the performance of of drilling engineers, workovers and drilling tool replacement,
the models in predicting new data (Lorena and de drilling string vibrations, washouts, and alterations in

1149
J. Geophys. Eng. 15 (2018) 1146 M Anemangely et al

geological layers can further contribute to errors in the data Table 3. Pseudo-code of the SFS-dominated plus-l-take-r method.
acquired at a well.
1. Initialization: number of features to be selected (d) is speci-
In order to remove noise from the data studied in the
fied. The selected feature set is considered to be empty. The
present research the Savitzky–Golay (SG) smoothing filter
best root mean square error (Best_RMSE) is assumed to be
(Savitzky and Golay 1964) was used. This method involves
infinity. Values of l and r are set.
applying a polynomial function (similar to equation (1)) to
remove the effect of noise from data, with the results of this 2. The following steps are performed while number of members
function ( x new , smoothed values) replacing initial (acquired) in the selected features set is less than d
2.1. For i=1 to l
values. In this method, a polynomial function of order n is

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024
2.1.1. For j=1 to number of existing features
fitted to a number of points within an interval on the basis of a 2.1.1.1. Add jth feature to the selected features set
least squares error. The number of points within this interval 2.1.1.2. Evaluate the selected feature set using the
should be odd and beyond the order of the polynomial. An neural network
increase in the order of the polynomial or a decrease in the 2.1.1.3. If the obtained RMSE from the evaluation is
number of points within the interval can further preserve data less than Best_RMSE, replace Best_RMSE by RMSE
structure and lower the extent to which data is smoothed. On and record the subscript of the corresponding feature,
the other hand, a decrease in polynomial order or an increase otherwise go to the next step
in the number of points within the interval eliminates part of 2.1.1.4. Eliminate the feature from the selected feature set and
the data structure and brings about further smoothing. let j=j+1
Therefore, determination of the proper polynomial order and 2.1.2. Add corresponding features to the recorded subscript in
step 1.1.3 to the selected features set and let i=i+1
the number of points within the interval is of paramount
2.2. Let Best_RMSE=infinity
significance. In order to obtain optimal values of the two 2.3. For i=1 to r
parameters, daily reports of the well under study and the 2.3.1. For j=1 to l
geologic setting of the field were investigated. According to 2.3.1.1. Take all features in the selected features set other than
the results, the optimal order of the polynomial and the the jth feature into a selected subset.
number of points within the interval for the drilling data 2.3.1.2. Evaluate the selected subset using neural network
were found to be 3 and 25, respectively. Undertaking the 2.3.1.3. If the obtained RMSE from the evaluation is less than
same analysis on petrophysical logs, the optimal order of Best_RMSE, replace Best_RMSE by RMSE and
the polynomial and the number of points within the interval record the subscript of the corresponding feature ( j),
were found to be 4 and 21, respectively. otherwise go to the next step
2.3.1.4. Let j=j+1
n
x new = å ci x i. (1 ) 2.3.2. Eliminate associated feature with the recorded subscript
from the set of selected features and let
i=0
RMSE_Best=infinity.

2.3. Feature selection

criterion are eliminated from the selected set. These steps are
Feature selection is classified under dimension reduction iterated until the number of the selected features reaches a
methods. In this method, of a total of P feature members, the predefined threshold. In cases where SBS dominates, the
d feature members of the largest contributions are selected mechanism is reversed, i.e. we begin with taking all features
for developing an estimator or classifier model (d<P). as selected and proceed to eliminate the features with lowest
Feature selection is very important because of three reasons: effect through successive iterations. As such, the algorithm
(1) accounting for each and any feature does not necessarily performs slower in the latter case as compared to former case
end up with a minimum error for the classifier or estimator where SFS dominates. On the basis of this conclusion, in the
model, (2) feature selection reduces the volume of calcula- present research, an SFS-dominated plus-l-take-r method was
tions, thereby enhancing the rate of data processing, and (3) used. Accordingly, l and r were set to 3 and 2, respectively. In
large volumes of memory are needed to take each and any this way, only one feature would be selected and added to the
feature into consideration. set of selected features at each step through the plus-l-take-r
In order to undertake feature selection among the para- method, so as to evaluate the effect of each feature in the
meters affecting ROP, the ‘plus-l-take-r’ method was used. estimator model. Table 3 shows the pseudo-code of the SFS-
This method is a combination of sequential forward selection dominated plus-l-take-r algorithm. In order to evaluate the
(SFS) and sequential backward selection (SBS); however it is selected features based on the root mean squares error
free of the disadvantages of the two systems, i.e. fixing the (RMSE) of estimations, the MLP was used (equation (2)).
selected (or eliminated) features. In the ‘plus-l-take-r’ Since the entire dataset was divided into training (70% of the
method, if l is greater than r, SFS dominates SBS, and entire dataset) and testing (30% of the entire dataset) subsets,
vice versa. The mechanism of this method when SFS dom- two values of the RMSE were determined each time the
inates is as follows: first, l members are added, one by one, to neural network was run. In order to obtain a single value of
the selected set in the order of the selection criterion. Then, r model error for each round of feature selection, equation (3)
members with the smallest contributions to the selection was used to combine the two error terms. In order to keep the

1150
J. Geophys. Eng. 15 (2018) 1146 M Anemangely et al

Table 4. A comparison among most popular training algorithms for neural networks.

Learning algorithm Principles of the algorithm Drawbacks

Back-propagation (BP) This algorithm uses rule chain to determine the effect of each - Slow convergence
weight in network structure with respect to derivatives of error - Trapping in local minima
function (Al Bulushi et al 2012). The process begins with cal- - The choice of proper value of
culating derivatives of the performance function at the last layer training rate
of the network and proceeds to propagate the derivatives back-
ward to the first layer of the network (Horne 1993,

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024
Haykin 1999).
BP with Momentum This method works similar to the error BP method, except that, in - The choice of optimal values of two
order to accelerate the convergence process, it not only sets parameters: training rate and momentum
weights according to error gradient, but also further varies the - Trapping in local minima
weights with μ.
Levenberg–Marquardt This method is a combination of decreasing gradient minimization - Trapping in local minima
and Gauss–Newton methods. When parameters are found to be
far away from optimality, this method acts similar to increasing
gradient, while it exhibits Gauss–Newtonian behavior when
parameters are near optimal.

model from being over-trained, training and testing error (Bishop 2007). Furthermore, progressive multilayer neural
weights (WTrain and WTest, respectively) were set to 0.45 and networks have been widely studied and applied in scientific
0.55, respectively. The used MLP neural network was com- problems. This type of neural network is ideally suitable for
posed of three hidden layers with 4, 5, and 3 neurons in the first, modeling relationships among a set of input variables (or
second and third hidden layers, respectively. The numbers of predictors) and one or more output variables (or responses)
hidden layers and neurons in each layer were determined fol- (Maimon and Rokach 2005). As such, the progressive MLP
lowing a trial and error approach. Being one of the most suc- was used to estimate the ROP in the studied well.
cessful training algorithms in terms of enhanced convergence The key task in developing a neural network is the
rate in multilayer artificial neural networks, the Levenberg– training process where values of the network parameters (i.e.
Marquardt (LM) algorithm was used to train the neural network weights and biases) are modified in a continuous process so as
(Hagan and Menhaj 1994). Due to random selection of initial to optimize the objective function. Generally, the objective
biases and weights of the neural network model, different function in the training process is the minimization of the
solutions were found each time the algorithm was run. In order associated error with the estimator or classifier model.
to avoid this, at each round of feature selection, the neural Numerous training algorithms have been proposed for this
network was run ten times and the lowest error associated with purpose, some of which are presented in table 4. Considering
the feature under consideration was considered.
the limitations faced by classical algorithms for determining
1 m optimal values of neural network parameters, COA and PSO
RMSE = å (T - O )2
n i=1
(2 )
algorithms were selected to train the prepared neural network.
RMSE model = WTrain ´ RMSE Train + WTest ´ RMSE Test. A trial and error approach was followed to determine number
(3 ) of hidden layers and number of neurons in each layer.
Accordingly, a MLP neural network with three hidden layers
containing 3, 4, and 3 neurons in the first, second, and third
layers, respectively, was developed. Of the total data points
2.4. MLP neural network and optimization algorithms selected in the feature selection phase, 70% were selected to
Artificial neural networks are inspired by the function of train the model, with the remaining 30% of data points being
biological neurons in the human brain. The main character- used to have the model tested.
istics of these networks is their ability to explore complex Many attempts have been made to replace conventional
nonlinear relationships among variables (Ali 1994), so that training algorithms in MLP neural networks with metaheur-
one can generalize a model trained by a particular problem to istic and communal intelligence algorithms. The cuckoo
other similar problems. Important steps toward constructing a search is a metaheuristic algorithm recently proposed by
neural network include data preparation, input variable Yang and Deb (2009). This algorithm is inspired by the
selection, and the choice of type and structure of neural net- special life of a bird called a cuckoo. Rajabioun (2011) pro-
work, transfer function, and training algorithm (Maimon and posed an optimizer version of this algorithm. A comparison
Rokach 2005). Among the different structures of neural net- between this algorithm and other optimization algorithms
works, the MLP network is more popular because of its indicated the higher superiority of its results over those of the
ability to estimate any function at adequate accuracy, pro- other algorithms (Kawam and Mansour 2012) while it tended
vided adequate hidden units are incorporated into it to offer faster convergence rates (Anemangely et al 2017a).

1151
J. Geophys. Eng. 15 (2018) 1146 M Anemangely et al

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024
Figure 3. Flowcharts of (a) the COA and (b) the PSO algorithm.

Figure 4. Recorded (in blue) and denoised (in red) petrophysical logs at the well under study.

However, compared to other algorithms, this algorithm populations of cuckoos (Npop). These initial cuckoos lay some
requires greater processing time. eggs (5–20 eggs) in nests of a number of host birds. Some of
Figure 3(a) shows a flowchart of the COA. Similar to other these eggs, which resemble well those of the host bird, find the
evolutionary algorithms, this algorithm starts with an initial opportunity to grow and become a mature bird. Other eggs,

1152
J. Geophys. Eng. 15 (2018) 1146 M Anemangely et al

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024
Figure 5. Recorded (in blue) and denoised (in red) mud logging data at the well under study.

however, are detected and destroyed by the host bird. The

surviving eggs are indeed an indication of suitability of the nests
in that region. The objective is to increase the population of
surviving eggs by finding suitable nests for laying eggs.
Cuckoos are always searching for suitable locations to lay their
eggs and increase the rate of surviving eggs. Once the surviving
eggs are grown to mature cuckoos, they form a community of n
cuckoos. Each community has its own habitat. The best habitat
across all cuckoo communities will be the destination of cuck-
oos in other communities. Then the cuckoos migrate toward this
best habitat and settle somewhere close to that. Considering the
number of eggs laid by each cuckoo and total number of eggs,
the maximum lay radius adjustment factor (α) and lower
(Varlow ) and upper (Varhi ) bounds of each decision variable, one
can define some egg laying radius (ELR) for each cuckoo
(equation (4)). Subsequently, the cuckoo starts to lay eggs
randomly in a number of nests within the laying radius. This
process continues until maximum situation or profit is achieved
and all cuckoos are collected in the same position. Figure 6. The p-values obtained from the t-test for each input under
investigation.
Number of current cuckoo’s eggs
ELR = a ´ following objective function was used in the present research.
Total number of eggs
´ (Varhi - Varlow). (4 ) Profit = - cost (habitat)
= - cost ([x1, x2, ¼, x Nvar ]) = - RMSE. (5 )

In order to solve an optimization problem, one should Undertaking a sensitivity analysis on the COA para-
form an array of decision variables. In an optimization pro- meters, the number of cuckoos was considered to be 30, the
blem with Nvar decision variables, a habitat is represented by algorithm was set to iterate for 40 iterations, and the ELR
a 1×Nvar array denoting the cuckoo’s current habitat. adjustment factor was assumed to be 1.
Associated profit with a habitat is obtained by evaluating its Figure 3(b) shows a flowchart of the PSO algorithm.
habitat array ( x1, x2, ¼, x Nvar ). Since COA is originally Using a population of particles called a particle swarm, PSO
designed to maximize an objective function, then the finds the best solution within a search space. This algorithm

1153
J. Geophys. Eng. 15 (2018) 1146 M Anemangely et al

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024
Figure 7. Results of applying the plus-l-take-r feature selection method Figure 8. A comparison of the reduction of RMSE in the training
(associated error with the models with different numbers of inputs). phases of the MLP-COA and MLP-PSO models.

Table 5. Results of applying the plus-l-take-r method on data of the where i = 1, 2, ¼, n , n is the number of particles in the
well under study (the features selected as input for the estimator swarm, w is the inertial weight (controls the recurrence of
model are marked in bold). particle velocity; Pedersen and Chipperfield 2010), c1 and c2
Number of input Selected variable(s) RMSE are positive coefficients referred to as personal and communal
training factors, respectively, and r1 and r2 are random
1 X8 0.1303
numbers in the interval [0, 1] (Coello et al 2007).
2 X8, X7 0.1164
3 X8, X7, X6 0.1126 According to the results of sensitivity analysis on the
4 X8, X7, X6 , X5 0.1071 parameters of this algorithm, the number of particles was set
5 X8, X7, X6, X5, X11 0.1038 to 50, the algorithm was set to iterate for a maximum of 200
6 X8, X7, X6 , X1, X5, X11 0.1030 iterations, and personal and communal training factors were
7 X8, X7, X11, X10 , X9 , X6 , X1 0.1022 both considered to be 2.
8 X8, X7, X11, X1, X10 , X5, X9 , X6 0.1022
9 X8, X7, X1, X11, X5, X10 , X3, X6 , X 4 0.1017
3. Results and discussion

starts with randomly initializing position of population in Figure 4 demonstrates a comparison between the petrophy-
search space (upper and lower bounds of decision variables) sical logs recorded at the well (in blue) along with denoised
and initializing the velocity between its lower (-Vmax ) and data (in red). As can be seen on this figure, denoised data has
upper (Vmax ) bounds. The positions determined for particles well preserved structure in the recorded data. Figure 5 pre-
are recorded as the best personal position (Pb) for them. Next, sents a comparison between recorded drilling data at the well
the position of each particle is evaluated using the objective (in blue) and denoised data (in red) using an SG filter.
function, and the particle with lowest value of objective According to the figure, trend of changes in the denoised data
function is selected as the best global position (Gb). At each is in good agreement with that of recorded data, i.e. noise
iteration, a new velocity (Vi (t + 1)) is defined for each par- effect has been minimized.
ticle (i) based on its previous velocity (Vi (t )) and the distance Following the research, a t-test was used to investigate
from the current position (xi (t )) to the best personal and the relationship between parameters and the objective para-
global positions (equation (6)). Consequently, the new posi- meter. Figure 6 demonstrates the obtained p-values from this
tion of the particle ( xi (t + 1)) is calculated based on its pre- test for all input parameters. Since the obtained p-values were
vious position and new velocity (equation (7)). In the next below 5%, it was concluded that it was very unlikely for the
step, the new position of the particles is evaluated using the relationships between these parameters and objective para-
objective function. The best personal and global positions are meter to be insignificant, i.e. the relationships were proved
updated by comparing the values of the new objective func- significant. The p-values could not be used as a criterion for
tion against previous global and personal best positions. selecting input parameters of highest effects on the objective
Iterations continue until a stopping criterion is met. parameter, but rather could be used as a criterion for elim-
Vi (t + 1) = wVi (t ) + c1r1(Pbi (t ) - xi (t )) inating the investigated parameter if the corresponding
+ c2 r2 (G b (t ) - xi (t )) (6 ) p-value exceeded 5%.
Table 5 presents the results of applying feature selec-
xi (t + 1) = xi (t ) + Vi (t + 1) (7 ) tion on the denoised data. Associated error values in table 5

1154
J. Geophys. Eng. 15 (2018) 1146 M Anemangely et al

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024
Figure 9. Cross plots of denoised actual data and estimated ROP using MLP-COA ((a) training phase, and (b) testing phase), and MLP-PSO
((c) training phase, and (d) testing phase).

Table 6. RMSE values and coefficients of determination obtained by Figure 8 shows the decreasing trend in error rate in the
estimating the ROP using the MLP-COA and MLP-PSO algorithms. training phases of the MLP-COA and MLP-PSO models with
Method Dataset RMSE r-square five input parameters. As is evident in the figure, compared to
the MLP-PSO model, the MLP-COA model provided a sig-
MLP-COA Train 0.060 0.939 nificantly higher rate of error reduction. Moreover, the gen-
Test 0.064 0.921 erated error by the MLP-COA error following 40 iterations
MLP-PSO Train 0.084 0.875
was much lower than that produced by the MLP-PSO model
Test 0.116 0.769
following as many as 200 iterations, indicating that the MLP-
PSO model was trapped in a local minimum. At final itera-
tions of both models, the slope of changes in error was seen to
are plotted in figure 7. In this study, since 11 parameters be nearly zero, indicating the proper selection of the number
were evaluated for ROP estimation, the applied feature of iterations for the models.
selection method chose nine parameters in a maximum Figure 9 shows cross plots of actual denoised and estimated
manner. As can be seen in this figure, with an increase in the values of ROP in training and testing phases for both models.
number of inputs, the error curve tends to decrease at a As is evident from this figure, the output results of MLP-COA
lower slope, so that changes in error will be negligible once were way better than those of the other model. The fitted line to
the number of input parameters into the model exceeds five. output data in the testing phase was close to the line defined as
As such, a model with five input parameters (rotary speed, Y=T; this shows that the model tends to produce lower
WOB, shear and compressional wave slowness values, amounts of error in the testing phase, indicating its good gen-
and flow rate) was selected to estimate the ROP in the eralizability. Table 6 reports the coefficients of determination
next stage. and RMSE values obtained in the training and testing phases of

1155
J. Geophys. Eng. 15 (2018) 1146 M Anemangely et al

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024
Figure 10. A comparison between results of the MLP-COA model trained with raw and denoised data against raw and denoised actual ROP.

Table 7. Selected variables and obtained coefficients for each variable using the MLR method.

Coefficient confidence intervals

Added variables Estimated coefficient Lower Upper Standard error t-statistics p-value
Intercept −17.637 −19.7713 −15.5026 1.0868 −16.228 1.5617e-49
X1 −0.3834 −0.6601 −0.1069 0.1408 −2.723 0.0067
X2 0.3987 0.2399 0.5575 0.0809 4.93 1.0599e-06
X3 −0.4478 −0.5388 −0.3568 0.0463 −9.6684 1.1234e-20
X4 0.0250 0.0199 0.0301 0.0026 9.6284 1.5765e-20
X5 0.6701 0.5604 0.7798 0.0559 11.997 6.196e-30
X1 X2 −0.0003 −0.0005 −0.0002 8.6877e-5 −3.6925 0.0002
X1 X5 0.0187 0.0061 0.0133 0.0064 2.9188 0.0036
X2 X5 −0.0169 −0.0241 −0.0098 0.0036 −4.6571 3.934e-06
X3 X 4 −0.0006 −0.0009 −0.0003 0.0002 −3.8022 0.0002
X3 X5 0.0238 0.0178 0.0297 0.0030 7.8626 1.7007e-14
X 4 X5 −0.0006 −0.0008 −0.0003 0.0001 −4.7264 2.8366e-06

Table 8. Results of implementing MLR onto selected features. Figure 10 presents a comparison between the MLP-
Adjusted COA model trained using five raw and denoised input
RMSE r-square r-square p-value parameters, and actual and denoised ROPs. As can be seen
in this figure, the model trained using denoised data
MLR model 0.105 0.804 0.801 7.22e-209
exhibited very good performance in terms of ROP predic-
tion. The results of the model trained with preprocessed
data could model well the trend of changes in ROP (noisy
both models. The difference between error values in the training and denoised), while the model trained with raw data
and testing phases of the MLP-COA model was indicative of showed large fluctuations within particular depth intervals.
the generalizability of and lack of overfitting in this model. No The RMSE of the training (0.1233) and testing (0.1319)
significant difference was found in the results between the two phases of the model trained with raw data further supports
models used in the training and testing phases, indicating the this conclusion. The difference in RMSE between the
superiority of the MLP-COA model. training and testing phases of this model increased the
A comparison between the results of MLP-COA and probability of the trained model been over-trained. This is
MLP-PSO models against those of the MLP neural network while the small difference in RMSE between training and
with LM as the training algorithm shows that, using meta- testing phases of the MLP-COA model trained with
heuristic optimization algorithms, one can significantly denoised data was indicative of high reliability of the
enhance neural networks in terms of accuracy. model.

1156
J. Geophys. Eng. 15 (2018) 1146 M Anemangely et al

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024
Figure 11. A comparison of the ROP estimation results among the MLP-COA, MLP-PSO, and MLR models along with actual denoised ROP values.

4. A comparison between MLP-COA, MLP-PSO, and figure and shown upon a comparison between tables 6 and 8,
stepwise regression MLP-COA and MLP-PSO were more accurate than MLR.
Meanwhile, MLP-COA provided better accuracy than those
Given the wide range of applications for regression methods in of the other two models.
previous research works, the MLR method was used to validate
the proposed MLP-COA and MLP-PSO models in this section.
For this purpose, the five parameters selected in the feature 5. Conclusion
selection phase were taken as input parameters into the
regression model. In model specification, the model type was In the present research, two hybrid methods, namely MLP-
set to interaction mode where the model included intercept, COA and MLP-PSO, were used to predict ROP in a vertical
linear terms and all products of pairs of distinct predictors well drilled into the Karanj Oilfield in southwestern Iran. For
(without squared terms). Since most of the models proposed to this purpose, drilling data and petrophysical logs at the well
predict the ROP are exponential, the model was considered to under study were collected.
be exponential (equation (8)). Five different criteria were used Implementation of an SG filter to reduce the effect of
to select input parameters: a change in the sum of squares of noise on data not only shortened the training time, but also
errors, an increase in the coefficient of determination, an provided the model with significantly higher accuracy. In the
increase in the adjusted coefficient of determination, a change next stage, the plus-l-take-r algorithm was applied to the
in the Akaike information (AIC), and a change in the Bayesian denoised data for feature selection. Results of implementing
information (BIC). Comparing the results based on these five this model illustrated that, with an increase in the number of
criteria, it was indicated that AIC was better than the other features in the model, the model tends to exhibit lower levels
criteria as far as simplicity and accuracy of the estimator model of error; however, the decrease in error was negligible in
were concerned. models with five or more input parameters. As such, the
following five parameters were selected to model the ROP
ROP = b e Input parameters. (8 ) using the MLP-COA and MLP-PSO hybrid models: rotary
speed, WOB, shear wave slowness, compressional wave
Table 7 shows the selected terms by applying this slowness, and flow rate.
method to the five input parameters. The obtained estimator The results of applying the two hybrid models, namely
model by applying this method is expressed in equation (9). MLP-COA and MLP-PSO, on preprocessed data were indi-
Table 8 reports the general results of implementing cative of the superior performance of these models over the
the model on data of the well under study. The obtained neural network trained using the LM training algorithm. A
p-value for this model showed that the model is very comparison of the results of these hybrid models in the
reliable. training phase showed that MLP-COA was of higher accur-
acy. Even though the model took longer to be trained, it
ROP = e-17.637e-0.3834x1 e0.3987x 2 e-0.4478x3 e0.025x 4
converged faster and gave the best solution within a lower
´ e0.6701x5 e-0.0003x1x 2 e0.0187x1x5 number of iterations as compared to MLP-PSO. Comparing
e-0.0169x 2 x5 e-0.0006x3 x 4 e0.0238x3 x5 e-0.0006x 4 x5 . (9 ) RMSE values of the two models in the training phase, the
MLP-PSO model was found to become trapped in local
Figure 11 demonstrates a comparison among the results minimum. The negligible difference in RMSE values between
of MLR, MLP-COA, and MLP-PSO models along the training and testing phases of the MLP-COA, rather than the
interval under investigation at the well. As is evident in this MLP-PSO model, indicated the high reliability of the MLP-

1157
J. Geophys. Eng. 15 (2018) 1146 M Anemangely et al

COA model, so that one can use it to predict the ROP in other Bourgoyne A T Jr and Young F S Jr 1974 A multiple regression
wells drilled into the considered field. A further comparison approach to optimal drilling and abnormal pressure detection
between results of the hybrid models and MLR method was Soc. Pet. Eng. J. 14 371–84
Cardiff M and Kitanidis P K 2010 Fitting data under omnidirectional
indicative of the superiority of the hybrid models when it noise: a probabilistic method for inferring petrophysical and
came to ROP prediction. It can be confidently stipulated that hydrologic relations Math. Geosci. 42 877–909
the proposed method can predict ROP at high accuracy in Coello C A C, Lamont G B and Van Veldhuizen D A 2007
future wells using the information provided from LWD tools Evolutionary Algorithms for Solving Multi-Objective Problems
or other techniques. vol 5 (New York: Springer)
Duan J, Zhao J, Xiao L, Yang C and Li C 2015 A ROP optimization

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024
approach based on improved BP neural network PSO Adv.
Swarm Comput. Intell. 9142 185–93
Garćia L P F, de Carvalho A C and Lorena A C 2013 Noisy data set
Acknowledgments identification Int. Conf. on Hybrid Artificial Intelligence
Systems (Springer) pp 629–38
National Iranian South Oil Company (NISOC) is gratefully Hagan M T and Menhaj M B 1994 Training feedforward networks
acknowledged for providing useful data and valuable support. with the Marquardt algorithm IEEE Trans. Neural Networks 5
989–93
Hamrick T R 2011 Optimization of operating parameters for
minimum mechanical specific energy in drilling PhD Thesis
References West Virginia University
Haykin S 1999 Neural network: a comprehensive foundation (Tom
Robbins)
Abtahi A 2011 Bit Wear Analysis and Optimization for Vibration Horne B G 1993 Progress in supervised neural networks Signal
Assisted Rotary Drilling (VARD) using Impregnated Diamond Process Mag. IEEE 10 8–39
Bits Masters Abstracts Int. vol 50 (Japan Association for Huang G-B, Wang D H and Lan Y 2011 Extreme learning machines:
Earthquake Engineering) a survey Int. J. Mach. Learn. Cybern. 2 107–22
Al Bulushi N, King P R, Blunt M J and Kraaijveld M 2012 Artificial Kawam A A L and Mansour N 2012 Metaheuristic optimization
neural networks workflow and its application in the petroleum algorithms for training artificial neural networks Int. J.
industry Neural Comput. Appl. 21 409–21 Comput. Inform. Technol 1 156–61
Ali J K 1994 Neural Networks: a New Tool for the Petroleum Li W, Zhao X, Li Y, Ji Y, Peng H, Liu L and Yang Q 2015
industry? European Petroleum Computer Conf. (Society of Laboratory investigations on the effects of surfactants on rate
Petroleum Engineers) (https://doi.org/10.2118/27561-MS) of penetration in rotary diamond drilling J. Pet. Sci. Eng. 134
Amar K and Ibrahim A 2012 Rate of penetration prediction and 114–22
optimization using advances in artificial neural networks, a Lorena A C and de Carvalho A C 2004 Evaluation of noise reduction
comparative study Proc. 4th Int. Joint Conf. on Computer techniques in the splice junction recognition problem Genet.
Intelligence pp 647–52 Mol. Biol. 27 665–72
Anemangely M, Ramezanzadeh A and Tokhmechi B 2017a Ma T, Chen P and Zhao J 2016 Overview on vertical and directional
Determination of constant coefficients of Bourgoyne and drilling technologies for the exploration and exploitation of
Young drilling rate model using a novel evolutionary deep petroleum resources Geomech. Geophys. Geo-Energy
algorithm J. Min. Environ. 8 693–702 Geo-Resources 2 365–95
Anemangely M, Ramezanzadeh A and Tokhmechi B 2017b Shear Maimon O and Rokach L 2005 Data Mining and Knowledge
wave travel time estimation from petrophysical logs using Discovery Handbook vol 2 (Berlin: Springer)
ANFIS-PSO algorithm: a case study from Ab-Teymour Mendes J R P, Fonseca T C and Serapião A 2007 Applying a genetic
Oilfield J. Nat. Gas Sci. Eng. 38 373–87 neuro-model reference adaptive controller in drilling
Ansari H R, Hosseini M J S and Amirpour M 2017 Drilling rate of optimization World Oil 228 29–36
penetration prediction through committee support vector Moraveji M K and Naderi M 2016 Drilling rate of penetration
regression based on imperialist competitive algorithm prediction and optimization using response surface
Carbonates and Evaporites 32 205–13 methodology and bat algorithm J. Nat. Gas Sci. Eng. 31
Arabjamaloei R and Shadizadeh S 2011 Modeling and optimizing rate 829–41
of penetration using intelligent systems in an iranian southern oil Paiaman A, Al-Askari M and Salmani B 2009 Effect of Drilling
field (Ahwaz oil field) Pet. Sci. Technol. 29 1637–48 Fluid Properties on Rate of Penetration Nafta pp 129–34
Bahari A and Baradaran Seyed A 2009 Drilling cost optimization in Pedersen M E H and Chipperfield A J 2010 Simplifying particle
a hydrocarbon field by combination of comparative and swarm optimization Appl. Soft Comput. 10 618–28
mathematical methods Pet. Sci. 6 451–63 Quinlan J R 1986 The effect of noise on concept learning Mach.
Bahari A and Baradaran Seyed A 2007 Trust-region approach to find Learn. An Artif. Intell. Approach 2 149–66
constants of Bourgoyne and Young penetration rate model in Rahimzadeh H, Mostofi M and Hashemi A 2011 A new method for
Khangiran Iranian gas field Latin American & Caribbean determining Bourgoyne and Young penetration rate model
Petroleum Engineering Conf. (Society of Petroleum constants Pet. Sci. Technol. 29 886–97
Engineers) (https://doi.org/10.2118/107520-MS) Rajabioun R 2011 Cuckoo optimization algorithm Appl. Soft
Bezminabadi S N, Ramezanzadeh A, Jalali S-M E, Comput. 11 5508–18
Tokhmechi B and Roustaei A 2017 Effect of rock properties Raji W O, Gao Y and Harris J M 2017 Wavefield analysis of
on rop modeling using statistical and intelligent methods: a crosswell seismic data Arab. J. Geosci. 10 217
case study of an oil well in Southwest of Iran Arch. Min. Sci. Savitzky A and Golay M J E 1964 Smoothing and differentiation of
62 131–44 data by simplified least squares procedures Anal. Chem. 36
Bishop C 2007 Pattern Recognition and Machine Learning 1627–39
(Information Science and Statistics) 2nd edn (New York: Sultan M A and Al-Kaabi A U 2002 Application of neural network
Springer) to the determination of well-test interpretation model for

1158
J. Geophys. Eng. 15 (2018) 1146 M Anemangely et al

horizontal wells SPE Asia Pacific Oil and Gas Conf. and Yang X-S and Deb S 2009 Cuckoo search via Lévy flights Nature &
Exhibition (Society of Petroleum Engineers) (https://doi.org/ Biologically Inspired Computing, 2009. NaBIC 2009. World
10.2118/77878-MS) Congress on (IEEE) pp 210–4
Tian-Shou M and Ping C 2015 Development and use of a downhole Zhang J, Li J, Hu Y and Zhou J Y 2012 The identification method of
system for measuring drilling engineering parameters Chem. igneous rock lithology based on data mining technology
Technol. Fuels Oils 51 294–307 Advanced Materials Research vol 466 (Trans Tech
Wang R Y, Storey V C and Firth C P 1995 A framework for analysis Publication) pp 65–9
of data quality research IEEE Trans. Knowl. Data Eng. 7 Zhou H-W 2014 Practical Seismic Data Analysis (Cambridge:
623–40 Cambridge University Press)

Downloaded from https://academic.oup.com/jge/article/15/4/1146/5204266 by National Institute of Technology Rourkela user on 05 March 2024

1159