Sunteți pe pagina 1din 20

Journal of Petroleum Science and Engineering 175 (2019) 604–623

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering


journal homepage: www.elsevier.com/locate/petrol

Application of hybrid artificial neural networks for predicting rate of T


penetration (ROP): A case study from Marun oil field
Seyed Babak Ashrafia, Mohammad Anemangelyb, Mohammad Sabaha,
Mohammad Javad Ameria,∗
a
Department of Petroleum Engineering, Amirkabir University of Technology (Tehran Polytechnic), 424 Hafez Avenue, Tehran, 15875-4413, Iran
b
School of Mining, Petroleum and Geophysics Engineering, Shahrood University of Technology, Shahrood, Iran

ARTICLE INFO ABSTRACT

Keywords: Rate of Penetration (ROP) can be considered as a crucial factor in optimization and cost minimization of drilling
Rate of penetration operations. In order to predict ROP with satisfactory precision, some hybrid artificial intelligence models were
Artificial neural network developed in the present research. To achieve this, dataset including petrophysical logs and drilling data was
Feature selection collected from a vertical well drilled in Marun oil field where is in southwest corner of Iran. Then, overall noise
Evolutionary algorithms
of the collected data was reduced with implementing Savitzky–Golay (SG) smoothing filter. In the next stage, a
feature selection method with application of genetic algorithm was employed for choosing the unrivaled features
and lessening the input vectors. Regarding results of the feature selection method, eight parameters including
weight on bit, bit rotational speed, pump flow rate, pump pressure, pore pressure, gamma ray, density log and
shear wave velocity were identified as the most influential parameters in drilling rate. After that, eight hybrid
artificial neural networks (ANN) were developed and trained by four evolutionary algorithms. These algorithms
are genetic algorithm (GA), particle swarm optimization (PSO), biogeography-based optimizer (BBO) and im-
perialist competitive algorithm (ICA). After model developments, to provide an objective assessment of per-
formances of the proposed hybrid models, their results were compared with results of conventional ANN models
and two multiple regression methods (NLMR and LMR) using different performance indices such as root mean
square error (RMSE), variance account for (VAF), and performance index (PI). Results showed that PSO-multi-
layer perception (PSO-MLP) and PSO-radial basis function (PSO-RBF) neural networks with RMSE of 1.12 and
1.4, respectively yielded the highest performance in prediction comparing to results of other developed models.
Concluding remark is that hybrid ANNs exhibit much more efficiency and reliability than conventional ANNs
and regression methods in terms of ROP prediction, indicating superiority of evolutionary algorithms over back
propagation (BP) methods in the ANNs training.

1. Introduction 2010).
Over the past few decades, many researchers tried to formulate
Rate of penetration (ROP) is a direct measurement of drilling job these parameters and develop some models (Warren, 1987; Winters
progression as it shows how much time it takes to drill a finite interval et al., 1987; BINGHAM, 1965; Hareland and Rampersad, 1994;
of depth. To monitor and enhance drilling operations, drilling engineers Motahhari et al., 2010). These physics-based models are associated with
must consider ROP which itself relies upon many other independent some limitations such as the requirement for auxiliary data (bit prop-
parameters such as geological and geotechnical parameters, operational erties, mud properties, bit design, etc.), utilizing empirical coefficients
factors, and the machine characteristics. However, there is not a and their high dependency to lithology (since the empirical coefficients
transparent relationship between ROP and mentioned parameters, heavily rely on lithology) (Hegde et al., 2017, 2018; Mendes et al.,
rendering the behavior of ROP to be non-linear and difficult to predict 2007), which make them useless for prediction of drilling rate. There-
(Akgun, 2007; Darbor et al., 2017). Nevertheless, impact of different fore, other methods should be considered to approximate the ROP with
parameters on drilling rate can be modeled in various correlations such least possible errors. Novel solutions, in particular, like application of
as non-linear, complex and stochastic relationships (Moradi et al., smart systems, data mining methods, evolutionary algorithms and their


Corresponding author.
E-mail address: Ameri@aut.ac.ir (M.J. Ameri).

https://doi.org/10.1016/j.petrol.2018.12.013
Received 31 July 2018; Received in revised form 24 November 2018; Accepted 5 December 2018
Available online 07 December 2018
0920-4105/ © 2018 Elsevier B.V. All rights reserved.
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

hybrids models have led into a robust performance in function ap- et al., 2017), and enhancing oil recovery and production (Le Van and
proximation. Also, a broad range of uncertainty can be endured by Chon, 2017; Xue et al., 2014; Shan et al., 2018) are examples of ANN
these approaches; Due to this fact, they have recently become highly implementations.
popular in various research fields (Yilmaz and Kaynar, 2011). As it is obvious from the literature, many researchers utilized the
A lot of studies have been conducted in the drillling industry using neural networks in ROP estimation. However, several scholars (Lee
artificial intelligence methods. Jacintoa et al. (2013) (Jacintoa et al., et al., 1991; Wang et al., 2004; Armaghani et al., 2017) reported that
2013) employed dynamic evolving neural-fuzzy inference system using back propagation (BP) algorithms to tune parameters of neural
(DENFIS) and Bayesian inference approach to predict and optimize the networks has some disadvantages such as slow learning rate and trap-
ROP. Bataee et al. (2014) (Bataee et al., 2014) used artificial neural ping in local minima. Utilizing evolutionary algorithms for training
networks and analytical methods for predicting the drilling rate and neural networks instead of conventional (like BP) algorithms can pos-
optimizing the parameters affecting ROP. Basarir et al. (2014) (Basarir sibly have many effectiveness of enhancing prediction performance in
et al., 2014) investigated the performance of linear regression, non- its favor.
linear regression and ANFIS model in terms of ROP prediction. Their The objective of this study is to develop some hybrid neural net-
results showed superiority of ANFIS method over regression methods. works that are trained by four different evolutionary algorithms in-
Bodaghi et al. (2015) (Bodaghi et al., 2015) established a relationship cluding particle swarm optimization (PSO), genetic algorithm (GA),
between drilling variables and rate of penetration employing support imperialism competitive algorithm (ICA) and Biogeography-based op-
vector regression (SVR) which was optimized by cuckoo search algo- timizer (BBO) for predicting drilling rate. Eventually, the performance
rithm and genetic algorithm. Hegde et al. (2015) (Hegde et al., 2015) of developed hybrid models are compared with NNs trained by BP al-
predicted the rate of penetration by application of statistical learning gorithm to give an objective assessment of evolutionary algorithms’
techniques such as trees, bagged trees and random forests which were performance in tuning neural networks parameters. Also, the perfor-
used for a data set with nine predictors. Moraveji and Naderi (2016) mance of two different ANNs were compared.
(Moraveji and Naderi, 2016) evaluated the impact of drilling para- The workflow of this study is demonstrated in Fig. 1 which involves
meters on ROP using response surface methodology and established a 5 main stages of data collection and up-scaling, noise reduction, feature
mathematical relationship between drilling rate and those parameters selection, developing proposed models and evaluation of the results.
with the greatest influence on ROP. Kahraman (2016) (Kahraman,
2016) found that neural network models are more reliable than the
regression models for predicting the drilling penetration rate. Jiang and 2. Data collection
Samuel (2016) (Jiang and Samuel, 2016) trained a neural network with
application of ant colony optimization (ACO) algorithm eventuating the In the present research, collected data of a vertical well drilled into
optimum values for drilling rate. Shi et al. (2016) (Shi et al., 2016) Marun oil field in southwest of Iran was used. Fig. 2 shows some in-
utilized a number of smart systems such as extreme learning machine, formation about location and stratigraphy of the area. The database
upper-layer solution aware model and neural network. Their outcomes consists of 1000 data points which are collected from Asmary formation
showed that all of these approaches are adequately proper techniques (one of primary reservoirs of Marun oil field), and covers 500 m of
to predict the ROP. Khandelwal and Armaghani et al. (2016) drilling depth (3000 m–3500 m). The available data are comprised of
(Khandelwal and Armaghani, 2016) used multiple regression, neural two different sources, mud logging and well logging (petrophysical)
network (NN) and hybrid genetic algorithm (GA-NN) models to esti- data. Mud logging information encompasses weight on bit (WOB), bit
mate drilling rate index (DRI) based on rock's physical properties. Their rotational speed (BRS), bit flow rate (BFR), pump pressure (PP) and
results showed that GA-NN technique yields better performance in mud weight (MW) while petrophysical information includes compres-
prediction compared to other implemented models. Bezminabadi et al. sional wave velocity (Tp ), shear wave velocity (Ts ), density log, gamma
(2017) (Bezminabadi et al., 2017) have estimated ROP employing ray (GR), neutron porosity (NP) logs and formation pore pressure (Pp ).
multiple nonlinear regression (MNR) and artificial neural network. In It should be noted that two parameters of bit type (roller cone bit) and
their study, physical and mechanical properties of rocks have been in- formation type (Asmary) were constant in the selected range of depth
volved with operational parameters to improve performance of pre- (3000 m–3500 m), so these parameters have not been considered as
dicting models. Ansari et al. (2017) (Ansari et al., 2017) examined ROP input vectors in developing the relevant models.
modeling utilizing committee support vector regression based on im- On the other hand, due to collection of data from the two different
perialist competitive algorithm (CSVR-ICA). Eskandarian et al. (2017) sources, they were recorded at disparate sampling rate (for petrophy-
(Eskandarian et al., 2017) applied data mining methods including sical logs measurements taken at 15.24 cm intervals, and for mud log-
random forest and monotone multi-layer perceptron (MON-MLP) to ging data measurements taken at 50 cm intervals). Hence, it was ne-
estimate the ROP. They implemented “fscaret” package (Szlek and cessary to implement an up-scaling method to contend with this
Mendyk, 2015) for feature ranking and determining the most influen- difficulty. In this approach, petrophysical data points were averaged
cing variables on ROP. Hegde et al. (2017) (10) used physics-based and within each 50 cm depth interval. This averaging method made it
data-driven models in ROP prediction which ended up with the con- possible to present the well logging and mud logging data in a single
clusion that data-driven models showed better fitness compared to dataset with a same depth interval of 50 cm.
traditional models. In a study conducted by Anemangely et al. (2018) Table 1 gives a list of statistical information about eleven in-
(Anemangely et al., 2018), two hybrid ANNs were developed by using dependent parameters and the dependent variable (ROP). Range of
particle swarm optimization (PSO) and cuckoo optimization algorithm changes, unit and average values are of some characteristics that pre-
(COA) as training functions to provide accurate prediction of ROP. sented in this Table. Furthermore, Fig. 3 Shows the relation (based on
Yavari et al. (2018) (Yavari et al., 2018) investigated superiority of regression coefficient) between independent variables and the depen-
different smart systems such as ANFIS and ANN over recently estab- dent variable (i.e. drilling rate), and the related p-values. Since the
lished analytical ROP models. Their results showed that developed calculated p-values do not exceed 0.05, it is concluded that effect of
smart systems are shown much more reliable results than analytical proposed input parameters on drilling rate is significant. However, p-
models for drilling rate prediction. Regarding ANNs, this methods have values is not a sufficient criterion for selecting most influential para-
been used in a variety of different sections in the oil industry. Divisions meters in ROP prediction, but it could be beneficial to use this value for
like characterizing reservoir-related properties (Vaferi et al., 2014, detecting and removing parameters with negligible effect on ROP
2015, 2016; Ghaffarian et al., 2014), defining lithological properties which are associated with p-value of more than 0.05.
and analyzing source rock characteristics (Martin et al., 2017; Bolandi

605
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Fig. 1. Schematic diagram of workflow applied to the ROP-prediction models developed (WOB = weight on bit; BRS = Bit Rotation Speed; BFR =Bit Flow Rate; PP
= Pump Pressure; MW = Mud Weight; Tp = compressional sonic velocity; Ts = shear sonic velocity; GR = Gamma Ray; NP = Neutron Porosity; Pp = Pore
Pressure).

2.1. Noise reduction 1964; Baba et al., 2014). In addition, signal's original properties will
maintain after smoothing process using this filter. In this approach, the
Noise of measurements is a common factor that affect most of raw effect of noise on data is mitigated by a polynomial function (i.e. initial
data (Wang et al., 1995). Scientists in statistical and machine learning values are replaced with much more smoothed values). SG polynomial
fields spot this as a phenomenon with different detrimental effects. A function of order n is fitted to a number of points which sets with an
large number of them are agree that learning process can be affected by interval founded upon the least squares error. Number of the points
data which are contaminated by noise. Furthermore, noisy data can within this interval should be odd and less than the order of the poly-
affect learning process and increase learning time (Quinlan, 1986). nomial. An increase in the order of polynomial or a decrease in the
They also assert that noisy data can upset the balance between para- number of points within the interval can further retain data structure
meters and eventually lead to incorrect rules in machine learning and make the extent to which data is smoothed lower. On the other
models as well as reducing overall function-approximation performance hand, the opposite brings about additional smoothing by wiping out a
(Garcia et al., 2015; Lorena and de Carvalho, 2004). What is more, number of the data. Hence, determination of proper polynomial order
values of environmental properties are often affected by noisy mea- and number of points within the interval is of paramount significance.
surements acknowledged by some geophysicists (Cardiff and Kitanidis, In order to obtain optimum values of these two parameters, polynomial
2010). Exploring source of the noise is of critical importance in dealing order and number of points within the interval, daily reports of the
with this incident. Some frequent cases can be said such as drilling tools studied well and geological setting of the field were evaluated. Ac-
replacement, drill-string vibrations, washouts, work-overs, and also cording to the results, optimal order of polynomial and number of
shifting and changes in geological layers which can impose errors to the points within the interval for the drilling data were found to be 3 and
acquired data. 11, respectively. Same analysis on petro-physical logs showed the
There are various methods such as low-pass filter, moving average, amount of these parameters would be 4 and 11, respectively. Fig. 4
wavelet transform, and Savitzky–Golay (SG) smoothing filter in order to demonstrates a comparison between the data recorded of the well (in
reduce the noise from data (Vaferi and Eslamloueyan, 2015; Rudin red) along with denoised data (in black) using SG filter. According to
et al., 1992; Savitzky and Golay, 1964). In this research, noise reduction the figure, trend of changes in the denoised data is in good agreement
was performed using Savitzky–Golay (SG) smoothing filter. The SG with that of recorded data, i.e. noise effect has been minimized.
technique is the common smoothing method that was widely used in
earth science for denoising petrophysical logs (Savitzky and Golay,

606
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Fig. 2. Schematic map showing the location of the Marun oil field and the stratigraphy of the study area. The Marun oil field is hosted by the Asmary Formation.

Table 1
Statistical information of input and output parameters.
Coded factor Parameter Unit Minimum Maximum Average Standard deviation

F1 Neutron Porosity (NP) – 0.3 0.43 0.35 0.022


F2 Density kg / m3 2212 2424 2375 0.026
F3 Shear Wave Velocity (Ts ) us /ft 226.31 319.82 261.28 17.710
F4 Compressional Wave Velocity (Tp ) us /ft 100.48 121.45 108.3 4.105
F5 Gamma Ray (GR) GAPI 70 117 102.48 5.984
F6 Weight on Bit (WOB) Kgf 612.25 10086.53 4879.47 4.982
F7 Bit Rotational Speed (BRS) RPM 117.29 143.69 135.58 4.597
F8 Pump Pressure (PP) MPa 23.56 26.41 25.38 56.47
F9 Bit Flow Rate (BFR) m3/s 0.0544 0.0581 0.0572 5.4374
F10 Mud Weight (MW) kg / m3 1680.0 1720.0 1696.0 0.0196
F11 Pore Pressure (Pp ) kg / m3 1195 1598 1476 88.6
F12 Rate of Penetration (ROP) m/hr 2.27 35.12 24.86 4.3172

607
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Fig. 3. Pair plot to analyze relationship of data.

3. Feature selection influential features (Anemangely et al., 2017). Conversely, wrapped


methods which employ sequential or heuristic search algorithms
Feature selection method is a problem of picking out a subset of the achieve more robust solutions, particularly in interacting with a large
most influential d parameters from a total set of D features according to number of features (Chandrashekar and Sahin, 2014; Vafaie and Imam,
results of an optimization criterion. Shortening the training time, ease 1994). Accordingly, in this study, genetic algorithm (GA), which is a
of model interpretation and reduction in over-fitting are of the most heuristic search algorithm, was implemented to determine indis-
important benefits of applying feature selection method (James et al., pensable inputs for ROP estimation.
2013). Regarding feature selection using GA, firstly, 12 initial populations
There are several methods for applying feature selection and containing a set of feasible solutions are generated. The generated po-
choosing the most influential independent parameters. Some of these pulations include 11 members (chromosomes), each of which is a so-
approaches are filter methods, embedded methods, unsupervised lution of the desired problem. Actually, every member is indicative of
learning techniques, sequential and heuristic search algorithms different combinations of input vectors that are defined as encoded
(Chandrashekar and Sahin, 2014; Law et al., 2004; Davoudi and Vaferi, values. The fitness of these members (combinations) are determined
2018; Gholami et al., 2018). Since the difference between values of using a cost function (objective function). A MLP-NN with two hidden
prediction and observed data should be as small as possible, residual layers was implemented as the cost function of genetic algorithm and
error is the only important factor in selecting parameters. So, feature optimum number of neurons for each hidden layer was determined
selection methods that are based on correlation coefficient, which is a through a trial and error approach. In other words, different archi-
subgroup of filter methods, are incompetent in selecting the most tectures of MLP-NN were developed using all of the available features

608
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Fig. 4. Recorded and denoised data of the well under study.

and their performance were evaluated based on MSE obtained through of MLP-NN composed of two hidden layers with 6 and 5 neurons in first
10-folds cross validation method. and second layer, respectively. In the next step, the members were
In 10-folds cross validation method, as it is evident in Fig. 5, the ranked depending on MSE values resulted from predicted and measured
entire data is split into 10 non-overlapped segments. One of these drilling rate. Again, the 10-fold cross validation method was also im-
segments is chosen as the testing data, and the remaining 9 folds are all plemented to determine the MSE value of each member and avoid oc-
considered as the training set. Then, the neural network is trained using currence of over fitting in MLP_NN training. After that, the most sui-
all the training data for 100 times and the lowest error is recorded. This table members (combinations of inputs) were subjected to specific
procedure is performed for 10 times, and each time one of the folds is genetic algorithm operators namely crossover, reproduction and mu-
considered as the testing set while the rest of folds are acting as the tation. These operators were performed to create new populations and
training set. Eventually, MSE values for all 10 testing sets are averaged guide the search toward the most suitable combinations of input vari-
and reported as the performance related to MLP-NN structure. ables.
According to results (Fig. 6), it is found that the best configuration It is important to note that all of the process described above can be

609
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Fig. 5. Schematic diagram of the 10-fold cross validation method.

4. Methods

4.1. Multi-layer perceptron neural network (MLP-NN)

Multi-layer perceptron neural network as one of the most common


and practical type of ANN is a branch of artificial intelligence (AI)
(Moghadassi et al., 2009; Gardner and Dorling, 1998). This network can
satisfactory estimate values for non-linear correlations (Esfe et al.,
2015a). The structure of a MLP-NN comprises three main sections, an
input layer, hidden layer(s), and an output layer. A relationship is es-
tablished between input and output parameters through hidden layer(s)
(Lashkarbolooki et al., 2012). Although number of neurons in input and
output layers are exactly the same as the number of input and output
parameters, respectively, number of hidden layers and also number of
neurons in each hidden layer can be chosen optimally (Ramchoun et al.,
2016). In each neuron, inputs are multiplied with a weight and summed
Fig. 6. Number of neurons in the two hidden layers of the developed MLP with
with a value called bias. Then, their summation are subjected to a
MSE values for feature selection.
transformation through a function called transfer function (activation
function). Outputs of each layer are taken as inputs for the next layer
Table 2 (Esfe et al., 2015b; Heidari et al., 2016). The output of MLP neural
Applying feature selection on different number of input vectors and associated network can be explained as in Eq. (1):
RMSE values.
Nk 1
Number of inputs Selected inputs MSE
Yjk = Fk Wijk Yi (k 1) + Bik
1 F8 19.4481 i=1 (1)
2 F11, F6 9.6154
3 F11, F6 , F7 6.6564 Where Yjk and Bik are the neuron j's output from k's layer and bias
4 F11, F9 , F6 , F8 5.3824
for neuron j in layer k, respectively. Wijk represents the weight that is
5 F11, F9 , F8 , F6 , F7 3.6123
6 F7 , F6 , F9 , F11, F2 , F8 3.2761 chosen on a random basis in the initial stage of training procedure. Fk is
7 F2 , F8 , F7 , F11, F6 , F9 , F3 3.1329 the transfer function which may be considered in many different forms
8 F6 , F5 , F9 , F11, F8 , F3 , F2 , F7 2.8951 such as identity function, binary step function, binary sigmoid, bipolar
9 F5 , F3 , F2 , F1, F9 , F7 , F6 , F8 , F11 2.8589 sigmoid, Gaussian, and linear functions (Fausett, 1994). Back-propa-
10 F8 , F2 , F1, F7 , F3 , F5 , F4 , F11 , F9 , F6 2.8488
gation (BP) algorithms are conventionally used for training MLP-NNs.
11 F1, F5 , F7 , F2 , F6 , F11, F4 , F3 , F9 , F10 , F8 2.8472
Some of these BP algorithms are Scaled Conjugate Gradient (SCG),
Levenberg–Marquardt (LM), Gradient Descent (GD) and Resilient back
performed to detect the best combination of features for different propagation (RP) (Ghoreishi and Heidari, 2013; Moghadassi et al.,
number of inputs. Therefore, the feature selection was conducted for 11 2011; Yetilmezsoy et al., 2011).
rounds with different number of inputs (from 1 input to 11 inputs)
which provides more understanding about the effect of number of in- 4.2. Radial basis function neural network (RBF-NN)
puts on overall MSE. Table 2 presents results of applying feature se-
lection for different number of inputs and their best combinations. In RBF-NNs began to function as a remedy in the late 1980s.
order to give a visual perception, associated errors are plotted in Fig. 7. Practicability of this type of neural network in dealing with arbitrarily
Regarding this figure, by increasing the number of input variables, the scattered data, easily generalizing to several dimensional space, and
curve of error tend to have a less-marked slope. Different stages with providing spectral accuracy have made it a particularly appropriate
regard to mentioned feature ranking method are delicately shown as a alternative for MLP-NNs (Elsharkawy, 1998; Lashkenari et al., 2013).
flowchart in Fig. 8. Other advantages of this neural network are the ability of accurately
In conclusion, changes in errors would be negligible and only lead distinguishing patterns for non-linear models and the feasibility of
to an increase in calculation time once the number of input parameters being trained in a single direct procedure which is unlike the process in
exceeds eight. Hence, the model with the 8 input parameters was se- MLP-NN (Venkatesan and Anitha, 2006). Structure of RBF-NN is highly
lected for ROP approximation in subsequent model development. similar to MLP-NN. Fundamentally, RBF-NNs has only one hidden layer
which consists of several nodes called RBF units.
Every RBF_NN include two tuning parameters representing location

610
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Fig. 7. Flow chart of feature ranking method.

of basis function's center and its width (spread). Distance between these Where, wT is the transposed weight, N is the number of neurons in
centers and input data are measured in the RBF units. The network hidden layer and (xi ) is the basis function which is typically in
reaches its highest performance when the measured distance becomes Gaussian form:
zero and degrades by increasing the distance. While RBF-NN has only a
(x c ) 2
single hidden layer, two sets of weights are exist, one for connection
(x i ) = e 2s 2 (3)
between hidden and input layers and the other for connection between
hidden and output layers. A non-linear function is used to transform the In Eq. (3), s is spread of Gaussian basis function and c represents the
input area into hidden area. Following this, the output of the network center of RBF unit.
are resulted from the output layer which is always linear (Wu et al.,
2012; Park and Sandberg, 1991). The output of the network can be
4.3. Genetic algorithm (GA)
shown in the following form (Eq. (2)):
Genetic algorithm (GA) as an evolutionary algorithm was firstly
N introduced by John Holland (Goldberg, 1989; Holland, 1992). GA is an
f (x i ) = wT (x i ) inquiry based optimization method inspired by natural selection or
j=1 (2) compatibility. This population based algorithm simulates the existed

611
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Fig. 8. Different number of inputs versus corresponding MSE values.

process in a natural system where only the strongest participant can asrand (0,1) . c1 and c2 are the speeds that regulate the length when going
survive. GA produces an initial population containing possible solutions towards the most optimal particles of the swarm, and w is the weight.
and recombines them in a way to steer the search toward high potential Higher weight value means bigger step size, resulting in more ex-
areas of desirable answer. Every of those possible solution is encoded as ploration behavior. If the weight value is low, the step size is small and
a chromosome, and a value called fitness is assigned to them by an exploitation behavior is more dominant. In order to have a high ex-
objective function (cost function). The high and low fitness values are ploration behavior at the beginning of optimization and the exploita-
representative of the better solution for maximization and minimization tion behavior gradually increase through iterations, a damping ratio is
problems, respectively. Strong members of each population are trans- defined and multiplied by weight value at each iteration.
ferred to a new population matrix undergoing some changes by specific
genetic operators such as crossover, reproduction, and mutation (Ab 4.5. Biogeography-based optimizer
Wahab et al., 2015). Crossover is recombination of two parents that are
elements of one generation which combine to generate new member for Simon (2008) (Simon, 2008) proposed a new evolutionary algo-
next generation. Reproduction selects the most compatible candidates rithm called BBO for the first time. The fundamental inspiration of this
of each generation, and mutation is another operator of GA which is algorithm is got from assessing geographical distribution of biological
analogous to natural biological mutation, altering genes with new un- organisms over time and place. Like other evolutionary algorithms,
expected genes (Grefenstette, 2013). main motivation of BBO is the natural processes. A number of search
agents are employed by BBO, and are called habitats (Wang et al.,
4.4. Particle swarm optimization (PSO) 2014a,b,c). These habitats are functioning the same as chromosomes in
GAs. A vector of habitants representing the variables of the problem is
Particle Swarm Optimization (PSO) is a population based stochastic assigned to each habitant in BBO algorithm (similar to genes in a GA).
procedure initially put forward by Eberhart and Kennedy (1995) Furthermore, overall fitness of a habitat can be determined as term of
(Eberhart and Kennedy, 1995). PSO implements an ordinary me- habitat suitability index (HSI), so the higher the HSI, the more fit the
chanism representing swarm behavior of birds flocking and fish habitat. The fittest habitants are then subjected to some specific BBO
schooling to guide the particles to search the optimal solutions. PSO operators namely immigration, emigration, and mutation. In nature,
was described with separation, alignment and cohesion by Del Valle these concepts lead to a balance between various ecosystems. In other
et al. (2008) (Del Valle et al., 2008). Separation avoid crowding local words, nature has a tendency to provide the general stability of various
flock-mates. While alignment is moving toward the mean direction of geographical areas. The BBO algorithm uses this concept to improve all
local flock-mates, moving toward the mean position is the cohesion. habitats’ HSI, and brings about development of the initial random so-
Three years after first implementation of PSO, inertia weight was in- lutions for a particular problem (Mirjalili et al., 2014).
troduced by Shi and Eberhart (1998) (Shi and Eberhart, 1998). This
term regulate the impact of velocity which controls the exploration and
4.6. Imperialist competitive algorithm
the exploitation behaviors of particle, and can be seen in following
formulas (Eq. (4) and Eq. (5)):
Another evolutionary technique is Imperialist competitive algo-
rithm (ICA) which is based on human being's sociopolitical evolution
vidt+ 1 = w. vidt + c1. rand (0,1). (pidt x idt) + c2. rand (0,1). pgd
t
x idt (Atashpaz-Gargari and Lucas, 2007). ICA like the other algorithms is
based on populations in which it starts with initial population named
(4) countries or colonies. Best countries (imperialists) in optimization are
those with the lowest cost. Therefore, the countries divide into two
x idt+ 1 = x idt + vidt+ 1 (5)
categories of colony and imperialist which together form an empire.
Where in the above equations, and
vidt are particle velocity and
x idt Basic operations in ICA are assimilation, revolution and competition.
particle position respectively, i is the particle index, d is the dimension The imperialist with more power than other ones try to possess the
in the search space, and t is the iteration number. Pi refers to the best colonies of weaker empires and take their control. This process is called
position of particle i and Pg represents the best position of the neighbors imperialist competitive process. In this procedure, the imperialists with
of particle i . The random values between 0 and 1 are showed more power will be aggravated in the power and the vice versa for the

612
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

weak ones. The empires that losses all of their colonies are demolished
and it is assumed that only one empire will remain in an ideal world
after passing a few decades. In other words, as it turned out, the most
powerful empire have all countries which are known as colonies (86).
Another case that makes ICA to stop is when no enhancement occurs in
an empire after a few decades (Ansari et al., 2017).

5. Development of ROP predictive models

This section presents developments of several ROP function ap-


proximations. As mentioned earlier, eight parameters including weight
on bit, pump pressure, bit rotational speed, bit flow rate, pore pressure,
density, shear wave velocity and gamma ray were recognized as the
most influential parameters on drilling rate. So, these variables were
utilized as input parameters for model development. Procedures of
developing the proposed models i.e. ANN, PSO-ANN, ICA-ANN, BBO-
ANN, GA-ANN, and regression models are described in the following
sections. Fig. 9. Number of neurons in the two hidden layers of the developed MLP with
MSE values for the ROP prediction.
5.1. Simple ANNs
determinant tuning parameter in developing these types of ANNs. Here
Process of developing two types of ANN (MLP and RBF) are de- a trial and error approach is presented to determine the optimum value
scribed in this section to predict the ROP. Firstly, data normalization of this parameter. However, different optimization algorithms can be
were performed with the aim of enhancing accuracy of models using used for this purpose (Najafi-Marghmaleki et al., 2017). Various
Eq. (6): structures of RBF-NN have been developed by changing number of
x in = 2 ×
xi xmin
1 neurons in the hidden layer, and their performances were evaluated
x max xmin (6) according to the average MSE obtained through cross validation
method. Fig. 11 shows prediction performance of developed RBF-NNs
In Eq. (6), i shows the number of parameters, and x max and x min
versus number of neurons in their hidden layer. It is obvious that using
represent maximum and minimum values of x i , respectively. In the next
40 neurons in hidden layer can be considered as optimum number of
stage, for developing and evaluating the models, 10-fold cross valida-
RBF units in RBF-NN.
tion method as described in section 3 was implemented. Among all of
the back propagation learning algorithms, Levenberg-Marquardt (LM)
algorithm was selected to train the MLP-NN. According to the litera- 5.2. Hybrid ANNs
ture, many scholars have confirmed efficiency of the LM algorithm in
approximation of geotechnical engineering problems (Armaghani et al., As mentioned earlier, each MLP has included some weights and
2017; Ornek et al., 2012; Ceryan et al., 2013). As it was highlighted by biases which should be optimized by an algorithm to achieve minimum
several researchers, a MLP-NN with two or more hidden layers lead to error in objective function. Here, aforementioned evolutionary algo-
greater accuracy in extremely complex problems (Pham et al., 2017; rithms are applied to adjust weights and biases of MLP-NN. To reach
Surkan and Singleton, 1990; Openshaw and Openshaw, 1997; Li and this goal, a set of randomly selected weights and biases were defined in
Yeh, 2002). Therefore, two hidden layers were considered in MLP-NN a matrix as follows:
to provide sufficient prediction performance. By evaluating perfor- [w1, w2, w3, wn, b1, b , b3, bn] (7)
mance of different transfer (activation) functions, sigmoid and purelin
was selected as transfer function of hidden layers and output layer, Where w and b refer to weight and bias, respectively. The generated
respectively. The number of neurons in input and output layer were matrix is called in various names in different evolutionary algorithm;
considered depending on the input and output variables. As it is stated populations in GA, swarms in PSO, countries in ICA, and species in
before, the performance of MLP-NN is significantly influenced by BBO. Each row of this matrix (member) presents a solution for the
number of neurons in the hidden layer(s). In this study, these para- optimization problem, and number of columns are equal to sum of
meters were determined through a trial and error procedure by evalu- number of weights and biases. After constructing the relevant matrix for
ating the performance of different network architecture (different each algorithm, members of matrix were ranked based on the defined
number of neurons in both hidden layers) based on MSE (average MSE objective function. In the next step, a new matrix including most fea-
for 10 testing folds). MSE value versus number of neurons in both sible solutions (weights and biases) is created using specific operators
hidden layers are shown in Fig. 9. According to the results, the optimum which are diverse in each evolutionary algorithm. Regarding the de-
case of 4 neurons in the first and 6 neurons in the second hidden layer velopment of hybrid RBF-NN, there are three tuning parameters. These
was chosen as the network structure which yield the minimum MSE. parameters include weights, centers and spread, and are optimized by
Fig. 10 illustrates the structure of developed MLP-NN. A summary of application of the same approach used for adjusting weights and biases
the properties of the developed MLP-NN are listed in Table 3. In ad- in MLP-NN.
dition, properties of evolutionary algorithms used for training the ANNs Maximum number of iterations for all evolutionary algorithms were
are mentioned in Table 4. considered as 300 and 250 during training process of MLP-NNs and
Turning to simple RBF-NN training, Gaussian function was selected RBF-NNs, respectively. In following, the size of generated matrix (po-
as the basis function of RBF-NN. After determining centers of the RBF pulation size) was set to 80 for all of the optimization algorithms im-
units, the weights are tuned by Gradient-Descent (GD) back-propaga- plemented for training ANNs. It is important to note that the archi-
tion algorithm. The centers were chosen by application of a clustering tecture of hybrid ANNs (number of hidden layers, number of neurons,
method based on the K-means algorithm, which is a commonly used transfer functions and basis function) were considered as same as the
and accurate cluster analysis technique in data mining (Hartigan and structure of ANNs trained using back propagation algorithm (see sec-
Wong, 1979). Maximum number of neurons (MNN) is a substantially tion 4.1). Fig. 12 illustrate the convergence curves of meta-heuristic

613
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Fig. 10. Structure of developed MLP-NN for ROP prediction.

Table 3 estimated values versus the measured data for simple MLP-NN and
Summary of architecture of the developed MLP to predict ROP. hybrid MLP-NNs. Obviously, the values of regression coefficient for
Network type Multi-layer perceptron (MLP) developed hybrid MLP-NNs are highly desirable, indicating that the
Training function Levenberg-Marquardt (LM) proposed models yield reasonably accurate predictions for drilling rate.
Number of layers 3 Also, this figure reveals that prediction performance of hybrid MLP-NNs
Nodes in 1st hidden layer 4 are far higher than of MLP-NN trained by LM back propagation algo-
Transfer function of 1st hidden layer TANSIG
Nodes in 2nd hidden layer 6
rithm. Regression plots with regard to RBF-NN and hybrid RBF-NNs are
Transfer function of 2nd hidden layer TANSIG shown in Fig. 14. It is noticeable that hybrid RBF-NNs exhibit much
Neurons in output layer 1 better performance in prediction than RBF-NN trained by GD back
Transfer function of output layer PURELIN propagation algorithm. Moreover, relative deviations for developed
Performance evaluation MSE
ANNs and hybrid ANNs are shown in Figs. 15 and 16. According to
results, the lowest relative deviations are related to ANNs that use PSO
algorithm as training method, indicating the impressive power of this
Table 4
evolutionary algorithm in training ANNs.
Properties of training algorithms in MLP and RBF neural networks.
There are also various statistical performance indices which can be
Algorithm Parameters Value for MLP Value for RBF utilized to quantitatively assess performance of these models. In this
GA Selection Method Roulette Wheel Roulette Wheel
study, variance account for (VAF), root mean square error (RMSE),
Crossover Uniform (p = 1) Uniform (p = 1) performance index (PI), and regression coefficient (R2) are used to
Mutation Uniform Uniform compare the prediction capability of the developed models as they were
(p = 0.1) (p = 0.05) employed by Yılmaz and Yuksek (2008) (Yılmaz and Yuksek, 2008),
Mutation Rate 0.1 0.1
Basarir et al. (2014) (Basarir et al., 2014), and Boyacioglu and Avci
Selection Pressure (Roulette 3 2
wheel) (2010) (Boyacioglu and Avci, 2010):
PSO Cognitive Constant 0.5 0.5 1
p
Social Constant 2.5 2.5 1 2
Inertia Weight 5 5 RMSE = (yr zr )2
p (8)
Inertia Weight Damping 0.6 0.6 r=1
Ratio
BBO Keep Rate 1 1 VAF
PI = R + RMSE
Immigration Rate [0,1] [0,1] 100 (9)
Emigration Rate [0,1] [0,1]
Mutation Probability 0.01 0.01 var (yr z r )
ICA Number of Empires 15 15 VAF = 1 × 100
Assimilation Coefficient 2 2
var (yr ) (10)
Revolution Probability 0.3 0.3 p
Revolution Rate 0.1 0.1 r=1 r
(y zr )2
Colonies Mean Cost 0.2 0.1
R2 = 1 p
(
r=1 r
y yr , mean )2 (11)
Coefficient
In the preceding equations, z is output of the model, y is actual
output (target) and p is the number of data points in database. The
training algorithms used for training the both MLP and RBF neural accuracies of the proposed models are distinctive referring to the small
networks. In addition, properties of used algorithms are listed in details values of RMSE and high values of R2, PI and VAF. Table 5 indicates
in Table (5). The optimum value of these listed parameters was de- values of RMSE, VAF, PI and R2 for the developed models. Regarding
termined using trial and error approach. this table, PSO-MLP yielded the highest values of R2, VAF and PI and
the lowest RMSE corresponding to this model, showing the best per-
6. Results and discussion formance among other models in ROP prediction. The second most
accurate model is PSO-RBF that presents slightly lower prediction ef-
In this section, the predicting performance of developed models is ficiency compared to PSO-MLP. GA-MLP and GA-RBF come in third and
evaluated to choose most reliable model and analyze whether evolu- fourth level, respectively, in terms of prediction performance. Overall,
tionary algorithms can be a proper alternative to conventional back according to the results, PSO and GA algorithms perform much better
propagation methods for training. Fig. 13 presents regression plot of than other implemented algorithms in training ANNs.

614
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Fig. 11. MSE of the ROP prediction achieved by the RBF network versus the number of neurons in the hidden layer (MNN).

Fig. 12. Convergence of the hybrid neural networks to the optimum solution (minimum RMSE): (a) MLP, (b) RBF.

615
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Fig. 13. Cross plot of measured versus predicted values of ROP for five developed machine-learning models: (a) MLP-PSO; (b) MLP-GA; (c) MLP-ICA; (d) MLP-BBO;
and, (e) MLP.

Comparing different results (RMSE, PI, VAF and R2) of hybrid ANNs MLP and simple RBF model, respectively. This low performance can be
in Table 7, it is clear that MLP-NN exhibits more effectiveness in ROP attributed to locally search behavior of BP algorithms which often leads
prediction than RBF-NN when using PSO or GA as training function. In to fail in finding optimum parameters of ANNs. In fact, it is highly
more detail, PSO-MLP and PSO-RBF having RMSE values of 1.12 and possible that a BP algorithm converge into a local minimum, while
1.4 for testing data, respectively, outperform the GA-MLP and GA-RBF, evolutionary algorithms such as those proposed in this study are suf-
with RMSE values of 1.43 for the former and 1.52 for the latter. Con- ficiently able to avoid happening this phenomenon.
versely, employing other evolutionary algorithms such as ICA and BBO In addition, a comparison between the de-noised drilling rate
as training function leads to better performance of the RBF-NN than of measurements and the results of the MLP-PSO model trained using the
the MLP-NN in terms of ROP prediction. pre-processed data are shown in Fig. 17-a. The model trained using the
Interestingly, ANNs trained by BP algorithms do not yield satisfac- de-noised data displays very good prediction performance and properly
tory results compared with hybrid models in ROP prediction. The va- follows the trend of changes observed in the ROP measured data in the
lues of 1.68 and 1.76 are yielded as RMSE for testing data of simple studied depth interval (3000 m–3500 m). However, figure for the model

616
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Fig. 14. Cross plot of measured versus predicted values of ROP for five developed machine-learning models: (a) RBF-PSO; (b) RBF -GA; (c) RBF eICA; (d) RBF eBBO;
and, (e) RBF.

trained using raw data (Fig. 17-b) exhibits large fluctuations and the Khandelwal et al., 2017). Finding this relationship is the goal of mul-
corresponding results are not in good accordance with the measured tiple regression method. Two types of multiple regression method are
drilling rate compared with trends in Fig. 17-a. exist, linear and non-linear multiple regression method. A linear re-
lationship between variables can be established with application of
linear multiple regression (LMR), and non-linear multiple regression
7. A comparison between stepwise multiple regression methods
(NLMR) can provide a non-linear relationship between inputs and
and hybrid ANNs
output(s) parameters (Armaghani et al., 2016).
In this research, both of the LMR and NLMR methods were con-
Regression method can statistically establish some relationships
sidered to probe any existed relationship between input and output
between different variables in a problem. This relationship is stated as a
variables and also validate the results of proposed hybrid neural net-
function called regression function which presents relationship between
works. In development stage, the eight parameters that were selected in
independent and dependent parameters (Bahrami et al., 2011;

617
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Fig. 15. Relative deviations of measured versus predicted values of ROP for five developed machine-learning models-The horizontal axis is data record index number
for testing data set: (a) MLP-PSO; (b) MLP-GA; (c) MLP-ICA; (d) MLP-BBO; and, (e) MLP.

feature selection section were taken as input parameters into the re- equation provides more accurate results than LMR model.
gression models. Using the established data, equations for NLMR and
LMR models were obtained that are shown in Eq. (7) and Eq. (8), re- ROP = 1946.1 1968 × F2 4.58 × F3 + 14.71 × F5 13.77 × F6
spectively. The given terms for developed NMLR and LMR are pre- + 6.52 × F7 2.42 × F8 + 19.69 × F9 1039.5 × F11
sented in Tables 6 and 7, respectively. Table 8 also reports general
1.85 × F2 × F5 2.49 × F2 × F7 + 0.67 × F2 × F8
results of implementing the NLMR and LMR models on data of the well
under study. The R2 and RMSE for testing data reveal the superiority of 0.008 × F3 × F6 + 0.006 × F3 × F9 0.007 × F5 × F7
the developed hybrid neural networks models in predicting ROP com- 0.01 × F5 × F9 0.02 × F6 × F7 0.004 × F6 × F8
pared to linear and non-linear regression models. Confirming the fact
+ 0.05 × F6 × F9 3.45 × F6 × F11 0.005 × F7 × F8
that PSO-MLP and PSO-RBF have statistically quite impressive results
than two employed regression models for the dataset extracted from the + 0.02 × F7 × F9 + 0.003 × F8 × F9 + 0.22 × F8 × F11
studied well. Moreover, comparing the results of both regression + 1.0184 × F9 × F11 0.0009 × F32 0.032 × F62 0.009 × F72
models indicates that the relationship between independent parameters 0.0002 × F82 0.02 × F92 196.03 × 2
F11 (12)
and dependent parameter (ROP) is completely non-linear as NLMR

618
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Fig. 16. Relative deviations of measured versus predicted values of ROP for five developed machine-learning models-The horizontal axis is data record index number
for testing data set: (a) RBF-PSO; (b) RBF -GA; (c) RBF eICA; (d) RBF eBBO; and, (e) RBF.

ROP = 25.336 × F2 + 0.12148 × F5 + 0.87147 × F6 0.045999 × F8 noise-free data to select the features with highest influence on ROP and
also to shorten the input vector. Accordingly, eight variables including
+ 0.51205 × F9 + 50.046 × F11 305.62 (13)
weight on bit, bit rotational speed, bit flow rate, pump pressure, pore
pressure, gamma ray, density log and velocity of sonic wave were se-
8. Conclusion lected as model's inputs. Considering these eight parameters and using
all 1000 data points, hybrid MLP and RBF neural networks were de-
In order to predict the ROP, a database consisting of 1000 data veloped using evolutionary algorithms including PSO, GA, ICA, and
points was collected from mud logging and petro-physical logs of a well BBO as learning algorithms. With the aim of giving better assessment of
drilled in Marun oil field. At first, Savitzky–Golay filter as was applied the hybrid models' performance, simple ANNs (training with a back-
on the extracted data to reduce the effect of noise and to develop the propagation method) and multiple regression methods were developed,
predictive models with high accuracy. In the next stage, a feature and the results of them were compared to the results of hybrid models.
ranking method based on heuristic search algorithm was applied on the Main achievements of this study are outlined as follows:

619
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Table 5 • ANNs (MLP and RBF) trained by PSO and GA meta-heuristic algo-
Statistical indicators of ROP-prediction performance accuracy for the simple rithms showed higher performances than other developed models,
ANNs and hybrid ANNs. indicating power of these algorithms in tuning parameters of ANNs
Model Data subset RMSE R VAF % PI in this specific problem.

MLP-PSO Test 1.12 0.933 93.39 0.7752


• MLP-NN exhibits more effectiveness in ROP prediction than RBF-NN
in condition of using PSO or GA as training function. However,
Train 0.9668 0.9496 94.96 0.9573
applying other evolutionary algorithms such as ICA and BBO as
MLP-GA Test 1.43 0.8931 89.44 0.4023
Train 1.23 0.9172 91.73 0.6387 training function leads to better results for RBF-NN than for MLP-NN
MLP-ICA Test 1.67 0.85 85.19 0.102 in terms of predicting ROP.

MLP-BBO
Train
Test
1.37
1.8
0.8976
0.8563
89.8
85.72
0.4657
−0.0209
• Hybrid ANNs are much more accurate than ANNs trained using back
propagation algorithms (LM algorithm for MLP and GD algorithm
Train 1.464 0.8778 87.79 0.3473
Simple MLP Test 1.68 0.83 83.75 0.0632
for RBF). It can be concluded that BP algorithms fail in discovering
Train 1.59 0.866 86.61 0.2034 global minimum and probability of trapping in local minima is high
RBF-PSO Test 1.4 0.8775 87.86 0.4093 for these algorithms.

RBF-GA
Train
Test
1.36
1.52
0.9028
0.8819
90.29
88.28
0.4856
0.2946
• Implementing the NLMR and LMR models on the established data
showed that these models cannot be considered as reliable ap-
Train 1.41 0.8908 89.12 0.4206
RBF-ICA Test 1.5 0.8601 86.81 0.2902 proaches for prediction of ROP. Confirming the fact that hybrid
Train 1.47 0.886 88.65 0.3487 ANNs yield much more accurate predictions than regression
RBF-BBO Test 1.61 0.8579 85.92 0.17 methods.
Simple RBF
Train
Test
1.5
1.76
0.878
0.8107
87.86
81.46
0.3084
−0.0539 • A comparison between the results of PSO-MLP trained using pre-
Train 1.69 0.84 84.93 0.0718
processed data and raw data showed that the noise reduction ap-
proach (SG filter) implemented in this study is very effective in in-
creasing accuracy of predictive models.

Fig. 17. Measured versus predicted ROP versus data record index numbers for the MLP-PSO model applied to the overall data subset.

620
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Table 6 References
Selected variables and obtained coefficients for each variable using NLMR
method. Ab Wahab, M.N., Nefti-Meziani, S., Atyabi, A., 2015. A comprehensive review of swarm
optimization algorithms. PLoS One 10 (5), e0122827.
added variables Estimated coefficients Standard error T-statistics
Akgun, F., 2007. Drilling rate at the technical limit. Int. J. Petrol. Sci. Technol. 1, 99–118.
Anemangely, M., Ramezanzadeh, A., Tokhmechi, B., 2017. Shear wave travel time esti-
Intercept −1946.1 1397.6 −1.3925 mation from petrophysical logs using ANFIS-PSO algorithm: a case study from Ab-
F2 −1968 288.84 −6.8134 Teymour Oilfield. J. Nat. Gas Sci. Eng. 38, 373–387.
F3 −4.5808 0.87679 −5.2245 Anemangely, M., Ramezanzadeh, A., Tokhmechi, B., Molaghab, A., Mohammadian, A.,
F5 14.713 3.0742 4.7861 2018. Drilling rate prediction from petrophysical logs and mud logging data using an
F6 −13.774 7.4501 −1.8488 optimized multilayer perceptron neural network. J. Geophys. Eng. 15 (4), 1146.
F7 6.5228 5.1346 1.2704 Ansari, H.R., Hosseini, M.J.S., Amirpour, M., 2017. Drilling rate of penetration prediction
F8 −2.4241 0.55334 −4.3808 through committee support vector regression based on imperialist competitive al-
F9 19.687 3.9658 4.964 gorithm. Carbonates Evaporites 32 (2), 205–213.
F11 −1039.5 301.41 −3.4487 Armaghani, D.J., Mohamad, E.T., Hajihassani, M., Yagiz, S., Motaghedi, H., 2016.
F2 F5 −1.8473 0.26677 −6.9247 Application of several non-linear prediction tools for estimating uniaxial compressive
strength of granitic rocks and comparison of their performances. Eng. Comput. 32 (2),
F2 F7 −2.487 0.89643 −2.7743
189–206.
F2 F8 0.67474 0.081699 8.2588
Armaghani, D.J., Mohamad, E.T., Narayanasamy, M.S., Narita, N., Yagiz, S., 2017.
F3 F6 −0.0080211 0.001093 −7.3389 Development of hybrid intelligent models for predicting TBM penetration rate in hard
F3 F9 0.0057337 0.00099838 5.743 rock condition. Tunn. Undergr. Space Technol. 63, 29–43.
F5 F7 −0.0066956 0.0028502 −2.3491 Atashpaz-Gargari, E., Lucas, C. (Eds.), 2007. Imperialist Competitive Algorithm: an
F5 F9 −0.01037 0.0030816 −3.365 Algorithm for Optimization Inspired by Imperialistic Competition. Evolutionary
F6 F7 −0.025033 0.0054302 −4.6099 Computation, 2007 CEC 2007 IEEE Congress on. IEEE.
F6 F8 −0.0041653 0.0012895 −3.23 Baba, K., Bahi, L., Ouadif, L., 2014. Enhancing geophysical signals through the use of
F6 F9 0.045603 0.010533 4.3294 Savitzky-Golay filtering method. Geofisc. Int. 53 (4), 399–409.
F6 F11 −3.4493 1.0739 −3.212 Bahrami, A., Monjezi, M., Goshtasbi, K., Ghazvinian, A., 2011. Prediction of rock frag-
F7 F8 −0.0052585 0.00052566 −10.004 mentation due to blasting using artificial neural network. Eng. Comput. 27 (2),
F7 F9 0.024154 0.0061148 3.95 177–181.
Basarir, H., Tutluoglu, L., Karpuz, C., 2014. Penetration rate prediction for diamond bit
F8 F9 0.0033364 0.00059644 5.5938
drilling by adaptive neuro-fuzzy inference system and multiple regressions. Eng.
F8 F11 0.21645 0.071884 3.0111
Geol. 173, 1–9.
F9 F11 1.0184 0.47716 2.1344 Bataee, M., Irawan, S., Kamyab, M., 2014. Artificial neural network model for prediction
F32 −0.00094814 0.00020179 −4.6986 of drilling rate of penetration and optimization of parameters. J. Jpn. Petrol. Inst. 57
F62 −0.031539 0.01055 −2.9894 (2), 65–70.
−0.0086858 0.003356 −2.5881 Bezminabadi, S.N., Ramezanzadeh, A., Jalali, S.-M.E., Tokhmechi, B., Roustaei, A., 2017.
F72
Effect of rock properties on ROP modeling using statistical and intelligent methods: a
F82 −0.0002487 3.4068 × 10 5 −7.3003 case study of an oil well in southwest of Iran. Arch. Min. Sci. 62 (1), 131–144.
F92 −0.020562 0.0030681 −6.7018 BINGHAM, G., 1965. A new approach to interpreting rock drillability. TECHNICAL
2 −196.03 34.971 −5.6055 MANUAL REPRINT. Oil Gas J. 93, 1965.
F11
Bodaghi, A., Ansari, H.R., Gholami, M., 2015. Optimized support vector regression for
drilling rate of penetration estimation. Open Geosci. 7 (1).
Bolandi, V., Kadkhodaie, A., Farzi, R., 2017. Analyzing organic richness of source rocks
Table 7 from well log data by using SVM and ANN classifiers: a case study from the Kazhdumi
formation, the Persian Gulf basin, offshore Iran. J. Petrol. Sci. Eng. 151, 224–234.
Selected variables and obtained coefficients for each variable using LMR Boyacioglu, M.A., Avci, D., 2010. An adaptive network-based fuzzy inference system
method. (ANFIS) for the prediction of stock market return: the case of the Istanbul stock ex-
change. Expert Syst. Appl. 37 (12), 7908–7912.
added variables Estimated coefficients Standard error T-statistics
Cardiff, M., Kitanidis, P.K., 2010. Fitting data under omnidirectional noise: a probabilistic
method for inferring petrophysical and hydrologic relations. Math. Geosci. 42 (8),
Intercept −305.62 21.157 −14.445 877–909.
F2 −25.336 3.9365 −6.436 Ceryan, N., Okkan, U., Kesimal, A., 2013. Prediction of unconfined compressive strength
F5 0.12148 0.016315 7.446 of carbonate rocks using artificial neural networks. Environmental earth sciences 68
F6 0.87147 0.037062 23.514 (3), 807–819.
F8 −0.045999 0.0027053 −17.003 Chandrashekar, G., Sahin, F., 2014. A survey on feature selection methods. Comput.
F9 0.51205 0.024267 21.1 Electr. Eng. 40 (1), 16–28.
F11 50.046 2.2713 22.034 Darbor, M., Faramarzi, L., Sharifzadeh, M., 2017. Performance assessment of rotary
drilling using non-linear multiple regression analysis and multilayer perceptron
neural network. Bull. Eng. Geol. Environ. 1–13.
Davoudi, E., Vaferi, B., 2018. Applying artificial neural networks for systematic estima-
Table 8 tion of degree of fouling in heat exchangers. Chem. Eng. Res. Des. 130, 138–153.
Results of implementing NLMR onto selected features. Del Valle, Y., Venayagamoorthy, G.K., Mohagheghi, S., Harley, R.G., Hernandez, J., 2008.
Particle Swarm Optimization: Basic Concepts, Variants and Applications in Power
Regression method RMSE R2 Adjusted R2 Systems.
A new optimizer using particle swarm theory. Micro Machine and Human Science. In:
Non-linear Model 1.91 0.81 0.804 Eberhart, R., Kennedy, J. (Eds.), MHS'95, Proceedings of the Sixth International
Linear Model 2.84 0.569 0.566 Symposium on; 1995. IEEE.
Modeling the properties of crude oil and gas systems using RBF network. In: Elsharkawy,
A.M. (Ed.), SPE Asia Pacific Oil and Gas Conference and Exhibition. Society of
Petroleum Engineers.
Acknowledgments Esfe, M.H., Wongwises, S., Naderi, A., Asadi, A., Safaei, M.R., Rostamian, H., et al., 2015a.
Thermal conductivity of Cu/TiO2–water/EG hybrid nanofluid: experimental data and
The authors thank National Iranian oil Company (NIOC) for pro- modeling using artificial neural network and correlation. Int. Commun. Heat Mass
Tran. 66, 100–104.
viding the mud Logging data. Esfe, M.H., Saedodin, S., Sina, N., Afrand, M., Rostami, S., 2015b. Designing an artificial
neural network to predict thermal conductivity and dynamic viscosity of ferromag-
netic nanofluid. Int. Commun. Heat Mass Tran. 68, 50–57.
Appendix A. Supplementary data Eskandarian, S., Bahrami, P., 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. Petrol. Sci. Eng. 156, 605–615.
Supplementary data to this article can be found online at https://
Fausett, L.V., 1994. Fundamentals of Neural Networks: Architectures, Algorithms, and
doi.org/10.1016/j.petrol.2018.12.013.

621
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Applications. Prentice-Hall Englewood Cliffs. neural networks for prediction of high pressure vapor-liquid equilibrium. Australian
Garcia, L.P., de Carvalho, A.C., Lorena, A.C., 2015. Effect of label noise in the complexity Journal of Basic and Applied Sciences 3 (3), 1851–1862.
of classification problems. Neurocomputing 160, 108–119. Moghadassi, A., Hosseini, S.M., Parvizian, F., Al-Hajri, I., Talebbeigi, M., 2011. Predicting
Gardner, M.W., Dorling, S., 1998. Artificial neural networks (the multilayer percep- the supercritical carbon dioxide extraction of oregano bract essential oil.
tron)—a review of applications in the atmospheric sciences. Atmos. Environ. 32 Songklanakarin J. Sci. Technol. 33 (5).
(14–15), 2627–2636. Moradi, H., Bahari, M.H., Naghibi Sistani, M.B., Bahari, A., 2010. Drilling rate prediction
Ghaffarian, N., Eslamloueyan, R., Vaferi, B., 2014. Model identification for gas con- using an innovative soft computing approach. Sci. Res. Essays 5.
densate reservoirs by using ANN method based on well test data. J. Petrol. Sci. Eng. Moraveji, M.K., Naderi, M., 2016. Drilling rate of penetration prediction and optimization
123, 20–29. using response surface methodology and bat algorithm. J. Nat. Gas Sci. Eng. 31,
Gholami, E., Vaferi, B., Ariana, M.A., 2018. Prediction of viscosity of several alumina- 829–841.
based nanofluids using various artificial intelligence paradigms-Comparison with Motahhari, H.R., Hareland, G., James, J., 2010. Improved drilling efficiency technique
experimental data and empirical correlations. Powder Technol. 323, 495–506. using integrated PDM and PDC bit parameters. J. Can. Petrol. Technol. 49 (10),
Ghoreishi, S., Heidari, E., 2013. Extraction of epigallocatechin-3-gallate from green tea 45–52.
via supercritical fluid technology: neural network modeling and response surface Najafi-Marghmaleki, A., Barati-Harooni, A., Tatar, A., Mohebbi, A., Mohammadi, A.H.,
optimization. J. Supercrit. Fluids 74, 128–136. 2017. On the prediction of Watson characterization factor of hydrocarbons. J. Mol.
Goldberg, D., 1989. Genetic Algorithms in Optimization, Search and Machine Learning. Liq. 231, 419–429.
Addison-Wesley, Reading. Openshaw, S., Openshaw, C., 1997. Artificial Intelligence in Geography.
Grefenstette, J.J., 2013. Genetic Algorithms and Their Applications: Proceedings of the Ornek, M., Laman, M., Demir, A., Yildiz, A., 2012. Prediction of bearing capacity of
Second International Conference on Genetic Algorithms. Psychology Press. circular footings on soft clay stabilized with granular soil. Soils Found. 52 (1), 69–80.
Drag-bit model including wear. In: Hareland, G., Rampersad, P. (Eds.), SPE Latin Park, J., Sandberg, I.W., 1991. Universal approximation using radial-basis-function net-
America/Caribbean Petroleum Engineering Conference. Society of Petroleum works. Neural Comput. 3 (2), 246–257.
Engineers. Pham, B.T., Bui, D.T., Pourghasemi, H.R., Indra, P., Dholakia, M., 2017. Landslide sus-
Hartigan, J.A., Wong, M.A., 1979. Algorithm AS 136: a k-means clustering algorithm. ceptibility assesssment in the Uttarakhand area (India) using GIS: a comparison study
Journal of the Royal Statistical Society Series C (Applied Statistics) 28 (1), 100–108. of prediction capability of naïve bayes, multilayer perceptron neural networks, and
Using trees, bagging, and random forests to predict rate of penetration during drilling. In: functional trees methods. Theor. Appl. Climatol. 128 (1–2), 255–273.
Hegde, C., Wallace, S., Gray, K. (Eds.), SPE Middle East Intelligent Oil and Gas Quinlan, J.R., 1986. The effect of noise on concept learning. Mach. Learn.: An artificial
Conference and Exhibition. Society of Petroleum Engineers. intelligence approach 2, 149–166.
Hegde, C., Daigle, H., Millwater, H., Gray, K., 2017. Analysis of rate of penetration (ROP) Ramchoun, H., Amine, M., Idrissi, J., Ghanou, Y., Ettaouil, M., 2016. Multilayer per-
prediction in drilling using physics-based and data-driven models. J. Petrol. Sci. Eng. ceptron: architecture optimization and training. IJIMAI 4 (1), 26–30.
159, 295–306. Rudin, L.I., Osher, S., Fatemi, E., 1992. Nonlinear total variation based noise removal
Hegde, C., Daigle, H., Gray, K.E., 2018. Performance comparison of algorithms for real- algorithms. Phys. Nonlinear Phenom. 60 (1–4), 259–268.
time rate-of-penetration optimization in drilling using data-driven models. SPE J. Savitzky, A., Golay, M.J., 1964. Smoothing and differentiation of data by simplified least
23 (05). squares procedures. Anal. Chem. 36 (8), 1627–1639.
Heidari, E., Sobati, M.A., Movahedirad, S., 2016. Accurate prediction of nanofluid visc- Shan, L., Cao, L., Guo, B., 2018. Identification of flow units using the joint of WT and
osity using a multilayer perceptron artificial neural network (MLP-ANN). LSSVM based on FZI in a heterogeneous carbonate reservoir. J. Petrol. Sci. Eng. 161,
Chemometr. Intell. Lab. Syst. 155, 73–85. 219–230.
Holland, J.H., 1992. Genetic algorithms. Sci. Am. 267 (1), 66–73. A modified particle swarm optimizer. Evolutionary Computation Proceedings. In: Shi, Y.,
Jacintoa, C.M., Freitas Filho, P.J., Nassarb, S.M., Roisenbergb, M., Rodriguesb, D.G., Eberhart, R. (Eds.), IEEE World Congress on Computational Intelligence, the 1998
Limab, M.D., 2013. Optimization models and prediction of drilling rate (ROP) for the IEEE International Conference on; 1998. IEEE.
Brazilian pre-salt layer. Chem. Eng. 33. Shi, X., Liu, G., Gong, X., Zhang, J., Wang, J., Zhang, H., 2016. An efficient approach for
James, G., Witten, D., Hastie, T., Tibshirani, R., 2013. An Introduction to Statistical real-time prediction of rate of penetration in offshore drilling. Math. Probl Eng. 2016.
Learning. Springer. Simon, D., 2008. Biogeography-based optimization. IEEE Trans. Evol. Comput. 12 (6),
Optimization of rate of penetration in a convoluted drilling framework using ant colony 702–713.
optimization. In: Jiang, W., Samuel, R. (Eds.), IADC/SPE Drilling Conference and Neural networks for bond rating improved by multiple hidden layers. Neural Networks.
Exhibition. Society of Petroleum Engineers. In: Surkan, A.J., Singleton, J.C. (Eds.), 1990 IJCNN International Joint Conference
Kahraman, S., 2016. Estimating the penetration rate in diamond drilling in laboratory on; 1990. IEEE.
works using the regression and artificial neural network analysis. Neural Process. Szlek, J., Mendyk, A., 2015. Package ‘fscaret’.
Lett. 43 (2), 523–535. Feature selection methods: genetic algorithms vs. greedy-like search. Vafaie, H., Imam,
Khandelwal, M., Armaghani, D.J., 2016. Prediction of drillability of rocks with strength I.F. (Eds.), Proceedings of the International Conference on Fuzzy and Intelligent
properties using a hybrid GA-ANN technique. Geotech. Geol. Eng. 34 (2), 605–620. Control Systems.
Khandelwal, M., Faradonbeh, R.S., Monjezi, M., Armaghani, D.J., Majid, M.Z.B.A., Yagiz, Vaferi, B., Eslamloueyan, R., 2015. Hydrocarbon reservoirs characterization by co-in-
S., 2017. Function development for appraising brittleness of intact rocks using ge- terpretation of pressure and flow rate data of the multi-rate well testing. J. Petrol. Sci.
netic programming and non-linear multiple regression models. Eng. Comput. 33 (1), Eng. 135, 59–72.
13–21. Vaferi, B., Gitifar, V., Darvishi, P., Mowla, D., 2014. Modeling and analysis of effective
Lashkarbolooki, M., Hezave, A.Z., Ayatollahi, S., 2012. Artificial neural network as an thermal conductivity of sandstone at high pressure and temperature using optimal
applicable tool to predict the binary heat capacity of mixtures containing ionic li- artificial neural networks. J. Petrol. Sci. Eng. 119, 69–78.
quids. Fluid Phase Equil. 324, 102–107. Vaferi, B., Eslamloueyan, R., Ayatollahi, S., 2015. Application of recurrent networks to
Lashkenari, M.S., Taghizadeh, M., Mehdizadeh, B., 2013. Viscosity prediction in selected classification of oil reservoir models in well-testing analysis. Energy Sources, Part A
Iranian light oil reservoirs: artificial neural network versus empirical correlations. Recovery, Util. Environ. Eff. 37 (2), 174–180.
Petrol. Sci. 10 (1), 126–133. Vaferi, B., Eslamloueyan, R., Ghaffarian, N., 2016. Hydrocarbon reservoir model detec-
Law, M.H., Figueiredo, M.A., Jain, A.K., 2004. Simultaneous feature selection and clus- tion from pressure transient data using coupled artificial neural network—wavelet
tering using mixture models. IEEE Trans. Pattern Anal. Mach. Intell. 26 (9), transform approach. Appl. Soft Comput. 47, 63–75.
1154–1166. Venkatesan, P., Anitha, S., 2006. Application of a radial basis function neural network for
Le Van, S., Chon, B.H., 2017. Evaluating the critical performances of a CO2–Enhanced oil diagnosis of diabetes mellitus. Curr. Sci. 91 (9), 1195–1199.
recovery process using artificial neural network models. J. Petrol. Sci. Eng. 157, Wang, R.Y., Storey, V.C., Firth, C.P., 1995. A framework for analysis of data quality re-
207–222. search. IEEE Trans. Knowl. Data Eng. (4), 623–640.
The effect of initial weights on premature saturation in back-propagation learning. Neural Wang, X., Tang, Z., Tamura, H., Ishii, M., Sun, W., 2004. An improved backpropagation
Networks. In: Lee, Y., Oh, S.-H., Kim, M.W. (Eds.), IJCNN-91-Seattle International algorithm to avoid the local minima problem. Neurocomputing 56, 455–460.
Joint Conference on; 1991. IEEE. Wang, G., Guo, L., Wang, H., Duan, H., Liu, L., Li, J., 2014a. Incorporating mutation
Li, X., Yeh, A.G.-O., 2002. Neural-network-based cellular automata for simulating mul- scheme into krill herd algorithm for global numerical optimization. Neural Comput.
tiple land use changes using GIS. Int. J. Geogr. Inf. Sci. 16 (4), 323–343. Appl. 24 (3–4), 853–871.
Lorena, A.C., de Carvalho, A.C., 2004. Evaluation of noise reduction techniques in the Wang, G.-G., Gandomi, A.H., Alavi, A.H., 2014b. An effective krill herd algorithm with
splice junction recognition problem. Genet. Mol. Biol. 27 (4), 665–672. migration operator in biogeography-based optimization. Appl. Math. Model. 38
Martin, K.G., Totten, M.W., Raef, A., 2017. Characterization of a reservoir ooid shoal (9–10), 2454–2462.
complex and Artificial Neural Networks application in lithofacies prediction: mis- Wang, G.-G., Gandomi, A.H., Alavi, A.H., Hao, G.-S., 2014c. Hybrid krill herd algorithm
sissippian St. Louis formation, Lakin fields, western Kansas. J. Petrol. Sci. Eng. 150, with differential evolution for global numerical optimization. Neural Comput. Appl.
1–12. 25 (2), 297–308.
Mendes, J.R.P., Fonseca, T.C., Serapião, A., 2007. Applying a genetic neuro-model re- Warren, T., 1987. Penetration rate performance of roller cone bits. SPE Drill. Eng. 2 (01),
ference adaptive controller in drilling optimization. World Oil 29–36. 9–18.
Mirjalili, S., Mirjalili, S.M., Lewis, A., 2014. Let a biogeography-based optimizer train Roller bit model with rock ductility and cone offset. In: Winters, W., Warren, T., Onyia, E.
your multi-layer perceptron. Inf. Sci. 269, 188–209. (Eds.), SPE Annual Technical Conference and Exhibition. Society of Petroleum
Moghadassi, A., Parvizian, F., Hosseini, S., 2009. A new approach based on artificial Engineers.

622
S.B. Ashrafi et al. Journal of Petroleum Science and Engineering 175 (2019) 604–623

Wu, Y., Wang, H., Zhang, B., Du, K.-L., 2012. Using radial basis function networks for 73–100.
function approximation and classification. ISRN Applied Mathematics 2012. Yetilmezsoy, K., Ozkaya, B., Cakmakci, M., 2011. Artificial intelligence-based prediction
Xue, Y., Cheng, L., Mou, J., Zhao, W., 2014. A new fracture prediction method by com- models for environmental engineering. Neural Netw. World 21 (3), 193.
bining genetic algorithm with neural network in low-permeability reservoirs. J. Yilmaz, I., Kaynar, O., 2011. Multiple regression, ANN (RBF, MLP) and ANFIS models for
Petrol. Sci. Eng. 121, 159–166. prediction of swell potential of clayey soils. Expert Syst. Appl. 38 (5), 5958–5966.
Yavari, H., Sabah, M., Khosravanian, R., Wood, D., 2018. Application of an adaptive Yılmaz, I., Yuksek, A., 2008. An example of artificial neural network (ANN) application
neuro-fuzzy inference system and mathematical rate of penetration models to pre- for indirect estimation of rock parameters. Rock Mech. Rock Eng. 41 (5), 781–795.
dicting drilling rate. Iranian Journal of Oil & Gas Science and Technology 7 (3),

623

S-ar putea să vă placă și