Вы находитесь на странице: 1из 9

Applied Soft Computing 48 (2016) 50–58

Contents lists available at ScienceDirect

Applied Soft Computing

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

A new approach for forecasting OPEC petroleum consumption based

on neural network train by using flower pollination algorithm
Haruna Chiroma a,e,∗ , Abdullah Khan b , Adamu I. Abubakar c , Younes Saadi a ,
Mukhtar F. Hamza d , Liyana Shuib a , Abdulsalam Y. Gital f , Tutut Herawan a
Faculty of Computer Science and IT, University of Malaya, Malaysia
Institute of Business and Management Science, Faculty of Information Data Base Agriculture University Peshawar, Pakistan
Faculty of Information and Communication Technology, International Islamic University Malaysia, Kuala Lumpur, Malaysia
Department of Mechatronics Engineering, Faculty of Engineering, Bayero University, Kano, Nigeria
Federal College of Education (Technical), Department of Computer Science, Gombe, Nigeria
Abubakar Tafawa Balewa University, Mathematical Science Programme, Bauchi, Nigeria

a r t i c l e i n f o a b s t r a c t

Article history: Petroleum is the live wire of modern technology and its operations, with economic development being
Received 13 July 2015 positively linked to petroleum consumption. Many meta-heuristic algorithms have been proposed in
Received in revised form 23 May 2016 literature for the optimization of Neural Network (NN) to build a forecasting model. In this paper, as an
Accepted 28 June 2016
alternative to previous methods, we propose a new flower pollination algorithm with remarkable balance
Available online 30 June 2016
between consistency and exploration for NN training to build a model for the forecasting of petroleum
consumption by the Organization of the Petroleum Exporting Countries (OPEC). The proposed approach is
compared with established meta-heuristic algorithms. The results show that the new proposed method
Flower pollination algorithm
Neural network
outperforms existing algorithms by advancing OPEC petroleum consumption forecast accuracy and con-
Accelerated particle swarm optimization vergence speed. Our proposed method has the potential to be used as an important tool in forecasting
Organization of the petroleum exporting OPEC petroleum consumption to be used by OPEC authorities and other global oil-related organizations.
countries (OPEC) This will facilitate proper monitoring and control of OPEC petroleum consumption.
Energy © 2016 Elsevier B.V. All rights reserved.
Petroleum consumption

1. Introduction to perform better than the genetic algorithm (GA), particle swarm
optimization (PSO) [47], and artificial bee colony (ABC) [42] in solv-
Recently, a new meta-heuristic algorithm called the Flower ing optimization problems.
Pollination Optimization Algorithm (FP) with remarkable balance Artificial Neural Network (NN)is referenced as one of the most
between consistency and exploration was proposed by [47]. The powerful techniques ever established [27] and is heavily applied in
switch probability or proximity probability is used to switch soft computing because of its effectiveness and robustness. The NN
between common global pollination to intensive local pollination. tolerates missing and erroneous values and any association can be
The effectiveness and robustness of the PFA can be attributed to modeled with arbitrary accuracy. In NN, new cases can automat-
the following two reasons: (1) Insect pollinators can travel in long ically be accommodated by updating the learning of the NN. The
distances which enable the flower pollination (FP) to avoid local NN is considered as one of the most reliable and promising com-
landscape to search in a very large space. (2) The FP ensures that putational intelligence techniques [31]. The FP can be applied to
similar species of the flowers are consistently chosen which guar- train NN to create a model for the forecasting of Organization of the
antee fast convergence to the optimal solution. The FP is found Petroleum Exporting Countries (OPEC) petroleum consumption.
Human activities in today’s modern world would not be possi-
ble without energy. Energy is the live wire of modern technology
∗ Corresponding author at: Department of Artificial Intelligence, University of and its operations, with economic development being positively
Malaya, 50603 Pantai Valley, Kuala Lumpur, Malaysia. linked to energy. As such, a lack of energy can make human life very
E-mail addresses: freedonchi@gmail.com, freedonchi@yahoo.com (H. Chiroma), difficult or even impossible [40]. About 2/3 of the world’s energy
abdullahdirvi@gmail.com (A. Khan), 100adamu@gmail.com (A.I. Abubakar), needs are supplied by petroleum and gas [6]. The consumption of
younessaadi@gmail.com (Y. Saadi), emukhtarfh@gmail.com (M.F. Hamza),
petroleum is positively related to economic growth [17], meaning
liyanashuib@um.edu.my (L. Shuib), asgital@yahoo.com (A.Y. Gital),
tutut@um.edu.my (T. Herawan).
that the more developed an economy is, the more the petroleum

1568-4946/© 2016 Elsevier B.V. All rights reserved.
H. Chiroma et al. / Applied Soft Computing 48 (2016) 50–58 51

consumption will increase. Conversely, if economic activities do sibility of being trapped in a premature convergence in view of the
not improve, the consumption of petroleum will reduce accord- fact that the PSO optimization evolve mostly by evaluation of self-
ingly. The world’s petroleum consumption growth is estimated at position with neighboring positions and global positions for all the
1.3 million bbl/d forthe year 2013, with an average of 90.5 million particles as a single pattern. The diversification of result in GA by
bbl/d used in the same year. On the other hand, in 2014 and 2015 crossover and mutation typically led to large number of unneces-
the consumption growth was estimated at 1.0 million bbl/d and 0.9 sary iterations which hider efficiency in problem solving [49]. The
million bbl/d, respectively [15]. performance of GA typically decreases because of the GA insuffi-
About 1/3 of the global petroleum supply is provided by 6 coun- cient balance between exploration and exploitation. The GA has
tries from the Organization of the OPEC, including Saudi Arabia, strong exploitation and weak exploration [25]. The PSO also lacks
Iran, Iraq, Kuwait, United Arab Emirates, and Qatar [14]. Five of sufficient balance between exploitation and exploration [35]. The
these OPEC countries are among the world’s top 5 countries with ABC lacks balancing between exploration and exploitation because
the highest proven reserves of petroleum. The combined petroleum its solution search shows that the ABC exploration is good while
reserves for these six OPEC countries comprise over 60% of the exploitation is poor [23].
global proven petroleum reserves [7]. The OPEC produces as well However, a newly proposed FP outperform the performance of
as exports petroleum to other parts of the world, which shows that the GA, ABC and PSO in terms of accuracy and convergence speed
OPEC has played a significant role in fueling the world’s economic because of its remarkable balance between consistency and explo-
activities in previous decades [28]. ration.
The OPEC petroleum consumption has increased 7 times within To advance forecasting accuracy and convergence speed, deviate
40 years, to 8.5 million bbl/d. Petroleum consumed by OPEC is from local minima and optimized the NN weights and biases, we
nearly as much as that consumed by the People’s Republic of China. proposed to train the NN by using FP for the forecasting of OPEC
This represents1/5 of the OPEC petroleum production. The rapid petroleum consumption.
consumption of petroleum by OPEC could possibly pose a chal- The remaining sections of the paper are organized as fol-
lenge to OPEC’s capability to increase the export of petroleum to lows. Section 2 presents the proposed methodology, comprising
other countries that rely on OPEC to supply petroleum. The growth a description of FP, NN and the hybrid of FP and NN. Section 3
of petroleum consumption by OPEC is estimated at 5.1% annually, describes OPEC petroleum consumption data. Section 4 provides
which is growing faster than OPEC’s income growth of 3.3% [19]. experiments set up and Section 5 presents results and discussion
It was found that predicting petroleum (energy) is required for followed by the concluding remarks and future research in Section
the proper management of supply, demand, production, invest- 6.
ment analysis, analysis of revenue generation, and management
of research related to energy. A better tomorrow depends on 2. The proposed method
the accuracy of today’s decisions [16]. The petroleum consump-
tion, production, prices, demand, etc. are characterized by their This section discusses the basic theoretical background of the FP
high uncertainty level. The uncertain nature of energy makes and NN required to understand our proposed method of forecasting
conventional methods unfit for forecasting it [22,8]. However, OPEC oil consumption. The section contains two major subsections:
meta-heuristic algorithms have extensively been mentioned in lit- the basic FP operation, and how it functions to achieve the desired
erature for solving complex problems. Meta-heuristic algorithms goal. A brief description of the NN model and reference for further
are more powerful than conventional mathematical programming details on the description of NN is given, since it is extensively dis-
[37], because the diversification feature of meta-heuristic algo- cussed in the literature. The theoretical background of the FP and
rithms ensures that the search space is thoroughly searched while NN hybridization is described in detail.
the intensification guarantees that the best among current solu-
tions is selected. Meta-heuristic algorithms are capable of solving
large problems more robustly and faster [41]. 2.1. Flower pollination algorithm
A number of meta-heuristic algorithms have been mentioned
in literature, such as GA, ABC, PSO, firefly, etc. [18]. Some meta- Yang [47] emulated the characteristics of biological flower
heuristic algorithms have been used for solving optimization pollination in flowering plants to develop the flower pollination
problems [13], such as training NN to build a model for oil con- algorithm based on the following rules:
sumption prediction. The ability of meta-heuristic algorithms to
improve convergence speed and accuracy, and prevent BPNNs 1. The global pollination processes are biotic and there is cross pol-
from being trapped in local minima has motivated researchers to lination, in which pollen transporting pollinators perform the
apply such algorithms to build a model for oil consumption predic- Levy flight.
tion. For example, Chiroma et al. [9] proposed a model for energy 2. Local pollination is viewed as ambitious self-pollination.
consumption in Greece based on the co-active neuro fuzzy infer- 3. Reproduction probability is considered flower constancy, which
ence system and Azadeh et al. [4] trained an NN based on GA is proportional to the resemblance between two flowers con-
searches to build a model to forecast electrical energy consump- cerned.
tion in Turkey. Padmakumari et al. [32] applied fuzzy NNs (FNN) 4. The switching probability controls both the local and global pol-
because the FNN combines the power of NNs and fuzzy logic to lination p ∈ [0,1]. Local pollination can have a fraction p that
create a robust synergistic model for the forecasting of long-term is significant in the entire pollination process due to physical
energy consumption. The NN parameters were optimized through proximity and wind.
ABC to forecast oil consumption in 4 different countries in the Mid-
dle East [10]. The oil demand in Iran was estimated using GA and The plant can possess multiple flowers and every flower patch
PSO [2]. An excellent review of the applications of meta-heuristic typically emits millions or even billions of pollen gametes in real-
algorithms in oil consumption modeling is in Ref. [40]. However, life pollination. To simplify the proposed algorithm development, it
the FP has demonstrated improved performance over established is assumed that each plant has a single flower and each flower emits
meta-heuriatic algorithms [11]. only a single pollen gamete. This results in the elimination of the
Previous literature mainly rely on GA, ABC and PSO to create a need to differentiate a pollen gamete, plant or solution to a prob-
model for the forecasting of oil consumption. The PSO has the pos- lem. Basically, a solution xi to a problem are equivalent to a flower
52 H. Chiroma et al. / Applied Soft Computing 48 (2016) 50–58

and pollen gamete. The major stages in FP design are global and Begin
local pollination. In global pollination, the flower pollen is moved
by pollinators, e.g. insects. Pollen can move over long distances, Initial NN structure
since insects typically fly over long distance ranges. This process
guarantees the pollination and reproduction of the fittest solution,
represented as g∗ . The flower constancy can be represented as: Initialization of the FP algorithm
t+1 t

xi = xi + L g∗ − xi , (1)
Enter each pollen as
where xit ,
t, g∗ , and L are pollen i or solution vector xi , iteration, weights to the network
the best solution found in the current generation or iteration, and
strength of pollination (step size), respectively. The Lèvy flight rep-
resentsthe movement of insects, thus, L > 0 from a Lèvy distribution Sample data
is given as follow:
 () sin(/2) 1 Compute MSE
L∼ s 1+ , (s  so > 0) , (2)

where  () represents the standard gamma function, and the dis-
tribution is valid for large steps s > 0. The symbol ∼ in Eq. (2) means No
Is the FPNN computed
the random variable obeys the distribution on right hand side; that for all samples?
is samples should be drawn from the distribution. From rule 2, the
local pollination and flower constancy are expressed as: Yes
xit+1 = xit + xr xit − xkt , (3) Compute the fitness

where xjt and xkt represent pollen from different flowers of the same
plant species, thus mimicking flower constancy in a limited neigh-
Is the fitness computed
borhood. The xr is the scald parameter which is drawn from the for all pollen
The current iteration
uniform distribution in [0,1], where uniform distribution values number is added by one
ranges between 0 and 1. The p is used to switch between common Yes
global pollination and intensive local pollination. The effective- Update the pollens global optimum fitness
ness of the PFA can be attributed to the following two reasons: based on the fitness of each pollen
(1) Insect pollinators can travel long distances, which enables the
FP to avoid the local landscape and have a very large search space
(explorations); (2) Ensure similar flower species are consistently Update each nest and position
and frequently chosen, which guarantees swift convergence to the
optimal solution (exploitation).

No Does the current iteration number

2.2. Back propagation neural network
reach maximum iteration or target
The NN has been extensively referred to in literature for solving
non-linear problems. The attention attracted by the NN is proba- Yes
bly attributed to the ability to model complex relationships in a
dataset that other conventional statistical models may find dif- End
ficult to model. An NN comprises at least three (3) layers with
Fig. 1. The proposed FPNN flowchart.
neurons distributed across the layers. Typically, the hidden and out-
put neurons have activation functions. Each neuron in the hidden
and output layers is associated with a threshold and weights in ∂ek
where  and are the learning rate and gradient of the error
addition to bias. Data are presented to the NN through the input ∂Wk
neurons and are propagated into the network structure. Compu- function to be minimized, respectively. The gradient-decent BP is
tation is performed by the hidden and output neurons, whereas still applied in many NN programs, though the BP is no longer
the input neurons only accept inputs without computation. The considered an optimal and efficient learning algorithm for NN train-
weights and bias of NNs are iteratively modified during NN train- ing [51].The NN have been used in energy forecasting [3]. The NN
ing to minimize the error functions, such as Mean Square Error is extensively discussed in literature, which is why only a brief
(MSE) [21]. The most widely employed NN learning algorithm is description is provided. However, a detailed discussion onNNis-
the Back-Propagation algorithm (BP), which is a gradient-descent given in [51,21].
technique of minimizing an error function developed in Ref. [36].
Werbos [43] popularized the algorithm among the research com- 2.3. The proposed training of neural network by using flower
munity. The synaptic weight (W) in a BPNN learning algorithm can pollination algorithm
be updated using Eq. (4) as follows:
The basic FPNN flowchart are shown in Fig. 1. The procedure of
Wk+1 = Wk + Wk , (4)
the propose FPNN algorithm are presented in Fig. 2. In the proposed
where, k is the iteration in a discrete time and the current weight FPNN algorithm, each best pollen represents a possible solution
adaptation is represented by Wk expressed as: (i.e., the initial weight space and corresponding biases for NN opti-
mization). The weight optimization problem and population size
∂ek represent the quality of the solution. In the first epoch, the best ini-
Wk = − , (5)
∂Wk tial weights and biases are initialized with FP. Subsequently, those
H. Chiroma et al. / Applied Soft Computing 48 (2016) 50–58 53

where V 1 (x) , V 2 (x) , V 3 (x) . . ...V n (x) in Eq. (10) represents

the average performance for each epoch.
The pollen replicates the MSE and it is found after all inputs
are processed for each pollen population. Thus, the flower pollen
gamete xj is computed as shown in Eq. (10):
xj = Min V 1 (x) , V 2 (x) , V 3 (x) , . . ., V n (x) . (10)

The remaining MSEs are considered other pollen gametes. A new

solution xi t+1 for global pollination i is generated using Lèvy flight,
according to Eq. (11):
xi t+1 = xi t + xr L g∗ − xi t , (11)

where xi t is pollen i. A new solution xi t+1 for local pollination i is

generated as follows:
xi t+1 = xi t + xr xj t − xk t . (12)

The movement of the other flower pollen gametes xi toward sxj

can be drawn from Eq. (13).
xi t + xr xj t − xk t ifxr [0, 1] > p˛
X= . (13)
xi t otherwise

The flower pollen gamete scan move from xi towards xj through

Lèvy flight, as:
⎧ ⎛ ⎞

⎨ ⎜ Uj ⎟
xi t + levy () ∼xr ⎜ ⎟ . (X − Xbest ) ifxr [0, 1] > p˛
∇ Xi = ⎝ 1 ⎠ ,

⎩ |Vj | 
xi t otherwise

where ∇ Vi is a small movement of xi towards xj . The weight and

bias for each layer is adapted according to Eqs. (15) and (16), respec-
Fig. 2. The pseudo- code of the FPNN.
Wx n+1 = Wx n − ∇ Xi (15)
weights are passed on to the NN. The weights in NN are computed Bx n+1
= Bx − ∇ Xi
and compared with the best solution. In the second cycle, the FP
updates the weights with the best possible solution. The FP contin- 3. OPEC oil consumption dataset
ues searching the best weights until the last cycle/epoch of the NN
is reached or the target minimum MSE is reached. The error can be The dataset for OPEC petroleum consumption in thousands
computed as: of barrels per day (tb/d) are obtained from the EIA of the US
E = (Tt − Yt ) , (6) Department of Energy, which is freely available online on the
organization’s official website [15]. Dataset are collected for the
where Yt is the output and Tt is the actual data of the NN, when domestic petroleum consumption of each of the 12 OPEC mem-
the tth input is presented.The performances index for the NN is ber countries, including Algeria, Angola, Ecuador, Iran, Iraq, Kuwait,
computed as: Libya, Nigeria, Qatar, Saudi Arabia, UAE, and Venezuela. Also, the
1 R total petroleum consumption for OPEC is collected, given that it is
VF (x) = E T .E, (7) the target forecasting value. The dataset are collected for a yearly
2 t=1
basis from 1980 to 2012 as shown in Fig. 3. The window for data
where E = (Tt − Yt ) and R in Eq. (7) represented thetotal number of collection in this study is determined according to dataset avail-
inputs patterns. The MSE of the performance index is calculated as ability [1]. The dataset are normalized to improve the accuracy and
shown in Eq. (8) convergence speed [34,30]. Hair et al. [20] stated that successful
N forecasting requires a set of independent variables to form a pos-
t=1 F (x)
V (x) = , (8) itive correlation. As a result, we used a correlation coefficient to
investigate the relationship that exists in the dataset depicted in
where Yr is the output of the network when the tth input neti is Fig. 3. The correlation coefficient results are presented in Table 1.
presented. The E1 = (Tr − Yr ) is the error for the tth input, V (x) in Table 1 shows there is no correlation with a negative value. A
Eq. (8) is the average performance, VF (x) in Eq. (7) is the perfor- correlation with a negative value would indicate that the variables
mance index and Pi is the number of flower populations in the ith do not move in the same direction but in the opposite direction.
iteration. At the end of each epoch, the list of the average sum of Thus, the attributes cannot determine the direction fluctuation of
square errors of the ith iteration in MSE can be expressed as: OPEC petroleum consumption. However, the attributes indicated in
  Table 1 form a significant positive relationship between the domes-
MSEi = V 1 (x) , V 2 (x) , V 3 (x) , . . ., V n (x) (9)
tic petroleum consumption of OPEC member countries and total
54 H. Chiroma et al. / Applied Soft Computing 48 (2016) 50–58

4.2. Initialization of the FPNN parameter settings

Parameters are controlled in meta-heuristic algorithm execu-

tion to yield the optimal solution within a few convergence rounds.
However, a systematic way of varying the parameters is an unre-
solved problem [46]. Similar to other meta-heuristic algorithms,
the FP is sensitive to the parameters used. In this study, the param-
eters for FP are set to p = 0.15, and the number of pollen gametes
is 20, as realized through several initial experiments. The NN also
requires initial parameter settings. The number of inputs and out-
put neurons are set to 12 and 1 respectively, because the number
of OPEC countries is 12 and the total petroleum consumption of
OPEC is 1. The number of hidden layers is set to 1 in view of the fact
that one hidden layer is enough to approximate any function with
Fig. 3. Petroleum consumption from the OPEC member countries and total for OPEC
(1980–2012). arbitrary accuracy [33]. The activation functions at the hidden and
output layers are nonlinear and linear as recommended [5].
Table 1
The number of hidden layer neurons is set to 7, which is adopted
Correlation relationship between OPEC oil consumption and individual member from [8]. The initial structure of the NN is (12- 7 -1 with bias). The
countries. weight and bias of the NN are initialized by the FP(12- 7 -1 with bias)
Country OPEC Country OPEC
(FPNN). Having 20 FPNN (12- 7 -1 with bias) with different weight
and bias initiated by FP. The maximum iteration for the generations
Algeria 0.976** Libya 0.756**
is 1000 and the target error is set to0.00001. The NN is very sensitive
Angola 0.940** Nigeria 0.580**
Ecuador 0.976** Qatar 0.937** to the initial weights values. Thus, to give intelligent weights values
Iran 0.983** Saudi Arabia 0.988** to NNs from FP, the NNs are trained with the weight value of FP
Iraq 0.969** United Arab Emirate 0.968** for each iteration. At each iteration the NN request for the new
Kuwait 0.952** Venezuela 0.929** update weights value from the FP. To guarantee convergence, the
Correlation is significant at the 0.01 level (2-tailed). convergence is made to stop when the target error is reached or the
maximum iteration is exhausted. Otherwise the NN will continue
training until the stopping criterion is satisfied.
OPEC petroleum consumption. Essentially, they can influence the
direction fluctuation of OPEC petroleum consumption in view of the
fact that they move in the same direction. Therefore, the petroleum 4.3. Running the proposed FPNN and the comparison algorithms
consumption of each of the 12 OPEC member countries are the
inputs and the total consumption for OPEC is the target output of The FPNN is independently run 25 times for each of the data
the proposed FPNN model. partition ratios, because meta-heuristic algorithms are run more
than once to compute the mean, best and worst results since meta-
4. Experiment setup heuristic algorithms are not deterministic [39]. For the purpose of
evaluating the effectiveness of the proposed FPNN, we have chosen
In this section, the configurations of the machine used to con- GA, ABC because they are meta-heuristic algorithms well estab-
duct our experiments are outlined. The fitness function adopted lished in the literature [12] and BP for optimizing NN weights and
for the study with justification is explained. In addition, initial biases. Also, accelerated particle swarm optimization (APSO) is cho-
parameter settings of the FPNN, running of the proposed FPNN, sen over particle swarm optimization (PSO) because APSO is an
and the comparison algorithms such as GA, APSO, ABC, and NN are improved version of PSO and it has better performance than the PSO
discussed. [44]. The ABC-trained NN (ABCNN), GA-trained NN (GANN), APSO-
trained NN (APSONN) and BP-trained NN(BPNN) are used for OPEC
4.1. System configuration and fitness function petroleum consumption forecasting. We believed the comparison
is fair because we compared the propose method with evolutionary
The FPNN model proposed in this study is used to forecast OPEC algorithms. Typically, if an NN is trained with an evolutionary algo-
petroleum consumption. The performance of FPNN is compared rithm, the fair comparison is to train the NN with other evolutionary
with that of GANN, ABCNN, accelerated PSO (APSO) NN, and BPNN. algorithms for evaluation purpose. This is in line with the literature
The FPNN is coded and implemented in MATLAB R2012b on a [26,44]. Therefore, the propose FP in our study is compared with
computer with the following configuration: Intel Core (TM) 2 Quad, GA, ABC and APSO in training the NN. The parameter settings of the
CPU 2.33 GHz, RAM 2 GB, 32-bit operating system. All experiments comparison algorithms are as follows: APSO: population size = 20,
are conducted on the specified machine. The FPNN requires ini- probability = 0.17, gamma = 0.95, alpha = 0.2, beta = 0.5, lower and
tial parameter settings to start running. The MSE is used as the upper band = [−5,5]. GA: population size = 20, crossover rate = 1,
fitness function to measure the FPNN’s accuracy in forecasting mutation rate = 0.1, lower and upper band = [−5,5]. ABC: colony
OPEC petroleum consumption. The MSE gives an estimated error size = 50, maximum cycle = 1000, dimension = 200, lower and upper
between actual OPEC petroleum consumption and that forecasted band = [−5,5]. The forecasting results of ABCNN, GANN, APSONN,
by the FPNN. The closer MSE is to 0, the better the accuracy is. MSE of and BPNN are compared with those of the proposed FPNN in both
0 indicates perfect forecasting. MSE is chosen because it is superior training and testing datasets.
to other performance indicators, such as normalized mean square
error, root mean square error, mean absolute error, etc. in measur- 5. Results and discussion
ing the performance of different algorithms on the same dataset
[33]. In our experiments, the performance of FPNN is compared In this section, the results obtained from the experiment set
with other established bio-inspired algorithms, thus the choice of up described in Section 4 are presented and discussed. The per-
MSE is appropriate for the experiments. formance of the proposed PFNN and that of the APSONN, ABCNN,
H. Chiroma et al. / Applied Soft Computing 48 (2016) 50–58 55

Table 2 Table 4
Training convergence time (seconds) for the algorithms on OPEC petroleum. Convergence time (seconds) for the algorithms on the OPEC petroleum consumption
test dataset.
Partition Algorithm Mean Best Worst
Partition Algorithm Mean Best Worst
90–10 FPNN 17.318478781 5.3441124 25.4477997
GANN 21.599760985 5.6450988 32.4089472 90–10 FPNN 16.374652086 4.2851508 24.5865379
ABCNN 1078.348204784 245.2569203 1644.0717490 GANN 21.546019869 5.5007970 32.4436726
APSONN 1147.936081480 1137.4864590 1160.9948420 ABCNN 1083.564687316 244.6992528 1654.8122530
BPNN 71.13991031 70.1956293 72.0825834 APSONN 596.943101968 589.8458523 604.0199192
80–20 FPNN 24.408422456 4.3132490 38.1623239 BPNN 67.33254157 66.3757263 68.2773495
GANN 21.853397747 5.5581175 32.9160964 80–20 FPNN 17.586660387 4.3132490 26.6075585
ABCNN 1098.291276980 240.1563437 1681.8133600 GANN 20.971776450 5.4818216 31.4890396
APSONN 622.424642788 614.5872761 630.6467917 ABCNN 1216.911415512 239.0465134 1880.4877670
BPNN 67.15956576 66.2133564 68.0796050 APSONN 842.041400112 825.7943131 855.6750413
70–30 FPNN 23.729440302 4.4629256 36.9146311 BPNN 67.84842447 66.7092589 68.9203720
GANN 29.823770399 5.5102011 46.2545622 70–30 FPNN 22.392703036 4.2923900 34.6760238
ABCNN 1183.652207924 240.8519360 1877.0641990 GANN 41.764004765 5.4640657 66.4537831
APSONN 627.689326384 620.1968321 635.1368825 ABCNN 1147.526381320 243.4676544 1868.1281420
BPNN 68.34625147 67.4067120 69.2892311 APSONN 744.068250620 736.8487364 751.2289768
60–40 FPNN 22.224060973 4.2621414 34.4074315 BPNN 75.20721331 74.2544014 76.1364554
GANN 23.541494887 5.6191962 35.7152157 60–40 FPNN 22.234225182 4.3096664 34.3950287
ABCNN 1120.666526924 242.0332384 1873.4761260 GANN 24.202904022 5.6057096 36.8271391
APSONN 679.589133940 671.8437457 687.3790149 ABCNN 1140.213973352 247.3804678 1877.0641990
BPNN 68.42942644 67.4332304 69.3962184 APSONN 734.756476676 726.1853597 744.0882833
50–50 FPNN 17.607291640 4.2019747 26.6877923 BPNN 70.92859406 69.8426321 72.0394524
GANN 23.816470953 5.5257098 36.2400019 50–50 FPNN 18.987609346 4.2095121 29.0942405
ABCNN 1198.852470176 240.0067170 1873.4761260 GANN 22.312732609 5.5565065 33.6778900
APSONN 752.682990404 744.8411809 760.8418091 ABCNN 1084.688633140 243.1838017 1656.3609250
BPNN 78.95930054 77.8704024 80.1682141 APSONN 587.010445164 579.9539146 594.0554508
BPNN 94.20479797 93.2177303 95.2055563

Table 3
Training accuracy (MSE) for the algorithms on the OPEC petroleum consumption Table 5
dataset. Accuracy (MSE) for the algorithms on the OPEC oil consumption test dataset.

Partition Algorithm Mean Best Worst Partition Algorithm Mean Best Worst

90–10 FPNN 0.000397234 0.0002137 0.0006857 90–10 FPNN 0.000409957 0.0000301 0.0011717
GANN 0.009676973 0.0045722 0.0141054 GANN 0.042849027 0.0202376 0.0771107
ABCNN 0.002709705 0.0012946 0.0057576 ABCNN 0.002421257 0.0000816 0.0092400
APSONN 0.003671513 0.0015747 0.0054657 APSONN 0.000045900 0.0000459 0.0000459
BPNN 0.001467548 0.0006741 0.0098920 BPNN 0.017557917 0.0032995 0.0420978
80–20 FPNN 0.000271476 0.0001299 0.0003811 80–20 FPNN 0.003152259 0.0001299 0.0060656
GANN 0.043583819 0.0032442 0.0689826 GANN 0.117393978 0.0616088 0.1545841
ABCNN 0.006194090 0.0016712 0.0099670 ABCNN 0.003678235 0.0009783 0.0057676
APSONN 0.008613166 0.0080576 0.0090576 APSONN 0.000738990 0.0006430 0.0008483
BPNN 0.114863750 0.0619405 0.1739274 BPNN 0.012365519 0.0062444 0.0314930
70–30 FPNN 0.003156222 0.0001577 0.0061697 70–30 FPNN 0.000520163 0.0001113 0.0018554
GANN 0.043021778 0.0064269 0.0736905 GANN 0.088926722 0.0520731 0.1134958
ABCNN 0.004191528 0.0006319 0.0099670 ABCNN 0.004173302 0.0007348 0.0099670
APSONN 0.003712283 0.0018223 0.0056822 APSONN 0.002791996 0.0016316 0.0046316
BPNN 0.149054708 0.0656022 0.2049988 BPNN 0.038997484 0.0140656 0.1022774
60–40 FPNN 0.000207407 0.0001246 0.0002454 60–40 FPNN 0.000556461 0.0001250 0.0029893
GANN 0.055158568 0.0046934 0.0987970 GANN 0.063814055 0.0435508 0.0899280
ABCNN 0.002105742 0.0006641 0.0057676 ABCNN 0.005073869 0.0008388 0.0099670
APSONN 0.000834024 0.0006990 0.0009800 APSONN 0.003782785 0.0025000 0.0057046
BPNN 0.116267763 0.0635650 0.2016471 BPNN 0.230197988 0.0320143 0.4416329
50–50 FPNN 0.001296049 0.0001615 0.0023688 50–50 FPNN 0.001371345 0.0006851 0.0019121
GANN 0.064654815 0.0030837 0.1123689 GANN 0.047616710 0.0354461 0.0656102
ABCNN 0.003715591 0.0002860 0.0099280 ABCNN 0.003970256 0.0008388 0.0099670
APSONN 0.000499143 0.0002328 0.0008333 APSONN 0.003313023 0.0015770 0.0057770
BPNN 0.157062536 0.0927028 0.2485662 BPNN 0.051838101 0.0269026 0.1589041

GANN and BPNN are compared for evaluation purpose. The values consistent algorithm performance and robustness keeping in view
of the OPEC petroleum consumption forecasted by the proposed that searching for optimal solutions is complex due to real world
FPNN and the actual OPEC petroleum consumption are depicted problems being mostly accompanied by uncertainties. Therefore,
and discussed. Lastly, the policy implication of our proposed focus is not only on the optimal solution, but solution robustness
approach is discussed. are also needed in engineering design and industry. Therefore, real
life application of solutions requires solutions that are both optimal
5.1. Performance of the proposed FPNN on training and test and robust [45]. The 2nd column indicates the algorithms with their
dataset corresponding performance indicator values. Columns 3rd, 4th and
5th are the mean, best, and worst values of the convergence time in
The results of the performance of our proposed algorithm and seconds and accuracy in MSE computed over 25 independent runs
comparison algorithms are presented in Tables 2–5. The first col- for each of the algorithms.
umn represents the partition ratio used to evaluate the algorithms’ The forecast accuracy and convergence time of FPNN for the
performance. The data partition ratio influences algorithm per- OPEC petroleum consumption dataset as described earlier is com-
formance [48]. As a result, different ratios were used to ensure pared with the results of ABCNN, GANN, APSONN, and BPNN in both
56 H. Chiroma et al. / Applied Soft Computing 48 (2016) 50–58

Fig. 4. FPNN forecast OPEC petroleum consumption for 3 years. Fig. 5. FPNN forecast OPEC petroleum consumption for 6 years.

training (see Tables 2 and 3) and testing (see Tables 4–5) dataset.
An algorithm result must be compared on both datasets because
more than one factor must be considered and the most signifi-
cant is the accuracy in both the training and testing datasets [24].
It is clearly shown in Tables 2–5 that the proposed FPNN offers
promising results (bold) in the training and testing cases of the
experiments. The proposed FPNN appears to perform better than
ABCNN, GANN, APSONN, and BPNN in both forecasting accuracy
and convergence time. The FP is able to successfully avoid local
minima to improve the predictive accuracy and convergence speed
over the ABC, GA, APSO and BPNN. There are two possible explana-
tions for the likely reason why FPNN outperforms ABCNN, GANN,
APSONN, and BPNN: (1) The ability of the FP to explore very large
search spaces possibly allows the FP to locate the best weights and
biases, hence improving accuracy. (2) The capability of the FP to Fig. 6. FPNN forecast OPEC petroleum consumption for 9 years.
frequently choose the same flower of the same species may pos-
sibly help the FA to consistently converge faster than the ABC, GA,
It can be observed from Tables 2–5 that ABCNN and APSONN
takes very long time to converge to an optimal solution compared
to GANN and BPNN. The ABC and APSO take a long period of time,
probably due to the slow convergence of ABC and APSO as a result
of their poor exploration and exploitation capabilities. The perfor-
mance accuracy of APSONN is closer to that of FPNN than GANN,
ABCNN and BPNN as observed, probably because of operation simi-
larity between the FP and APSO. Thus, GANN, ABCNN, APSONN, and
BPNN do not perform consistently. The best algorithm is expected
to converge to the best solution after fewer times. Therefore, FPNN
can be considered the best algorithm among the 5 being compared.
Moreover, FPNN can be deemed robust, since it is able to exhibit
similar performance in all possible data partition ratios as well as
maintain consistent performance throughout the experiments. This
Fig. 7. FPNN forecast OPEC petroleum consumption for 12 years.
approach added enhanced accuracy and convergence speed to the
approaches already discussed in the literature.

5.2. The OPEC petroleum consumption forecasted by the proposed


Typically in the literature, the proposed algorithm is applied to

forecast the actual data based on the historical data reserved for
testing to show the benefits and effectiveness of the proposed algo-
rithm as done in [50,38,29]. As such, the FPNN is applied to forecast
OPEC petroleum consumption based on the data reserved for test-
ing to show the forecast ability of the FPNN. Figs. 4–8 depicted
FPNN forecast of OPEC petroleum consumption for different num-
ber of years. The OPEC petroleum consumption forecast by FPNN
are very close to the actual values as shown in Figs. 4–8. This indi-
cated the accuracy and robustness of the FPNN as it can forecast Fig. 8. FPNN forecast OPEC petroleum consumption for 16 years.
the OPEC petroleum consumption across 3, 6, 9, 12 and 16 years.
H. Chiroma et al. / Applied Soft Computing 48 (2016) 50–58 57

This means that the FPNN is accurate, promising and robust, which ness of OPEC petroleum consumption. This study contributed an
is required for applications in real life settings as explained earlier. alternative method of NN training with improved accuracy, conver-
The FPNN model proposed in our study is not meant as a sub- gence speed and robustness over the methods already discussed in
stitute for human experts or existing models. However, FPNN can literature. The approach proposed in our research for OPEC oil con-
serve as a complementary model to human expertise and various sumption forecasting can be helpful to the OPEC administration to
previous models to provide additional information for forecasting efficiently monitor, control, formulate, adapt existing policies and
OPEC petroleum consumption. The FPNN proposed in the study implement OPEC petroleum consumption policies.
has the ability to forecast OPEC petroleum consumption, provided In the future, we recommend researchers to hybridized flower
the oil consumption of the OPEC countries is known. FPNN may pollination algorithm and cuckoo search algorithm to train recur-
advance the generalization ability of existing methods of forecast- rent neural network. The performance of the proposed algorithm
ing petroleum consumption. should be modified to incorporate the capability of exploring Mete-
orological and Oceanographic, oil and gas big data datasets.
5.3. The policy implication
The proposed approach of forecasting OPEC petroleum con-
sumption can be of help to the OPEC management for efficient This work is supported by University of Malaya High ImpactRe-
formulation, adaption of existing policies and implementation of search Grant No vote UM.C/625/HIR/MOHE/SC/13/2 from Ministry
OPEC petroleum consumption policies. This would be possible of Higher Education Malaysia.
because the accurate forecasting of energy consumption affects
capital investment, the quality of the environment, revenue anal- References
ysis, and market research management while continually insuring
supply [16]. [1] M.O. Adetutu, Energy efficiency and capital-energy substitutability: evidence
Economy growth is related to the increase in petroleum con- from four OPEC countries, Appl. Energy 119 (2014) 363–370.
[2] E. Assareh, M.A. Behrang, M.R. Assari, A. Ghanbarzadeh, Application of PSO
sumption, which implies that the more the economy grows, the (particle swarm optimization) and GA (genetic algorithm) techniques on
more petroleum is consumed. This can increase carbon dioxide demand estimation of oil in Iran, Energy 35 (12) (2010) 5223–5229.
emissions, thereby contributing to global warming. The OPEC can [3] H.B. Azad, S. Mekhilef, V.G. Ganapathy, Long-term wind speed forecasting and
general pattern recognition using neural networks, IEEE Trans. Sustain.
use our proposed approach to properly measure estimates of Energy 5 (2) (2014) 546–553.
petroleum consumption by OPEC, so that necessary policies can [4] A. Azadeh, S.F. Ghaderi, S. Tarverdian, M. Saberi, Integration of artificial neural
be formulated to carefully regulate the consumption of petroleum networks and genetic algorithm to predict electrical energy consumption,
Appl. Math. Comput. 186 (2) (2007) 1731–1741.
by OPEC. Policy accuracy heavily depends on the accuracy of the
[5] M.H. Beale, M.T. Hagan, H.B. Demuth, Neural Network ToolboxTM User’s
model used to forecast the petroleum expected to be consumed. Guide, The MathWorks, Inc., Natick, 2013.
Monitoring fluctuations in the consumption of petroleum is [6] B.N. Behmiri, J.R.P. Manso, How crude oil consumption impacts on economic
growth of Sub-Saharan Africa? Energy 54 (2013) 74–83.
possible with forecasted petroleum consumption values. Other
[7] BGR (Bundesanstalt fur Geowissenschaften und Rohstoffe). Reserves,
organizations related to petroleum activities, such as the Inter- resources and availability of energy resources 2010. Annual Report, 2009.
national Energy Agency, Organization for Economic Cooperation [8] H. Chiroma, S. Abdul-Kareem, S.A. Muaz, A. Khan, E.N. Sari, T. Herawan, Neural
and Development, Energy Information Administration and British network intelligent learning algorithm for inter-related energy products
applications, in: Advances in Swarm Intelligence, Springer International
Petroleum may can find our model useful for monitoring OPEC Publishing, 2014, pp. 284–293.
petroleum consumption due to their influence on the global energy [9] H. Chiroma, S. Abdulkareem, E.N. Sari, Z. Abdullah, S.A. Muaz, O. Kaynar, H.
market. It is estimated that by the year 2030, OPEC countries Shah, T. Herawan, et al., Soft computing approach in modeling energy
consumption, in: B. Murgante (Ed.), ICCSA 2014, Part VI, LNCS 8584, Springer
will consume 50% more energy than estimated by the Interna- International Publishing, Switzerland, 2014, pp. 770–782.
tional Energy Agency, Energy Information Administration, British [10] H. Chiroma, S. Abdulkareem, S.A. Muaz, A.I. Abubakar, E. Sutoyo, M. Mungad,
Petroleum, and OPEC itself [19]. This signifies that other energy- Y. Saadi, E.N. Sari, T. Herawan, An intelligent modeling of oil consumption, in:
Advances in Intelligent Informatics, Springer International Publishing, 2015,
related organizations have an interest in monitoring the activities pp. 557–568.
of OPEC with regards to petroleum consumption. The FPNN have [11] H. Chiroma, N.L.M. Shuib, S.A. Muaz, A.I. Abubakar, L.B. Ila, J.Z. Maitama, A
the potential to be developed into software with a graphical user review of the applications of bio-inspired flower pollination algorithm,
Procedia Comput. Sci. 62 (2015) 435–441.
interface that can serve as an important tool in the forecast-
[12] P. Civicioglu, E. Besdok, A conceptual comparison of the Cuckoo-search,
ing of OPEC petroleum consumption for the authorities in OPEC, particle swarm optimization, differential evolution and artificial bee colony
International Energy Agency, Energy Information Administration, algorithms, Artif. Intell. Rev. 39 (4) (2013) 315–346.
[13] V. D’Orangeville, M.A. Mayers, M.E. Monga, M.S. Wang, Efficient cluster
Organization for Economic Cooperation and Development, and
labeling for support vector clustering, IEEE Trans. Knowl. Data Eng. 25 (11)
British Petroleum. (2013) 2494–2506.
[14] EIA, Independent Statistics & Analysis U.S Energy Information Administration.
http://www.eia.gov, 2014 (accessed 16.12.14).
6. Conclusions and future work [15] Energy Information Administration of the US Department of Energy, 2014.
[Online]. Available: http://www.eia.gov/.
[16] L. Ekonomou, Greek long-term energy consumption prediction using artificial
In this research, an alternative approach to improving OPEC neural networks, Energy 35 (2) (2010) 512–517.
petroleum consumption forecasting accuracy and convergence [17] European Statistics. http://ec.europa.eu/eurostat, 2014.
speed is proposed. This is a different choice for proper monitor- [18] J.I. Fister, X.-S. Yang, I. Fister, J. Brest, D. Fister, A brief review of
nature-inspired algorithms for optimization, Elektrotehniski Vestnik 80 (3)
ing and controlling OPEC petroleum consumption. In this paper, (2013) 116–122.
NN training based on the FP algorithm is proposed rather than [19] D. Gately, N. Al-Yousef, H.M. Al-Sheikh, The rapid growth of OPEC s domestic
training with commonly used meta-heuristic algorithms to build oil consumption, Energy Policy 62 (2013) 844–859.
[20] F.J. Hair, W.C. Black, J.B. Babin, R.E. Anderson, Multivariate Data Analysis,
a model for OPEC petroleum consumption forecasting. Established Pearson Prentice Hall, New Jersey, 2010.
meta-heuristic algorithms such as GA, ABC, APSO and BP are used [21] S. Haykin, Neural Networks, 2nd ed., Prentice Hall, New Jersey, 1999.
to train the NN for the purpose of comparing the results with [22] M.A. Kaboudan, Compumetric forecasting of crude oil prices, in: The
Proceedings of IEEE Congress on Evolutionary Computation, Seoul, South
those of our proposed method. It was found that the method pro-
Korea, 2001, pp. 283–287.
posed in this paper performs better than the GA, ABC, APSO and BP [23] J. Kennedy, J.F. Kennedy, R.C. Eberhart, Y. Shi, Swarm Intelligence, Morgan
in terms of convergence speed, forecasting accuracy and robust- Kaufmann, San Fransisco, CA, 2001.
58 H. Chiroma et al. / Applied Soft Computing 48 (2016) 50–58

[24] Y. Jin, A comprehensive survey of fitness approximation in evolutionary solutions for predicting temperature settings in a continuous annealing
computation, Soft Comput. 9 (2005) 3–12. furnace, Appl. Soft Comput. 35 (2015) 13–28.
[25] S.H. Jung, Queen-bee evolution for genetic algorithms, Electron. Lett. 39 (6) [39] M. Subotic, M. Tuba, N. Bacanin, D. Simian, Parallelized cuckoo search
(2003) 1. algorithm for unconstrained optimization, in: Proceedings of the 5th WSEAS
[26] D. Karaboga, B. Akay, C. Ozturk, Artificial bee colony (ABC) optimization Congress on Applied Computing Conference, and Proceedings of the 1st
algorithm for training feed-forward neural networks, in: Modeling Decisions International Conference on Biologically Inspired Computation, World
for Artificial Intelligence, Springer, Berlin, Heidelberg, 2007, pp. 318–329. Scientific and Engineering Academy and Society (WSEAS), 2012, pp. 151–156.
[27] H. Ma, Y. Wu, A grey forecasting model based on BP neural network for crude [40] L. Suganthi, A.A. Samuel, Energy models for demand forecasting—A review,
oil production and consumption in China, in: Proceedings of IEEE Computer Renew. Sustain. Energy Rev. 16 (2) (2012) 1223–1240.
Society Third International Symposium on Information Processing, Qingdao, [41] E.G. Talbi, Metaheuristics: From Design to Implementation, John Wiley &
China, 2010, pp. 105–109. Sons, Hoboken, NJ, USA, 2009.
[28] K.I. Matsumoto, V. Voudouris, D. Stasinopoulos, R. Rigby, C. Di Maio, Exploring [42] R. Wang, Y. Zhou, Flower pollination algorithm with dimension by dimension
crude oil production and export capacity of the OPEC Middle East countries, improvement, Math. Probl. Eng. (2014), http://dx.doi.org/10.1155/2014/
Energy Policy 48 (2012) 820–828. 481791, 9 pages.
[29] S. Mollaiy-Berneti, Optimal design of adaptive neuro-fuzzy inference system [43] P. Werbos, The Roots of the Backpropagation: From Ordered Derivatives to
using genetic algorithm for electricity demand forecasting in Iranian industry, Neural Networks and Political Fore-Casting, John Wiley and Sons, Inc, New
Soft Comput. (2015) 1–10, http://dx.doi.org/10.1007/s00500-015-1777-3. York, 1993.
[30] K. Movagharnejad, B. Mehdizadeh, M. Banihashemi, M.K. Sheikhi, Forecasting [44] X.-S. Yang, S. Deb, S. Fong, Accelerated particle swarm optimization and
the differences between various commercial oil prices in the Persian Gulf support vector machine for business optimization and applications, in:
region by neural networks, Energy 36 (2011) 3979–3984. Networked Digital Technologies, Springer, Berlin, Heidelberg, 2011, pp.
[31] S.P. Niculescu, Artificial neural networks and genetic algorithms in QSAR, J. 53–66.
Mol. Struct. THEOCHEM 622 (1) (2003) 71–83. [45] X.-S. Yang, S. Deb, Cuckoo search: recent advances and applications, Neural
[32] K. Padmakumari, K.P. Mohandas, S. Thiruvengadam, Long term distribution Comput. Appl. 24 (1) (2014) 169–174.
demand forecasting using neuro fuzzy computations, Int. J. Electr. Power [46] X.-S. Yang, Cuckoo search and firefly algorithm: overview and analysis, in:
Energy Syst. 21 (5) (1999) 315–322. X.S.I. Yang (Ed.), Cuckoo Search and Firefly Algorithm, Springer International
[33] G.Z. Peter, B.E. Patuwo, M.Y. Hu, A simulation study of artificial neural Publishing, 2014, pp. 1–26.
networks for nonlinear time-series forecasting, Comput. Oper. Res. 28 (2001) [47] X.S. Yang, Flower pollination algorithm for global optimization,
381–396. Unconventional Computation and Natural Computation (2012) 240–249.
[34] C. Quek, M. Pasquier, N. Kumar, Novel recurrent neural network-based [48] L. Yu, W. Shouyang, K.K. Lai, A neural-network-based nonlinear
prediction system for option trading and hedging, Appl. Intell. 29 (2008) metamodeling approach to financial time series forecasting, Appl. Soft
138–151. Comput. 9 (2009) 563–574.
[35] J. Riget, J.S. Vesterstrøm, A Diversity-Guided Particle Swarm Optimizer-the [49] S. Yu, K. Wang, Y.M. Wei, A hybrid self-adaptive Particle Swarm
ARPSO, 2, Dept. Comput. Sci., Univ. of Aarhus, Aarhus, Denmark, Tech. Rep., Optimization–Genetic Algorithm–Radial Basis Function model for annual
2002. electricity demand prediction, Energy Convers. Manag. 91 (2015) 176–185.
[36] D.E. Rumelhart, J.L. McClelland, Parallel Distributed Processing Explorations in [50] L. Yu, W. Dai, L. Tang, J. Wu, A hybrid grid-GA-based LSSVR learning paradigm
the Theory of Cognition, vol. 1, MIT Press, Cambridge, MA, 1986. for crude oil price forecasting, Neural Comput. Appl. (2015) 1–23, http://dx.
[37] M.G. Sahab, V.V. Toropov, A.H. Gandomi, Traditional and modern structural doi.org/10.1007/s00521-015-1999-4.
optimization techniques—theory and application, in: A.H. Gandomi, X.Y. Yang, [51] Y.H. Zweiri, J.F. Whidborne, L.D. Sceviratne, A three-term backpropagation
S. Talatahari, A.H. Alavi (Eds.), Metaheuristic Applications in Structures and algorithm, Neurocomputing 50 (2003) 305–318.
Infrastructures, Elsevier, Waltham, MA, 2013 (Chapter 2).
[38] A. Sanz-Garcia, J. Fernandez-Ceniceros, F. Antonanzas-Torres, A.V.
Pernia-Espinoza, F.J. Martinez-de-Pison, GA-PARSIMONY: A GA-SVR approach
with feature selection and parameter optimization to obtain parsimonious