Received: 28 June 2017 Revised: 6 February 2019 Accepted: 10 February 2019 Published on: 13 March 2019

DOI: 10.1002/joc.6037

RESEARCH ARTICLE

Long-term modelling of wind speeds using six different heuristic

artificial intelligence approaches
Saman Maroufpoor1 | Hadi Sanikhani2 | Ozgur Kisi3 | Ravinesh C. Deo4 |
5
Zaher Mundher Yaseen

1
Irrigation and Reclamation Engineering
Department, Faculty of Agriculture, University of Wind speed is an essential component that needs to be determined accurately, espe-
Tehran, Tehran, Iran cially over long-term periods for various engineering and scientific purposes
2
Water Engineering Department, Faculty of including renewable energy productions, structural building sustainability and
Agriculture, University of Kurdistan,
others. In this study, six different heuristic methods: multi-layer perceptron artifi-
Sanandaj, Iran
3
cial neural networks, (ANN), adaptive neuro-fuzzy inference system (ANFIS) with
Faculty of Natural Sciences and Engineering, Ilia
State University, Tbilisi, Georgia grid partition (GP), ANFIS with subtractive clustering (SC), generalized regression
4
Centre for Sustainable Agricultural Systems and neural networks (GRNN), gene expression programming (GEP) and multivariate
Centre for Applied Climate Sciences, School of adaptive regression spline (MARS) are developed to model monthly wind speeds
Agricultural, Computational and Environmental using meteorological input information. The atmospheric pressure, temperature,
Sciences, University of Southern Queensland,
Springfield Central, Queensland, Australia
relative humidity and rainfall values are obtained from Jolfa and Tabriz meteoro-
5
Sustainable Developments in Civil Engineering
logical stations, Iran, and are used to build the proposed predictive models.. Differ-
Research Group, Faculty of Civil Engineering, ent statistical indicators are computed to evaluate and comprehensively assess the
Ton Duc Thang University, Ho Chi Minh City, performance of the six heuristic methods. Over the testing phase, the ANFIS-GP
Vietnam
and GRNN models are seen to exhibit the highest predictive performance for the
Correspondence
Ozgur Kisi, Faculty of Natural Sciences and
Jolfa and Tabriz stations, respectively. That is, the maximum coefficient of deter-
Engineering, Ilia State University, Tbilisi 0162, mination are found to be 0.874, 0.858, 0.850, 0.849, 0.847 and 0.826, for the
Georgia. GRNN, ANFIS-GP, ANFIS-SC, ANN, GEP and MARS models, respectively, for
Email: [Link]@[Link]
Jolfa station, respectively, revealing the superiority of GRNN over the five counter-
part models. The results show the generalization capability of the tested heuristic
artificial intelligence techniques for both study stations, and therefore could be
explored for windspeed prediction and various decisions made in regards to climate
change studies.

KEYWORDS

gene expression programming, multivariate adaptive regression spline, neural

networks, neuro-fuzzy, prediction, wind speed

1 | INTRODUCTION (Liu et al., 2015a). Long-term (e.g., monthly or seasonal)

prediction of wind speed can also help us to devise a precise
In recent years, the prediction of wind speed has become and reliable evaluation of the wind speed characteristics,
very important for the long-term climate change studies, especially for renewable energy applications for any area
wind energy applications. In their practical implementations, where WECS needs to be installed (Olaofe and Folly, 2012).
accurate and reliable wind speed forecasting can reduce the Therefore, the development and validation of novel
possibility of the wind power breakdown and protect the approaches for windspeed prediction is a paramount task that
security of the wind energy conversion systems (WECS) is undertaken by scientists and engineers to help explore the

Int J Climatol. 2019;39:3543–3557. [Link]/journal/joc © 2019 Royal Meteorological Society 3543

3544 MAROUFPOOR ET AL.

future potential of wind resource as a clean energy option The primary motivation of the current study is to investi-
including the desire to combat climate change and study cli- gate the feasibility of six different heuristic, artificial intelli-
mate shifts and variability patterns. gence (AI) methods to model wind speeds over two
Wind speed forecasting methods can be divided into relatively isolated case study regions located in Iran region.
physical, statistical and computational or intelligent (heuris- The methods applied in this study are based on an ANN,
tic) models. Every type of model has own merits and weak- ANFIS-GP, ANFIS-SC, GRNN, GEP and MARS algorithm.
nesses. The purpose of statistical methods is to explain the The applied predictive models are built based on four causal
relationships between historical and present time series of climatological variables including the atmospheric pressure,
wind speed data at the location of interest by the general temperature, relative humidity and rainfall data sets. In addi-
recursive methods (Tascikaraoglu and Uzunoglu, 2014). In tion, the modelling procedure aimed to integrated the sea-
sonality component of the wind speed as additive predictor
such an application, Erdem and Shi (2011) found that the
variable to the climatological information used to model
efficacy of autoregressive moving average-based approach
windspeeds. Given the uniqueness of the methods in their
to the prediction of the tuple of wind speed and its direction.
own right, such applying approaches based on different
However, in these conventional statistical methods, the wind
types of learning intelligent algorithms that are expected to
speed data are assumed to be normally distributed. Notwidh-
extract data features quite differently from one another, the
standing this, it is clear that the real wind speed series are
utilization of the prescribed predictive models for windspeed
not normally distributed. To address this limitation of the
prediction is considered an interesting research endeavour
statistical methods, in the past few decades, a number of data for the present study sites.
intelligent approaches such as the adaptive neuro-fuzzy
inference system (ANFIS), artificial neural networks
(ANNs) and genetic programming (GP), have been success- 2 | M AT E RI A L S AN D M E T H O DS
fully been applied for time series-based wind speed forecast-
ing. For example, some studies applied the wavelet-GA 2.1 | Analysis of meteorological variables
(Genetic Algorithm)-MLP (Multi-Layer Perceptron) model
In this study, meteorological input variables measured at
and the wavelet-PSO (Particle Swarm Optimization)-MLP
Jolfa and Tabriz stations, East Azarbaijan province, Iran
model for the realization of high-precision multi-step wind
were applied into the six heuristic models. The input vari-
speed prediction (Troncoso et al., 2015).
ables including wind speed, temperature, atmospheric pres-
In studies elsewhere, the work of Liu et al. (2015b)
sure, relative humidity and rainfall which were intrinsincly
applied the Ada-boost Algorithm (ABA) to build an intelli-
related to the objective variable (i.e., wind speed), have been
gent framework including several MLP neural networks for collected for the period 1996–2010. The geographical coor-
wind speed prediction. Kong et al. (2015) applied reduced dinates of the study stations are shown in Table 1. Figure 1
support vector machines with feature selection for wind illustrates the location of each study station. Azerbaijan's cli-
speed prediction. In one experiment, a total of eight models mate is generally hot and dry, but due to the topographical
including Persistent Model, ARIMA model, Multiple Linear diversity, the climates are different. The province has been
Regression and five different neural networks were used to under the influence of cold winds from the north and moist
predict the wind speed 1-hr trial. A comparison showed that winds from Siberia, the Black Sea, the Mediterranean and
the BP (Back Propagation) neural network model displayed the Atlantic ocean. The time variation graphs of each meteo-
optimal results (Su et al., 2014). Zhang et al. (2012) utilized rological variable are shown in Figure 3. The figure shows
the PSO algorithm to optimize the first-order adaptive coeffi- the relationships between the temperature, relative humidity
cient (FAC) and the second-order adaptive coefficient (SAC) and rainfall for Jolfa and Tabriz meteorological station, but
for wind speed prediction. Li and Shi (2010) used three dif- for wind speed and atmospheric pressure two stations show
ferent ANNs for wind speed prediction. Data intelligent different characteristics.
models in this study included the adaptive linear element Tabriz is famous with agriculture. It should be noted
neural network, the BP neural network (BPNN) and the here, the land use of studied regions included irrigated farm-
radial basis function neural network (RBFNN). Moreover, land, rainfed farmland, salty and flood plains rangeland and
Haque et al. (2012) utilized a performance analysis of short- forest (see Figure 2).
term wind speed prediction method based on heuristic
TABLE 1 Geographic of the study stations
models formulated through a RBFNN, a BPNN, and an
ANFIS modelling framework. However, the application of Station name

data-intelligent models for windspeed prediction in the pre- Statistics Jolfa Tabriz
Latitude (
E) 35.56 38.05
sent study region is yet to be explored and a comparison of
Longitude (
N) 45.36 46.17
different windspeed prediction models has also not
Elevation (m, ASL) 736 1,364
been made.
MAROUFPOOR ET AL. 3545

FIGURE 1 Location of the test stations

FIGURE 2 The land use map of studied regions [Colour figure can be viewed at [Link]]

As observed from Figure 3, the seasonality pattern of the atmospheric pressure, the maximum and minimum value are
wind speed and thus a fifth input variables has been included seen to occur in Jolfa (0.94 bar) and Tabriz (0.85 bar) sta-
to the four meteorological variables to incorporate the wind tions, respectively.
speed seasonality component. The effect of periodicity was In Table 2, the correlation matrix of the meteorological
investigated by adding a component α which takes values data used in the study is shown. For both stations, there
between 1 and 12 according to the month of the year to be appears to be a high rate of correlation, evidenced by the
forecasted. Here α indicates the periodicity component coefficient, between the wind speed data and various meteo-
(Sanikhani and Kisi, 2012; Sanikhani et al., 2018). rological variables used in the heuristic models. Wind speed
The maximum and minimum monthly wind speed occurs demonstrates a positive correlation with atmospheric temper-
in Jolfa station with value of about 6.43 and 0.36 m/s. For ature for both stations, while it is negatively correlated with
3546 MAROUFPOOR ET AL.

150
Tabriz Jolfa

Rainfall (mm)
100

50

0
0 20 40 60 80 100 120 140 160 180

40 Tabriz Jolfa
Temperature (0C)

30

20

10

0
0 20 40 60 80 100 120 140 160 180
-10

100 Tabriz Jolfa

Relative humidity (%)

80

60

40

20

0
0 20 40 60 80 100 120 140 160 180

Tabriz Jolfa
Atmospheric Pressure (bar)

0.95

0.9

0.85

0.8
0 20 40 60 80 100 120 140 160 180

7 Tabriz Jolfa
Wind speed (m/s)

6
5
4
3
2
1
0
0 20 40 60 80 100 120 140 160 180

Month of the years (1996-2010)

FIGURE 3 Mean monthly variation of meteorological variables used in this study [Colour figure can be viewed at [Link]]

atmospheric pressure and relative humidity data. For Jolfa 2.2 | Artificial neural network
station, the correlations are higher than those of Tabriz sta- ANN processing system, inspired by the human brain, has
tion. Among the different input variables analysed for these own functions for feature extraction (Haykin, 1999). ANN
stations, the rainfall data do not seem to have a high correla- has been developed as a computational model of human per-
tion with the wind speed data, which seems to suggest differ- ception or neural biology, and a neural network, detected by
ent climatic conditions at the two study locations. their structure, which represents the pattern of relationship
MAROUFPOOR ET AL. 3547

TABLE 2 Correlation matrix of the meteorological variables for Tabriz 2.3 | Adaptive neuro fuzzy inference system (ANFIS)
and Jolfa stations
Fuzzy computing system is based on fuzzy set theory, fuzzy
Wind Atmospheric Relative
Variables Speed Pressure Temperature Humidity Rainfall rules and fuzzy reasoning. The combination of fuzzy sys-
Tabriz tems with neural networks (resulting in an ANFIS model) is
Wind speed 1 −0.772 0.719 −0.718 −0.166 a powerful predictive tool (Jang and Sun, 1995). The neuro-
Atmospheric 1 −0.559 0.524 0.00004 fuzzy system is related to Takagi Sugeno model, the weights
pressure of the neural network are equivalent to the parameters of the
Temperature 1 −0.924 −0.297 fuzzy system. A sample architecture of an ANFIS model is
Relative 1 0.411 shown in Figure 4b. This structure was first introduced by
humidity
Jang (1993). Although ANN is a powerful tool for modelling
Rainfall 1
various natural processes that contain some form of feature
Jolfa data set, it also has some degree of deficiency. The combina-
Wind speed 1 −0.825 0.791 −0.790 −0.130 tion of neural network and fuzzy logic results in an
Atmospheric 1 −0.869 0.841 −0.115 improved speed, fault tolerance and better compliance than a
pressure standalone neural networks or a fuzzy logic model (Setlak,
Temperature 1 −0.895 0.006 2008). The proposed ANFIS model is thus a multi-layer neu-
Relative 1 0.155 ral network-based fuzzy system which has a total of five
humidity
layers in their structure. In this continuous structure, input
Rainfall 1
and output nodes delegate the input states and output
response separately, and also in the hidden layer, the nodes
between the hidden neurons. ANN model includes the function acts as membership functions (MFs) with a set of
method of specifying the connection weights and the activa- rules. This can overcome the challenges of the feed-forward
tion functions within the hidden neuronal layer (Fausett, multilayer network (Slim, 2006). ANFIS can be designed
1994). It consists of an input layer, hidden layers and an out- with various model identification approaches of the Sugeno
put layer. The input variables into the ANN model for wind model, for example, using the grid partitioning (GP) or using
speed estimation in this study are the atmospheric pressure, the subtractive clustering (SC). In the ANFIS-GP model,
temperature, relative humidity, rainfall and wind speed sea- input space is partitioned into rectangular sub-spaces by uti-
sonality (Figure 4a). These layers are able to analysis the lizing various local fuzzy regions. The least square method
data in the interconnecting processing elements known as regarding the MFs type and grid partition is thus utilized for
the neurons or nodes (Kalogirou, 2001). a computation of the fuzzy sets and model parameters
The benefits of ANN model compared to the other statis- (Abonyi et al., 1999). For applying the ANFIS-GP model,
tical models are: the number of input variables ought to be under 6, which
accords to various literature data (Sanikhani et al., 2012). In
1. Their application does not need a past knowledge of the this paper, fiveinput variables are utilized for wind speed
processes in view of their black-box property. prediction and thus the ANFIS-GP can be applied effec-
2. They have the natural specification of non-linearity since tively. For more information about the ANFIS model, one
neurons activate a non-linear filter called an activation can refer to the work of Jang (1993).
function
3. They adjust to the changes in problematic situations for 2.4 | Gene expression programming (GEP)
enhances feature extraction.
Gene expression programming (GEP) algorithm has been
4. They are able to acknowledge numerous inputs having
inspired from biological evolution. It was introduced by Fer-
diverse specifications (Rahnama and Nouri, 2008).
reira (2001), as an extension of the GP model (Koza, 1992).
GEP is able to preserve a few properties of the GA
In this study, different backpropagation algorithms com- (Goldberg, 1989), where the chromosomes in the GP model
prised of the Levenberg–Marquardt (lm), gradient descent act as representatives from the trees. However, GEP is an
(gd), gradient descent with adaptive learning rate (gda), gra- improved version of the GP model as it presents as a linear
dient descent with momentum and adaptive learning rate fixed-length string model containing the merits of the GP
back- propagation (gdx) and scaled conjugate gradient (scg) model.
equations have been utilized in the problem of monthly wind In GEP, the individuals are encoded by the chromosomes,
speed prediction. The ANN method was selected in this which consist of genes structurally organized in head and the
study based on its very efficient skills as a well-established tail where each gene is represented by the genotype. With this
heuristic method. To determine the optimal number of neu- formulation of the algorithm, the GEP model becomes an
rons, trial and error method is used in the ANN model. expression of the phenotype and is therefore, referred to as
3548 MAROUFPOOR ET AL.

(a)
Neuroni
Wi1
X1
Wi2
X2
Wi3
X3
Activation Function
Xn Win Wi0

Input Weights Summation Activation Output

Atmospheric Pressure

Relative Humidity

Temperature

Rainfall

Seasonality Component

(b) Atmospheric Temperature Relative Seasonality

Rainfall Component
Pressure Humidity

Layer 1 Layer 2 Layer 3 Layer 4 Layer 5

П N 1
Atmospheric A1
Pressure П N
2
A2
П N
3
П N
B1 4
Temperature П
B2
П

Relative C1
f
Σ
Humidity
C2

C3
Rainfall D1
D2

F1 П N
Seasonality
Component П N
F2
П N

Premise parameters Conseqent parameters

FIGURE 4 (a) ANN model architecture for wind speed prediction, (b) first-order Sugeno ANFIS architecture (Jang, 1993) [Colour figure can be viewed at
[Link]]

the gene expression tree (ET). To create the chromosomes, function. Various genetic operators are used within and
the genes are connected with each other by means of the between the chromosomes. Evolution is applied until it
available linking functions. Some of these individuals forms reached a satisfactory standard (Ferreira, 2006).
the specimen population, which undergoes an evolution by The application of the GEP model for solving practical
calculating the expression from each chromosome, using pre- problems has been investigated quite significantly in recent
defined genetic operators and computation of the the fitness years (Ustoorikar and Deo, 2008; Kisi and Guven, 2010;
MAROUFPOOR ET AL. 3549

Create Chromosomes of Initial Population parameters applied are summarized as: Number of chromo-
somes: 30, number of genes: 3, head size: 8, mutation rate:
0.044, inversion rate: 0.1, one point recombination rate: 0.3,
Express Chromosomes
two point recombination rate: 0.3, gene transposition rate:
0.1, insertion sequence transposition rate: 0.1, root insertion
Execute Each Program
sequence transposition: 0.1, gene recombination rate: 0.1.
The mentioned parameters valued obtained using trial and
Evaluate Fitness
error during modelling procedure.
Terminat
Iterate or Terminate? End

Iterate 2.5 | Multivariate adaptive regression spline

Keep Best Program
Friedman (1991) introduced the multivariate adaptive regres-
sion spline (MARS) model. MARS is a generalization of the
Select Programs
stepwise regression or modified regression trees (Hastie
et al., 2009). In the univariate form of the MARS model, the
Replication response variable (y) is estimated using a specific pattern
called the basis function (i.e., mirror functions or hockey
Mutation stick functions) of the independent variable (x). This follows
(Friedman, 1991):
IS transposition


x−t x>t
h 1 ðx Þ = ðx − t Þ + = ð1Þ
RIS transposition 0 x≤t
Reproduction


Gene Transposition t−x x<t
h2 ðxÞ = ðt − xÞ + = ð2Þ
0 x≥t
1-Point Recombination

where t is the node constant and + indicates the positive sec-

2-Point Recombination
tion. In practice, the t node is one of the observations (i.e., x)
and h1(x) and h2(x) are hockey stick functions where both of
Gene Recombination
them are the reflected pair in the node t. Figure 6 shows the
Create Chromosomes of Initial Population basis functions for t = 0.5.
In a MARS model, the C value acts as a collection to
FIGURE 5 The flowchart of the gene expression algorithm consider all of the reflected pairs where {x1, x2, x3, …, xn}
(Ferreira, 2006) values are observed as
 
C = ðx – t Þ + , ðt – x Þ + 8t 2 fx1 , x2 , x3 , …, xn g ð3Þ
Zakaria et al., 2010; Khatibi et al., 2014). In the present
study, the soft computing program known as GeneXpro The general structure of a univariate MARS regression
Tools 4.0 was applied to predict the wind speed data. A for estimating y accords to Equation (4) where hi is one of
flowchart of the GEP model is shown in Figure 5. The the functions of C:

X
2m
0.6
x - ….. 0.5…. Y = f 2m ðxÞ = β0 + βi hi ðxÞ ð4Þ
i=1
0.5
In Equation (4), the coefficients of β0 and β1 are calcu-
0.4
Basic function

lated based on the sum of square errors. The development of

0.3 the MARS model consists of three stages (Friedman, 1991):
0.2
1. determination of the MARS structure (i.e., creating an
0.1
initial pattern by the forwarding method);
0 2. the modification of the initial MARS model (i.e., revi-
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
sion of the initial patterns); and
Node
3. the choice of the optimal MARS model structure with
FIGURE 6 Basis function for t = 5 (Ferreira, 2006) an optimal number of basis functions
3550 MAROUFPOOR ET AL.

TABLE 3 Testing results of the ANN training algorithm for the Tabriz and 2.6 | Generalized regression neural networks
Jolfa stations
Generalized regression neural network (GRNN) formulated
Model Iteration RMSE,
Training algorithm structure number (m/s) by Specht (1991), contains networks are classified in the cat-
Tabriz egory of probabilistic neural networks. GRNN, like other
Levenberg–Marquardt 5-2-1 1,000 0.439 probabilistic neural networks, needs only a fraction of the
Gradient descent 5-8-1 50,000 0.444 training samples that required by the backpropagation neural
Gradient descent with adaptive 5-2-1 50,000 0.439 network model. Therefore the use of a probabilistic neural
learning rate network model is especially advantageous due to its ability
Gradient descent with momentum 5-8-1 50,000 0.444
to converge to the underlying function of the input data with
Gradient descent with momentum 5-2-1 50,000 0.444
and adaptive learning rate
only a few training samples. This makes GRNN a very
Scaled conjugate gradient 5-2-1 1,000 0.443
impressive tool to perform forecasts and for a comparisons
Jolfa of system efficiency in real operations. More detailed infor-
Levenberg–Marquardt 5-4-1 1,000 0.498 mation about the GRNN can be found in Specht (1991).
Gradient descent 5-8-1 50,000 0.570
Gradient descent with adaptive 5-9-1 50,000 0.502
learning rate 2.7 | Evaluation of model performance
Gradient descent with momentum 5-2-1 50,000 0.571
For the model evaluation, eight distinctive statistical evalua-
Gradient descent with momentum 5-5-1 50,000 0.507
and adaptive learning rate
tion criteria including the determination coefficient (R2), root
Scaled conjugate gradient 5-8-1 1,000 0.541
mean square error (RMSE), mean absolute error (MAE), rel-
ative root mean square error (RRMSE), relative mean abso-
RMSE: root mean square error.
lute error (RMAE), Nash-Sutcliffe model efficiency
The different patterns (or features) are fitted over the coefficient (NS), Willmott Index (WI) and Legates and
input-target data in the MARS model's structure and the McCabe index (LMI) were used. The mentioned indices can
modification step is applied to calculate the sum of the be ascertained by the accompanying equations:
square error (SSE). The final MARS model is chosen in the
model selection stage based on the lowest generalized cross
validation (GCV) statistic. TABLE 5 Testing results of different GEP basic and linking functions for
the parse tree for Tabriz and Jolfa stations
SSEj
GCVj = vm j = 1, 2, …, 2n – 2 ð5Þ Definition RMSE (m/s)
1− nj Tabriz

In Equation (5), mj is the number of basis functions, SSEj F1 +, −,×, 0.448

is the sum of squares error, v is the smoothing parameter, F2 +, −,×,,lnx, ex 0.486

pffi pffi
and n is the number of observations in the training dataset. F3 +, −,×,, 3 , , x3, x2 0.519
F4 +, −,×,,lnx, ex, p ffi p ffi , x3 , x2
3 , 0.437
F5 +, −, ×, ,lnx, e , p
x ffi pffi, x3, x2, sinx, cosx,
3 , 0.460
Arctg x
TABLE 4 Testing results of different ANFIS-GP structures for Tabriz and Linking function
Jolfa stations Addition 0.472
Type MFs Model structure RMSE (m/s) Multiplication 0.469
Tabriz Subtraction 0.476
Gaussian (gaussmf) 2-2-2-2-2 0.566 Division 0.460
Two Gaussian (gauss2mf) 2-2-2-2-2 0.483 Jolfa
Generalized bell (gbellmf) 4-2-2-2-2 0.619 F1 +, −,×, 0.634
dsig 2-2-2-3-2 0.436 F2 +, −,×,,lnx, ex 0.713
pffi pffi
psig 2-2-2-2-2 0.437 F3 +, −,×,, 3 , , x3, x2 0.742
pffi pffi
Pi-shaped (pimf) 2-2-2-2-2 0.577 F4 +, −,×,,lnx, ex, 3 , , x3, x2 0.615
Jolfa pffi pffi
F5 +, −, ×, ,lnx, ex, 3 , , x3, x2, sinx, cosx, 0.523
Gaussian (gaussmf) 2-2-2-3-3 0.631 Arctg x
Two Gaussian (gauss2mf) 2-3-2-2-3 0.595 Linking function
Generalized bell (gbellmf) 2-2-2-2-3 0.548 Addition 0.546
dsig 2-3-2-3-3 0.620 Multiplication 0.523
psig 2-3-2-3-3 0.616 Subtraction 0.534
Pi-shaped (pimf) 2-3-2-2-3 0.719 Division 0.646
MAROUFPOOR ET AL. 3551

(a)

(b)
Constants Sub-ET 1 Sub-ET 2 Sub-ET 3
Sub-ET 1 Atan
3Rt
C0 = -8.509
C1 = -8.567 Atan
d1
Sub-ET 2 + + Atan
C0 = 1.555
Atan
C1 = 7.717
Sin C1
Sub-ET 3 Cos Atan
+
C0 = -6.472
Atan
C1 = 2.547
Sqrt Atan
Functions
+ X2 C0 Ln
d5 d4 C0 d5
Sin, Atan
Cos, x3, d4
d5
x2, Ln, C1 d4
, , Terminals
d1= Temperature
ex
d3= Relative Humidity
d4 = Atmospheric Pressure
d5= Seasonality Component

FIGURE 7 Expression trees of the best GEP models (a) Tabriz, (b) Jolfa) [Colour figure can be viewed at [Link]]
3552 MAROUFPOOR ET AL.

TABLE 6 Testing results of the optimal ANN, GRNN, ANFIS-GP, ANFIS-SC, GEP, and MARS models for Jolfa and Tabriz stations

Training algorithm R2 RMSE (m/s) RRMSE (%) MAE (m/s) RMAE (%) NSE WI McV
Tabriz
ANN 0.889 0.498 24.58 0.392 25.87 0.834 0.957 0.611
ANFIS-GP 0.863 0.548 27.05 0.452 28.51 0.800 0.947 0.552
ANFIS-SC 0.873 0.523 25.82 0.405 28.97 0.817 0.952 0.598
GEP 0.861 0.523 25.82 0.432 30.68 0.817 0.947 0.572
GRNN 0.874 0.537 26.51 0.428 26.13 0.807 0.852 0.576
MARS 0.885 0.514 25.41 0.431 24.99 0.823 0.898 0.573
Jolfa
ANN 0.832 0.439 15.08 0.361 12.33 0.539 0.877 0.346
ANFIS-GP 0.804 0.436 14.96 0.357 12.135 0.546 0.873 0.353
ANFIS-SC 0.821 0.437 15.01 0.354 11.91 0.543 0.874 0.359
GEP 0.789 0.437 14.99 0.345 11.62 0.544 0.873 0.375
GRNN 0.817 0.426 14.61 0.345 11.43 0.571 0.829 0.375
MARS 0.821 0.436 14.96 0.357 12.02 0.546 0.837 0.353

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN 2 = 132 data points) while the remaining 4 years of data
i = 1 ðWO − WM Þ
RMSE = ð6Þ (i.e., 48 data) were utilized for testing the heuristic models.
N
Six different backpropagation training algorithms, based
PN on lm, gd, gda, gdx and scg with different iterations, were
jWiO − WiM j
MAE = i = 1 ð7Þ utilized in computing the ANN model weights and biases.
N
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P ffi Tables 3 provides the test results of the optimal ANN model
N
ðWiO − WiM Þ2 derived for each study station. In this table, the model struc-
i=1
N
RRMSE = O ð8Þ ture denoted as 5-2-1 indicates the ANN model composed of
W
PN   a total of five inputs corresponding to the monthly atmo-
jWiO − WiM j
i=1 WiO spheric pressure, temperature, relative humidity rainfall and
RMAE = ð9Þ wind speed seasonality, two hidden layer and one output
N
PN layer nodes. Various hidden nodes (ranging from a value of
ðWiO − WiM Þ2
NS = 1 − PiN= 1 ð10Þ 1 to 10) were investigated, and the best neuronal arrange-
 2
i = 1 ðWiO − WO Þ ment that provided the least RMSE in the testing stage was
PN 2 chosen. For the hidden and output nodes, the tangent sig-
i = 1 ðWiO − WiM Þ
WI = 1 − PN ð11Þ moid and linear activation functions were utilized, since
ðjWiM − W  O j + jWiO − W
 O j Þ2
i=1
these equations provided better results than the other transfer
PN
jWiO − WiM j functions. According to the results of Tables 3, it is clear that
LMI = 1 − PiN= 1 ð12Þ
 the scaled conjugate gradient and gradient descent with
i = 1 jWiO − WO j
momentum and adaptive learning rate backpropagation algo-
The N is the number of data set, WiO is the mean of
rithms have better performances than the other training algo-
observed wind speed, WiM and WO denote the wind speed rithms in the testing stage for both Tabriz and Jolfa stations.
values produced by various models and the observed Different number of MFs of Gaussian (gaussmf), two
monthly wind speed values, respectively. Gaussian (gauss2mf), pi-shaped (pimf), difference between
two sigmoidal (dsig), product of two sigmoidal (psig) and
generalized bell (gbellmf) were used for the ANFIS-GP
3 | AP PLICA TION A ND RES ULT S models and the iteration number was set to 100 since more
than this value provided not effect on model's results. The
In this paper, the application of ANN, ANFIS-GP, ANFIS- optimal number of MFs for both stations that provided the
SC, GEP, GRNN and MARS models are made for the pre- minimum RMSE in the testing stage was chosen for the
diction of monthly wind speed at two locations in Iran. The ANFIS-GP model. In order to deduce the best result, various
inputs used in the applied models for each of the study ANFIS-GP model structures were compared in Table 4. In
station are: atmospheric pressure, temperature, relative the table, the architecture denoted as 2-2-2-3-2 indicates an
humidity, rainfall and wind speed seasonality over monthly ANFIS-GP model comprised of 2, 2, 2, 3 and 2 Gaussian
time-scales. The models were trained (i.e., calibrated) by MFs for the inputs, atmospheric pressure, temperature, rela-
using 11 years of data (comprised of 11 years × 12 months tive humidity, rainfall and wind seasonality, respectively. It
MAROUFPOOR ET AL. 3553

(a) (b)

FIGURE 8 The percentage frequency of the absolute prediction error (m/s) for the ANN, ANFIS-GP, ANFIS-SC, GEP, MARS, and GRNN models with
periodicity included as predictor variable. (a) Tabriz, (b) Jolfa [Colour figure can be viewed at [Link]]

is clear from the table that the ANFIS-GP model with psig
and Generalized bell MF attains the highest performance in V monthly mean wind speed = sin − 8:5092 − 8:567 − a
the testing stage according to the RMSE and R2 statistics for rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
3 pffiffiffi 1:553
Jolfa and Tabriz stations, respectively. Different radii values × cos ð7:717d Þ − e + e +
e
were also tested, and the optimal radii value was determined × ½a tan ða tan ða tan ……ð ln ðd ÞÞÞÞ ð14Þ
to be 0.61 for the ANFIS-SC model.
In the above equations, the parameters a, c, d and e are
For the GEP models, different basis and linking func-
temperature, relative humidity, atmospheric pressure and
tions were utilized and the best one that yielded the lowest
wind speed seasonality, respectively. According to equa-
RMSE in the testing stage was chosen. The testing perfor-
tions, it is clear that the GEP model has not used the rainfall
mance of the GEP models are reported in Table 5. It is obvi-
(at both stations) and relative humidity (at Jolfa station). This
ously seen from the table that the best performance was
is due to the ability of GEP to automatically select inputs.
provided by the GEP model with an additional linking func-
This GEP's capability can prevent the multi-collinearity
tion for both study stations. It is important to mention that
problem.
the F5 operators was chosen for both study stations. Figure 7
The statistical evaluation criteria for the optimal heuristic
shows the ET for the optimal GEP model.
models are presented in Table 6 for Tabriz and Jolfa stations.
The equations of the optimal GEP model for Tabriz
Based on the tabulated results, the minimum absolute error
(Equation 13) and Jolfa (Equation 14) stations are as follows:
metrics are indicated for the implementation of ANN and
V monthly mean wind speed GRNN models, that are found to be superior compared to
ða + 7:927 + 2c + 7:927dÞ the other heuristic models for Tabriz and Jolfa stations,
= p ffiffiffiffiffiffiffiffiffiffi h i
d × ðb + d Þ
3 respectively. However, as a general overview, the applied
104:620 + d 2 × ðe − 6:061Þ2 + ðc + eÞ a + 4:090
a
models are seen to perform reliably for the investigated case
ð13Þ study regions.
3554 MAROUFPOOR ET AL.

5
y = 0.7915x + 0.2582 5 y = 0.7661x + 0.354 5 5
y = 0.8969x - 0.0762 y = 0.8861x - 0.0773
R2 = 0.8322 R2 = 0.8038 R2 = 0.8633
R2 = 0.8887
4 4 4
4
predicted

predicted

predicted
predicted
3 3
3
3
2 2
2
2
1 1
ANN-Tabriz ANN-Jolfa ANFIS-GP-Jolfa
ANFIS-GP-Tabriz
1
1 0 0
1 2 3 4 5
1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
observed
observed observed observed

5 5 5 5
y = 0.7686x + 0.3347 y = 0.7573x + 0.3876 y = 0.8963x - 0.076 y = 0.8152x + 0.1242
R2 = 0.8214 R2 = 0.7895 R2 = 0.8725 R2 = 0.8613
4 4
4 4

predicted
predicted

predicted

predicted
3
3 3
2 2

2 2
1 1
ANFIS-SC-Tabriz GEP-Tabriz ANFIS-SC-Jolfa GEP-Jolfa
1 1 0 0
1 2 3 4 5 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
observed observed observed observed

5 5 y = 0.7397x + 0.4397 5 5 y = 0.7787x + 0.1568

y = 0.7889x + 0.2765 y = 0.843x + 0.0175 R2 = 0.8744
R2 = 0.8169 R2 = 0.8852
R2 = 0.8211
4 4
4 4

3 3

predicted
predicted

predicted

predicted

3 3
2 2

2 2
1 1
MARS-Tabriz MARS-Jolfa MARS-Jolfa
GRNN-Tabriz
1 1 0 0
1 2 3 4 5 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
observed observed observed observed

FIGURE 9 The scatterplots of predicted and observed values of wind speed for the heuristic models [Colour figure can be viewed at [Link]]

Another way to verify the accuracy of the applied heuris- accuracy in prediction the monthly wind speed data for the
tic predictive models, is the assessment of the frequency Jolfa station when compared with the Tabriz station.
error encountered in different error brackets where more reli- Although it is not absoliutely clear, a plausible reason of this
able understanding can be extracted from this analysis. result may be the fact that there was a higher correlation of
Figure 8a and b demonstrates the frequency analysis for all each meteorological input data and the actual wind speed
the applied predictive models and for both meteorological data for the Jolfa station.
stations over the testing phase. Each bar in these figures
indicates the percentage counts. It is apparent from the
4 | C O NC L U S IO N
graphs that the ANN and GRNN models present consistently
better performance. Figure 9 shows another important graph-
The purpose of the present study was to investigate the
ical prediction evaluator, as a scatter plot between the accuracy of six different heuristic methods including the
observed and predicted values of the tested wind speeds. For ANN, ANFIS-SC, ANFIS-GP, GEP, GRNN and MARS
Tabriz station, it is evident that the highest determination models in the prediction of monthly wind speed data based
coefficient was attained for the ANN model (R2 = 0.83); on monthly meteorological input variables. Monthly mete-
whereas, for Jolfa station, the results achieved the highest orological input data comprised of atmospheric pressure,
R2 = 0.88 using the GRNN model. Figures 10 and 11 show temperature, relative humidity and rainfall in addition to
the temporal variations of the best heuristic models for Jolfa the wind speed seasonality pattern collected from two sta-
and Tabriz stations. It is apparent from this figure that the tions, Iran, were used as the input attributes for the applied
estimates of windspeed from ANN and GRNN model are models. The results of all heuristic methods were com-
closer to the observed values than those predictions from the pared to each other with respect to the root mean-squared
other models for the Jolfa and Tabriz stations, respectively. error, coefficient of determination, mean absolute error, rel-
However, all six heuristic models generally provide better ative RMSE, relative mean absolute error, Nash-Sutcliffe
MAROUFPOOR ET AL. 3555

observed ANN ANFIS-GP ANFIS-SC GEP GRNN MARS

Jolfa
5

4
wind speed (m/s)

0
1 5 9 13 17 21 25 29 33 37 41 45 49
month

FIGURE 10 Testing results of the best ANN, ANFIS-GP, GRNN, MARS, ANFIS-SC, and GEP models-Jolfa station [Colour figure can be viewed at
[Link]]

observed ANN ANFIS-GP ANFIS-SC GEP GRNN MARS

5
Tabriz

4
wind speed (m/s)

1
1 5 9 13 17 21 25 29 33 37 41 45 49
month

FIGURE 11 Testing results of the best ANN, ANFIS-GP, GRNN, MARS, ANFIS-SC and GEP models-Tabriz station [Colour figure can be viewed at
[Link]]

model efficiency coefficient, Willmott Index and LMI sta- and Tabriz stations, respectively. For Jolfa station, the opti-
tistics. Results were also evaluated in terms of visual mal ANN model attains a lower RMSE error compared to
assessment of observed and predicted windspeed at both the optimal ANFIS-GP, ANFIS-SC, GEP, GRNN and
study locations. The comparison of the results in this study MARS models by approaximatly 9.2, 5.7, 5.7, 7.5 and
demonstrated that the performance of the ANN and GRNN 3.9%, respectively. In the Tabriz station, however, the opti-
models was better than the other heuristic models applied mal GRNN model led to a reduction in the RMSE over
in the monthly wind speed prediction problem for Jolfa the ANN, ANFIS-GP, ANFIS-SC, GEP and MARS model
3556 MAROUFPOOR ET AL.

by similar percentage 2.3. The present results suggested Fausett, L. (1994) Fundamentals of neural networks: architectures, algorithms,
and applications. Englewood Cliffs, NJ: Prentice Hall.
that the ANN and GRNN models are useful heuristic Ferreira, C. (2001) Algorithm for solving gene expression programming: a new
modelling tools that can be successfully applied in model- adaptive problems. Complex Systems, 13(2), 87–129.
ling the monthly wind speed data. Ferreira, C. (2006) Gene Expression Programming: Mathematical Modeling by
an Artificial Intelligence, 2nd edition. Berlin: Springer-Verlag.
These models can also be further explored for their util- Friedman, J.H. (1991) Multivariate adaptive regression splines. The Annals of
ity in wind speed prediction for developing renewable Statistics, 19, 1–67.
energy solutions, energy sustainability studies and studying Goldberg, D.E. (1989) Genetic Algorithms in Search Optimization and Machine
Learning, Vol. 412. Reading Menlo Park: Addison-Wesley.
climate change patterns. In spite of the very good accuracy
Haque, A.U., Mandal, P., Kaye, M.E., Meng, J., Chang, L. and Senjyu, T. (2012)
of the present heuristic models, a follow-up study could A new strategy for predicting short-term wind speed using soft computing
improve the practicality of the developed models by evaluat- models. Renewable and Sustainable Energy Reviews, 16(7), 4563–4573.
ing their performance for short-term (i.e., minute, hourly or Hastie, T., Tibshirani, R. and Freidman, J. (2009) The Element of Statistical
Learning: Data Mining, Inference, and Prediction, 2nd edition. New York,
daily) prediction horizons, as these are more important for NY: Springer 552 pp.
real-time implementation of models in energy applications. Haykin, S. (1999) Neural Networks: A Comprehensive Foundation. Upper Sad-
Furthermore, incorporating land surface conditions informa- dle River, NJ: Prentice-Hall. Inc.
Jang, J.S.R. (1993) ANFIS: adaptive-network-based fuzzy inference system.
tion is highly recommended to be investigated as an extrenal IEEE Transactions on Systems, Man, and Cybernetics, 23(3), 665–685.
predictor for wind speed patterns. Jang, J.S.R. and Sun, C.T. (1995) Neuro-fuzzy modeling and control. Proceed-
It should be noted that various methodologies have been ings of the IEEE, 83(3), 378–406.
Kalogirou, S.A. (2001) Artificial neural networks in renewable energy systems appli-
applied for wind speed simulation including physical, statis-
cations: a review. Renewable and Sustainable Energy Reviews, 5(4), 373–401.
tical and artificial intelligence models. The feasibility of the Khatibi, R., Ghorbani, M.A., Naghipour, L., Jothiprakash, V., Fathima, T.A. and
physical models are based on the usage of weather data in Fazelifard, M.H. (2014) Inter-comparison of time series models of lake levels
which associated with several limitations of data collection, predicted by several modeling strategies. Journal of Hydrology, 511, 530–545.
Kisi, O. and Guven, A. (2010) A machine code-based genetic programming for
required topography information and further validation suspended sediment concentration estimation. Advances in Engineering Soft-
tasks. Hence, the exploration of more robust and reliable ware, 41(7), 939–945.
intelligence models for such climatological application is Kong, X., Liu, X., Shi, R. and Lee, K.Y. (2015) Wind speed prediction using
reduced support vector machines with feature selection. Neurocomputing,
always the motivation of the climatology scientists. The AI 169, 449–456.
models have evidenced their potential in capturing the com- Koza, J.R. (1992) Genetic programming: on the programming of computers by
plex and high stochasticity phenomena of wind speed pattern means of natural selection, Vol. 1. Cambridge, MA and London: MIT Press.
Li, G. and Shi, J. (2010) On comparing three artificial neural networks for wind
data. Although, AI model are featured by several merits;
speed forecasting. Applied Energy, 87(7), 2313–2320.
however, there are several disadvantages are associated with Liu, H., Tian, H.Q. and Li, Y.F. (2015a) Comparison of new hybrid FEEMD-
those models such as over fitting, sensitivity to the internal MLP, FEEMD-ANFIS, wavelet packet-MLP and wavelet packet-ANFIS for
optimization parameters, and their applicability to wide wind speed predictions. Energy Conversion and Management, 89, 1–11.
Liu, H., Tian, H.Q., Li, Y.F. and Zhang, L. (2015b) Comparison of four Ada-
range of data set. Thus, the research studies on this field is boost algorithm based artificial neural networks in wind speed predictions.
still ongoing for examination. Energy Conversion and Management, 92, 67–81.
Olaofe, Z.O. and Folly, K.A. (2012) Statistical analysis of wind resources at Dar-
ling for energy production. International Journal of Renewable Energy
Research, 2(2), 250–261.
ACKNOWLEDGEMENTS
Rahnama, M.B. and Nouri, M. (2008) Developing of Halil River rainfall runoff
The authors indicate their gratitude and appreciation for the model using conjuction of wavelet transform and ANN. Research Journal of
Environmental Sciences, 5, 2.
Iran Meteorological Organization for providing the climato-
Sanikhani, H. and Kisi, O. (2012) River flow estimation and forecasting by using
logical data. The authors also thank the reviewers and the two different adaptive neuro-fuzzy approaches. Water Resources Manage-
Editor-in-Chief for their constructive comments that have ment, 26(6), 1715–1729.
improved the quality of this paper. Sanikhani, H., Kisi, O., Nikpour, M.R. and Dinpashoh, Y. (2012) Estimation of
daily pan evaporation using two different adaptive neuro-fuzzy computing
techniques. Water Resources Management, 26(15), 4347–4365.
ORCID Sanikhani, H., Deo, R.C., Yaseen, Z.M., Eray, O. and Kisi, O. (2018) Non-tuned
data intelligent model for soil temperature estimation: a new approach. Geo-
Hadi Sanikhani [Link] derma, 330, 52–64.
Ozgur Kisi [Link] Setlak, G. (2008) The fuzzy-neuro classifier for decision support. International
Ravinesh C. Deo [Link] Journal of Information Theory and Applications, 15, 21–26.
Slim, C. (2006) Neuro-fuzzy network based on extended Kalman filtering for
Zaher Mundher Yaseen [Link]
financial time series. World Academy of Science, Engineering and Technol-
7137 ogy, 22, 134–139.
Specht, D.F. (1991) A general regression neural network. IEEE Transactions on
Neural Networks, 2(6), 568–576.
REFERENC ES Su, Z., Wang, J., Lu, H. and Zhao, G. (2014) A new hybrid model optimized by
Abonyi, J., Andersen, H., Nagy, L. and Szeifert, F. (1999) Inverse fuzzy-pro- an intelligent optimization algorithm for wind speed forecasting. Energy
cess-model based direct adaptive control. Mathematics and Computers in Conversion and Management, 85, 443–452.
Simulation, 51(1), 119–132. Tascikaraoglu, A. and Uzunoglu, M. (2014) A review of combined approaches
Erdem, E. and Shi, J. (2011) ARMA based approaches for forecasting the tuple for prediction of short-term wind speed and power. Renewable and Sustain-
of wind speed and direction. Applied Energy, 88(4), 1405–1414. able Energy Reviews, 34, 243–254.
MAROUFPOOR ET AL. 3557

Troncoso, A., Salcedo-Sanz, S., Casanova-Mateo, C., Riquelme, J.C. and

Prieto, L. (2015) Local models-based regression trees for very short-term
wind speed prediction. Renewable Energy, 81, 589–598.
How to cite this article: Maroufpoor S,
Ustoorikar, K. and Deo, M.C. (2008) Filling up gaps in wave data with genetic Sanikhani H, Kisi O, Deo RC, Yaseen ZM. Long-term
programming. Marine Structures, 21(2), 177–195. modelling of wind speeds using six different heuristic
Zakaria, N.A., Azamathulla, H.M., Chang, C.K. and Ghani, A.A. (2010) Gene
artificial intelligence approaches. Int J Climatol.
expression programming for total bed material load estimation–a case study.
Science of the Total Environment, 408(21), 5078–5085. 2019;39:3543–3557. [Link]
Zhang, W., Wu, J., Wang, J., Zhao, W. and Shen, L. (2012) Performance analy-
sis of four modified approaches for wind speed forecasting. Applied Energy,
99, 324–333.