ArticlePDF Available

Forecasting of short-term wind speed at different heights using comparative forecasting approach

Authors:

Abstract

The forecasting of wind speed with high accuracy has been a very significant obstacle to the enhancement of wind power quality, for the volatile behavior of wind speed makes forecasting difficult. In order to generate more reliable wind power and to determine the best model for different heights, wind speed needs to be predicted accurately. Recent studies show that soft computing approaches are preferred over physical methods because they can provide fast and reliable techniques to forecast short-term wind speed. In this study, a MLP (Multilayer Perceptron) neural network and an ANFIS (Adaptive Neural Fuzzy Inference System) are utilized to both forecast wind speed and propose the best model at heights of 30, 50 and 60 meters. It is obvious that various internal and external parameters for soft computing methods have paramount importance for forecasting. In order to analyze the impact of these parameters new wind speed data was collected from a wind farm location. Miscellaneous models were created for every wind turbine elevation by adjusting the parameters of soft computing methods in order to improve wind speed forecasting errors. The experimental results demonstrate that elevation of collected wind speed data significantly affects the wind speed forecasting. Our experimental results reveal that although behavior of wind speed for every height appears identical there is no single model to predict wind speed with the best accuracy. Therefore, every model for the soft computing methods shall be modified for every particular wind turbine height so that wind speed forecasting accuracy is improved. In this way, the approaches perform with fewer errors and models can be used to predict wind speed and power at different heights.
Turk J Elec Eng & Comp Sci
() :
© TÜBİTAK
doi:10.3906/elk-1601-213
Turkish Journal of Electrical Engineering & Computer Sciences
http://journals.tubitak.gov.tr/elektrik/
Researc h Article
Forecasting of short-term wind speed at dierent heights using a comparative
forecasting approach
Emrah KORKMAZ1,2,, Ercan İZGİ1,, Salih TÜTÜN3,
1Department of Electrical Engineering, Yıldız Technical University, İstanbul, Turkey
2Department of Electrical and Computer Engineering, Binghamton University, Binghamton, NY, USA
3Department of Systems Science and Industrial Engineering, Binghamton University, Binghamton, NY, USA
Received: 19.01.2016 Accepted/Published Online: 29.04.2018 Final Version: ..201
Abstract: The forecasting of wind speed with high accuracy has been a very signicant obstacle to the enhancement
of wind power quality, for the volatile behavior of wind speed makes forecasting dicult. In order to generate more
reliable wind power and to determine the best model for dierent heights, wind speed needs to be predicted accurately.
Recent studies show that soft computing approaches are preferred over physical methods because they can provide fast
and reliable techniques to forecast short-term wind speed. In this study, a multilayer perceptron neural network and an
adaptive neural fuzzy inference system are utilized to both forecast wind speed and propose the best model at heights of
30, 50, and 60 m. It is obvious that various internal and external parameters for soft computing methods have paramount
importance for forecasting. In order to analyze the impact of these parameters, new wind speed data were collected from
a wind farm location. Miscellaneous models were created for every wind turbine elevation by adjusting the parameters of
soft computing methods in order to improve wind speed forecasting errors. The experimental results demonstrate that
elevation of collected wind speed data signicantly aects the wind speed forecasting. Our experimental results reveal
that although behavior of wind speed for every height appears identical there is no single model to predict wind speed
with the best accuracy. Therefore, every model for the soft computing methods shall be modied for every particular
wind turbine height so that wind speed forecasting accuracy is improved. In this way, the approaches perform with fewer
errors and models can be used to predict wind speed and power at dierent heights.
Key words: Forecasting, wind energy, soft computing methods, time series analysis
1. Introduction
Nowadays, most countries rely heavily on fossil fuels to generate their own electricity. Power plants based on
fossil fuels, however, cause many environmental problems. These power plants have emitted large amounts of
greenhouse gases. As a result, many people face a sharply high risk of breathing problems, cancer, and heart
attacks [1]. Thus, worldwide many people are seeking out new energy sources that will produce cleaner energy.
Even though nonrenewable energy sources (e.g., coal, petroleum, and natural gas) are available in most
of the world, these sources will get more expensive because of restricted reserves [2,3]. On the other hand,
renewable energy is clean, environmentally friendly, and inexhaustible. Over the last decade renewable electricity
generation capacity has increased signicantly, but this capacity has still not been enough to replace the energy
capacity that comes from fossil fuels. If we want to replace fossil fuels with renewable energy sources, then more
Correspondence: emr3234@hotmail.com
This work is licensed under a Creative Commons Attribution 4.0 International License.
1
KORKMAZ et al./Turk J Elec Eng & Comp Sci
research is needed concerning their negative sides, and better renewable energy technologies must be developed.
This could well lead to people having cheaper and cleaner energy.
Wind energy is one of the fastest growing forms of renewable energy. For instance, over the last 10 years
in Turkey, installed wind power capacity has increased approximately 10% each year [4]. Despite this rapid
growth of wind power, it still does not look like a reliable energy source that can meet future demands of the
electricity grid. One of the most important reasons for this is the unreliability issue: wind speed proles are
irregular [5]. In this regard, it is essential to make accurate wind speed estimates in order to develop a more
reliable structure for wind power plants.
In the present paper, a comparative forecasting approach based on soft computing methods is proposed
to improve the prediction of short-term wind speed at dierent heights. It is well known that soft computing
methods outperform other methods and can achieve better results in short-term wind speed forecasting [6].
Therefore, we utilized algorithms of ANFIS and MLP neural network to predict wind speed with minimum
errors. In order to achieve this goal, we created dierent models for each forecasting method and compared
these models based on forecasting errors so that wind speed estimation error can be diminished.
This paper presents current and future research aimed at the development of comprehensive wind speed
forecasting for dierent wind turbine heights. The study is performed as experimental work in order to guide
wind speed forecasting researchers who might utilize soft computing methods. With these methods, a large
number of model parameters shall be optimized and the best models can be determined for realistic wind speed
forecasting. However, the authors are not familiar with publications featuring experimental studies for wind
speed prediction at dierent heights using soft computing methods. As a result, we present a comparative
forecasting approach in order to forecast short-term wind speed at dierent heights for the same location. Since
wind speed has nonlinear behavior, the features of models of soft computing methods should be particularly
distinguished for dierent wind turbine height.
Furthermore, many studies on soft computing based wind speed forecasting were conducted at height of
10 m. It is obvious that wind speed data at 10 m is not sucient for selection of wind turbine location and
not feasible for wind energy estimation and so we collected new data directly from a data logger located at a
wind farm. Therefore, the present study aims to enhance soft computing based wind speed prediction results
regarding their use in industrial applications for choosing dierent wind turbine heights of 30, 50, and 60 m.
The rest of this paper is organized as follows. Section 2 describes related work. In Section 3 we present
the materials and methodologies used in our problem formulations. In Section 4 we introduce our results to
show the proposed models predict wind speed at dierent heights with outperformance. Finally, we provide a
conclusion with the best models to forecast wind speed in Section 5.
2. Related work
Dierent forecasting methods are used in most research studies. Generally speaking, all forecasting techniques
can be described under three main approaches: physical, statistical, and hybrid methods. The most used
physical method is the numerical weather prediction method (NWP) developed by experts in meteorology. The
main purpose of NWP is to dene atmospheric phenomena using mathematical models. This can be done
using large amounts of weather data that represent every small region. It is quite hard to develop a perfect
weather prediction because it needs many calculations, and these calculations need supercomputers. Even when
supercomputers are used to predict weather, however, the calculation time is very long. Although NWP needs
long calculation times, it is more eective for long-term prediction. Therefore, many industries and government
2
KORKMAZ et al./Turk J Elec Eng & Comp Sci
agencies are extensively using the NWP technique. Furthermore, it has been very important in allowing military
operations to forecast weather. However, the NWP technique is quite ineective for short-term wind forecasting
because wind speed has a very high variation in a short period of time [7].
Recent publications have preferred using statistical methods over physical methods. Statistical methods
can be developed in which historical values of wind speed are utilized. Statistical prediction methods can be
classied into two main approaches: (1) time-series models (such as autoregressive and moving average models)
and (2) soft computing models (such as articial neural networks (ANNs), fuzzy logic). Recent publications
demonstrate that genetic algorithms (GAs), support vector machines (SVMs), and Kalman ltering (KF) models
are integrated into gray-box methodologies in order to reduce prediction errors. A brief review of recent studies
on wind speed prediction is introduced in the following paragraphs.
Based on the time series models, short-term wind speed prediction is performed by linear and nonlinear
autoregressive models [8]. Wind speed prediction models dier regarding their prediction intervals ranging from
10 min to 1 h. According to [9], KF can be successfully implemented for short-term predictions and the authors
propose a new KF method instead of using the standard KF in order to yield better prediction results. In
[10], the researchers develop a new technique using heteroscedastic support vector regression diminishing the
uncertainty of short-term wind speed. The wind speed prediction results are obtained for wind power plants in
China and predicted for 30, 60, and 120 min in the future with errors above 10% MAPE.
Today, more and more researchers are using soft computing algorithms based on ANNs to make short-
term forecasts [11–14]. This is because ANNs use linear assumptions and they are more eective in modeling
the nonlinearity relationship of wind speed data [15].
The literature on this issue includes many publications. These can be summarized in the following
studies. Cadenas and Rivera [16] used the ANN model to predict short-term wind speed. The results of their
study show that the two-layer and three-neuron models for the training and testing stages give satisfactory
accuracy for short-term forecasting. The implementation of ANN to forecast wind speed in [17] demonstrates
that ANN models predict wind speed with acceptable accuracy (8.9% MAPE and correlation of 0.9380 m/s).
In the study conducted by De Giorgi et al. (the wind farm model in southern Italy), wind energy prediction
is made by autoregressive moving average (ARMA), ve dierent ANN models, and the neuro-fuzzy inference
system (ANFIS) [18]. For predictions of 1, 3, 6, and 12 h, multilayer perceptron (MLP) performance appears
better than other methods and gives a short calculation time.
Another study is conducted by [19], where the authors predict daily wind speed using an MLP neural
network and propose to utilize meteorological parameters (e.g., temperature, air pressure, solar addition, and
altitude) as input variables. However, the collected data are obtained for only two dierent altitudes that are
very close to each other (14.5 m and 18.5 m). The recently published article by [20] presents a short-term wind
speed prediction technique that demonstrates that wind speed can be forecasted easily using ANNs, but the
maximum lead time for the measured data must be 14 h.
In [7], several dierent ANFIS models were used to predict very short-term wind speed. The data set is
prepared by a 21-month time series using 2.5-min intervals. In that study, the ANFIS model estimates results
in less than 4% MAPE. Approaches using BPNN, RBFNN, and ANFIS short-term wind speed forecasts, which
are 1-h-ahead and 3-h-ahead, are assessed in [21]. These forecasting techniques are combined with a similar day
(SD) approach. ANFIS models based on the SD approach are more successful in transforming historical data
into wind speed forecasts.
3
KORKMAZ et al./Turk J Elec Eng & Comp Sci
Furthermore, in the literature, hybrid approaches are able to predict wind speed with high accuracy.
Hybrid models can consist of not only physical and statistical techniques but also can only be developed by
dierent statistical methods. In [22], two new hybrid approaches, known as the ARIMA-ANN and ARIMA-
Kalman lter models, are found suitable for wind speed prediction. Guo et al. [23] presented a case study
using a hybrid forecasting method with a back propagation neural network (BPNN) and seasonal adjustment.
The results show that rather than using only BPNN a hybrid technique must be used to improve prediction
performance.
In spite of that, in the literature, the performance of ANFIS and MLP approaches is not evaluated in
terms of dierent heights in a particular wind farm location. In order to propose or establish a reliable soft
computing method for short-term wind forecasting in the wind power industry, multivariate soft computing
models must be evaluated considering their use at dierent heights. The main contribution of the present paper
is to provide researchers with a comprehensive analysis of the eects of various heights on short-term wind
speed forecasting by using ANFIS and MLP approaches.
3. Materials and methods
3.1. Adaptive neural fuzzy inference systems approach
ANFIS methodology was rst introduced by Roger Jang in 1993 [24]. It aims to combine fuzzy logic and ANN
methods. Figure 1 provides a simple description of the ANFIS architecture, which has two inputs and one
output. In the rst layer (called the “input layer”), input values are transferred to the second layer. The
second layer, known as the “fuzzication layer”, enables nodes to change output depending on the membership
function. The parameters are dened for membership functions (µ(e.g., bell-shaped) as given by the equations
below:
A1 A2 B1 B2
X Y
N
W1
N
-
X
Y
X
Y
Layer1
Layer2
Layer3
Layer4
Layer5
Lay
Σ
ππ
er 6
f
W2
-
W1W2
W2
-f2
W1
-f1
Figure 1. Architecture of a two input ANFIS model.
4
KORKMAZ et al./Turk J Elec Eng & Comp Sci
O2,i =µAi(x)(1)
O2,j =µBj2(y)(2)
µ(x) = exp 1
2xc
σ2,(3)
where cand σare parameters that correspond to the mean and standard deviation of the membership function,
respectively. In general, these parameters are referred to as premise or antecedent parameters.
In the third layer, each neuron represents a single fuzzy rule. The output of each node can be calculated
by the ring power of each fuzzy rule as Eq. (4).
O3,k =µAk(x)µBk(y) = ωk, k = 1,2(4)
The fourth layer can be described as a normalization layer. The normalized ring level can be expressed as the
ratio of the kth ring power of the sum of all ring powers.
O4,k =ωk
ωk
= ¯ωk, k = 1,2(5)
In the fth layer, in order to calculate the results of the fuzzy logic rules, weighted result values for each node
are calculated by the following formula, where α, βγ are constant values:
O5,k = ¯ωk(αk+βkx+γky), k = 1,2(6)
Finally, in the last layer, the sum of each output value received from previous layers is calculated and found as
f.
3.2. MLP articial neural network approach
ANNs are a commonly used technique in dierent tasks from process monitoring, fault diagnosis, and adaptive
human interference to articial intelligence based on atmospheric processes and computers [25]. The MLP is a
widely used type of neural network and usually called a feed-forward neural network. It consists of an input
layer, one or more hidden layers, and an output layer. Basically, the MLP solves the complex relationship
between input vector and output vector using connections of weighted layers. To solve a relationship of this
complexity the MLP needs training sets that consist of both input and output data. The MLP uses delta
learning, which is based on the least square method. This learning methodology consists of two training steps:
feed-forward and backpropagation.
The feed-forward learning is usually called feed forward because feedbacks among the nodes do not
appear. With a feed-forward neural network, the connection between the ith and jth neuron can be described
by the weight coecient ωij [26]. The output of each node for n input neuron and m hidden neuron can be
represented by the following equation:
yi=fH
vi+
n
j=1
ωij xj
,(7)
5
KORKMAZ et al./Turk J Elec Eng & Comp Sci
where fHis called the activation function of a node and Viis the threshold coecient that corresponds to the
weight coecient of each jth neuron. This coecient is called the bias value when xjequals 1. All connection
weights and bias values must be initially assigned random values. In the training process, these values will be
adjusted by the network to nd the best output results.
The sigmoid function, which varies between 0 and 1, is used as the activation function of a node, as given
in the following equation:
fH(s) = 1
1 + es(8)
The supervised adaption process changes weight coecients and bias values between predicted and target
outputs [26]. By minimizing the objective function with a training algorithm, these values can be obtained.
The objective function can be calculated as
e=
n
i=1
(tiyi)2(9)
In the case of backpropagation training, a Bayesian approach was utilized to accomplish better t, minimum
error, and minimum number of patterns and weights of the network. This process can be demonstrated by the
following equations:
ω(k+1)
ij =ω(k)
ij λ∂e
∂ωij (k)
(10)
v(k+1)
i=v(k)
iλ∂e
∂vi(k)
,(11)
where is λa constant learning rate. The learning process is repeated by many iterations so that the MLP
network may memorize the training data. Therefore, generalization between input and output patterns can be
eliminated.
3.3. Forecasting of wind speed at dierent heights
The steps used to predict accurate wind speeds as seen in Figure 2 can be summarized as follows:
(1) Use the ANFIS method and select the most accurate models separately for all heights.
(2) Use the MLP-based method and select the most accurate models separately for all heights.
(3) Do a comparison between the MLP and ANFIS models for every height and determine the best models
and wind speed prediction values.
The Silivri region in İstanbul was selected in order to implement wind speed forecasting methods. The
latitude and longitude of the recorded area is N 04108.377’ and E 02819.110’, respectively and the site
elevation is 210 m. Silivri is in the Marmara region, which has very high energy potential. It has been thus
a very investable area for the wind energy sector [27]. Wind speeds were measured at heights of 30 m, 50 m,
and 60 m via NRG #40C cup anemometers. The wind speed data was recorded as 10 min samples between
February 2009 and March 2010, and the total data has 56,548 points.
6
KORKMAZ et al./Turk J Elec Eng & Comp Sci
S elect the Key
Features
Initialize
Determine
Parameters of
the Models
Choose Method
Utilize the
ANFIS
Approach
Utilize the
MLP-ANN
Approa ch
Compare the
Bes t Mode ls
h>60
S elect the Key
Features
End NoYes
Figure 2. The ow chart of the framework for forecasting wind speed.
The recorded dataset is used as inputs for all ANFIS and ANN models. Firstly, the dataset was split
into training (50%), evaluation (25%), and testing (25%) sets. The training set is used to train parameters of
the ANFIS and ANN models. The evaluation dataset is used to evaluate measures of forecasting accuracy by
tuning its parameters. The testing set is utilized to show error between real and forecasted wind speed values.
In order to solve wind forecasting problems using ANFIS and ANN methods, the input of each model
must be chosen correctly. The most common way to choose wind vectors is time-series methodology. Hence,
this methodology was used to determine model inputs in the present study. For instance, if one wants to predict
wind speed at v(t+ 1) time lag, the inputs of time series must be chosen using m previous measurements (such
as v(tm), v (tm+ 1), v (tm+ 2) . . . v (t)) . In this particular wind forecasting study, the number of
previous observations was changed for both methodologies, and then for the selected wind speed input number,
each model was named Model 1, Model 2, Model 3, and Model 4.
3.3.1. Models of adaptive neural fuzzy inference systems
In the present study, many dierent ANFIS architectures are compared to nd the best wind forecasting
accuracy. In order to use ANFIS methods, the type of membership function must be decided on. The
membership function is considered as linear and constant. After many experiments, we settled on 2 and 3
membership functions as the appropriate number. Furthermore, since membership function type changes are
linear and constant, the number of epochs for each model must be selected. A lot of research has shown that a
large number of epochs usually does not improve the accuracy of models or signicantly increase the forecasting
time [28]. Therefore, it is adjusted to 10 and 100 in the structure of ANFIS. All of the parameter changes above
were implemented in all combinations individually.
3.3.2. Models of MLP articial neural networks
In order to test MLP based neural networks, the characteristic parameters of each model must be chosen so
that the number of input neurons is changed to 1, 2, 3, and 4. For all MLP models, hidden layers are used
7
KORKMAZ et al./Turk J Elec Eng & Comp Sci
and the output neuron number is assigned to 1. The error between computed and desired wind speed values
is selected as 1 ×10 5instead of 0 to stop training in a specic time. The number of epochs is selected as
50. The reason for the selection of a xed number of iterations and errors is that after a certain number these
values do not aect the results of estimated wind speed values.
3.4. Performance metrics for forecasting results
It is also essential to compare the performance of models in terms of forecasting accuracy. Unfortunately, there
is no unique metric to evaluate models as a universal standard [29]. Thus, the performance of models must be
evaluated by using dierent metrics such as determination of coecient (R2), root-mean-square error (RMSE),
and mean absolute percentage error (MAPE).
R2= 1 σE
σY
(12)
RMSE =
1
N
n
i=1
(TiYi)2(13)
M AP E =1
n
n
i=1
TiYi
Yi
×100 (14)
The values of σYand σEindicate the standard deviation of observation values and error values, which is
calculated by each actual and forecasted wind speed dierence, and Tiand Yivalues show observation and
forecast values, respectively. The number of the dataset is also symbolized as nin the equations.
The coecient of determination shows the accuracy of predicted values. This value is expected to be
between 0 and 1. To have the best forecast model, the coecient of determination value should approach 1. The
RMSE is known as the standard deviation of the estimation errors in regression analyses. It can be found by
calculating the square root of the average of quadratic errors. The last metric is chosen as MAPE to calculate
forecast error. Since it is indicated by percentage expression, it gives the idea to everyone to understand the
calculated error simply. In the present study, the RMSE and the MAPE are minimized. In many papers, it
goes nearly to zero, but in this particular research some large numbers might be seen due to the large number
of test datasets used.
4. Results for wind speed prediction at dierent heights
Forecasting methods ensure decision makers develop stronger knowledge for better decision making [30,31]. The
purpose of the present research is to analyze the applicability of soft computing methods such as ANFIS and
ANN. In order to address this issue, we formed dierent soft computing models based on ANFIS and ANN and
then compared the best models in each category. The same validation data through forecasting measures (R2,
RMSE, and MAPE) are used to evaluate and compare the proposed models. In order to obtain experimental
wind speed forecasting results MATLAB is utilized in developing various soft computing models.
4.1. Evaluation of the models and results
In order to compare the performance of models in terms of forecasting accuracy, both algorithms for every
height were performed and the forecasting results are tabulated in Tables 1 through 6. In the ANFIS results,
8
KORKMAZ et al./Turk J Elec Eng & Comp Sci
when 3 input membership functions are chosen for all models, the prediction values are less accurate. The best
predicted wind speed values are found when the MF number equals 2. For predicted wind speed values at 30
m, the 4 input model gives the fewest errors and higher R2values except MAPE. The lowest RMSE obtained
is 0.5942 m/s and the highest R2values obtained is 0.9753 m/s. At 50 m, Model 3 gives better wind speed
prediction results, both in the test period and the validation period. Table 3 shows that estimated values are
more accurate for Model 1 and Model 4. At 60 m, wind speed values are more irregular because the wind speed
changes sharply at this height. Thus, when the number of input models increase, poor forecasted results are
observed, and so Model 1 gives a better MAPE. However, during the testing period Model 3 would produce
better results in terms of R2and RMSE values. In order to demonstrate performance of the best models,
Figure 3 is shown to analyze best-tting lines for predicted and actual wind speed at 30 m, 50 m, and 60 m,
respectively. In addition, Figure 4 demonstrates that the best models predict wind speed with perfect accuracy
at all heights.
Table 1. Wind speed forecasting errors for 30 m.
No. Inputs Output Test Validation
MF no. Epoch no. MF type R2RMSE MAPE R2RMSE MAPE
1 2 10 Linear 0.9749 0.5992 8.9115 0.9791 0.6715 9.3041
2 2 100 Linear 0.9739 0.6103 8.9902 0.9789 0.6745 10.145
3 2 100 Constant 0.9752 0.595 8.972 0.9791 0.6715 10.238
4 2 100 Constant 0.9753 0.5942 8.945 0.9792 0.6702 10.648
Table 2. Wind speed forecasting errors for 50 m.
No. Inputs Output Test Validation
MF no. Epoch no. MF type R2RMSE MAPE R2RMSE MAPE
1 2 10 Linear 0.9762 0.61 8.0197 0.9808 0.6838 8.9543
2 2 10 Linear 0.9739 0.6387 8.1602 0.9808 0.6836 10.3151
3 2 100 Constant 0.9766 0.6052 8.1222 0.9848 0.6846 10.309
4 2 100 Constant 0.9765 0.6094 8.1273 0.9809 0.6833 10.3887
Table 3. Wind speed forecasting errors for 60 m.
No. Inputs Output Test Validation
MF no. Epoch no. MF type R2RMSE MAPE R2RMSE MAPE
1 2 10 Linear 0.9779 0.6125 8.0716 0.9821 0.69 9.22
2 2 100 Constant 0.9778 0.6139 8.4279 0.9816 0.7044 11.649
3 2 100 Constant 0.9781 0.6096 8.1991 0.9816 0.7011 10.436
4 2 100 Constant 0.9199 1.1721 17.015 0.928 1.3866 19.424
The wind speed was forecasted by dierent MLP models. As can be seen from Tables 4 and 5, Model
4 performs better than the other models. The amount of input wind data signicantly aects wind speed
estimation. According to the 30 m and 50 m values, input model 4 performs much better than other models
9
KORKMAZ et al./Turk J Elec Eng & Comp Sci
0 5 10 15 20 25 30
Predicted Value
0
5
10
15
20
25
30
Actual Value
0 5 10 15 20 25 30
Predicted Value
0
5
10
15
20
25
30
Actual Value
0 5 10 15 20 25 30
Predicted Value
c) 60 meters
a) 30 meters
b) 50 meters
0
5
10
15
20
25
30
Actual Value
Figure 3. Correlation between actual and predicted values for the ANFIS testing stage.
with 8.8431% and 8.042% MAPE. The results in Table 6 demonstrate that the accuracy improves when the
input number of models is increased. Figure 5 shows best-tting lines for predicted and actual wind speed at
30 m, 50 m, and 60 m, respectively. The predicted and actual values are almost the same as shown in Figure
6. Figure 6 demonstrates that the best models predict wind speed with perfect accuracy at all heights.
10
KORKMAZ et al./Turk J Elec Eng & Comp Sci
0 5000 10000 15000
Time (min)
0
5
10
15
20
25
Wind Speed (m/s)
Predicted Value
Actual Value
c) 60 meters
a) 30 meters b) 50 meters
0 5000 10000 15000
Time (min)
0
5
10
15
20
25
Wind Speed (m/s)
Predicted Value
Actual Value
0 5000 10000 15000
Time (min)
0
5
10
15
20
25
Wind Speed (m/s)
Predicted Value
Actual Value
Figure 4. Comparison between prediction and actual results for ANFIS models.
4.2. Discussion
In the present study, it is noted that although the behavior of wind speed for all heights seems to be very similar,
the prediction of wind speed is substantially inuenced by changing collected data containing wind speed at
various heights. As seen in the forecasting wind speed results, it is not possible to say that a single model might
present good prediction results for all heights. Furthermore, a detailed experimental analysis was carried out
due to eects of the various internal and external parameters on the wind speed forecasting. Tables 1–3 show
11
KORKMAZ et al./Turk J Elec Eng & Comp Sci
Table 4. Wind speed forecasting errors for 30 m.
No. Test Validation
R2RMSE MAPE R2RMSE MAPE
1 0.9773 0.6398 10.3683 0.9803 0.7709 18.0652
2 0.9543 0.8883 13.0279 0.9601 1.0407 18.9147
3 0.9357 1.0456 15.003 0.9435 1.2262 18.8005
4 0.975 0.597 8.8431 0.9792 0.6698 9.5798
Table 5. Wind speed forecasting errors for 50 m.
No. Test Validation
R2RMSE MAPE R2RMSE MAPE
1 0.9123 1.1713 18.0428 0.9222 1.3786 24.5106
2 0.9293 1.051 15.4362 0.9349 1.2211 20.4918
3 0.9507 0.8777 12.3616 0.9582 1.0107 15.6393
4 0.9763 0.6083 8.042 0.981 0.6809 9.3231
Table 6. Wind speed forecasting errors for 60 m.
No. Test Validation
R2RMSE MAPE R2RMSE MAPE
1 0.9773 0.6398 10.3683 0.9803 0.7709 18.0652
2 0.9543 0.8833 13.0279 0.9335 1.0407 18.9147
3 0.9357 1.0456 15.003 0.9435 1.2262 18.8004
4 0.9187 1.1805 16.9167 0.9271 1.3957 18.0814
that increasing the input number of ANFIS models does not always demonstrate better estimation errors for
all heights. In spite of that, for the best estimation results, the input number of wind speed data is usually
between 4 and 6 [15].
The constant type output MF usually shows better performance for all models. There is a certain rule
in the literature that increasing the epoch number of ANFIS models decreases the estimation errors. In the
current study, this is not obtained for every model because it depends on the output MF type. When output MF
decides linear function, better results can be seen with a lower number of epochs. However, it is also observed
that linear output MF function with big epoch numbers signicantly aects the calculation time of short-term
wind speed prediction. Thus, in order to obtain realistic wind speed predictions and propose reliable ANFIS
models, epoch number must be limited to the satised values especially for the linear MF output functions.
From the MLP-ANN prediction results, we observed that with respect to the amount of input wind
speed data MLP-ANN models illustrate the same patterns except at the height of 60 m. The 60 m wind
speed forecasting results show that various numbers of input wind speed data are required for all heights to
accomplish better prediction results. Furthermore, the height of the measured wind speed may signicantly
aect the results. For instance, while MLP-ANN gives sucient results for heights of 30 m and 50 m, it produces
very undesirable results when 60 m wind data are used for the MLP-ANN approach.
12
KORKMAZ et al./Turk J Elec Eng & Comp Sci
0 5 10 15 20 25
Predicted Value
0
5
10
15
20
25
Actual Value
Outputs vs. Targets, R=0.98935
Data Points
Best Linear Fit
0 5 10 15 20 25 30
Predicted Value
0
5
10
15
20
25
30
Actual Value
Outputs vs. Targets, R=0.99046
Data Points
Best Linear Fit
0 5 10 15 20 25 30
Predicted Value
0
5
10
15
20
25
30
Actual Value
Outputs vs. Targets, R=0.98955
Data Points
Best Linear Fit
c) 60 meters
b) 50 metersa) 30 meters
Figure 5. Correlation between actual and predicted values for MLP neural network testing stage at a) 30 m, b) 50 m,
c) 60 m.
In the evaluation and testing stages, both methods are compared according to the above tables. Two
methods are compared and then the best models are determined for each selected height according to three
dierent evaluation metrics. The error values are compared in Table 7. In this table, only testing errors are
shown because the dierences between actual and last estimated wind speed values are shown at this stage.
The evaluation stage optimizes parameters of articial intelligence algorithms. As seen in Table 7, according to
13
KORKMAZ et al./Turk J Elec Eng & Comp Sci
0 5000 10000 15000
Time (min)
0
5
10
15
20
25
Wind Speed (m/s)
Predicted Value
Actual Value
c) 60 meters
a) 30meters b) 50 meters
0 5000 10000 15000
Time (min)
0
5
10
15
20
25
Wind Speed (m/s)
Predicted Value
Actual Value
0 5000 10000 15000
Time (min)
0
5
10
15
20
25
Wind Speed (m/s)
Predicted Value
Actual Value
Figure 6. Comparison between prediction and actual results for MLP neural network models.
the results at heights of 30 and 50 m, MLP-ANN provides better forecasting results for only the MAPE metric.
Additionally, while MLP-ANN performs as the poorest model with above 10% error, the ANFIS approach gives
satisfactory results for all heights in terms of R2and RMSE values. Therefore, we can say that ANFIS models
perform with reliable accuracy for all heights.
14
KORKMAZ et al./Turk J Elec Eng & Comp Sci
Table 7. Comparison of performances of methods.
Height Method Model no. Test stage
R2RMSE MAPE
30 m ANFIS 4 0.9753 0.5942 8.945
MLP 4 0.975 0.597 8.8431
50 m ANFIS 3 0.9766 0.6052 8.1222
MLP 4 0.9763 0.6083 8.042
60 m ANFIS 3 0.9781 0.6096 8.1991
MLP 1 0.9773 0.6398 10.3683
5. Conclusions
Wind power generation has grown considerably in recent years; however, it is still unreliable as a main grid
supplier because variability in wind speed dramatically aects the predicted wind energy. Since wind speed
at various heights exhibits very dierent behavior, forecasted wind speed may not be calculated accurately.
The primary signicance of this study is that we ll a research gap for short-term wind speed prediction at
dierent heights. Recent research has shown that the MLP-ANN and ANFIS methods can successfully predict
short-term wind speed variation. However, wind speed predictions can be done at dierent heights and thus
model performances may be dierent at various heights.
In light of the results and discussion presented so far in this paper, it is clear that MLP-ANN and
ANFIS based algorithms can be used to predict wind speed 10 min in advance. The forecasting models vary
by the number of wind data inputs. MLP-ANN and ANFIS give satisfactory results for short-term wind speed
prediction with approximately under 10% MAPE and 0.6398 m/s RMSE. In light of the results of the best
models, we see that no single model works best for all heights. Thus, it is dicult to say with certainty that
one model works better for every height. Moreover, at all heights, the ANFIS approach gives better prediction
results. The best ANFIS models for wind prediction show that 3 and 4 inputs give the most accurate prediction
results. The MLP-ANN algorithm with 3 and 4 input wind speed models shows better performance for 30 and
50 m; however, a 1 input wind speed model enables better forecasting results for 60 m.
In conclusion, for the rst time, dierent ANFIS and MLP-ANN models are proposed to understand
the behavior of wind speed and to obtain the most accurate speed forecast at dierent heights. For dierent
heights, the proposed model can be used to dene the best location of wind turbines and to forecast irregular
wind energy. Short-term wind energy forecasting can be improved by using these models to enhance wind power
quality.
Acknowledgments
This research is part of a master’s thesis “Wind Power Calculation by Using Forecasted Wind Speed with Soft
Computing Method at Dierent Heights”. Therefore, we would like to appreciate everyone who contributed to
the successful realization of this research.
15
KORKMAZ et al./Turk J Elec Eng & Comp Sci
References
[1] Markandya A, Wilkinson P. Electricity generation and health. Lancet 2007; 370: 979-990.
[2] Rahman MM, Mostaz SB, Paatero JV, Lahdelma R. Extension of energy crops on surplus agricultural lands: a
potentially viable option in developing countries while fossil fuel reserves are diminishing. RenewSust Energ Rev
2014; 29: 108-119.
[3] Tutun S, Chou CA, Canıyılmaz E. A new forecasting framework for volatile behavior in net electricity consumption:
a case study in Turkey. Energy 2015; 93: 2406-2422.
[4] Turkish Wind Energy Association (TWEA). Turkish Wind Energy Statistics Report 2015.
[5] Şahin AD. Progress and recent trends in wind energy. Progr Energ Combust 2004; 30: 501-543.
[6] Lei M, Shiyan L, Chuanwen J, Hongling L, Yan Z. A review on the forecasting of wind speed and generated power.
Renew Sust Energ Rev 2009; 13: 915-920.
[7] Potter CW, Negnevitsky M. Very short-term wind forecasting for Tasmanian power generation. IEEE T Power Syst
2006; 21: 965-972.
[8] Lydia M, Kumar SS, Selvakumar AI, Kumar GEP. Linear and non-linear autoregressive models for short-term wind
speed forecasting. Energ Convers Manage 2016; 112: 115-124.
[9] Zuluaga CD, Álvarez MA, Giraldo E. Short-term wind speed prediction based on robust Kalman ltering: an
experimental comparison. Appl Energ 2015; 156: 321-330.
[10] Hu Q, Zhang S, Yu M, Xie Z. Short-term wind speed or power forecasting with heteroscedastic support vector
regression. IEEE T Sustain Energ, 2016; 7: 241-249.
[11] Noorollahi Y, Jokar MA, Kalhor A. Using articial neural networks for temporal and spatial wind speed forecasting
in Iran. Energ Convers Manage 2016; 115: 17-25.
[12] Youse M, Hooshyar D, Youse M, Khaksar W, Sahari KSM, Alnaimi FBI. An articial neural network hybrid
with wavelet transform for short-term wind speed forecasting: a preliminary case study. International Conference
on Science in Information Technology (ICSITech); 27–28 October 2015; Yogyakarta, Indonesia: pp. 95-99.
[13] Singh R, Sahay KB, Srivastava SA. Short-term wind speed forecasting of Oak Park Weather Station by using
dierent ANN algorithms. In Smart Grid Technologies-Asia (ISGT ASIA); 3–6 November 2015; Bangkok, Thailand:
pp. 1-6.
[14] Huang SH, Mu KM, Lu PY, Tsao CY, Leu YG, Chou LF. The application of neural network in wind speed
forecasting. In Networking, Sensing and Control (ICNSC), IEEE 12th International Conference; 9–11 April 2015;
Taipei, Taiwan: pp. 366-370.
[15] Tascikaraoglu A, Uzunoglu M. A review of combined approaches for prediction of short-term wind speed and power.
Renew Sust Energ Rev 2014; 34: 243-254.
[16] Cadenas E, Rivera W. Short-term wind speed forecasting in La Venta, Oaxaca, México, using articial neural
networks. Renew Energ 2009; 34: 274-278.
[17] Fadare DA. The application of articial neural networks to mapping of wind speed prole for energy application in
Nigeria. Appl Energ 2010; 87: 934-942.
[18] De Giorgi MG, Ficarella A, Tarantino M. Error analysis of short-term wind power prediction models. Appl Energ
2011; 88: 1298-1311.
[19] Ramasamy P, Chandel SS, Yadav AK. Wind speed prediction in the mountainous region of India using an articial
neural network model. Renew Energ 2015; 80: 338-347.
[20] Ghorbani MA, Khatibi R, Fazelifard MH, Naghipour L, Makarynskyy O. Short-term wind speed predictions with
machine learning techniques. Meteorol Atmos Phys 2016; 128: 57-72.
16
KORKMAZ et al./Turk J Elec Eng & Comp Sci
[21] Haque AU, Mandal P, Kaye ME, Meng J, Chang L, Senjyu T. A new strategy for predicting short-term wind speed
using soft computing models. Renew Sust Energ Rev 2012; 16: 4563-4573.
[22] Liu H, Tian HQ, Li YF. Comparison of two new ARIMA-ANN and ARIMA-Kalman hybrid methods for wind speed
prediction. Appl Energ 2012; 98: 415-424.
[23] Guo ZH, Wu J, Lu HY, Wang JZ. A case study on a hybrid wind speed forecasting method using BP neural network.
Knowl-Based Syst 2011; 24: 1048-1056.
[24] Jang JSR. ANFIS: adaptive-network-based fuzzy inference system. Systems, IEEE T Syst Man Cyb 1993; 23:
665-685.
[25] Izgi E, Öztopal A, Yerli B, Kaymak MK, Şahin AD. Short–mid-term solar power prediction by using articial neural
networks. Sol Energy 2012; 86: 725-733.
[26] Svozil D, Kvasnicka V, Pospichal J. Introduction to multi-layer feed-forward neural networks. Chemometr Intell
Lab 1997; 39: 43-62.
[27] Ozcira S, Bekiroglu N, Agcal A. A study on feasibility of wind energy production in Silivri region by using laboratory
setup. Renewable Energy Research and Applications (ICRERA); 20–23 October 2013; Madrid, Spain: pp. 1175-
1179.
[28] Taşcıkaraoğlu A, Uzunoğlu M. Dalgacık dönüşümü ve yapay sinir ağları ile rüzgar hızı tahmini. Elektrik Elektronik
Bilgisayar Sempozyumu (FEEB 2011); 5–7 October 2011; Elazığ, Turkey: pp. 106-111.
[29] Li G, Jing S. On comparing three articial neural networks for wind speed forecasting. Appl Energ 2010; 87:
2313-2320.
[30] McNelis PD. Neural Networks in Finance: Gaining Predictive Edge in the Market. Burlington, MA, USA: Elsevier
Academic Press, 2005.
[31] Cankurt S, Subaşı A. Tourism demand modelling and forecasting using data mining techniques in multivariate time
series: a case study in Turkey. Turk J Elec Eng & Comp Sci 2016; 24: 3388-3404.
17
... Son yıllarda, yenilenebilir enerjiye bağlı elektrik üretim kapasitesi önemli ölçüde artış göstermiştir. Ancak bu kapasite hala fosil yakıtlardan kaynaklı enerji kapasitesinin yerini alamamıştır [2]. Fosil yakıtlara dayalı enerji üretim tesisleri birçok çevre sorununa neden olmaktadır. ...
... Çünkü kısa vadeli rüzgar hızının tahmini ancak yazılım gibi hızlı ve güvenilir teknikler ile gerçekleştirilebilmektedir. Depolanan enerjiyi programlamak ve ana şebeke ile olan çalışmayı optimum seviyede yönetebilmek için makine öğrenmesi algoritmaları büyük öneme sahiptir [2,5] [3]. Ramesh ve arkadaşları tarafından yapılan çalışma ile geleneksel istatistiksel yöntem ve geri yayılma ağı (BPN), genelleştirilmiş regresyon sinir ağı (GRNN) ve radyal tabanlı fonksiyon ağları (RBFN) gibi yapay sinir ağları kullanarak kısa süreli rüzgar hızı tahmininin yapılması amaçlanmıştır. ...
Conference Paper
Full-text available
Bu çalışmada rüzgar hızının ve buna bağlı olarak rüzgardan elde edilebilecek enerjinin Aşırı Öğrenme Makineleri (AÖM) algoritması tarafından tahmin edilmesi amaçlanmıştır. Bu amaçla, rüzgar hızını etkileyebilecek dört adet parametre seçilmiştir. Bunlar sırasıyla maksimum sıcaklık, minimum sıcaklık, basınç ve bağıl nem’dir. Öncelikle bu dört parametre ile rüzgar hızı arasındaki ilişkinin belirlenmesi amacıyla korelasyon analizleri yapılmıştır. Ardından, eğitim amaçlı veri seti oluşturulmuştur. Dört parametreyi içeren ve 2018 yılının tüm aylarına ait 365 satırdan oluşan veriler eğitim verisi olarak belirlenmiştir. 2019 yılının Nisan ayının ilk 16 günlük verileri ise test verisi olarak belirlenmiştir.Rüzgar hızı değerleri1 ile 5 arasında matematiksel olarak sınıflandırılmıştır. Daha sonra, AÖM burada oluşturulan 365 adet verinin 335 tanesini eğitim verisi olarak kalan 30 tanesini ise test verisi olarak kullanmıştır.Sonucun güvenilirliği açısından 12-kez çapraz doğrulama yöntemi kullanılmıştır. Sınıflandırmada sine, sigmoid ve hardlimit aktivasyon kodları kullanılmıştır. 8 farklı nöron sayısında çalışma yapılmıştır. Çapraz doğrulama sonucunda en yüksek başarım elde edilen aktivasyon kodu ve nöron sayısı gerçek test verileri olan 2019 yılı Nisan ayının ilk 16 günlük verilerine uygulanmıştır.AÖM tarafından tahmin edilen rüzgar hızısınıflandırma değerleri ile matematiksel sınıflandırma değerleri karşılaştırılarak AÖM’nin başarısı ölçülmüştür. Son olarak, gerçek rüzgar hızı verilerine göre her bir gün için saatlik ortalama rüzgar enerjisi üretim değerleri ile AÖM tarafından yapılan tahmini sınıflandırma sonuçlarına denk gelen saatlik ortalama rüzgar enerjisi üretim değerleri karşılaştırılmıştır.
Article
Full-text available
Hourly wind speed forecasting is presented by a modeling study with possible applications to practical problems including farming wind energy, aircraft safety and airport operations. Modeling techniques employed in this paper for such short-term predictions are based on the machine learning techniques of artificial neural networks (ANNs) and genetic expression programming (GEP). Recorded values of wind speed were used, which comprised 8 years of collected data at the Kersey site, Colorado, USA. The January data over the first 7 years (2005–2011) were used for model training; and the January data for 2012 were used for model testing. A number of model structures were investigated for the validation of the robustness of these two techniques. The prediction results were compared with those of a multiple linear regression (MLR) method and with the Persistence method developed for the data. The model performances were evaluated using the correlation coefficient, root mean square error, Nash–Sutcliffe efficiency coefficient and Akaike information criterion. The results indicate that forecasting wind speed is feasible using past records of wind speed alone, but the maximum lead time for the data was found to be 14 h. The results show that different techniques would lead to different results, where the choice between them is not easy. Thus, decision making has to be informed of these modeling results and decisions should be arrived at on the basis of an understanding of inherent uncertainties. The results show that both GEP and ANN are equally credible selections and even MLR should not be dismissed, as it has its uses.
Article
Full-text available
11 In this study, multiple linear regression (MLR), multilayer perceptron (MLP) 12 regression, and support vector regression (SVR) are used to make multivariate tourism 13 forecasting for Turkey. Comparison of forecasting performances in terms of relative 14 absolute error (RAE) and root relative squared error (RRSE) measurements shows that 15 SVR model gives a better performance. This paper firstly is a comparative study of data 16 mining techniques based on multivariate regression modelling with monthly data points 17 to forecast the tourism demand and secondly focuses on Turkey region. Both MLP and 18 SVR methods are widely employed in the variety forecasting problems. However, the 19 application of them for multivariate forecasting in the context of tourism demand has 20 not been widely explored. Lastly, the results obtained in this study provide information 21 for researchers interested in applying data mining techniques to the tourism demand 22 forecasting and help the policy makers, government bodies, investors and managers for 23 their regularization, planning and investments by the way of the accurate tourism 24 demand forecasting.
Article
Over the past few years, significant progress has been made in wind power generation worldwide. Because of the turbulent nature of wind velocity, the management of wind intermittence is a substantial field of research in the wind energy sector. This paper presents an investigation of this problem in two parts, the prediction of wind speed in both temporal and spatial dimensions, using artificial neural networks (ANNs). ANNs are novel methods applicable in modeling of complicated systems such as wind speed which generally investigated by a large amount of registered data exemplifying the behavior of.
Article
Wind speed forecasting aids in estimating the energy produced from wind farms. The soaring energy demands of the world and minimal availability of conventional energy sources have significantly increased the role of non-conventional sources of energy like solar, wind, etc. Development of models for wind speed forecasting with higher reliability and greater accuracy is the need of the hour. In this paper, models for predicting wind speed at 10-min intervals up to 1 h have been built based on linear and non-linear autoregressive moving average models with and without external variables. The autoregressive moving average models based on wind direction and annual trends have been built using data obtained from Sotavento Galicia Plc. and autoregressive moving average models based on wind direction, wind shear and temperature have been built on data obtained from Centre for Wind Energy Technology, Chennai, India. While the parameters of the linear models are obtained using the Gauss–Newton algorithm, the non-linear autoregressive models are developed using three different data mining algorithms. The accuracy of the models has been measured using three performance metrics namely, the Mean Absolute Error, Root Mean Squared Error and Mean Absolute Percentage Error.
Article
This book explores the intuitive appeal of neural networks and the genetic algorithm in finance. It demonstrates how neural networks used in combination with evolutionary computation outperform classical econometric methods for accuracy in forecasting, classification and dimensionality reduction. McNelis utilizes a variety of examples, from forecasting automobile production and corporate bond spread, to inflation and deflation processes in Hong Kong and Japan, to credit card default in Germany to bank failures in Texas, to cap-floor volatilities in New York and Hong Kong.
Article
Wind speed or wind power forecasting plays an important role in large-scale wind power penetration due to their uncertainty. Support vector regression, widely used in wind speed or wind power forecasting, aims at discovering natural structures of wind variation hidden in historical data. Most current regression algorithms, including least squares support vector regression (SVR), assume that the noise of the data is Gaussian with zero mean and the same variance. However, it is discovered that the uncertainty of short-term wind speed satisfies Gaussian distribution with zero mean and heteroscedasticity in this work. This kind of task is called heteroscedastic regression. In order to deal with this problem, we derive an optimal loss function for heteroscedastic regression and develop a new framework of ν-SVR for learning tasks of Gaussian noise (GN) with heteroscedasticity. In addition, we introduce the stochastic gradient descent (SGD) method to solve the proposed model, which leads the models to be trained online. Finally, we reveal the uncertainty properties of wind speed with two real-world datasets and test the proposed algorithms on these data. The experimental results confirm the effectiveness of the proposed model.
Article
Usually, a typical wind turbine system cannot generate electricity power consistently, because the power outputs are heavily depended on wind speed. However, terrain, temperature, humidity and other factors can also affect wind speed. Therefore, wind power forecasting is a complex, multi-dimensional, and highly non-linear system. Neural network is able to learn the relationship between system inputs and outputs without mathematical conversion, and perform complex non-linear mapping, data classification, knowledge processing, and so forth. In addition, neural network also has the ability of parallel processing to reduce computing time, so it is suitable for wind power forecasting. The purpose of this paper is to use neural network technology to design a wind power forecasting system. Moreover, the efficiency analysis of the proposed wind power forecasting system in Kinmen farm is described. Finally, we use MATLAB to implement the proposed wind power forecasting system in Kinmen farm, which is capable of forecast within 48-hours ahead.
Article
The use of wind energy for power electric systems attempts to reduce the dependence on fuel-based energy. With the aim of generating electrical power based on wind energy, it becomes necessary to model and predict wind speed. Wind speed observations are packed with outliers what makes it difficult to propose accurate predictors. This paper presents an experimental comparison of three different methods for making a Kalman filter robust to outliers in the context of one-step-ahead wind speed prediction. Two wind speed databases were used to test the predictive performance of the algorithms. The performance for all the methods is measured in terms of skewness and kurtosis for the predicted signal. The algorithms discussed worked efficiently in a sequential approach, and outperformed the standard Kalman filter.