Application of Artificial Neural Network for Short

Term Wind Speed Forecasting

Tarlochan Kaur, Sanjay Kumar

Ravi Segal

Electrical Engineering Department

PEC University of Technology
Chandigarh, India
tarlochan.kaur@gmail.com
sanjnitham@gmail.com

Energy Management
GE Bangaluru
Bangaluru, India
Ravi.Segal@ge.com

Abstract The wind speed forecasting is important for managing

issues arising because of intermittency and variability of wind in
wind power integration in electrical power systems. The main
objective of this work is short-term forecasting of wind speed on
the basis of historical time series meteorological data. Nonlinear
data driven model is used here for forecasting the wind speed as it
is a powerful and adaptive modeling tool for forecasting purpose.
Five different ANN models are developed. The minimum mean
square error (MSE) for these models are found on the basis of
network performance. The developed model with least error can
be used for forecasting wind speed for predicting power
generation from wind turbines within error of 30% (as per
CERC) .
KeywordsANN; wind speed forecasting; time series

I.

INTRODUCTION

Wind Energy being a clean and renewable power source

should be exploited more and more so as to have larger share in
the energy basket of India. As on March 2015, the contribution
of power from wind resoursce is only 8.6%, i.e 23.4 GW in the
total installed capacity of power which stands at 263.66 GW
[Source: MNRE GoI; CEA Statistics]. One of the key aspect
associated with wind powers optimal integration into
electricity grid is the management of its variability. According
to Indian Electricity, the wind generators are responsible for
forecasting their generation with an accuracy of 70%. Since,
wind power generation is a direct function of cube of wind
speed. Wind speed is variable. The best way to deal with the
wind intermittency and variability is to forecast wind speed.
Therefore, wind speed modeling and forecasting are very
important for wind power applications.
The wind speed estimation, specifically, the short-term
forecast is vital for scheduling, controlling and dispatching the
actual power generated [1]. There has been widespread
diffusion of wind energy for electricity generation in recent
years in developing countries like India very fast manner in
recent years. As renewable energy resources are expanding, it
has been a focus of many researchers [2-6]. Increased wind
energy applications, made necessary to integrate wind energy
with conventional power grid. So proper planning of utilities is
more significant to reduce excess capacity [7]. However, wind
speed is commonly reflected as most challenging parameters to
be forecasted due to haphazard fluctuations. On the basis of

978-1-4673-6660-1/16/$31.00 2016 IEEE

various applications, wind forecasting is classified on the

different time scale zone, the short-term of few seconds to 30
minutes ahead, short-term 30 minutes to 6 hours ahead,
medium term 6 hours to 1 day ahead and long term 1 day to 1
week or more ahead.
In the literature many methods have been reported such as
physical models, spatial correlation models, and mathematical
models, artificial intelligence models for the forecasting of
wind speed [8, 9]. Physical models are employed for the long
term forecasting and derived from some physical parameters
such as temperature, pressure, terrain shape and geographic
location. [10-12]. In spatial correlation models, time series of
the estimated points and its neighboring sites are taken into
account to forecast the wind speed at different sites [13-16].
Statistical models always utilize statistical equations to forecast
the wind speed [17, 18]. Artificial Neural Network (ANNs)
models are also used for wind speed forecasting. ANNs have
wide applications in renewable, agriculture, medical,
engineering, space dynamics and manufacturing industries etc.
because of its ability to handle nonlinear complex problem
easily. Sfetsos [19], estimates mean hourly wind speed using
ten minutes time series data. The root mean square errors were
found and it is four times lower than other models which were
based on mean hourly data. Cadenas [20], presented the short
term wind speed forecasting ANN hourly time series model.
The performance indices for this model were mean squared
error and mean absolute error values of 0.0016 and 0.0399,
respectively.
The 3-layered ANN model, feed-forward, back-propagation are
used for the monthly mean daily wind speed data. The input
parameters latitude, longitude and altitude and the month of the
year were used whereas the monthly mean wind speed was
used as the output. Minimum mean absolute percentage error
(MAPE) and correlation coefficient (r) was found. The
estimated values of wind speed using ANN are presented in the
form of monthly maps to visualize monthly map of wind
energy potential for different locations within Nigeria [21].
Two feed forward neural networks are compared to estimate
the hourly wind speed in a coastal region. Nearest and natural
neighbor spatial interpretation method along with the inverse
distance and square distance weighted average interpolation
methods are used and the results are verified for the area of

II.

DATA COLLECTION AND ANALYSIS

A. Analysis of Meterological Parameters Varation for Wind

Speed
This study involves forecasting wind speed for site in Satara
(Lat 1742 N , 7402 E ) in Maharashtra. The data from July
2013 to October 2013 at every 10 minutes intervals for wind
speed, wind direction, output power, terrain (latitude and
longitude). The data pertaining to temperature, pressure and
humidity was collected. Regression analysis was done to find
correlation between wind speed and temperature, pressure and
humidity. Figure 1-3 shows variation of wind speed variations
with temperature, pressure and humidity respectively. As wind
speed is dependent on these parameters, they are taken as input
parameters to the ANN model.

B. Analysis of Wind Speed, Wind Direction and Output

Power Variations
For 1 hour the wind speed variation is not appreciable but
wind power variation for the same period is significant
because a small change in wind speed causes a significant
change in wind power. Therefore 3 wind speeds i.e
(i) wind speed (t),
(ii) wind speed (t+20),
(iii) wind speed (t+40))
were taken as input to the ANN model to predict wind speed
(t+60).

Temperature (C)

Time (Hrs.)

Wind Speed (m/sec)

Variation of Temperature ( July to November 2013)

Wind Direction ()

Fig. 4. Wind speed variation for 1 hour (16 October 2013, 10-11am)

Time (Hrs.)

Time (Hrs.)
Fig. 5. Wind direction variation for 1 hour (16 October 2013, 10-11am)

Temperature (C)
Fig. 2. Variation of Speed with Temperature for July to November 2013

Form the above plot it was observed that wind speed changes
with change in pressure, so pressure should be an input to the
ANN model.

Output Wind Power

(KW)

Fig. 1.

The variation of wind speed with respect to humidity is also

plotted. It is also considered as input to the ANN model.

Wind Speed (m/sec)

study [22]. Ramasamy [23] proposed an artificial neural

network model for 11 location of the western Himalayan state
having mean absolute percentage error and correlation
coefficient 4.55% and 0.98 respectively.

Wind Speed (m/sec)

Time (Hrs.)
Fig. 6. Wind Direction variation for 1 hour (16 October 2013, 10-11am)

III.

Humidity (%)
Fig. 3. Variation of Speed with Humidity for July to November 2013

METHODLOGY

A. Time Series Based ANN Model

The input vector and target vector for the wind forecasting was
decided on this basis. The input matrix having 4x1 order for
every hourly interval with input data wind speed (t), wind

speed (t+20), wind speed (t+40), wind direction, , temperature

(t), pressure (t), humidity (t). The target vector is of order 1x1
matrices for every hourly interval having output data wind
power at interval (t+60). Using the command nnstart the
neural network toolbox for wind forecasting, time series
prediction which predicts the future value based on previous
hour values obtained. Nonlinear input-output time series
problem two series are involved, an input series x(t) and an
output/target series y(t). Here to predict values of y(t) from
previous values of x(t), but without knowledge of previous
values of y(t). This input output model can be written as
Y (t) = f(x (t 1), ..., x(t d))
(1)
For wind forecasting input-output mapping was done because
we want wind power to be the function of wind speed and the
variable of atmosphere and we do not require any past value of
wind power as input. The present wind power should be a
function of present wind power and we do not require any
delay. The trainbr and Levenberg-Marquardt optimization
algorithm is used for the training, testing and validation. This
technique minimizes a combination of squared errors and
weights, and then determines the correct combination so as to
produce a network that generalizes well.
B. Wind Prediction Models with Selected Input
Different ANN models are developed using Artificial Neural
Network nonlinear input-output (NIO) time series. The change
has been made in hidden layer neuron, layers and iteration for
each model to check the forecasting accuracy. The
performance of each model is evaluated by mean square error
(MSE) with respect to number of epochs. The epochs are one
comprehensive MSE plot of training, testing and validation
data. The network is considered as trained network model with
minimum MSE.
C. Statistical Performance Parametrs for Forecast Accuracy
In order to find the accurate model, mean square error (MSE)
parameter was employed for model performance and model
comparison.
This statistical error measures can be defined as

MSE =

1
n

e
t =1

2
t

(2)

The performance is evaluated by plotting mean square error

(MSE) with respect to number of epochs. The performance plot
shows MSE in training, testing and validation data. The
network with minimum MSE in validation is called as the
trained ANN model. The training of network spontaneously
stops when validation error ends, indicating an improvement in
validation of data samples. Training a number of times will
give different results due to random initialization of connection
weights.
The proposed algorithm to calculate the mean square error
(MSE) is shown in Fig. 7.

Different Input Variables

Development of Database for

different ANN

Training & Testing of Data

Change
of
Hidden
Layer
Neurons

Change the Inputs/

Combination of
Inputs

Designing of
Different Models

No
Training & Testing of
Different Models
Is MSE
with in
tolerance?

Calculation of Mean
square error (MSE)

Selection of Most Suitable

Models from ANN
Yes
Wind Speed
Fig. 7. Proposed algorithm for wind speed forecasting

IV.

RESULTS AND DISCUSSIONS

To find the prediction accuracy, five ANN models are

developed (ANN1, ANN2, ANN3, ANN4 and ANN5). In the
ANN1 model trainbr algorithm is used with two hidden layers
having 21 neurons in each layer is developed. In the ANN2
model only the algorithm (Levenberg-Marquardt algorithm) is
changed and hidden layers and numbers of neurons are kept
same.It was observed that performance MSE has increased. In
ANN3 model the hidden layers are changed and mean square
error is evaluated. For ANN4 model 3 layer network having 30
neurons in each layer using trainbr algorithm are used for
training purpose. In ANN5 the network is trained same as that
of ANN4 only iterations are changed from 200 to 1000 and
then MSE is evaluated. The ANN1 model is best predicted
model with minimum MSE. All the models and their
performance are compared in the table I.
TABLE I.
Attribute
Input
Neuron
Number
Hidden
layer
neuron
Output
Neuron

ANN MODELS

Model 1

Model 2

Model 3

Model 4

Model 5

1ann10

2ann19

3ann30

4ann40

5ann50

10

19

30

40

50

Best Validation Performance is 0.95816 at epoch 58

3

10

Train
Validation
Test
Best

MSE AND REGRESSION VALUES FOR THE ANN MODELS

Models
1ann10
2ann19
3ann30
4ann40
5ann50
1ann10
2ann19
3ann30
4ann40
5ann50

Testing

Training

Regression
.83
.92
.86
.85
.86
.87
.93
.88
.87
.86

MSE
2.70
1.13
2.47
2.41
2.35
2.34
1.15
2.14
2.17
2.16

For ANN1 model the performance represent that mean square

error is minimum for 2 layers network at epoch 200. Error set
for the test and validation has comparable characteristics and
best performance has happened near epoch 200 in fig. 8. The
response output for the time series are shown in fig. 9. The
relation between outputs and targets value of the ANN1 model
are represented by correlation coefficient (R). The value of R
fluctuates between 1 and 0 shows input and output relationship
bonding. The value of R should more nearer to 1, shows good
relationship between input and output.

Mean Squared Error (mse)

TABLE II.

10

10

10

10

10

20

30
40
64 Epochs

50

Fig. 9. Variation of MSE of Model 2

18
Actual Measured Value
Model Output

16

Validation: R=0.94113

12
10
8
6
4
2

10
Target

16
14
12

14

Data
Fit
Y=T

12
Wind spedd (m/s)

14

Data
Fit
Y=T

Output ~= 0.88*Target + 0.8

Output ~= 0.87*Target + 0.89

Training: R=0.93487
16

60

10

10

0
0

15

10
Target

1000

2000

3000
4000
Number of data samples

5000

6000

7000

15

16
14

Test: R=0.92373

All: R=0.93409

Data
Fit
Y=T

Data
Fit
Y=T

Output ~= 0.87*Target + 0.9

Output ~= 0.85*Target + 1.1

Fig. 10. Comparison of the real data with the result of Model 2

12
10
8
6
4
2
5

10
Target

15

16
14

V. CONCLUSIONS

12
10
8
6
4
2
5

10
Target

Fig. 8. Regression plots of Model 2

15

The variability of wind speed is an important aspect to be

managed in order to enhance share of wind power in the total
mix of generated power in India. Therefore prediction of wind
speed assumes large significance. This paper presents the
nonlinear input-output (NIO) time series model to forecast
wind speed. In the test the data obtained from NREL has been
presented to the ANN and the output has been calculated.
Diverse configuration of ANNs were generated and compared
through error measures for ensuring the accuracy of the chosen
model.. The 2ann19 model of 19 hidden layers 4 inputs and 1
output was the best for short term wind speed forecasting. The
result was compared with real data and showed very good
accuracy. CERC mandates forecasting wind generation with
an accuracy of 70%. So this model can be used for wind speed
forecasting. As observed, since the data is completely random,

therefore more data processing is likely to reduce the error

function further and refine the model.
REFERENCES
[1]

M. G. de Giorgi, A. Ficarella, and M. Tarantino,Assessment of the

benefits of numerical weather predictions in wind power forecasting
based on statistical methods, Energy, vol. 36, pp. 39683978, 2011.
[2] M.S.Roulston, Kaplan DT, Hardenberg J, Smith LA, Using mediumrange weather forecasts to improve the value of wind energy
production, Renewable Energy, vol. 28, pp. 585-602, 2003.
[3] B. Alboyaci,B. Dursun, Electricity restructuring in Turkey and the
share of wind energy production, Renewable Energy, vol.33, pp. 174958, 2008.
[4] M. Cordeiro, A. Valente, S. Leitao, Wind energy potential of the region
of Trasos- Montes and Alto Douro, Portugal, Renewable Energy,
vol.19, pp. 185-91,2000.
[5] Z.Y. Zhao, J. Hu, J. Zuo,Performance of wind power industry
development in China: a Diamond Model study, Renewable Energy,
vol. 34, pp. 2883-91, 2009.
[6] H. Liu, H.Q. Tian, C. Chen, Y. Li, A hybrid statistical method to
predict wind speed and wind power, Renewable Energy, vol.35, pp.
1857-61, 2010.
[7] R.E. Abdel-Aal, M.A. Elhadidy, S.M. Shaahid,Modeling and
forecasting the mean hourly wind speed time series using GMDH-based
abductive networks, Renewable Energy, vol. 34, pp. 1686-99, 2009.
[8] S. S. Soman, H. Zareipour, O. Malik, P. Mandal, A review of wind
power and wind speed forecasting methods with different time
horizons, In: IEEE north American power symposium, pp. 18, 2010.
[9] E. Cadenas, W. Rivera, Wind speed forecasting in three different
regions of Mexico, using a hybrid ARIMA-ANN model, Renewable
Energy, vol. 35, pp. 2732-2738, 2010.
[10] M. Negnevitsky and C.W. Potter, Innovative short-term wind
generation prediction techniques, In: Proceedings of the power systems
conference and exposition, pp. 60-52006..
[11] L. Landberg, Short-term prediction of the power production from wind
Farms, Wind Engineering and Industrial Aerodynamics, vol. 80, pp.
207-20, 1999.
[12] J.L. Torres, A. Garca, M. De Blas, A. De Francisco, Forecast of
hourly average wind speed with ARMA models in Navarre, Solar
Energy, vol. 79, pp. 65-77, 2005.

[13] K. Lalarukh, ZJ. Yasmin, Time series models to simulate and forecast
hourly averaged wind speed in Quetta, Pakistan. Solar Energy,vol. 61pp.
23-32, 1997.
[14] T.G. Barbounis, J.B. Theocharis, Locally recurrent neural networks for
wind speed prediction using spatial correlation, Information Sciences,
vol. 177, pp. 5775-97, 2007
[15] M.C. Alexiadis, P.S. Dokopoulos, H.S. Sahsamanoglou, Manousaridis
IM. Shorterm forecasting of wind speed and related electrical power,
Solar Energy, vol.63, pp. 61-68, 1998.
[16] U. Focken, M. Lange, K. Monnich, HP. Waldl, H. Georg Beyer, A.
Luig, Short-term prediction of the aggregated power output of wind
farms a statistical analysis of the reduction of the prediction error by
spatial smoothing effects, Wind Engineering and Industrial
Aerodynamics, vol. 90, pp. 231-46, 2002.
[17] L. Ma, S.Y. Luan, C.W. Jiang, H.L. Liu, Y. Zhang, A review on the
forecasting of wind speed and generated power, Renewable Sustainable
Energy Reviws, vol.13,pp. 15-20,2009.
[18] R.G. Kavasseri, K. Seetharaman, Day-ahead wind speed forecasting
using f-ARIMA models, Renewable Energy, vol. 34, pp. 88-93, 2009.
[19] A. Sfetsos, A novel approach for the forecasting of mean hourly wind
speed time series, Renewable Energy,vol. 27, pp. 163-74, 2002.
[20] D.A. Fadare, The application of artificial neural networks to mapping
of wind speed profilefor energy application in Nigeria, Applied Energy
vol. 87, pp.934942, 2010
[21] K. Philippopoulos, D. Deligiorgi, Application of artificial neural
networks for the spatial estimation of wind speed in a coastal region
with complex topography, Renewable Energy, vol. 38, pp. 75-82, 2012.
[22] P.Ramasamy, S.S. Chandel, A.K.Yadav, Wind speed prediction in
moutains region of India using artificial neural network, Renewable
Energy,vol.80, pp. 338-347, 2015.
[23] http://www.accuweather.com/en/in/india-weather