You are on page 1of 5

2019 2nd International Conference on Intelligent Computing, Instrumentation and Control Technologies (ICICICT)

Wind Speed Forecasting using Long Short Term


Memory Networks

Prabha P.P V. Vanitha Resmi R


Department of Electrical and Department of Electrical and Department of Electrical and
Eectronics Engineering, Eectronics Engineering, Eectronics Engineering,
Amrita School of Engineering, Amrita School of Engineering, Amrita School of Engineering,
Coimbatore Coimbatore Coimbatore
Amrita Vishwa Vidyapeedum, India Amrita Vishwa Vidyapeedum, India Amrita Vishwa Vidyapeedum, India
prabhapanakkal18@gmail.com vvanitha55@gmail.com resmidhun2010@gmail.com

Abstract— Globally, wind power industry is growing very wind speed. [10] dealt with seasonal wind speed forecasting
fast, but intermittency of the wind is causing difficulty in using ANFIS. This paper proposes Long Short Term Memory
integrating wind power to the grid due to issues such as power (LSTM) for the forecasting of wind speed.
scheduling, system control and dispatch, voltage regulation and
frequency response reserve, energy imbalance etc. So, accurate II. LONG SHORT TERM MEMORY NETWORK
wind speed prediction is recommended for scheduling the wind
LSTM is a model of the deep learning network, proposed
power to maintain stability in the power system. In this paper,
by Sepp hochetrier and Jurgen schmidhuber in 1997, which is
long short term memory network (LSTM) is proposed for
predicting wind speed one hour ahead by clustering the time used as a complex non-linear unit to construct a deep neural
series data into windy and non-windy months. LSTM model is network [7]. It can reflect the long term memory and also has
applied to forecast wind speed for four different sites and results the ability of the deep neural network. The LSTM network
are presented. helps in preserving the error that can be back propagated [7]
[3]. The information of the LSTM is contained in the gated
Keywords— one hour-ahead, long short term memory network, cell. Fig.1 shows the structure of the LSTM showing an input
Root Mean Square Error, seasonal forecast, Wind speed layer, an output layer and a memory unit.
forecasting.

I. INTRODUCTION
Fluctuations of wind result in lots of stability and reliability
issues in the grid. Unforeseen increase or decrease in the
amount of wind power generation may result in inefficient
market design, reserve capacity reduction, improper
scheduling, high operational cost for both wind power
generators and conventional generators and curtailment loss.
Forecasting the wind power generation well in advance
resolves many issues such as imbalance charges, efficient
maintenance and operations planning etc. Wind forecasting
methods are classified into four types- physical, statistical & Fig. 1. Stucture of LSTM [8]
computational and hybrid model [1] The physical method uses
the data from numerical weather predictors (NWP). The The LSTM cells are governed by the Equations (1) to (3)
statistical method includes the time series analytical methods i t  g(w xi x t  w hi s t 1  bi )
(1)
like auto regression and moving average methods. [2] applied
Forget Layer:
Auto Regressive Integrated Moving Average(ARIMA) for
wind speed forecasting. The computational method is based on f t  g(w xf x t  w hf s t 1  b f )
(2)
artificial intelligence, such as neural network, Fuzzy logic, Output layer:
ANFIS etc, which helps in increasing the wind forecasting o t  g(w xo x t  w ho  b o )
accuracy. The deep neural network is the advanced method in (3)
the artificial intelligence [4]. By using the networks of deep
neural networks like recurrent neural network, convolution Where,
network and stacked auto-encoders, the wind forecasting 𝑥𝑡 is the input to the LSTM unit at the present step t.
errors can be reduced. Deep learning is used mainly for the big 𝑠𝑡 is the hidden input of the LSTM at the present step t.
data analysis and it has the powerful capacity of learning big g denotes the gate function.
data and generalizing the big data [5]. The deep learning 𝑏𝑖 ,𝑏𝑜 , 𝑏𝑓 denotes the bias at each layer.
techniques like the stacked auto encoder and the deionizing 𝑊𝑥𝑖 , 𝑊𝑥𝑓 , 𝑊𝑥𝑜 , 𝑊ℎ𝑖 , 𝑊ℎ𝑓 , 𝑊ℎ𝑜 - Weights at each layer.
stacked auto encoder is used for the ultra-short term wind
speed forecasting [6]. [9] discusses about the application of
Neural network, ARIMA and hybrid models for predicting

978-1-7281-0283-2/19/$31.00 ©2019 IEEE 1310

Authorized licensed use limited to: Newcastle University. Downloaded on May 31,2020 at 04:18:11 UTC from IEEE Xplore. Restrictions apply.
2019 2nd International Conference on Intelligent Computing, Instrumentation and Control Technologies (ICICICT)

TABLE I. STATISTICS OF WIND DATA

Site Available Year Mean Minimum Maximum


years of of wind Wind Wind
data data speed Speed Speed
used (m/s) (m/s) (m/s)

Jagmin 2009- 2013 2011 6.839 0.4 23.1

Jogimatti 2011-2015 2012 8.544 0.383 18.7

Kayathar 2013 - 2013 6.647 0.28 22.86


2016

Lamba 2009-2014 2013 5.947 0.27 17.59

20

Wind Speed(m/s)
15
10
5
0
0 200 400 600 800
Time(hours)
Fig. 3. Variations of wind speed for jagmin for january 2012
Fig. 2. Flowchart for LSTM
20
The number of input neurons in each layer and number of
Wind Speed(m/s)

hidden layers influence training of neural network and hence 15


its accuracy [8]. The model becomes complex when the hidden 10
neurons and the neurons in each layer increase. If the number
of neurons are more, the error gets reduced but the training 5
time gets prolonged [8-9]. Fig. 2 shows the flow chart for the 0
LSTM network. For the coding of the LSTM network, the 0 200 400 600 800
input data is added in .csv format. Then, this data is shaped, Time(hours)
scaled and split as testing data and training data. The training
data is given to the LSTM network for its modelling and fitting Fig. 4. Variations of wind speed for jogimatti for january 2012
followed by the wind prediction using the testing data. If the
error calculated is less, then the data is written in the excel and 20
Wind Speed(m/s)

if the error is more, then tuning of LSTM is done by varying


the number of epochs or by changing the batch size or by 10
varying the LSTM units in the model
0
0 200 400 600 800
III. WIND SPEED FORCASTING USING LSTM Time(hours)
The LSTM model is developed for the calculation of wind
speed for one hour ahead forecasting. The wind data is Fig. 5. Variations of wind speed for kayathar for january 2013
collected from National Institute of Wind Energy (NIWE) link
for the four sites in India. The wind speed data is available at 20
Wind Speed(m/s)

a time interval of 10 minutes for each site, whose details are


given in Table 1. The pre-processing of data such as data 15
filling, data cleansing and data conversion is done using 10
Windographer.
5
0
0 200 400 600 800
Time(hours)
Fig. 6. Variations of wind speed for lamba for january 2013

978-1-7281-0283-2/19/$31.00 ©2019 IEEE 1311

Authorized licensed use limited to: Newcastle University. Downloaded on May 31,2020 at 04:18:11 UTC from IEEE Xplore. Restrictions apply.
2019 2nd International Conference on Intelligent Computing, Instrumentation and Control Technologies (ICICICT)

Fig. 3 to Fig. 6 show the variations of wind speed for The forecasting accuracy is calculated is using the root
different sites during the month of January. Fig. 7 and Fig. 8 mean square error (RMSE) given by equation (4)
show the wind speed trend for Jagmin in 2012 and 2013
respectively. The wind speed seasonal fluctuations can be 1 m 2
observed in consecutive years. For wind speed forecasting, the
RMSE  e t
M t 1
yearly wind speed is clustered into two groups as high wind (4)
months (April – October) and low wind months (January – The RMSE is calculated for the four sites Jagmin,
March & November-December). Fig. 9 shows the flow chart Jogimatti, Kayathar and Lamba during high wind months and
for the wind speed forecasting using LSTM method. The wind low wind months and shown in Table 2. The advantage of
speed data for a year is split into high wind and low wind seasonal clustering is that the prediction models gets more
months. Then the wind speed forecasting for one hour ahead similar inputs, thus resulting in better accuracy. Results show
is done taking 3 months data for training, fourth month data that the RMSE error is less for the high wind period and more
for testing and fifth month data for validation. for low wind period

TABLE II.RMSE FOR FOUR SITES


20
Wind Speed(m/s)

Sites RMSE (m/s)


15 High Wind Months Low Wind Months

10 Jagmin 0.5 0.72

5 Jogimatti 0.2 0.4

0 Kayathar 0.78 0.91


0 1 2 3 4 5 6 7 8 9 10 11 12
Months
Lamba 0.8 0.98

Fig. 7. Wind speed variation trend of jagmin 2012

IV. DEVIATION CHARGE CALCULATION BASED ON THE


20
REGULATION OF WIND SPEED FORECASTING
Wind Speed(m/s)

15 According to the CERC (Central Electricity Regulation


10
Committee), the wind power producers have to pay deviation
charges to the state board if they deviate from their scheduled
5 generation. The deviation payment is calculated as absolute
error, given by Equation (5).
0 actual generation (MW)  scheduled generation(MW)
actual error 
0 1 2 3 4 5 6 7 8 9 10 11 12 available capacity(MW)
(5)
Months For calculating wind power generation, Suzlon 2.1 MW
Fig. 8. Wind speed variation trend of jagmin 2013 wind turbine is considered, whose power curve is shown in
Fig. 10. Actual generation refers to wind power generation
corresponding to actual wind speed and scheduled generation
corresponds to that of predicted wind speed. Available
capacity corresponds to rated power of turbine selected.

2.5
CLUSTERING CLUSTERING
AS HIGH WIND AS LOW WIND 2
Wind Power (MW)

MONTHS MONTHS

1.5
TRAINING TRAINING
USING 3 USING 3
1
MONTHS DATA MONTHS DATA

0.5
FORCASTING FORECASTING
FOR 4th MONTH FOR 4th MONTH 0
0 2 4 6 8 10 12 14 16 18 20 22 24 26
VALIDATING VALIDATING
Wind Speed (m/s)
USING 5th USING 5th Fig. 10. Power curve for SUZLON 2.1 MW
MONTH DATA MONTH DATA

Fig. 9. Wind speed forecasting using LSTM

978-1-7281-0283-2/19/$31.00 ©2019 IEEE 1312

Authorized licensed use limited to: Newcastle University. Downloaded on May 31,2020 at 04:18:11 UTC from IEEE Xplore. Restrictions apply.
2019 2nd International Conference on Intelligent Computing, Instrumentation and Control Technologies (ICICICT)

TABLE III. REGULATON FOR MAHARASHTRA


800

Number Of Predicted
Sl Deviation
Absolute error
no charges(Rs) 600

Samples
1. <10% No penalty 400

2. 10 to 20 % 0.50 / per unit 200

3. 20 to 30% 1.50 / per unit 0


<10% 10 TO 20% 20 TO 30% >30%
4. >30% 1.75 / per unit Penalty Blocks
Fig. 11. No. of predicted samples coming under each penalty for jagmin
TABLE IV. REGULATON FOR KARNATAKA
800

Number Of Predicted Samples


Sl.
Absolute Error Deviation Charges(Rs)
No.
600
1 <15% No penalty
400
2 15 to 25% 0.50 Ps./kWh
200
3 25-35% 1.0/kWh
0
4 >35% 1.50/kWh <15% 15 TO 25% 25 TO 35% >35%
Penalty Blocks
TABLE V. REGULATON FOR GUJARAT
Fig. 12. No. of predicted samples coming under each penalty for jogimatti
Sl. No. Absolute Error Deviation Charges(Rs)
800
Number Of Predicted Samples

1 <12% No penalty 700


600
2 12 to 20% 0.35 Ps./kWh
500
400
3 20-28% .70 /kWh
300
4 >28% .1.05/kWh
200
100
0
TABLE VI. REGULATON FOR TAMILNADU
<12% 12 TO 20% 20 TO 28% >28%
Penalty Blocks
Deviation charges Fig. 13. No. of predicted samples coming under each penalty for Kayathar
Sl no Absolute error
(Rs)

1. <10% No penalty 800


Number Of Predicted

600
Samples

2. 10 to 20 % 0.50 / per unit


400
3. 20 to 30% 1.50 / per unit
200
4. >30% 2.75 / per unit
0
<12% 12 TO 20%20 TO 28% >28%
The deviation percentage of predicted samples using Penalty Blocks
LSTM is plotted for each site, which are show in Fig.11- Fig. 14. No. of predicted samples coming under each penalty for Lamba
Fig.14.

978-1-7281-0283-2/19/$31.00 ©2019 IEEE 1313

Authorized licensed use limited to: Newcastle University. Downloaded on May 31,2020 at 04:18:11 UTC from IEEE Xplore. Restrictions apply.
2019 2nd International Conference on Intelligent Computing, Instrumentation and Control Technologies (ICICICT)

Table 7 shows the penalty charges for a unit to be paid to the


state electricity board for a period of one month

TABLE VII. PENALTY CHARGES FOR EACH SITE IN JANUARY

Site Penalty Charges /Month(Rs)

Jagmin 43.5

Jogimatti 36.7

Kayathar 42

Lamba 11.5

V. CONCULSION
In this paper, LSTM model based on seasonal data is used
for one hour ahead wind speed forecasting in Jogimatti,
Jagmin, Kayathar and Lamba. The results shows that the
seasonal wind speed forecasting using LSTM model can
reduce the forecasting errors. The deviation charges are
calculated for a 2.1MW wind turbine based on the predicted
wind speed. It is seen that most of the blocks are coming under
no penalty region and only less charges need to be paid to state
load centre.
REFERENCES
[1] David Barbosa de Alencar, Carolina de Mattos Affonso, Roberto Célio
Limão de Oliveira, Jorge Laureano Moya Rodríguez , Jandecy Cabral
Leite and José Carlos Reston Filho (2017) Different models for
forecasting wind power generation: case study In: ENERGIES, vol 10,
2017, pp no.1-27
[2] Vishnupriyadharshini, A., Vanitha, V., Palanisamy, T.”Wind speed
forecasting based on statistical Auto Regressive Integrated Moving
Average (ARIMA) method”, International Journal of Control Theory
and Applications” Vol.9, No.15, pp. 7681-7690,2016.
[3] Shilong Yu(2018) Residual Learning and LSTM networks for wearable
human activity recognition In: preceedings of 37 th Chinese control
interference, Wuhan, China, K. Elissa, “Title of paper if known,”
unpublished.
[4] Hui Liu, Xiwei Mi, Yanfei Li (2018) Smart deep learning based wind
prediction model using wavelet decomposition, convolution network
and convolution long shirt term network In: energy conversion and
management 166, pp no. 120-131
[5] Qu Xiaoyun, Kang Xiaoning, Zhang Chao, Jiang Shuai, Ma Xiuda
(2016) Short-Term Prediction of Wind Power Based on Deep Long
Short-Term Memory In: 2016 IEEE PES Asia-Pacific Power and
Energy Conference - Xi'an – China, pp no. 1148-1152.
[6] Mahdi Khodayar, Okyay Kaynak, Mohammad E. Khodaya r(2017)
Rough Deep Neural Architecture for Short-Term Wind Speed
Forecasting In: IEEE Transactions On Industrial Informatics, Vol. 13,
No. 6, pp no. 2770-2780, December
[7] Anderson Tenório Sergio, and Teresa B. Ludermir (2015) Deep
Learning for Wind Speed Forecasting in North eastern Region of Brazil
In: 2015 Brazilian Conference on Intelligent Systems. pp. no.322-327.
2015
[8] https://medium.com/@kangeugine/long-short-term-memory-lstm-
concept-cb3283934359
[9] Nair, K.R., Vanitha, V., Jisma, M.,”Forecasting of wind speed using
ANN, ARIMA and Hybrid models”, International Conference on
Intelligent Computing, Instrumentation and Control Technologies,
ICICICT 2017, pp. 170-175.
[10] Delna Raphel, V.Vanitha (2018) Seasonal wind speed forecasting using
ANFIS In: 2018 Biennial International Conference on Power and
Energy Systems: Towards Sustainable Energy (PESTSE) (in press)

978-1-7281-0283-2/19/$31.00 ©2019 IEEE 1314

Authorized licensed use limited to: Newcastle University. Downloaded on May 31,2020 at 04:18:11 UTC from IEEE Xplore. Restrictions apply.

You might also like