Content uploaded by Nadeem Javaid

Author content

All content in this area was uploaded by Nadeem Javaid on Nov 04, 2019

Content may be subject to copyright.

Wind Power Forecasting Based

on Eﬃcient Deep Convolution

Neural Networks

Sana Mujeeb1, Nadeem Javaid1(B

),HiraGul

1, Nazia Daood1,

Shaista Shabbir2, and Arooj Arif1

1COMSATS University Islamabad, Islamabad 44000, Pakistan

nadeemjavaidqau@gmail.com

2Virtual University, Kotli 11100, Pakistan

http://www.njavaid.com/

Abstract. Due to the depletion of fossil fuel and global warming,

the incorporation of alternative low carbon emission energy generation

becomes crucial for energy systems. The wind power is a popular energy

source because of its environmental and economic beneﬁts. However, the

uncertainty of wind power, makes its incorporation in energy systems

really diﬃcult. To mitigate the risk of demand-supply imbalance by wind

power, an accurate estimation of wind power is essential. Recognizing

this challenging task, an eﬃcient deep learning based prediction model

is proposed for wind power forecasting. In this proposed model, Wavelet

Packet Transform (WPT) is used to decompose the wind power signals.

Along with decomposed signals and lagged inputs, multiple exogenous

inputs (calendar variable, Numerical Weather Prediction (NWP)) are

used as input to forecast wind power. Eﬃcient Deep Convolution Neural

Network (EDCNN) is employed to forecast wind power. The proposed

model’s performance is evaluated on real data of Maine wind farm ISO

NE, USA.

Keywords: Data analytics ·Wind power ·Demand side

management ·Energy management ·Forecasting ·Deep learning

1 Introduction

Due to the continual decrease in fossil fuel, the energy crisis has become crucial

[1]. To mitigate the energy crisis, regulative acts that encourage the utiliza-

tion of renewable energy are promoted worldwide. Among the renewable energy

resources, wind energy, as an alternative to traditional generation, has attracted

a lot of attention. The reason of popularity of wind power is its environment

friendly nature. Wind power has no carbon emission and therefore, helps in

reducing environmental pollution [2]. It is introduced worldwide as a way to

reduce greenhouse gas emission. According to the Global Wind Energy Council

[3], the cumulative capacity of wind power reached 486 GW across the global

c

Springer Nature Switzerland AG 2020

L. Barolli et al. (Eds.): 3PGCIC 2019, LNNS 96, pp. 47–56, 2020.

https://doi.org/10.1007/978-3-030-33509-0_5

48 S. Mujeeb et al.

market in 2016. Wind power is expected to signiﬁcantly expand leading to an

overall zero emission power system.

The wind power is not only environmental friendly, but it also has a low

investment cost (due to the developing technology) [4]. In the USA, the U.S.

Department of Energy target of renewable integration is responsible of providing

20% of the total energy through wind, by the year 2030 [5]. In this regard, the

Independent System Operators (ISOs) are producing signiﬁcant wind power and

increasing their wind generation.

It is acknowledged widely that accurate WPF signiﬁcantly reduces the risks

of incorporating wind power in power supply systems [6]. Generally, the WPF

results are in the deterministic form (i.e., point forecast). Reducing the fore-

casting errors of WPF is the focus of many researchers [7]. A point forecast is

the estimated value of future wind energy. However, the wind power is random

variable having a Probability Density Function (PDF), and point forecasts is

unable to capture the uncertainty of this random variable. This the limitation of

the point forecasts. Therefore, point forecasts have limited use in stability and

security analysis of power systems. To overcome the limitation of point forecasts,

deep learning methods are widely used in the ﬁeld of WPF and other electricity

related forecasting tasks [8–10]. Deep Neural Networks (DNN) have the inherent

property of automatic modeling of the wind power characteristics [11].

The wind power has a chaotic nature. Therefore, the incorporation of wind

power in power supply systems is a risky task. To mitigate this risk, wind power

forecasting is the most popular method. The wind power is forecasted using clas-

sical [9–17], statistical and artiﬁcial intelligent methods. In literature, there are

two types of wind power forecasting techniques: time series [12] and multivari-

ate [13]. The accuracy of wind power forecasting is very important to avoid the

demand supply imbalance. Therefore, researcher are still competing to improve

the wind power forecasting accuracy.

Convolution Neural Network (CNN) is a state-of-the-art deep learning

method. It is the CNN’s characteristic that it can extract the spatial features

automatically. CNN is the most popular method for extracting features from

the images and widely used in the ﬁeld of computer vision. The eﬃcient feature

extraction capability of CNN motivate us to exploit it for wind power forecast-

ing. CNN successfully extracts the spatio-temporal correlations in wind power

data.

2 Contributions

In this paper, we are concerned with the problem of predicting the wind power.

The contributions of this research work are listed below:

– For forecasting wind power, the Numeric Weather Prediction (NWP) are used

along with lagged wind power and Wavelet Packet decomposed (WPT) past

wind.

Eﬃcient Deep Convolution Neural Networks Based Wind Power Forecasting 49

Fig. 1. Overview of proposed system for wind power forecasting.

– A predictive Deep Convolution Neural Network (DCNN) model for accurate

wind power prediction is proposed, that employs an eﬃcient activation func-

tion and loss function in output layer (Fig.1).

3 Proposed Model

The proposed method for forecasting wind power generation and power man-

agement algorithm are discussed in this section.

3.1 Data Preprocess

The features and targets (wind power) are normalized using min-max normal-

ization (as shown in Eq. 1).

Xnor =Xi−min(X)

max(X)−min(X)(1)

The inputs to the forecast model are shown in Table 1. Three types of inputs are

given to the forecasting model are: (i) Numerical Weather Prediction (NWP):

dew point temperature, dry bulb temperature, wind speed, (ii) past lagged values

of wind power and (iii) wavelet packet decomposed wind power. The wavelet

decomposition is described in the next section.

3.2 Feature Engineering

The historical wind power signal is decomposed using Wavelet Packet Trans-

form (WPT). The WPT is a general form of the wavelet decomposition which

50 S. Mujeeb et al.

Table 1. Inputs to the forecast model.

Input Description

Dew point temperature Past NWP forecast

Dry bulb temperature Past NWP forecast

Wind speed Past NWP forecast

Lagged wind power 1 Wind power (t-24)

Lagged wind power 2 Wind power (t-25)

Decomposed wind power Wavelet decomposed past wind power

Hour Time of the day

Fig. 2. Wavelet packet tree with three levels.

performs a better signal analysis. WPT was introduced in 1992 by Coifman and

Wickerhauser [14]. Unlike, Discrete Wavelet Transform (DWT), the WPT wave-

forms or packets that are interpreted by three diﬀerent parameters: frequency,

position and scale (similar to the DWT). For every orthogonal wavelet func-

tion multiple wavelet packets are generated, having diﬀerent bases (as shown in

Fig. 2). With the help of these bases, the input signal can be encoded in such a

way that the global energy of signal is preserved and exact signal can be recon-

structed eﬀectively. Multiple expansions of an input signal be achieved using

WPT. The suitable most decomposition is selected by calculating the entropy

(e.g. Shannon entropy). The minimal representation of the relevant data based

on a cost function is calculated in WPT. The beneﬁt of WPT is its characteristic

of analyzing signal in diﬀerent temporal as well as spatial positions. For highly

nonlinear and oscillating signal like wind power DWT doesn’t guarantee good

results [15]. In WPT, both the approximation and detail coeﬃcients are further

decomposed into approximation and detail coeﬃcients as the level of tree goes

deeper. Wavelet packet decomposition operation can be expressed Eqs. 2and 3.

Eﬃcient Deep Convolution Neural Networks Based Wind Power Forecasting 51

For a signal ato be decomposed, two ﬁlters of size 2Nare applied on a.The

corresponding wavelets are h(n) and g(n).

W2n(a)=√2

2N−1

k=0

h(k)Wn(2a−k) (2)

W2n+1(a)=√2

2N−1

k=0

g(k)Wn(2a−k) (3)

Where, the scaling factor W0(a)=φ(a) and the wavelet function is W1(a)=

ψ(a).

After decomposing the past wind signals, the engineered features along with

NWP variables and lagged wind power are input to the proposed forecasting

model. The proposed forecasting model is discussed in the next section.

3.3 Eﬃcient DCNN

The inputs are given to the Eﬃcient Deep CNN (EDCNN) for predicting day-

ahead hourly wind power (24 values). Firstly, the functionality of trivial CNN is

discussed in this section. Secondly, the proposed method EDCNN is explained.

CNN works on the principle of visual system of human brain. CNN has an

excellent capability of extracting deep underlying features from data. The CNN

eﬀectively identify the spatially local correlations in data through convolution

operation. In the convolution operation, a ﬁlter is applied to a block of spatially

adjacent neurons and result is passed through an activation function. This output

of convolution layer becomes the input to next layer’s neuron. Thus, the input to

every neuron of a layer is the output of convolved block of previous layer. Unlike

ANN, the CNN training is eﬃcient due to the weight sharing scheme. Due to

the weight sharing, the learning eﬃciency improves. CNN is composed of three

altering layers: (i) convolution layer, (ii) sampling layer and (iii) fully connected

layer. The convolution operation can be explained by following Eq. 4.

Suppose, X = [x1,x

2,x

3,...,x

n] are the training samples and C =

[c1,c

2,c

3,...,c

n] is the vector of corresponding targets. nis the number of train-

ing samples. CNN attempts to learn the optimal ﬁlter weights and biases that

minimize the forecasting error. CNN can be deﬁned as:

Ym

i=f(wm⊗Xm

i+bm) (4)

Where, i = [1, 2, ..., n] and m = [1, 2, ..., M]. mis the number of layer to be

learned. The ﬁlter weights of the mth layer is denoted by wm.bmrepresents the

corresponding biases, ⊗refers to the convolution operation. f(·) is the nonlinear

activation function. Ym

iis the feature map generated by sample Xiat layer m.

The proposed forecasting method EDCNN, there are eleven layers: three con-

volution layers, three max pooling layers, two batch normalization layers, three

ReLU (Rectiﬁed Linear Unit) layers, one modiﬁed fully connected layer and

52 S. Mujeeb et al.

modiﬁed output layer (Enhanced Regression Output Layer (EROL)). Function-

ality of two layers are modiﬁed, in order to improve the forecasting performance

of EDCNN.

According to the ANN literature, there is no standard way to choose an

optimal activation function. However, its a well-known fact that machine learning

methods have an excellent optimization capability of any model or function.

On basis of these facts, a modiﬁed activation function is employed in a hidden

layer. The proposed activation function is ensemble of three activation functions:

hyperbolic tangent, sigmoid and radial base function.

TH =ex−e−x

ex+e−x(5)

σ=ex

1+ex(6)

φ(x, c)=φx−c(7)

F=(TH +σ+φ)

3(8)

In the proposed output layer EROL, a modiﬁed objective function is embedded.

The objective is to minimize the absolute percentage error between the forecasted

values and actual targets. The objective can be expressed as 9:

min Loss =L(w, Xi,c

i)+F(Yk,c) (9)

Where, L(w, Xi,c

i) is the forecasting error or loss from sample Xi,andF(Yk,c)

represents the objective. The input to objective function F(·) are the feature

maps generated at the kth layer and care their respective targets. The objective

function is expressed as 10:

F=1

n

n

i=1

Yi−ci

Yi

100 (10)

0 50 100 150

Time (H)

0

200

400

600

800

Wind Power (MW)

Spring

0 50 100 150

Time (H)

0

200

400

600

Wind Power (MW)

Summer

0 50 100 150

Time (H)

0

200

400

600

Wind Power (MW)

Autumn

0 50 100 150

Time (H)

0

500

1000

Wind Power (MW)

Winter

Fig. 3. Wind power of all four seasons of a year.

Eﬃcient Deep Convolution Neural Networks Based Wind Power Forecasting 53

Where, Yiis the output at the output layer and ciis the desired or actual target.

4 Results and Analysis

The proposed algorithms are implemented using MATLAB R2018a on a PC with

core i3 processor and 4 GB RAM.

4.1 Data Description

The three year hourly data of wind power is taken from ISO New England’s

wind farm located in Maine [16]. The weather parameter, i.e., wind speed data

is taken from Maine weather station data repository.

4.2 Wind Power Analysis

The wind power is directly proportional to the wind speed. The wind speed vary

from season to season. In Maine, USA the wind speed is eﬀected by seasonality.

In Fig. 3, the one day wind power of all the four seasons, is shown. The wind

power in autumn is higher compared to other seasons. The reason behind this is

the fast winds in coastal area of Maine, where the wind turbines are installed.

4.3 Performance Evaluation

For performance evaluation of wind power forecasting, three evaluation indica-

tors are used: MAE and NRMSE and MAPE (Fig. 4and Table 2).

0 5 10 15 20 25

Time (H)

100

200

300

400

500

Wind Power (MW)

Spring

Observed

ECNN

SELU CNN

CNN

0 5 10 15 20 25

Time (H)

200

400

600

800

Wind Power (MW)

Summer

Observed

ECNN

SELU CNN

CNN

0 5 10 15 20 25

Time (H)

600

700

800

900

1000

Wind Power (MW)

Autumn

Observed

ECNN

SELU CNN

CNN

0 5 10 15 20 25

Time (H)

0

200

400

600

Wind Power (MW)

Winter

Observed

ECNN

SELU CNN

CNN

Fig. 4. All seasons predictions of wind power.

54 S. Mujeeb et al.

Table 2. MAPE and NRMSE of proposed and compared methods.

Method Season MAPE NRMSE MAE

CNN Spring 8.42 2.34 3.34

Summer 8.23 2.27 3.24

Autumn 7.9 2.65 3.36

Winter 8.1 2.71 2.89

SELU CNN Spring 3.47 0.12 3.1

Summer 3.62 0.13 3.3

Autumn 3.45 0.12 3.4

Winter 3.27 0.17 3.2

EDCNN Spring 2.67 0.092 2.4

Summer 2.43 0.096 2.24

Autumn 2.56 0.085 2.67

Winter 2.62 0.094 2.18

Table 3. Diebold-Mariano test results at a 95% conﬁdence level.

Method Season Diebold-Mariano

EDCNN vs. CNN Spring

Summer

Autumn

Winter

EDCNN vs. SELU CNN Spring

Summer

Autumn -

Winter

4.4 Diebold-Mariano Test

The aforementioned error indicator are utilized for accuracy comparison of fore-

casting models. However, the lesser error or higher accuracy of a model doesn’t

guarantee its superiority over other models. A model is better as compared to

another model, if the diﬀerence between their accuracies is statistically signif-

icant. Diﬀerent statistical tests are used to validate the signiﬁcance of models,

such as Friedman test [17], error analysis [18], Diebold-Mariano (DM) test [19],

etc. To validate performance of the proposed forecasting model EDCNN, a well-

known statistical test DM is used. Diebold and Mariano propose the classical

Diebold-Mariano statistical test in 1995 [19]. DM is widely used for validation

of wind power forecasting [20].

A vector of values that are to be forecasted are [y1,y

2, ..., y

n]. These values

are predicted by two forecasting models: M1and M2. The forecasting errors of

these models are [εM1

1,ε

M1

2, ..., ε

M1

n] and [εM2

1,ε

M2

2, ..., ε

M2

n]. A covariance

Eﬃcient Deep Convolution Neural Networks Based Wind Power Forecasting 55

loss function L() and diﬀerential loss are calculated in DM as 11 [21]:

dM1,M

2

t=L(εM1

t)−L(εM2

t) (11)

DM is applied to the forecasting results of EDCNN and two compared methods:

CNN and SELU CNN [13]. The results of DM test with conﬁdence level of 95%

are shown in Table 3. The check marks are shown at the places where the perfor-

mance of EDCNN is signiﬁcantly better as compared to the comparable method.

If the forecasting accuracy is not signiﬁcantly improved, hyphen is placed. The

performance of proposed forecaster EDCNN is compared with standard CNN

and SELU CNN [13]. The predictive analysis are performed for all four seasons

of a year.

5 Conclusion

In this paper, the problem of predicting wind power generation is considered. In

order to take part in the daily market that regulates the supply and demand in

the Maine electricity system. A deep-learning technique EDCNN is developed

to accurately predict the hourly day-ahead wind power on the Maine wind farm

data. The numeric results validates the eﬃciency of proposed model for wind

power forecasting.

References

1. Zhao, Y.N., Ye, L., Li, Z., Song, X.R., Lang, Y.S., Su, J.: A novel bidirectional

mechanism based on time series model for wind power forecasting. Appl. Energy

177, 793–803 (2016)

2. Jong, P., Kiperstok, A., Sanchez, A.S., Dargaville, R., Torres, E.A.: Integrating

large scale wind power into the electricity grid in the Northeast of Brazil. Energy

100, 401–15 (2016)

3. Global Wind Energy Council. GWEC Global Wind Report (2016)

4. U.S. Department of Energy, Staﬀ Report to the Secretary on Electricity Markets

and Reliability (2017)

5. U.S. Department of Energy, 20% Wind energy by 2030: increasing wind energy’s

contribution to US electricity supply, Energy Eﬃciency and Renewable Energy

(EERE) (2008)

6. Chen, Z.: Wind power in modern power systems. J. Mod. Power Syst. Clean Energy

1(1), 2–13 (2013)

7. Haque, A.U., Nehrir, M.H., Mandal, P.: A hybrid intelligent model for deterministic

and quantile regression approach for probabilistic wind power forecasting. IEEE

Trans. Power Syst. 29(4), 1663–1672 (2014)

8. Kazmi, H.S.Z., Javaid, N., Imran, M.: Towards energy eﬃciency and trustfulness

in complex networks using data science techniques and blockchain. MS thesis.

COMSATS University Islamabad (CUI), Islamabad, Pakistan, July 2019

9. Zahid, M., Javaid, N., Rasheed, M.B.: Balancing electricity demand and supply in

smart grids using blockchain. MS thesis. COMSATS University Islamabad (CUI),

Islamabad, Pakistan, July 2019

56 S. Mujeeb et al.

10. Bano, H., Javaid, N., Rasheed, M.B.: Electricity price and load forecasting using

enhanced machine learning techniques. MS thesis. COMSATS University Islam-

abad (CUI), Islamabad, Pakistan, July 2019

11. Juban, J., Siebert, N., Kariniotakis, G.N.: Probabilistic short-term wind power

forecasting for the optimal management of wind generation. In: Power Tech, 2007

IEEE Lausanne, pp. 683–688. IEEE (2007)

12. Wang, H.Z., Li, G.Q., Wang, G.B., Peng, J.C., Jiang, H., Liu, Y.T.: Deep learning

based ensemble approach for probabilistic wind power forecasting. Appl. Energy

15(188), 56–70 (2017)

13. Torres, J.M., Aguilar, R.M.: Using deep learning to predict complex systems: a

case study in wind farm generation. Complexity 2018, 10 (2018)

14. Coifman, R.R., Wickerhauser, M.V.: Entropy-based algorithms for best basis selec-

tion. IEEE Trans. Inf. Theory 38(2), 713–8 (1992)

15. Burrus, C.S., Gopinath, R., Guo, H.: Introduction to Wavelets and Wavelet Trans-

forms: A Primer. Prentice Hall, Upper Saddle River (1997)

16. ISO NE Market Operations Data. https://www.iso-ne.com. Accessed 20th Jan

2019

17. Derrac, J., Garcia, S., Molina, D., Herrera, F.: A practical tutorial on the use of

nonparametric statistical tests as a methodology for comparing evolutionary and

swarm intelligence algorithms. Swarm Evol. Comput. 1(1), 3–18 (2011)

18. Martin, P., Moreno, G., Rodriguez, F., Jimenez, J., Fernandez, I.: A hybrid app-

roach to short-term load forecasting aimed at bad data detection in secondary

substation monitoring equipment. Sensors 18(11), 3947 (2018)

19. Diebold, F.X., Mariano, R.S.: Comparing predictive accuracy. J. Bus. Econ. Stat.

13, 253–63 (1995)

20. Chen, H., Wan, Q., Wang, Y.: Reﬁned Diebold-Mariano test methods for the eval-

uation of wind power forecasting models. Energies 7(7), 4185–4198 (2014)

21. Lago, J., De Ridder, F., De Schutter, B.: Forecasting spot electricity prices: deep

learning approaches and empirical comparison of traditional algorithms. Appl.

Energy 221, 386–405 (2018)