A Review on Wind Power Forecasting Models for

improved Renewable Energy Integration

S.Rajendra Prasad1, M.Gopichand Naik2
Electrical Department, Andhra University, Visakhapatnam, India
1
[email protected]
2
[email protected]

Abstract— This paper gives an overview of the research that has true for a number of reasons. First and foremost, on the basis
been done on wind power forecasting models, which may be used of market pricing, suitable incentives in the form of attractive
to aid in the optimum integration of Renewable Energy (RE) into market prices are supplied on energy imbalance charges. In
electric power networks. Aside from the economic advantages, addition, making an accurate prediction may aid in the
the variable nature of wind energy production has a variety of
negative consequences for the electric grid system, including
development of well-functioning hour ahead or day-ahead
stability, reliability, and the capacity to plan for future markets. [4]-[7] go into more detail about some of the issues
operations, among other things. As a result, precise forecasting of that were talked about. Getting better at predicting how much
wind power output is critical for grid stability and security, as wind power there will be has big economic and technical
well as for the promotion of large-scale wind power. To ensure benefits. [4] proposes a probabilistic method for figuring out
the accuracy of wind energy forecasts, a variety of conventional, how much energy wind generators will use if they make bad
artificial intelligence, and hybrid methodologies have been predictions. Case studies show that these costs can be as high
developed. The simplicity and robustness of time-series-based as 10% of the total income from selling wind energy. a short-
methods have made them popular for forecasting applications.
term probabilistic forecast of wind power is shown and
Artificial Neural Networks (ANNs) and Fuzzy Logic have
recently been supported by several researchers for forecasting discussed in [5]. There, an optimal bidding strategy is
because of their flexibility. This review covers the performance of discussed that takes into account the uncertainty in the
many wind power forecasting models that are classified forecasts and how to make the best bids.
according to their categories. It is also offered a critical
examination of contemporary studies, which includes statistical On a wide scale, the numeric weather predictor (NWP) model
model and machine learning models that are based on historical combined with a physical flow model and a statistical
data. Aspects of this study that are taken into consideration
include the advantages and disadvantages of various forecasting
forecasting approach might save 15 to 25 percent of the fossil
models, including hybrid models, as well as performance fuel now used compared to the persistence model[6]. A cost
matrices used in assessing the forecasting model. In addition, the model is developed for the possibility of using a wind farm's
possible advantages of model optimization are examined in detail output for ancillary services such as reactive power support,
as well. primary frequency support, power oscillation damping support,
and so on, in the event that the wind farm's output remains
Keywords— Wind Power Forecasting, Renewable Energy unutilized due to forecast errors, as reported in [7]. In this
Integration, Statistical Model, Machine Learning Model, model, the larger wind prediction error result in higher
Artificial Intelligence and Hybrid Model
payments to wind farms for their reactive power supply,
which is owing to the increased lost opportunity cost
I. INTRODUCTION associated with higher wind forecast errors.
The Variability of the wind is the most difficult challenge to
implementing wind energy as a stable and self-sufficient According to the most current accessible published studies,
source of electric power. For large-scale wind penetration to this paper presents an in-depth analysis of wind power and
be successful solutions must be found to numerous problems wind speed forecasting techniques. It is the primary
including competitive market designs, real-time grid contribution of this work that the literature on wind
operations, standards of interconnection, ancillary service power/speed forecasting is classified based on the forecasting
requirements and costs, quality of power, transmission system horizons and time-scales used in their forecasting is classified.
capacity and future upgrades, stability and reliability of the A review article with this particular topic does not yet exist in
power system, and optimal reductions in greenhouse gas the literature, despite the fact that there are multiple review
emissions of the entire power system, which is typically papers accessible on wind power/speed forecasts (e.g., [8],
determined by the optimal amount of wind penetration into [9]), to the best of the authors' knowledge. A significant
the system [1], [2]. Improved wind forecasting is widely contribution of this analysis is that it gives a clear comparison
recognized as a very effective approach for addressing many and assessment of each approach depending on the time
of these issues. Accurate wind forecasts, for example, are horizon for which it is designed.
always appealing in competitive power markets, and this is
 II. Wind Speed Versus Wind Power Day-ahead reserve
Wind turbine output power is dependent on wind speed, which setting, Unit commitment
fluctuates over time and is affected by regional landscape type, & economic dispatch
weather patterns, and seasonal fluctuations. The amount of Generator Online/Offline
wind power available and realistically wind power travelling Medium 72 hours to 7 Days Decisions, Operational
across the rotor blades per unit sweep area is defined as[10] term ahead Security in Day-Ahead
1 Electricity Market
Pav υ   ρt Aυ 3 (1)
2 Planning of the electricity
Long- 7 Days to 2 week or
1 generation, transmission
Preal υ   ρt Aυ 3 C p υ  (2) term more ahead
and distribution
2
Where v( ) denotes the ideal available wind power and
eal( ) denotes the wind turbines realistic power output in
IV. Classification of wind power forecasting models
Watts (W), ( ) denotes the time-varying air density, which
varies in response to the surrounding atmospheric pressure A large number of researchers have studied the many
and temperature. A is the sweep area of the blades in square approaches for predicting the output of wind energy. A
metres (m2), and v is the wind speed in m/s. power coefficient number of strategies and models for forecast wind energy
( ) is the ratio of wind turbines realistic power eal( ) output were suggested and developed. The following are the
divided by ideal available wind power v( ) at specific wind most widely used and important forecasting techniques, in
speed. is governed by the turbine's tip angle, blade design order of popularity.
and wind speed-to-rotor speed relationship. 0.593 is the
greatest power coefficient (Betz limit). However, in reality, A. Persistence model
this value is not possible. An estimate of the power coefficient
under different circumstances of operation was not provided. The forecasted wind power output is assumed to remain
All three models were compared using a power coefficient of constant at the same time the previous or following day in this
0.5. model. The expected forecasted wind energy output over the
As seen in equation 1 and 2, air density is a critical element next 24 hours is as follows:
impacting the quantity of wind energy produced by a wind
turbine. The connection between the temperature, air density P f t   Ppd t  (4)
and barometric pressure at the place is as follows: At the time t, Pf is the forecasted wind power output and Ppd is
the previous day wind power output at the same time[11].
 gh  When it comes to short-term forecasting, this model is most
 P    RT  often utilised for wind power and speed forecasts, particularly
ρ t     e (3)
 RT  for one-hour forward forecasting. The accuracy of this
Where ρ( ) is the time variable air density in kg/m3, T is the forecasting model is dependent on the stability of the
air temperature in K, is barometric pressure in Pa, g is the meteorological conditions.
gravity of Earth in 9.81m/s2, R is the specific gas constant for
B. Autoregressive moving average (ARMA) model
dry air in 287.058 J/(kg.K) and h is the hub height above
According to Karakus et al[21], statistical procedures for wind
ground level in m.
forecasting are based on historical data and meteorological
conditions and need just a single step. In Shukur and Lee[22]
III. Time Scale Classification of Forecasting & It’s created a time-series prediction model based on the
Applications autoregressive moving average (ARMA) and the Kalman
The categorization of wind forecasting technologies according filter (KF) techniques. According to Li and Hu[23], statistical
to their time scale is vague. However, as shown in Table I, models provide the greatest outcomes for situations involving
wind forecasting can be separated into four categories: short-term forecasting. In Equation 5, we see how the ARMA-
based technique is expressed.
TABLE I p q
Time-Scale Classification and It’s Applications xt  
i 1
 i X t i   e
j 1
j t i  k  et (5)
Time
Range Applications
Horizon where xt represents the forecasting parameter at time
Very Grid stability operations instant t, φ denotes the Auto Regression (AR) parameter, θ
Few seconds to 30
short and voltage regulation denotes the Moving Average parameter, k is the constant and
minutes ahead
term actions et represents the random white noise. p and q are the orders of
Maintenance planning of Auto Regression(AR) and Moving Average(MA)
Short- 30 minutes to 72
network lines, respectively[12].
term hours ahead
Congestion management,
The capability of the ARMA model to extract statistical
characteristics, as well as its use of the Box-Jenkins technique,
are the primary reasons for its widespread use and acceptance.
This model is commonly utilised for various forecasting i
ui  WjIj
models with an acceptable degree of accuracy, and an j 1
extension of it known as the AR integrated MA is often used
for these models (ARIMA). When using the ARIMA model,
an integrated portion is used to eliminate any nonstationary
information from the data. The most significant drawback of
the ARMA model is that the time series data must be
stationary nature. Fig.1: (a) Schematic diagram of an ANN structure, which consists of an input,
hidden, and output layers. (b) A mathematical model of an ANN cell.
The next step is to utilise the autocorrelation function (ACF)
and the partial autocorrelation function (PACF) to determine Utilizing the activation function, the network output is formed
an ARMA model that is appropriate for the data. The ACF is by adding up the weighted inputs. So the network's activation
defined as the proportion of Xt and Xt-1 autocovariance to the function serves as a squeeze function, transferring the input in
variance of a dependent variable as indicated in the form of an output. The fundamental ANN mathematical
formula is represented as
cov X t , X t k  N
ACF k  
Var X t 
(6) UN b  W
j 1
j Ij  (7)

where Cov denotes the covariance and Var denotes the There are many types of activation functions that may be used
variance. To assess the degree of association between Xt and to forecast wind power output. The most generally utilised are
Xt-k are associated when the effects of additional time lags Xt- the sigmoid, hyperbolic tangent sigmoid and Gaussian radial
1,...,Xt-k-1 are eliminated, the partial autocorrelation function basis functions since they are continuous, differentiable, and
(PACF) was utilised. provide non-linearity to the network. The two fundamental
TABLE II
Properties Of ACF And PACF
processes of a neural network are training and testing. The
network is taught during the training stage utilising the
AR(p) MA(q) ARMA(p,q) training input data set and learning algorithm. The learning
Cuts off algorithm of the NN attempts to map the input-output
ACF Tails off Tails off relationship by updating the synaptic weight values. It is
after lag q
necessary to compare the output generated by the network
Cuts off
PACF Tails off Tails off with the desired output in order to compute the error. Because
after lag p
of this, the weight and bias values of NN (which serve as a
compensator) are adjusted in response to the error. After that,
For AR and MA, the number of orders was determined by
the process is repeated till the desired output result is obtained.
examining the attributes of the ACF and PACF (Katchova,
As a result, during the testing stage, the network produces the
2017). The next stage is to identify the most suitable models
final output depending on the testing input data set and the
based on the results of the white noise test and the quality of
model weight value that was used. Several kinds of neural
fit. The next stage was to plot the ACF of residuals to see
networks (NNs) have been developed, each with a different
whether the residuals resembled white noise.
architectures and a different set of input-output mapping
techniques. The multilayer perceptron NN(MLPNN),
C. Artificial neural network (ANN)
multilayer feed-forward NN (MLFFNN), radial basis function
Because of the non-linearity of meteorological data, artificial NN (RBFNN), recurrent NN (RNN), and general regression
neural networks (ANNs) are frequently employed in most NN (GRNN) are the most often utilised neural networks. The
studies to forecasting wind power production. When a non- primary disadvantages of ANN are that they need a large
linear and complicated bonding between the data occurs amount of data during the training phase and overfitting[14].
between the data without any previous assumptions, ANN is
more appropriate than statistical approaches. The input, D. Adaptive network based Fuzzy Inference System(ANFIS)
hidden, and output layers, as well as the neurons and
connections, are the primary components of an ANN. In the adaptive neuro-fuzzy inference system (ANFIS), the
According to the illustration, the neuron cell is divided into fuzzy logic approach is combined with the neuro network
two sections. The first component is the "combination technology, which allows the fuzzy inference system to
function," which generates a value by adding all of the inputs benefit from the learning capabilities of neural networks. The
together. The second section is referred to as the "activation parameters of the membership function in an ANFIS are
function"[13]. adjusted using neuro-adaptive learning techniques, which are
implemented in the ANFIS. The parameters have an effect on
the form of the membership functions. When it comes to wind
energy forecasting, the structure of the neuro-fuzzy model where ψ (x) is the feature vector of inputs x; w ∈ Rn is a
may be represented as a particular multilayer feedforward weight vector, and b ∈ R is the bias term, which are estimated
neural network. by minimizing the regularized risk function [16,17]
Mamdani and Sugeno-type fuzzy inference techniques are two
N
of the most often utilised fuzzy inference methods. It is 1 1 2
recommended that the Sugeno-type approach be employed in
R C   C
N  L y , f  2 w
i 1
e i i (9)
the ANFIS in the review article since it is effective when used
in adaptive techniques and optimization (Sugeno-type fuzzy  y  f i   , if y i  f i   
Le  y i , f i    i  (10)
inference, 2017)[15]. As a result, the output membership  0 otherwise 
functions are restricted to constants and linear functions. The
maximum, minimum, and average wind speeds observed at an
elevation above ground were used as inputs to the model. Where ε-insensitive loss function, C and ε are the user-
ANFIS is the most generally used method because it is determined prescribed parameters, yi is the actual value at
computationally less costly, transparent, and delivers results period i, and fi is the forecasted value at period i. The loss will
that are as robust as statistical models. be equal to zero if the forecasted value is within the ε-tube.
The second term, ½ ||w||2 measures the flatness of the
E. Support Vector Machine(SVM) function.

SVM is a supervised machine-learning approach that is based F. HYBRID MODEL

on principle of structural risk minimization (SRM). SRM
reduces the expected risk of a upper bound to low value as When it comes to forecasting wind power generation in
possible. As a result, SVM may reduce the amount of error in different scenarios, the performance of a single model is not
the training data. Vapnik developed the SVM algorithm in accurate. This issue may be attributed to the limitations of a
order to solve the classification problem. SVM, on the other stand-alone technique. As a hybrid model, it's best to use two
hand, has been extended to the area of regression issues. or more different techniques together. In some forecasting
Support vector regression is the term used to describe the use applications, these models have been used to get better results
of SVM in time series regression (SVR). SVR is an because they have a better chance of accuracy. One of the
appropriate method in this situation since wind power main goals of these models is to look at how different
generation forecasting is a typical time series analysis issue. topologies can work together to improve the accuracy of
forecasts.

The use of a fuzzy inference model with RNN in wind power

forecasting has been shown effective. It has been decided to
apply a fuzzy inference model in this instance in order to
smooth out the meteorological data, which will be used to
forecasting wind energy production. Wind power forecasting
using a hybrid fuzzy-Genetic Algorithm(GA) forecasting
model [18] has also shown promising results. [18] In
numerous hybrid forecasting models, the wavelet transform
(WT) is employed with conventional or artificial intelligence
approaches in the prediction of wind power. In this scenario,
WT is used to de-noise the input data, which is then processed
further. The meteorological data collected after WT was
utilised as input for ANN and SVM-based forecasting models,
Fig.2. Changing nonlinear regression into linear regression. which accurately forecasting wind power generation with the
least amount of error.
Consider the following set of training data:; {(x1,y1), (x2,y2)
..........(xl,yl)} where xl  Rn in the input vector (meteorological Comparisons of forecasting methodologies employed in
variables data), and yl  Rn is the corresponding to the output various forecasting horizons are shown in Table-III. Although
several approaches are acceptable for short-term and medium-
value (wind power output). The estimation function f (x) is
term wind forecasting, only a handful have been tested for
represented as long-term forecasting.
Y f  f x   w  x   b (8) Table-III
Forecasting method/ model used in different forecasting horizon
 Method / Short Medium further wind energy installations. As a result, the accuracy
Long Term assessment of the wind power forecasting model is an
Model Term Term
Physical extremely important aspect of the forecasting process itself.
(Numeric When evaluating the accuracy of wind power forecasting
Not models, many assessment matrices have been applied. In the
Weather Applicable Applicable
Predictors Applicable forecasting model review and benchmarking process,
(NWP)) standardised performance metrics would be useful. According
to the table below, the mean square error [MSE], root mean
Not square error [RMSE], normalised root mean square error
ARMA Applicable Applicable
Applicable [nRMSE], mean absolute error [MAE], mean absolute
Not percentage error [MAPE], mean relative error [MRE] and
ARIMA Applicable Applicable
Applicable mean bias error [MBE] are all commonly used in evaluating
Support the accuracy of wind power forecasting models[19-20].
Not N
vector Applicable Applicable 1
machine
Applicable MSE 
N W
i 1
forecasted  Wtrue 2
Multiple Not
Applicable Applicable N
Regression Applicable 1
Exponential Not
RMSE 
N W
i 1
forecasted  Wtrue 2
Applicable Applicable
Smoothing Applicable
 N 
Fuzzy  1 
Inference Applicable
Not Not nRMSE  
 N
W i 1
forecasted  Wtrue 
2
  100 Wtruemax

Applicable Applicable  
system
N
Genetic Not 1
algorithm
Applicable Applicable
Applicable
MAE 
N W
i 1
forecasted  Wtrue

Neural Not N W forecasted  Wtrue

Applicable Applicable 1
Network Applicable MAPE 
N  Wtrue
 100%
Machine i 1
Applicable Applicable Applicable N
Learning 1 W forecasted  Wtrue

Persistence
Not
Applicable Applicable
MRE 
N 
i 1
Wtotal
 100%

Applicable N
1
Wavelet
Based Model
Not
Applicable
Applicable Applicable MBE 
N W
i 1
forecasted  Wtrue 
Fuzzy+ Where
Neural Not Not Wforecasting = The forecasted wind power at each time point,
Applicable Wtrue = The observed/measured wind power at each time point,
Network Applicable Applicable
model Wtrue(max) = The maximum observed wind power of this scale,
Wtotal = The wind installation capacity,
Fuzzy+GA Not Not N = The number of data sample for the time scale.
Applicable
model Applicable Applicable
Not Not
WT+ANN Applicable VI. CONCLUSION
Applicable Applicable
Due to the climate change and global warming, the integration
of renewable energy source like wind energy into the
V. Performance Evaluation of Wind forecasting Models electricity grid has risen in recent years as a study area of
concern. Since there is a pressing need to fulfil rising energy
Today's power grid system receives an increasing quantity demand while also mitigating climate change and stabilising
of electricity from wind energy systems, and this number is electric grid systems, precise forecasting of wind energy
growing on a regular basis. A consequence has been a output has become vital. Because of this, a plethora of
significant reliance on the unit commitment of wind power research have been undertaken on the subject from various
systems for the stability of grid networks. If the wind system angles. The current research presents the results of a thorough
fails to provide the agreed energy, the grid system will have and complete evaluation of the literature on direct wind power
issues in terms of supplying an acceptable amount of power. It forecasting models and approaches. This review paper begins
is thus critical to accurately forecasting the production of wind with an overview of wind energy production and the
energy in order to maintain grid stability and encourage importance of wind power forecasting. It has also been taken
into account the significance of the correlation of [4] A. Fabbri, T. G. S. Roman, J. R. Abbad, and V. H. M.
meteorological and wind power output data, as well as the Quezada,“Assessment of the cost associated with wind generation prediction
value of pre-processing model input data. It has been errors in a liberalized electricity market,” IEEE Trans. Power Syst., vol. 20,
no.3, pp. 1440- 1446, Aug. 2005.
discovered that when the inputs of a model have a high
[5] P. Pinson, C. Chevallier, and G. N. Kariniotakis, “Trading wind
correlation factor with the output, the accuracy of a model is
generation from short-term probabilistic forecasts of wind power,” IEEE
improved. Additionally, when pre-processed input data is Trans. Power Syst., vol. 22, no.3, pp.1148-1156, Aug. 2007.
employed, the accuracy of the predicting output is improved. [6] S. J. Watson, L. Landberg, and J. A. Halliday, “Application of wind speed
Following a discussion of input selection, a critical analysis of forecasting to the integration of wind energy into a large scale power system,”
wind power forecasting was carried out, which was based on IEE Proc. - Gener. Transm. Distrib., vol. 141, no. 4, pp.357-362, July 1994.
the categorization of the models from several viewpoints. The [7] N. R. Ullah, K. Bhattacharya, and T. Thiringer, “Wind farms as reactive
findings demonstrate that the outcomes of forecasting models power ancillary service providers — technical and economic issues,” IEEE
are dependent on the forecasting horizon, the amount of data Trans. Energy Convers., vol. 24, no. 3, pp. 661-672, Sept. 2009.
provided, and the technique used for the forecasting. A short- [8] M. Lei, L. Shiyan, J. Chuanwen, Liu Hongling, and Z. Yan, “A review on
term forecasting horizon was used in the development of the the forecasting of wind speed and generated power,” Renewable and
wind power forecasting model in the vast majority of the Sustainable Energy Reviews, vol. 13, no. 4, pp. 915-920, May 2009.
situations studied. Even though the forecast model parameters [9] A. Costa, A. Crespo, J. Navarro, G. Lizcano, H. Madsen, and E. Feitosa,
“A review on the young history of the wind power short-term prediction,”
are the same, the accuracy of the wind power forecasting
Renewable and Sustainable Energy Reviews, vol. 12, no. 6, pp. 1725-1744,
model changes depending on the forecasting horizon used to
Aug. 2008.
make the prediction. Future researchers in this subject will [10] C. W. Potter, and M. Negnevitsky, “Very short-term wind forecasting for
benefit from the literature evaluations of current works on Tasmanian power generation,” IEEE Trans. Power Syst., vol. 21, no. 2, pp.
direct forecasting of wind power output in this article, which 965- 972, May 2006.
will include the strengths and limits of the models as well as [11] M. Milligan, M. Schwartz, and Y. Wan, “Statistical wind power
recommendations for future studies in this field. It was studied forecasting models: Results for U.S. wind farms,” in Proc. Windpower, Austin,
and compared the performance of numerous modern TX, May 18–21, 2003, NREL/CP-500-33 956 Rep.
forecasting approaches against various parameters, including [12]Rajagopalan S, Santoso S. Wind power forecasting and error analysis
accuracy, dependability, computing cost, and complexity, and using the autoregressive moving average modeling. Power & Energy Society
the results were conclusive. According to recent research, General Meeting 2009PES'09 IEEE: IEEE; 2009. p. 1–6.
ANN and SVM-based forecasting models performed [13] Mellit A, Kalogirou SA. Artificial intelligence techniques for
admirably under rapidly changing and varied environmental photovoltaic applications: a review. Progress Energy Combust Sci
2008;34:574–632.
circumstances. Furthermore, the majority of the research used
[14]Ticknor, J. L., (2013). A Bayesian regularized artificial neural network
a variety of methodologies to create the forecasting model,
for stock market forecasting. Expert Systems with Applications, Vol. 40, pp.
which resulted in higher accuracy. A significant number of 5501-5506.
studies divided the anticipated day into distinct groups [15] Anon., n.d. Adaptive neural-fuzzy modeling. [Online]
depending on the weather conditions using a variety of Availableat:https://www.mathworks.com/help/fuzzy/adaptive-neuro-fuzzy-
methodologies and then constructed a forecasting model to inference-systems.html [Accessed 23 October 2017].
account for this classification. However, the range of the error [16] Zhang F, Deb C, Lee SE, Yang J, Shah KW. Time series forecasting for
that was found was very wide because of different weather building energy consumption using weighted Support Vector Regression with
conditions. To avoid making mistakes, the separate model for differential evolution optimization technique. Energy Build 2016;126:94–103.
each weather condition must work well. However, the cost [17] Hong W-C. Electric load forecasting by support vector model. Appl
and complexity of the model should be taken into account in Math Model 2009;33:2444–54.
this case. Furthermore, the performance matrices that are used [18] Chen S-M, Chang Y-C, Pan J-S. Fuzzy rules interpolation for sparse
fuzzy rule based systems based on interval type-2 Gaussian fuzzy sets and
to evaluate the forecasting models have been added. Results
genetic algorithms. IEEE Trans Fuzzy Syst 2013;21:412–25.
show that RMSE was used more often than other things.
[19] Mbuvha, R., (2017). Bayesian neural networks for short term wind
Furthermore, the benefits of model optimization were also power forecasting. Stockholm, Sweden: KTH Institute of Technology School
talked about. Findings show that the optimised algorithm of Computer Science and Communication.
made the model a lot more accurate at predicting the future. In [20] Utpal Kumar Das, et.al 2018: Forecasting of photovoltaic power
this process, GA is one of the best ways to improve things in generation and model optimization: A review. Renewable and Sustainable
this field. Energy Reviews 81 (2018) 912–928.
[21] Karaku，s O, Kuruo˘glu EE, Altınkaya MA. One-day ahead wind
REFERENCES speed/power prediction based on polynomial autoregressive model. IET
[1] Y-K Wu, and J-S Hong, “A literature review of wind forecasting Renew Power Generat. 2017;11(11):1430-1439.
technology in the world,” IEEE Power Tech 2007, Lausanne , pp. 504-509, 1- [22]Shukur OB, Lee MH. Daily wind speed forecasting through hybrid KF-
5 July 2007. ANN model based on ARIMA. Renew Energy. 2015;76:637-647.
[2] H. Lund, “Large-scale integration of wind power into different energy [23] Li C, Hu J-W. A new ARIMA-based neuro-fuzzy approach and swarm
systems,” Energy, vol. 30, no. 13, pp.2402-2412, Oct. 2005. intelligence for time series forecasting. Eng Appl Artif Intell. 2012;25(2):295-
[3] M. Negnevitsky, P. Johnson, and S. Santoso, “Short term wind power 308.
forecasting using hybrid intelligent systems,” IEEE Power Engineering
Society General Meeting 2007, pp.1-4, 24-28 June 2007.