Renewable and Sustainable Energy Reviews: Jaesung Jung, Robert P. Broadwater

Renewable and Sustainable Energy Reviews 31 (2014) 762–777
Contents lists available at ScienceDirect
Renewable and Sustainable Energy Reviews

journal homepage: www.elsevier.com/locate/rser
Current status and future advances for wind speed

and power forecasting
Jaesung Jung n, Robert P. Broadwater
Department of Electrical and Computer Engineering, Virginia Polytechnic Institute and State University, 302 Whittemore Hall, Virginia Tech,
Blacksburg, VA 24061-0111, USA
art ic l e i nf o a b s t r a c t
Article history: This paper presents an overview of existing research on wind speed and power forecasting. It first
Received 15 April 2013 discusses state-of-the-art wind speed and power forecasting approaches. Then, forecasting accuracy is
Received in revised form presented based on variable factors. Finally, potential techniques to improve the accuracy of forecasting
30 November 2013
models are reviewed. A full survey on all existing models is not presented, but attempts to highlight the
Accepted 29 December 2013
most promising body of knowledge concerning wind speed and power forecasting.
Available online 22 January 2014
& 2014 Elsevier Ltd. All rights reserved.
Keywords:
Wind speed
Wind power
Spatial correlation forecasting
Regional forecasting
Probabilistic forecasting
Offshore forecasting
Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763
2. Overview of current wind speed and power forecasting. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763
2.1. Wind speed and power forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763
2.1.1. Physical forecasting approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763
2.1.2. Statistical forecasting approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 764
2.1.3. Combination approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 768
2.2. Spatial correlation forecasting. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 768
2.3. Regional forecasting. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 769
2.4. Probabilistic forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 769
2.4.1. Parametric approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 770
2.4.2. Non-parametric approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 770
2.4.3. Using ensemble forecast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 770
2.4.4. Additional models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 771
2.5. Offshore forecasting. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 771
3. Accuracy of wind speed and power forecasting. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 771
3.1. Reference models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 771
3.1.1. Persistence model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 771
3.1.2. Weighted sum between persistence and the mean power production. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 771
3.1.3. OL-Persistence model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 771
3.2. Model performance evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 772
3.3. Forecast accuracy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 772
3.3.1. The accuracy by forecast time horizon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 772
3.3.2. The accuracy by location complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 772
n
Corresponding author. Tel.: þ 1 540 577 6396; fax: þ 1 540 231 3362.
E-mail address: [email protected] (J. Jung).
1364-0321/$ - see front matter & 2014 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.rser.2013.12.054
J. Jung, R.P. Broadwater / Renewable and Sustainable Energy Reviews 31 (2014) 762–777 763
3.3.3. The accuracy by seasonal variability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 772

3.3.4. The accuracy by weather condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 772
4. Improvement in forecasting performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 772
4.1. Kalman filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 772
4.2. Optimal combination. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773
4.3. Spatial smoothing effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773
4.4. Statistical downscaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773
4.5. Input parameter selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773
4.6. Power curve modeling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773
4.7. Forecasting parameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773
5. Conclusion and future works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 774
5.1. Discussion and conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 774
5.2. Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 774
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 775
Reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 775
1. Introduction presented in Section 4. Finally, conclusions and future works are

drawn in Section 5.
Wind power is one of the most rapidly growing renewable
energy sources, and is regarded as an appealing alternative to
conventional power generated from fossil fuel. This led to a 2. Overview of current wind speed and power forecasting
collaborative effort to achieve 20% of U.S. electricity supplied from
wind power by 2030 [1]. Although the integration of wind power 2.1. Wind speed and power forecasting
brings many advantages, high penetration of wind power provides
a number of challenges in power system operations and planning, The basic role of wind speed and power forecasting is to
mainly due to its uncertain and intermittent nature. In the provide information about the wind speed and power that can
electricity system the power supply must be equal to the power be expected in the next few minutes, hours, or days. Based on
demand at all times. However, the variation of wind power output power system operation requirements, the forecast can be divided
makes it difficult to maintain this balance. into four different horizons: very short-term (few seconds to
One of the possible solutions to the balance challenge is to 30 min), short-term (30 min to 6 h), medium-term (6–24 h), and
improve the wind speed and power forecasting. Research in the long-term (1–7 days) [10,11]. Very short-term forecasts are used
area of forecasting wind speed or the power produced by wind for turbine control and load tracking. Short-term forecasts are
farms has been devoted to the development of effective and utilized for preload sharing. Medium-term forecasts are used for
reliable tools and many different approaches have been proposed power system management and energy trading. Long-term fore-
and reviewed in [2–13]. Accurate forecasting tools reduce operat- cast are used for maintenance scheduling of the wind turbines.
ing costs and improve reliability associated with the integration of Research in the area of forecasting wind speed and the power
wind power into the existing electricity supply system [14–29]. produced by wind farms has been devoted to the development of
There are different users of wind speed and power forecasts. These effective and reliable tools and many different approaches have
users not only need point forecasts but also the uncertainty of the been proposed. These tools can be classified whether the terrain
forecast is essential for determining the size of the operating reserves information at the location is used as an input or not. Two
necessary to balance the generation with load. For the market mainstream approaches are the physical and the statistical
operator, the forecast of the total wind power production in a region approach. In some models a combination approach is used in an
is needed more than the individual wind plant forecast. For the system attempt to integrate the advantages of both approaches. In this
planner, it is essential to plan the installation of new wind turbines. section an overview of existing wind speed and power forecast
The desired forecasting capabilities typically vary among user groups. approaches is presented.
This paper divides the forecasting methods into five categories: wind
speed and power forecasting, spatial correlation forecasting, regional 2.1.1. Physical forecasting approach
forecasting, probabilistic forecasting, and offshore forecasting. The physical approach to forecasting, in contrast to statistical
A large amount of research has been directed toward the approach, uses the detailed physical description to model the on-
development of accurate and reliable wind speed and power site conditions at the location of the wind farm [7,30]. The basic
forecasting models and many different approaches have been operation of a physical approach is illustrated in Fig. 1.
developed. However it is difficult to draw conclusions as to which It carries out the refinement of the Numerical Weather Predic-
model is the best because each model in use has significant site tion (NWP) data to take into account the on-site conditions by the
dependencies. Thus, a forecasting model may perform well at its downscaling method, which are based on the physics of the lower
site, but this does not guarantee that the model will work well at atmospheric boundary layer. The downscaling method requires
another site. This paper discusses forecasting accuracy based on the detailed physical descriptions of the wind farms and their
variable factors that are used in the forecast. surroundings, including: description of the wind farm (wind farm
Furthermore, several potential techniques for improving fore- layout and wind turbine power curve, etc.) and description of the
casting accuracy have been reported in the literature. These are terrain (orography, roughness, obstacles, etc.). Then, the refined
also discussed and reviewed in this paper. wind speed data at the hub height of the wind turbines is plugged
The paper is organized as follows. Section 2 presents an overview into the corresponding wind power curve to calculate the wind
of existing wind speed and power forecasting approaches. In Section power production. If the on-line data is available, model output
3, the general forecasting accuracy is discussed. Some of the potential statistics are performed to reduce the error of the forecast.
techniques to improve the performance of forecasting models are Contrary to the statistical approach, the physical approach does
764 J. Jung, R.P. Broadwater / Renewable and Sustainable Energy Reviews 31 (2014) 762–777
represents the relation between wind power or speed forecasting

Start and explanatory variables including NWPs and on-line measured data
[7]. The general form of the model is illustrated in Fig. 2.
The statistical approach generally uses previous history data to
description of build the statistical model. This model uses NWP forecast for time
the wind farm tþk and on-line measurement at time t to forecast the present
Downscaling to Wind
Turbine Hub Height over the next few hours. It is easy to model and inexpensive.
NWP
However, contrary to the physical approach, the statistical
Forecasts
approach requires historical data to train the statistical model.
Many different approaches are employed [2,3,5,6,11,13]. Some of
the representative statistical approaches are reviewed in this
description of
Power Curve Model section.
the wind farm
2.1.2.1. Conventional statistical approach. In the conventional

statistical approach a time series model is applied to forecast
Model Output Statistics future wind power or speed. According to the forecasting process,
On-line Data
for Wind Power which was proposed by Box–Jenkins, this model is divided into
four main steps to make a mathematical model of the problem
including model identification, model estimation, model
diagnostics checking, and forecasting. Several types of time
series models may be considered, including autoregressive
Forecast Power Output model (AR), moving average model (MA), autoregressive moving
average model (ARMA), and autoregressive integrated moving
average model (ARIMA). The general form of the model is
p q
X t ¼ c þ εt þ ∑ φi X t i þ ∑ θi εt i ð1Þ
End i¼1 i¼1
where X t represents the forecasting parameter at time t, φi is the

Fig. 1. The physical approach to forecasting wind speed and power.
autoregressive parameter, θi is the moving average parameter, c is
the constant, and random variable εt is the white noise. This model
not require training input from historical data. However, acquiring represents the ARMA model having the autoregressive model of
the physical data is one of the main drawbacks of the approach. order p and the moving average model of order q (ARMA(p,q)). If
A number of physical approaches have been introduced the order of the moving average model (q) is zero, it represents the
[2,7,8,31,32]. The Prediktor is developed by the Risoe National autoregressive model of order p (AR(p)). If the order of the
Laboratory in Denmark. It uses Wind Atlas Analysis and Applica- autoregressive model (p) is zero, it represents the moving
tion Program (WAsP) and PARK program to take the local condi- average model of order q (MA(q)). The ARIMA model is a
tions into account by using the NWP forecast from High Resolution generalization of an ARMA model.
Limited Area Model (HIRLAM) [33–35]. The Previento, developed The ARMA model is used for wind power forecasting in U.S.
by the University of Oldenburg in Germany has a similar physical wind farms in [39,40]. Some of the models offer improvements
approach but uses a different NWP forecast from Lakelmodell of over the persistence model but some are not by changing the
the German Weather Service [36]. The LocalPred is developed by order of AR and MA. The performance of the model is highly
CENER – National Renewable Energy Centre in Spain. It involves dependent on the parameters of the model.
adaptive optimization of the NWP forecast, time series modeling, The Improved Time Series Method (ITSM) based on ARIMA is
meso-scale modeling with MM5, and power curve modeling [37]. used in [41]. The sub-wind speed series obtained from the wavelet
The eWind, developed by AWS TrueWind Inc. in the USA, has a decomposition is forecasted and then the aggregate calculation of
similar physical approach with Prediktor but uses a high-resolution the sub-series is performed to obtain the final wind speed forecast.
boundary layer model (ForeWind) as a numerical weather model The simulation results indicate that the proposed method
to take the local conditions into account [38]. improves the accuracy of forecasting over the classical time series
The physical approaches are based on the models using the model and the neural network model.
fundamental physical principles for conservation of mass, momentum, The Autoregressive Conditional Heteroscedastic (ARCH) model
and energy in air flows. These models address computational fluid is combined with the ARIMA model to consider the heteroscedas-
dynamics (CFD) for simulating the atmosphere. Although there are ticity effect of the residual series [42]. The ARIMA–ARCH model is
many CFD models available, they are all based on the same basic used to forecast the sub-wind speed series obtained from the
physical principles. They differ in how the grids are structured and wavelet decomposition. The final wind speed forecasting is the
scaled, and how the numerical computations are performed. sum of these forecast values. The results show that it can improve
In the majority of cases the statistical approaches provide good the accuracy of the forecasting. In [43], the ARIMA–ARCH model is
results in short-term, medium-term, and long-term forecasting. used to forecast wind speed itself. It is shown that the ARIMA–
However, in the very short-term and short-term horizon, the influ- ARCH model shows better performance than the ARIMA model.
ence of atmospheric dynamics becomes more important, and in these To deal with the non-stationary of wind speed, wind speed
cases the use of physical approaches becomes essential. data are transformed to a Gaussian distribution and standardized
to remove the non-stationary in [44]. The parametric AR model is
then applied to the hourly wind speed data and is updated during
2.1.2. Statistical forecasting approach the real-time operation by a recursive least squares algorithm. The
The alternative main approach for wind speed and power fore- method is validated by using the wind speed data from a wind
casting is based on statistical modeling. The statistical approach power site.
Fig. 2. The statistical approach to forecasting wind speed and power.
In [45], the ARMA model is used to forecast the tuple of wind ANN is one of the widely used statistical approaches for wind
speed and direction. Four different approaches are proposed and speed and power forecasts. It consists of an input layer, one or
their performances are compared; component, traditional-linked more hidden layers, and an output layer. Each layer has a number
ARMA, vector AR (VAR), and restricted VAR approaches. The of artificial neurons, and it uses a connectionist approach to
component approach is better than traditional-linked ARMA for connect the neurons to the neurons of the previous layer. This
wind direction forecasting but not for wind speed forecasting. The approach is able to model the complex non-linear relationship
VAR approach offers better performance in wind direction and between the input and output layers through a training and
close performance in wind speed comparing with traditional- learning process. This approach does not require explicit mathe-
linked ARMA. There is little difference in terms of forecasting matical expressions as used in the physical and statistical
performance between VAR and restricted VAR approaches. approaches reviewed previously. Furthermore, it has the ability
The AR model using a Bayesian approach is used to forecast the of self-learning, self-organizing and self-adaption.
wind speed in [46]. First, Box-Cox transformation is performed to In [47], the ANN methodology using the Markov Chain
correct the non-normality of the wind speed. Then, the Markov approach is proposed for very short-term wind speed forecast.
Chain Monte Carlo (MCMC) simulation is performed to estimate The Markov Chain is applied to modify the preliminary forecast
the AR model parameters, which can then be used to forecast wind obtained from ANN according to the long-term patterns. A data set
speeds. It shows the great potential and possible improvements of with 2.5 s resolution is used to evaluate the performance. The
a Bayesian framework. results show the effectiveness of the proposed method.
In summary, conventional statistical approaches are based on The ANN methodology using Particle Swarm Optimization
classical linear statistical models such as AR, MA, ARMA, and the (PSO) is proposed for short-term wind power forecast in [48].
Box–Jenkins approach, where Box–Jenkins is based on ARIMA or PSO is employed to select the input variable from the several
seasonally adjusted ARIMA models, also known as SARIMA mod- nearby locations. The proposed method shows the reduction of the
els. However, the references show that the forecasting accuracy forecast error comparing with ANN.
can be improved depending on the model parameters that are Furthermore, the forecast method composed of a new
selected. Enhanced Particle Swarm Optimization (EPSO) technique and
The conventional statistical approaches are mostly aimed at Modified Hybrid Neural Network (MHNN) is proposed to have
very short-term and short-term forecasting. These models are easy high learning capability and avoid the overfitting problem and
to formulate and are capable of providing timely forecasts. As trapping in local minima and dead bands in [49]. The effective
forecasting accuracy is improved, conventional statistical two-stage feature selection technique, enhanced Mutual Informa-
approaches are often used as a reference model. tion (MI), is applied to select the most informative candidate
inputs by filtering out both irrelevant and redundant candidate
inputs for the forecast method. The results confirm the validity of
the developed approach.
2.1.2.2. Artificial Neural Network approach. Another approach is In [50], the forecast method based on Ridgelet Neural Network
Artificial Neural Networks (ANN). The NWP forecasts and further (RNN) is proposed as the forecast engine. The inputs of the forecast
meteorological variables are transformed into the wind power or engine are selected by the Mutual Information (MI) based feature
speed forecast by ANN which has been trained by the large sets of selection component among the set of candidate inputs. A New
historical data in order to learn the dependence of the output on Differential Evolution (NDE) algorithm with novel crossover opera-
input variables. The general ANN approach for wind speed and tor and selection mechanism is proposed to train the RNN. The
power forecast is shown in Fig. 3. results show the effectiveness of the proposed method.
Input Output
Hidden Layer
Layer Layer
Wind Speed Wind Speed or Power Forecast
Wind Direction
Temperature
Fig. 3. ANN approach for wind speed and power forecast.
In [51], a comprehensive comparison study of typical three that the proposed approach outperforms the reference. Further-
types of ANN approach is performed; Feed Forward Back- more, a similar approach based on ANN using EMD is found
Propagation (FFBP), Radial Basis Function (RBF), and Adaptive in [57].
Linear Element (ADALINE) approaches. It is found that different Generalized Feed-forward Neural Network (GFNN) is used to
inputs and learning rates, as well as model structures, directly forecast the annual wind speed probability density distribution in
influence the forecast accuracy. Furthermore, none of the [58]. The same parameters required by the Weibull function are
approaches outperform others universally based on the evaluation used as inputs. The proposed method shows better accuracy to
criteria even for the same wind dataset. forecast the wind speed probability density distribution than the
Complex-Valued Neural Network (CVNN) is proposed for the Weibull function.
wind power forecast in [52]. The CVNN used the wind vector (both Multilayer Feed-forward Neural Network (MFNN) is used to
wind speed and direction) rather than real-valued data as an forecast a day-ahead wind power forecasting in [59]. The NWP
input. The results show the improvement in the forecast accuracy model is established by coupling the Global Forecasting System
than Real-Valued Neural Network (RVNN). (GFS) with the Weather Research and Forecasting (WRF) System.
In [53], an ANN approach in conjunction with the adaptive Furthermore, the Kalman filter is integrated to reduce the sys-
Bayesian learning and the Gaussian process approximation is tematic error from WRF. The proposed method is proved to be
proposed for very short-term wind power forecast. The Bayesian successful by the simulation.
framework is employed updating to calculate the posterior prob- In [60], Kernel Ridge Regression (KRR) with active learning is
ability distribution for the weights of the model. The Gaussian performed for wind speed forecast. The active learning acts as a
process approximation makes the calculation of this posterior smart sample selector capable to yield compact forecast models
probability possible to solve an integration problem of the Baye- and also to filter out noisy training samples. Three active learning
sian learning. The Bayesian framework provides not only point methods to construct the training set for KRR are proposed; Pool
forecast but also the interval about the forecast. of regressors (PAL), Distance from the closest training sample
A three-layered Feed-forward NN trained by the Levenberg– (DAL), and Residual regression (RSAL). The experimental results
Marquardt algorithm is considered to forecast the wind power show that smart collection of training samples can be of benefit for
sub-series by Wavelet Transform (WT) in [54]. The obtained wind speed forecast.
forecast sub-series reconstruct the future wind power series by ANN has been a good selection for wind speed and power
inverse WT. The proposed approach is compared with the refer- forecasting. ANN is able to model a complex non-linear relation-
ence to demonstrate its effectiveness regarding forecasting accu- ship and extract the dependence between the input and output
racy and computation time. Similarly, in [55], Adaptive Wavelet through the learning process. It is simple to construct and only
Neural Network (AWNN) is carried out to forecast the wind speed requires short development times, and does not require explicit
in stage-I. Then, Feed-forward Neural Network (FNN) is used to mathematical expressions. ANN can be designed by setting up the
transfer the wind speed forecast to the wind power forecast in network structure and then selecting a learning algorithm.
stage-II. The first step with ANN is to choose the network structure,
In [56], standard multilayer Feed-forward Neural Network where network structures derive from two basic classes: 1 – A
(FNN) is employed to forecast the wind speed. The non-linear feed-forward topology where information flows in only one
and non-stationary wind speed series is first decomposed into a direction, from inputs to outputs; 2 – A recurrent topology where
finite and often small number of Intrinsic Mode Functions (IMFs) information flows in both directions, from input to output and also
and one residual series to obtain the stationary data series using from output to input.
the Empirical Mode Decomposition (EMD). FNN is performed to The second step with ANN is to select the learning algorithm
forecast the decomposed sub-series except the high frequency, and to learn the proper response. This is achieved with three
whose input variables are selected by Partial Autocorrelation learning paradigms: 1 – Supervised learning that sets ANN para-
Function (PACF). Finally, the forecast sub-series are summed to meters from training data; 2 – Un-supervised learning that
reconstruct the forecast of wind speed series. The results show sets ANN parameters based on given data and a cost function; 3
– Reinforcement learning that sets ANN parameters, where data is computational costs from HGP. The ℓ1=2 regularizer is employed for
generated by interactions with the environment. Each learning the sparsification. The simulation results demonstrate the effective-
paradigm has many learning algorithms. Some commonly used ness of the proposed method.
algorithms include evolutionary methods, gene expression pro- In [66], the modified Taylor Kriging (TK) method for forecasting
gramming, simulated annealing, expectation-maximization, non- wind speed time series is proposed. The original TK method is
parametric methods, and particle swarm optimization. developed for the spatial estimation. In this paper, the original TK
method is properly modified for the time series forecasting. The
simulation results indicate that the proposed method outperforms
2.1.2.3. ANN–Fuzzy approach. The Fuzzy logic approach is a non-
the ARIMA method.
linear mapping of input variables into the output by using
The Grey forecasting model (GM) is presented for short-term
numerical and linguistic values Fuzzy rather than fixed and
wind speed forecast in [67]. The Grey model is designed for the
exact. It used soft linguistic variable (small, medium, and large)
system analysis being characterized by inadequate information.
and a truth variable that ranges in degree between 0 and 1. The
The GM(1,1) Grey model, a single variable first-order Grey model,
Fuzzy logic approach is employed when the system is difficult to
is used. The simulation results demonstrate the efficiency of the
model accurately but an inexact model is available, because it
proposed model.
allows to use the approximate values and incomplete or
In [68], two-stage hybrid network with Bayesian Clustering by
ambiguous data. However, using the Fuzzy logic approach alone
Dynamic (BCD) and Support Vector Regression (SVR) is proposed.
is not satisfying because of its weak learning ability.
First, BCD clusters the input training dataset into several subsets
The ANN–Fuzzy approach refers to combination of ANN and
with similar dynamical properties. Then, SVRs is used to model the
Fuzzy logic approaches. ANN is based on the learning and
training data in each subset having similar property. Experiments
connectionist structure that performs well in low-level computa-
and comparisons with the persistence method demonstrate the
tion with raw data. However, Fuzzy logic performs well with the
effectiveness and efficiency of the proposed method.
human-like reasoning in high-level computation. Therefore, the
The forecasting of three-dimensional wind field is considered
combination of two approaches offers a promising approach to
in [69]. To exploit the available second order statistics, the
forecast the wind speed and power by compensating their weak-
augmented quaternion statistics is employed and the adaptive
ness each other.
Widely Linear Quaternion Least Mean Square (WL-QLMS) algo-
In [61], ANN is combined with Fuzzy logic approach in order to
rithm for the forecasting is used to allow for both three-
optimize the best use of the NWPs. First, the ANN model provides
dimensional wind model and the fusion of atmospheric para-
the preliminary forecast of wind power based on NWPs. Then, the
meters. It is shown that additional information from the vertical
Fuzzy model estimates the quality of the forecasts provided by
wind speed and atmospheric parameters assist in providing better
NWPs. Finally, this is exploited by an ANN model to provide the
forecast.
final forecasts. The simulation results show the validity of the
In [70], the Bayesian Model Averaging (BMA) approach with
proposed method.
Markov Chain Monte Carlo (MCMC) is proposed for developing the
The hybrid Wavelet Transform (WT)–Particle Swarm Optimiza-
long-term wind speed distribution. The derived BMA probability
tion (PSO)–Adaptive Network based Fuzzy Inference System
density function (PDF) is an average of many candidate models
(ANFIS) approach (hybrid WPA approach) is proposed for short-
included in the model space weighted by their posterior prob-
term wind power forecasting in [62]. The wavelet transform is
abilities over the sample data. The case studies reveal that none of
performed to convert a wind power series in a set of constitutive
reference distribution outperforms others; however, the BMA
sub-series. The transformed sub-series are separately forecasted
distribution is always suitable to describe the wind speed dis-
by the ANFIS. The PSO is used to adjust the parameters of the
tribution with high accuracy.
membership function in ANFIS. The hybrid WPA approach outper-
The forecasting model integrating the concepts of structural
forms seven other reference approaches.
breaks and Bayesian inferences is proposed for very short-term
The Simultaneous Perturbation Stochastic Approximation
wind speed forecasting in [71]. It allows prior information about
(SPSA)-based Multilayer Feed-forward Neural Network (MFNN)
the wind speeds to be incorporated into the model and somehow
with Fuzzy inputs algorithm is presented in [63]. The SPSA is
boosts forecasting performance by considering structural breaks.
employed to achieve the training task of the MFNN. The inputs for
The computational results confirm the advantages of the proposed
the MFNN are modeled by triangle fuzzy numbers. A better
method. However, there are several limitations; for example, the
performance of the proposed method is shown than the reference
computing efficiency is not high and the size of training sample is
method. However, it cannot be used for multistep forecasting.
not large enough. Further research is needed to cure these
ANN–Fuzzy approaches are complementary tools in building
limitations.
intelligent forecasting systems. While ANN performs well in low-
In [72], First-order and Second-order Adaptive Coefficient (FAC
level computational structures, Fuzzy logic deals with reasoning
and SAC) methods are used to wind speed forecasting. Particle
incorporated in the higher-level computational structures. Mer-
Swarm Optimization (PSO) is employed to select the most suitable
ging of the ANN approach and the Fuzzy logic approach into one
parameter of FAC and SAC methods. Furthermore, Seasonal Expo-
integrated system offers a promising approach to building wind
nential Adjustment (SEA) is employed to decompose the wind
speed and power forecasting systems.
speed data into seasonal and trend components. Different combi-
nations of these techniques are simulated. The combination
2.1.2.4. Etc. A Gaussian-Process-based method for forecasting the obtains higher forecasting accuracy than the single FAC and SAC.
upper and the lower bounds of wind speed as well as the average Data mining approaches in wind energy system have more
is proposed in [64]. It makes use of the kernel machine technique and attention in recent years [12]. The time series models are built
the Bayesian estimation. The simulation results show that they with data mining algorithms for the short-term forecasting in [73].
provide the better forecasting in consideration of the uncertainty Five different data mining algorithms are employed; the Support
than the reference. Furthermore, the advanced Gaussian-Process- Vector Machine Regression (SVMreg), Multilayer Perceptron
based method, called the Sparse Heteroscedastic Gaussian Process (MLP), M5P tree, Reduced Error Pruning (REP) tree, and the
(SPHGP) is proposed in [65]. This method is developed to make up the bagging tree algorithms. The computational results of all time
misestimation of the noise level from standard GP and reduces the series models and data mining algorithms are compared and
discussed. Furthermore, this approach is applied to forecast the for monotonically wind speed changes while the persistence
wind power ramp rates in [74]. One of the disadvantage of the method is more suitable for random data with white noise. The
proposed approach is that the time series model uses its own simulation results demonstrate improvements in the combination
previously forecasted value. Therefore, the forecast accuracy approach.
decreases when the number of forecast steps increase. In [75], a In [83], two combination approaches, namely ARIMA–ANN and
data mining approach is expanded to forecast both short-term and ARIMA–SVM, are proposed. The ARIMA method is employed for
long-term period. Furthermore, an advanced model, the ANN modeling the linear characteristics and the ANN or SVM methods
approach, is employed. The simulation results show that the are employed for modeling the non-linear characteristics. The
model generated by an ANN outperforms all other models for results show that the combination is viable for forecasting both
both short-term and long-term forecast. wind speed and wind power. However, the combined approaches
The frequency domain approach is introduced for characteriz- do not always outperform the single method. A similar combina-
ing the wind speed patterns in [76]. This approach separates the tion approach is also presented in [84].
discrete waveform into a sum of sinusoids of different frequencies In [85], two combination approaches are proposed; ARIMA–
from wind speed patterns. It shows a potential to forecast future ANN and ARIMA–Kalman. ARIMA is used to decide the best
wind speed patterns with better accuracy than time domain. structure of the ANN model and to obtain the best initial
Furthermore, it also shows a potential to estimate the wind speed parameters of the Kalman model. Both combination approaches
patterns for a target location using available data from the have good performance, which can be applied to non-stationary
reference site. It is necessary to develop the appropriate forecast- wind speed forecasting.
ing system using the components obtained from the frequency The Bayesian Model Averaging (MBA) approach is proposed to
domain approach. combine the wind speed forecasts obtained from different ANN
models in [86]. Three different ANN models are considered;
Adaptive Linear Element (ADALINE) network, Back-Propagation
2.1.3. Combination approach (BP) network, and Radial Basis Function (RBF) network. The BMA
The basic idea of the combination approach is to combine approach weights individual forecasts based on their posterior
different approaches, retaining advantages of each approach. The model probabilities, with the better performing forecasts receiving
desire is to improve the forecast accuracy, but combining forecasts higher weights than the worse. The simulation results show the
does not always perform better than the best individual forecasts. effectiveness of the proposed combination approach.
However, in some cases it is viewed as less risky to combine In [87], the combination of the Soft Computing Models (SCMs)
forecasts than to select an individual forecast [77]. In this section, and Similar Days (SD) methods is considered. Three SCMs models
the combination forecasting method is only reviewed. The combi- are considered, including Back-Propagation Neural Network
nation of input data (NWPs) is discussed in Section 4. (BPNN), Radial Basis Function Neural Network (RBFNN), and
Several approaches have been developed based on the combi- Adaptive Neuro-Fuzzy Inference System (ANFIS). The average of
nation of various models. The ANN–Fuzzy approach discussed in a selected number of similar wind speed days from SD methods is
the previous section is one of the combination approaches. In [78], used for the input of SCMs and the application of SCM refines the
the final forecast is made through an adaptive linear combination results to obtain the final wind speed forecast. The simulation
of the alternative forecasts, where the weights given to each results show that all evaluated SCMs incur some level of perfor-
forecasting methods are based on their forecasting performance. mance improvement with the addition of the SD model. Further-
For the combination, nine different time series models are con- more, the SD–ANFIS model outperforms all other individual and
sidered by varying the model parameters. The proposed combina- combined models.
tion shows better performance than individual forecasts. The combination between Seasonal ARIMA (SARIMA) and Least
The Anemos project considers the combination method, called Square Support Vector Machine (LSSVM) is developed to forecast
Adaptive Exponential Combination (AEC) in [79]. In the first step, the average monthly wind speed in [88]. The SARIMA model is
several combination methods are used, AEC being one of them. In developed by using training data and the LSSVM model is devel-
the second step, the AEC method is used to combine the alter- oped to describe the residuals from the SARIMA model. The
native combinations of the first step. The results show the simulation results show that the proposed method is simple and
usefulness of the proposed method. efficient.
In [80], a Multiple Architecture System (MAS) for wind speed The main objective of the combination approach is to improve
forecasting is proposed. The different regression algorithms are performance by taking advantage of the strengths of each model-
considered as candidate models for making up the ensemble ing approach. Combinations can consist of physical and statistical
forecast; the Multiple Linear (MLR)-based regression, the Multi- or alternative statistical approaches. The advantage of combining
layer Perceptron (MLP) neural networks, Radial Basis Functions results from different modeling approaches is improved forecast
(RBF) neural networks, and Support Vector Machines (SVM) accuracy. Furthermore, the combination of models reduces risks
regression. Three combination strategies are explored; simple during extreme events such as storms, where some model types
average, weighted average, and non-linear fusion by means of an may have significant forecasting errors.
ANN method. The proposed combination strategies improve the
performance with respect to the single forecast.
In [81], two types of univariate statistical forecasting models 2.2. Spatial correlation forecasting
based on the Least Squares-Support Vector Machine (LS-SVM) are
proposed: the univariate LS-SVM model and the hybrid model of Typically, wind speed and power forecasting refers to the
ARIMA and LS-SVM model. The hybrid approach is designed to future, and the forecast horizon depends on power system opera-
capture both the linear ARIMA modeled and the non-linear tion requirements. Spatial correlation forecasting, on the other
patterns modeled by LS-SVM. However, there are no significant hand, is mostly used for characterizing the wind resource at the
improvements between LS-SVM and the hybrid model. site where sufficient information is not available but a neighboring
The combination approach between the Radial Basis Function measuring station is available. It is a useful indicator when
(RBF) neural networks and persistence methods is considered in performing wind energy potential assessments at sites without
[82]. The experimental results show that the RBF method is better wind measurements.
Several approaches can be found regarding the problem of reduction is achieved than when the wind direction signal is
spatial correlation forecasting. The most commonly used approach omitted.
is the Measure-Correlate-Predict (MCP) method [89]. It is widely Being different from traditional wind speed and power fore-
used for future wind integration studies. The method is used for casting, spatial correlation forecasting takes into account the
the long-term wind data at the target site using the wind data spatial relationship of wind speed and power from different sites.
obtained from the reference site. Four different MCP methods are The spatial correlation forecasting requires wind speed and power
introduced: Linear Regression, Matric MCP, Weibull Scale, and measurements from multiple spatial correlated sites, where the
Wind Index MCP model. It is found that all perform well, but the measurements often involve time delays. The measurements and
performance is of course dependent upon the condition of the their time delays adds complexity and cost to the implementation
available data. of spatial correlation forecasting. Spatial correlation forecasting
Furthermore, in [90], the MCP method is used to forecast the can either employ physical models, which take into account the
long-term wind speed and then, the Monte Carlo based numerical terrain information, or statistical models, which take spatial
simulation is performed to utilize the probability models at the correlation information into account.
target site. This model is used for assessing the uncertainty of
overall wind energy potential prior to the construction of wind 2.3. Regional forecasting
turbines. For the case study it was found that the method
effectively evaluated the expected annual energy production at Regional forecasting is to forecast aggregated power output
the site. from a region where a number of wind turbines are sited. The
The Bayesian hierarchical model is developed to characterize regional forecasting is required for the market operator to estimate
the wind resource at the desired location using the known the total wind power production in a region. It provides a faster
resource [91]. The hierarchical model has two levels. The wind approach than calculating the sum of the wind power forecast
speed data from the reference site is defined as the sum of a from each wind farm. Furthermore, it is more accurate because of
temporal, a spatial, and error components together at the first its spatial smoothing effects discussed in Section 4. Several
level. At the second level, a temporal and a spatial component are approaches are proposed for regional forecasting.
modeled as a first order random walk and a multivariate normal Linear up-scaling in Previento is used for the regional forecast-
distribution respectively. It is found that the Bayesian inference ing in [36]. The ratio between the fluctuation of the wind power
can be a useful tool in spatial correlation forecasting. output of the measured single site and the ensemble of the sites is
In [92], the wind speed at “upwind” remote sites can be used used for up-scaling. It shows how this method works practically in
for forecasting at “downwind” sites. The ANN approach is used to Germany.
forecast the wind speed at the target site using the “upwind” and In [97], the aggregated forecasting method using the distances
“downwind” site measurements. It is found that the short-term between wind speed forecasts for a set of selected coordinates to
wind speed forecasting is improved when the recent measure- calculate wind power forecasts in a region is proposed. Once the
ments from the reference site are used. distance is obtained, the wind power forecast is calculated by
In [93], the ANN approach is used to estimate the wind speed at comparing the new wind speed forecast with the stored historical
the target site using the wind speed at a strong correlation site wind speed forecasts and their corresponding total wind power
among nearby sites. A high value of the Sample Cross Correlation measured in the region. The accuracy of the proposed method is
Function (SCCF) indicates the strongest correlation site. The verified by comparing the results with the reference method.
simulation results indicate that higher SCCF values between two In [98], an ANN–Fuzzy approach based on the up-scaling
sites leads to better wind speed estimates. However, this method, approach is used to forecast the regional wind generation. Three
contrary to other methods, requires the measurements from both different approaches are introduced based on how much the
the target and reference sites to calculate their SCCF. NWPs and on-line SCADA data are available: FNN up-scaling,
A Takagi, Sugeno, and Kang (TSK) Fuzzy approach is proposed Cascaded model, and FNN cluster model. The results show that
to forecast the wind speed and power at the target site using the the overall performance is improved but it is difficult to beat the
wind speed and direction measured at neighboring sites in [94]. persistence model for the first look-ahead time. SCADA data is
The Genetic Algorithm (GA) based learning scheme is employed to necessary to enhance the performance in the first forecast horizon.
achieve the training task of the model. The performance is A finite state Markov Chain model is developed to forecast the
evaluated at various terrain cases and demonstrates significant aggregated power output from a wind farm in [99]. The model
improvements over the persistent model. takes into account both the spatial and temporal dynamics of wind
In [95], a method for long-term wind speed and power power output. For spatial dynamics, a rigorous step-by-step
forecasting is proposed by using the NWPs provided at nearby procedure to characterize the probability distribution of the
sites of the wind park. Three types of local recurrent neural aggregated power output is presented. The temporal dynamics is
network models, depending on the neuron dynamic model, are characterized using auto-regression analysis tools.
employed: the Infinite Impulse Response Multilayer Perceptron The models of regional forecasting are mostly based on the up-
(IIR-MLP), the Local Activation Feedback Multilayer Network (LAF- scaling approach. It is difficult to obtain the wind power measure-
MLN), and the Diagonal Recurrent Neural Network (DRNN). ments and NWPs from all the wind farms in a region. Therefore,
Furthermore, two learning schemes for updating the weights of the up-scaling approach is used to scale the available on-line
the model are employed; the Global Recursive Prediction Error measurements and NWPs up to the region. There is not much
(GRPE) and the Decoupled RPE (DRPE). It is shown that the work reported dealing with regional forecasting. When wind
proposed methods outperform the static and the persistence energy penetration increases significantly good solutions to regio-
models. nal forecasting will be needed.
The ANN approach is used to estimate the wind speed at a
target station in [96]. The wind speeds recorded at neighboring 2.4. Probabilistic forecasting
measuring stations are used as signals for the input layer of the
model. It is found that as the number of reference stations The forecasting models discussed so far provide the point
increase, the estimation errors tend to decrease. Furthermore, forecast for how much wind speed and power can be expected
when the wind direction is used as an input, greater error for a number of hours into the future. It is important to know not
only the point forecast but also its expected uncertainty. Knowing In [109], a probabilistic forecasting method based on the
the uncertainty of the forecast enables the system operator to Markov Chain model is proposed. This provides not only the
assess the risk of the point forecast. Different approaches can be estimates of future wind power generation but also the associated
found in the literature regarding the problem of probabilistic probability distributions. First Order Markov Chain (FOMC) and
forecast. In this section, some of the representative methods are Second Order MC (SOMC) are considered. The proposed models
discussed. are compared with the persistence model to evaluate forecast
errors.
In [110], the proposed probabilistic model is designed to
2.4.1. Parametric approach forecast eight quantiles of the wind power distribution using
If there is any underlying assumption on the distribution one multiple Radial Basis Function Neural Networks (RBFNNs). The
tries to model for the forecasting error, it is a parametric approach. main idea is to exploit all information influencing the forecasting
A Beta Probability Density Function (PDF) is used to model a errors in order to estimate the wind power density function.
PDF of the wind power forecast for the persistence model in [100]. The three uncertainties considered are the weather stability, the
First, the forecast results are divided into power classes or bins for structure of the forecasting model, and the NWP accuracy. The
modeling the distribution of power with the Beta PDF. The overall developed RBFNNs are trained with the Ordinary Orthogonal Least
forecast error PDF is obtained from adding up the calculated error Square (OLS) algorithm and their performance is improved using a
PDFs of all bins. The results show the usefulness of the proposed Particle Swarm Optimization (PSO) algorithm. The simulation
forecast error PDF for finding the optimal storage size. results show that the proposed method performs well in different
The Cauchy distribution is proposed as a model distribution for weather conditions and different terrains.
the forecast error for the persistence model in [101]. The shape of In [111], the sparse Warped Gaussian Process (WGP) model is
the distribution is found to change significantly with the forecast- formulated to provide short-term probability forecasts of the wind
ing horizon. The effect of the proposed wind power forecast power generation. It converts a non-Gaussian wind power series
uncertainty is illustrated. to a latent series, which is well modeled by a Gaussian process.
There is little literature available on the parametric approach Furthermore, a sparsification method is employed to reduce the
due to the difficulty of defining the stochastic nature of wind computational costs of the model. The simulation results validate
speed. the effectiveness of the proposed model.
Non-parametric approaches are used when it is not possible to
formulate the distribution of the forecast errors. This distribution-
2.4.2. Non-parametric approach free approach is appealing since it is difficult to define the
Contrary to the parametric approach, the non-parametric stochastic nature of wind. Furthermore, the wind speed and power
approach does not rely on any assumption about the distribution. forecast are very different depending on the forecast time horizon
A non-parametric approach based on kernel density estimation and location. Non-parametric approaches can be applied in this
is proposed to provide the complete wind power forecast dis- case, and as a result are suitable to estimate the uncertainty of the
tribution in [102]. The distribution of wind power is decomposed wind speed and power forecast.
into a continuous part corresponding to all occurrences of wind
speed having zero production and a discrete part corresponding to
the discrete probabilities of all production values. The method 2.4.3. Using ensemble forecast
levels with the references while being fast and producing the Ensemble forecasting methods are designed to take advantage
complete PDF. of ensemble NWPs. The ensemble NWPs are obtained by running
The kernel density estimation is expanded to a time-adaptive differently calibrated NWP models or by slightly varying the initial
model and to adopt different kernels for several types of variables conditions, in order to obtain a set of different forecasts [112].
in [103]. The kernel density estimation model based on the Different types of meteorological ensemble forecasts are provided
Nadaraya–Watson estimator is proposed to estimate the uncer- either by the European Centre for Medium-range Weather Fore-
tainty in short-term wind power forecasts. It includes the adoption casts (ECMWF) or the National Centre for Environmental Predic-
of specific kernels for the explanatory variables and the develop- tion (NCEP). A MultiScheme Ensemble Prediction System (MS-EPS)
ment of a time adaptive model. It confirms its value from testing is introduced to utilize the ensemble NWPs for short-term wind
with real data from two large wind farms in the U.S. and shows power forecasting [113].
better results based on the evaluation criteria in [104]. Further- In [114], skill forecasting based on ensemble forecasts of wind
more, kernel density estimation based on the time-adaptive power production is introduced to estimate the expected level of
quantile-copula estimator is presented in [105]. forecast uncertainty. Prediction risk indices based on the disper-
In [106], the generic methodology for assessing the forecast risk sion of wind power ensembles over a single forecast horizon or
is presented. First, confidence intervals based on the resampling over a set of successive look-ahead times are introduced. Then,
approach are calculated by considering the forecast horizon, the these indices are used to forecast the distribution of expected
power class and the cut-off risk. Furthermore, the Meteorological forecast errors as skill forecasts. The simulation results show that
Risk Index (MRI) evaluating the weather stability is used to tune the proposed skill forecasting can provide useful information on
the confidence intervals. The proposed method is generic and can the expected level of forecast uncertainty.
be applied to any wind forecasting models. However, it does not The wind power density forecasts using the calibrated and
seem to have very much influence on the quantile estimation in smoothed ensemble forecast are proposed in [115]. The systematic
[107]. The most influential variable is the forecasted wind power bias of the ensemble forecasts is corrected by the calibration
production. Furthermore, the results show increased uncertainty approach that also incorporates kernel smoothing, with para-
for westerly winds. Also, the effect of the horizon on the quantiles meters optimized using maximum likelihood. The proposed power
is minor. density forecasts provide more accurate results for wind power
Probabilistic wind power forecasts using Local Quantile Regres- density and point forecasting than the reference.
sion (LQR) is proposed in [108]. This approach does not require the The ensemble forecasts give a much better idea of what
distribution assumption and is easy to include more predictive weather events may occur in the future. By comparing different
information. The simulation results demonstrate its effectiveness. forecasts, it identifies the expected spread of weather conditions
and assesses the probability of particular weather events. When powerful winds. However, the offshore wind forecasting is still in
the forecasts vary significantly there is a lot of uncertainty. the early phases of the development. There are not many available
However, if forecasts are very similar, then more confidence can specific forecasting models for offshore in the literature. Further
be placed in the forecasts. The ensemble forecast is useful in both advances in offshore forecasting will help offshore wind penetra-
wind speed and power forecasting. tion levels.
Most forecasting models have been designed for onshore. For
2.4.4. Additional models the physical approach to offshore forecasting, the special meteor-
In [116], the relation between typical weather situations and the ological characteristic of the marine atmospheric boundary layer
corresponding wind speed forecast error is investigated to model has to be considered including the stability of the boundary layer,
the uncertainty of the forecast. First, the Cluster analysis using the wind–wave interaction, etc., [122]. However, the statistical
Principal Component Analysis (PCA) rather than complex meteorolo- approach does not need to consider the precise offshore condi-
gical information is used to group days with similar meteorological tions. Results from various offshore forecasting methods are
conditions into classes. Then, for each of these groups the forecast investigated in [123].
error is determined separately. It is found that different meteorological In [124], the evaluation of NWPs forecasts for offshore sites is
situations have significant differences in the forecasting error. performed. Then, for the physical forecasting approach, the special
In [117], it is necessary to assess both temporal and spatial meteorological characteristics of the marine boundary layer are
interdependences in order to evaluate the aggregate error for the considered to calculate the wind speed at the hub height of the
regional uncertainty estimates. In this manner, the spatial model- wind turbines by using the derived vertical wind profile. The
ing of the wind power forecast uncertainty is proposed, which effectiveness of the proposed method is demonstrated by compar-
is important for the optimal management of power flows in ing with measurements.
the large region. The improvement of the probabilistic forecasting In [61,114,125,126], some of the statistical approaches use
is discussed. offshore data as an input, but the specific technique for dealing
The spatio-temporal modeling of wind power forecast uncer- with offshore data is not discussed.
tainty considering the interdependence of both time and space is
proposed in [118]. It is shown that there is significant cross-
3. Accuracy of wind speed and power forecasting
correlation between forecast errors obtained from neighboring
areas with lags of a few hours. These cross-correlation patterns
3.1. Reference models
characterized by wind speed and direction show that a higher
prevailing wind results in a stronger dependency on remote
A number of reference models have been introduced as bench-
places; while a lower prevailing wind results in a stronger
mark models. These models may be used for initial testing of the
dependency on local places. A further study is needed for the
accuracy of new approaches to wind forecasting. A number of
generalization by the methodology.
existing reference models are reviewed in this section.
In [119], the uncertainty of wind power forecasting regarding
the meteorological situation is investigated. From the simulation
results, the uncertainty of wind speed forecasts is not dependent 3.1.1. Persistence model
on the magnitude of the wind speed forecast. Furthermore, the The persistence model is a commonly used benchmark model.
uncertainty is larger for low pressure situations than high pressure In this model the future wind power will be the same as occurred
situations. However, some sites do not show these differences. in the present time step as given by
A framework for evaluating the quality of probabilistic fore- P^ t þ kjt ¼ P t ð2Þ
casting of wind power is presented in [104]. The framework is
composed of measures and diagrams, with the aim of providing where P^ t þ kjt is the forecast at time t for the look ahead time k and
useful information on each of the properties including reliability, P t is the measurement at time t. This model performs well for
sharpness and resolution, and a unique skill score. The framework short forecasting horizons.
is applied for comparing the quality of two non-parametric
methods; adaptive quantile regression and adapted re-sampling. 3.1.2. Weighted sum between persistence and the mean power
The relevance and interest of the proposed framework are production
discussed. The persistence model performs well for the short-term
In [120], a methodology for using the variability in wind speed forecast. However, it is not reasonable to use the persistence
forecasts at multiple grid points in NWP systems is presented to model when the forecast time horizon is more than a few
forecast the uncertainty in future wind power generation. The hours. For this case, a new reference model has been proposed
basic concept is that a major source of uncertainty in NWP [127]. It is a weighted sum between the persistence and the mean
forecasts arises from potential misplacement of weather features. given as
The comparison between the proposed method and NWP ensem- P^ t þ kjt ¼ ak P t þ ð1 ak ÞP ð3Þ
bles is performed. It demonstrates good accuracy in forecasting
large forecast errors. where P t is the measurement at time t, P is the estimated mean,
A wide comparison is carried out between the ARMA models, and ak is the correlation coefficient between P t and P t þ k defined as
five kinds of ANN models, and Adaptive Network based Fuzzy ð1=NÞ∑N k ~ ~
t ¼ 1P t P t þ k
Inference System (ANFIS) models in [121]. The comparison looks at ak ¼ ð4Þ
ð1=NÞ∑N k ~ 2
many forecasting methods, time horizons and a performance t ¼ 1P t
analysis. A further analysis of the statistical distribution of normal- where
ized errors from the forecasting models is performed.
P~ t ¼ P t P ð5Þ
2.5. Offshore forecasting
3.1.3. OL-Persistence model
Nowadays, offshore wind has a lot of attention and is expected For regional forecasting, total regional power is seldom avail-
to be installed more in the future because of more frequent and able. Therefore, the On-Line (OL) Persistence model is introduced
for the regional forecasting as a benchmark model [32]. This model 3.3. Forecast accuracy
is the sum of the production of the representative wind farms with
measurements, scaled to the total wind power and is given as Using the advanced forecasting models introduced in Section 2,
it is possible to provide better forecasts of wind speed and power
reg P reg Mr
P^ t þ kjt ¼ M instwf ∑ P wf i;t ð6Þ than the reference models. However, it is very difficult to say
∑i ¼r 1 P inst;i i ¼ 1 which model is the best because of its site dependency. Although
the forecasting model performs well, it does not guarantee that
where P reg wf
inst is the total installed capacity in the region, P inst;i is the the model works well at another site. Therefore, how the general
installed capacity of the M r wind farms for which on-line mea-
forecast accuracy varies is discussed in this section.
sures are available, and P wf i;t is the power measured at those wind
farms at time t.
3.3.1. The accuracy by forecast time horizon
The forecast accuracy decreases when prediction length
3.2. Model performance evaluation increases. The MAE of short-term forecasts is typically in the range
of 5–15% and then the errors increase rapidly with an increase in
To determine the model performance in terms of the forecast- the forecast time horizon. Thus, the MAE is typically in the range
ing accuracy, several model accuracy measures are discussed in of 13–21% range for 1–2 days ahead and rises to 20–25% range
this section. after about 3 days [10].
The forecast error is the difference between the measurement
and the forecast and is given as
3.3.2. The accuracy by location complexity
et þ kjt ¼ P t þ k P^ t þ kjt ð7Þ In [79], the average value of the normalized MAE (NMAE) for
the 12 h forecast horizon is compared based on the terrain
Sometimes, the normalized forecast error is used. The forecast error complexity of the wind farm. Much smaller NAME values in a flat
is normalized by the installed capacity (P inst ) as terrain wind farm (low complexity location) are observed than in
1 complex terrains (high complexity location). There is a significant
εt þ kjt ¼ e ð8Þ increase of the NMAE value when the terrain complexity is
P inst t þ kjt
increased. Furthermore, offshore wind farms have slightly higher
To assess the quality of the forecast, several evaluation criteria NMAE values relative to the flat terrain wind farm.
are used; Bias, Mean Absolute Error (MAE), Mean Absolute Per-
centage Error (MAPE), Mean Square Error (MSE), Root Mean Square 3.3.3. The accuracy by seasonal variability
Error (RMSE), and Standard Deviation of Errors (SDE) [128,129], as In [130], the performance of the forecasting models is related to
given by the seasonal variability. Smaller forecasting errors are observed in
1 N winter than summer because of the higher level of wind speeds
Biask ¼ Avgðet þ kjt Þ ¼ ek ¼ ∑ e ð9Þ
N t ¼ 1 t þ kjt and the larger uncertainty in the summer stormy weather situa-
tions, in particular low pressure systems with fast moving frontal
1 N zones containing complicated wind patterns.
MAEk ¼ ∑ e ð10Þ
N t ¼ 1 t þ kjt
3.3.4. The accuracy by weather condition

1 N et þ kjt In [106], the forecasting error is compared as a function of the
MAPEk ¼ ∑ ð11Þ
N t ¼ 1 Pt þ k Meteorological Risk Index (MRI) evaluating the weather stability.
The forecasting error increases linearly when MRI increases (the
1 N higher MRI represents unstable weather regimes). Furthermore, in
MSEk ¼ ∑ ðe Þ2 ð12Þ [119], the forecasting uncertainty does not depend on the wind
N t ¼ 1 t þ kjt
speed but on the pressure. The forecasting error in the low
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pressure situation is larger than the high pressure situation. These
pffiffiffiffiffiffiffiffiffiffiffiffi 1 N
RMSEk ¼ MSEk ¼ ∑ ðe Þ2 ð13Þ patterns are also observed in [116].
N t ¼ 1 t þ kjt
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 N 4. Improvement in forecasting performance
SDEk ¼ Stdðet þ kjt Þ ¼ ∑ ðe ek Þ2 ð14Þ
N t ¼ 1 t þ kjt
Development activities are on-going in reducing the forecasting
where N is the total number of samples, and Avgð U Þ and Stdð U Þ error in wind speed and power forecasting, spatial correlation
denote the average and the standard deviation respectively. forecasting, regional forecasting, probabilistic forecasting, and off-
Furthermore, the RMSEk can be expressed by the combination shore forecasting. In this section, some of the potential options to
between Biask and the SDEk as help improve the forecasting accuracy are discussed.
RMSE2k ¼ Bias2k þSDE2k ð15Þ

4.1. Kalman filtering
In the model comparison, it is important to quantify the
improvements of the advanced model over the benchmark model. It is well known that the accuracy of NWPs has a major impact
In this manner, the improvement with respect to the reference on the forecasting of wind speed and power. Therefore, reducing
model is introduced as [128,129] the uncertainty of NWPs can lead to significant forecasting
improvements. To improve the accuracy of NWPs, the Kalman
EC ref ;k EC adv;k
ImpEC;k ¼ ð16Þ filtering algorithm is applied to filter out systematic errors. The
EC ref ;k Kalman filtering algorithm provides the statistically optimal esti-
where EC stands for the evaluation criteria such as MAE, RMSE, mate recursively by combining recent weighted observations
SDE, etc. which minimize the corresponding biases.
In [131], the improvements in wind speed forecasting with the 4.5. Input parameter selection
application of the Kalman filtering are investigated. Third-order
polynomial functions are implemented in the Kalman filtering The selection of the input parameters for the forecasting model
algorithm. For the evaluation, two NWP models having different is crucial to the performance of the forecasts. Several input
characteristics and horizontal resolutions are tested. The simula- parameters are investigated in the literature for the purpose of
tion results show that the forecasting accuracy is significantly reducing the forecasting error.
improved by using the filtered NWPs. The significant forecast In [146], the inclusion of off-site observations with NWP
improvement using the Kalman filtering can be also found in models can increase the forecast accuracy. Despite the inherent
[59,132,133]. large forecast error from NWPs, adding off-site observations leads
to increases in forecast accuracy. Furthermore, as the number of
reference stations providing the measurements increase, the
4.2. Optimal combination forecasting errors tend to decrease [96]. The improvements by
the off-site observations can be the motivation for developing
The combination of different NWPs or forecasting models takes spatial correlation forecasting which forecasts wind speed and
advantage from each model. The advantage of combining the power at target sites by using the observations from neighboring
models, of course, improves the forecast accuracy. Furthermore, measuring stations, as described in Section 2.2.
the combination can reduce the risk from extreme events. The Several meteorological parameters as inputs have been con-
combination of forecasting methods is discussed in Section 2, and sidered for improving the performance of the forecast. The
the optimal combination can significantly reduce forecast error. inclusion of wind direction with wind speed can reduce the
Here, the optimal combination of NWPs is discussed. forecasting error [96,116,118,126]. Sometimes the wind vector
In [134,135], a classification scheme is used to find optimal applying both wind speed and direction is used [45,52,69]. Besides
combinations for specific weather situations. The different NWP the obvious importance of the wind speed as an input parameter,
models are exploited by applying the optimal weights depending temperature and pressure can help improve forecasts [69,116,147].
on the prevailing weather conditions. It shows that the optimal In [148], the inclusion of the wind speed forecast at wind
combination of NWPs significantly outperforms others even if the turbine hub height shows significant improvements in forecast
single NWP is already very well optimized. The combination of accuracy.
different NWPs is performed based on the quality of individual Furthermore, transformed data rather than wind speed itself
forecasts in [125,136]. show potential for improving the forecast. Several transformation
In [137], the combination between measured data and multiple techniques are used: wavelet decomposition [41,42,54,55,62],
wind forecasting models is proposed by advanced machine learn- empirical mode decomposition [56,57], and frequency domain
ing techniques. The input sources are the wind forecasts profiles, approach [76].
derived by synoptic and local atmospheric models, and the actual
climatic variable evolution measured by local environmental 4.6. Power curve modeling
stations. The simulation results demonstrate the effectiveness of
the proposed data fusion algorithms. When converting the wind speed to wind power production,
inaccuracies in the non-linear relation leads to further errors.
Wind power and wind speed are related through a cubic relation-
4.3. Spatial smoothing effects ship. Thus, small variations in the wind speed result in much
larger deviations in the wind power. The use of the certified
Under operational conditions, regional forecasting is required manufacture power curves does not guarantee accurate conver-
for the system operator to estimate the total wind power produc- sions. Using an accurate relation between the wind speed and
tion. The fluctuations of the combined power output of distributed power can greatly minimize conversion errors when converting
wind farms are damped, which results in decrease in fluctuations the wind speed forecast to the wind power forecast.
of the regional power output compared to the forecast for single The conventional wind power curve model is extended to a
sites. This is called spatial smoothing effects. These effects provide wind farm model by the addition of the wind direction depen-
the motivation for developing regional forecasting models rather dence with the wind speed dependence. Such a two dimensional
than calculating the sum of the wind power forecast from single wind farm power curve model may help to provide more accurate
sites. Several forecasting models are discussed in Section 2.3. The relations between wind speed and power [50,140,142].
reduction of the forecast error is found in the literature [138,139]. Statistical techniques rather than the wind power curve may be
used to describe the non-linear relation between wind speed and
wind power. The polynomial regression is used to estimate the
4.4. Statistical downscaling wind power by using the explanatory variables such as wind
speed, direction, etc. Furthermore, non-linear hyperbolic functions
NWPs are primary inputs to forecasting models. So improving containing trigonometric terms which describe the non-linear
NWPs can help to improve wind speed and power forecasting. behavior are used [133]. Sometimes, advanced statistical
NWPs are only available for grid points which generally corre- approaches (ANN, Fuzzy logic, etc.) are used to estimate wind
spond to areas that are larger than the wind farms themselves. power from the various inputs [55]. However, such a relation does
Therefore, estimating the wind speed at the wind farm location not need to be described when the wind power is forecasted
can provide more accurate wind speed forecast as an input to the directly from the raw input data. However, it is required when
forecasting models. The basic idea of statistical downscaling is to forecasting the wind speed and then converting it to the power.
use higher resolution calculations by adding more complete
physics descriptions to capture more precise local conditions. 4.7. Forecasting parameter
The analysis results show that wind power forecasting using
the statistical down scaled NWPs can improve the performance One of the most important decisions in a wind power system
[140–142]. Furthermore, some of the advanced NWP models are is choosing which parameters will be used in the forecast. The
examined in [143–145]. forecaster can choose the wind speed alone or the wind speed and
direction or the wind vector as the forecasting parameters. Then, the height of the measuring mast, the cross correlation factor, or
the forecasting wind information is converted to the wind power other relative meteorological data. These models are more difficult
production. Or the forecaster can forecast the wind power produc- than typical time-ahead forecasting models because they require
tion directly. Furthermore, the forecaster can choose regional or measurements from many spatially correlated sites and the
probabilistic forecasting. Choosing the appropriate forecasting associated measurement delay times.
parameters can improve the performance [149]. Regional forecasting has become more important for market
operators to estimate the total wind power production in a region.
Furthermore, it has been shown that regional forecasts provide
5. Conclusion and future works faster and more accurate results than single farm forecasts because
of the spatial smoothing effects. However, few studies have looked
5.1. Discussion and conclusion into the particularities of regional forecasting. When wind energy
penetration increases significantly solutions to this problem will
Today0 s power system is facing growing challenges in main- be needed.
taining a secure and reliable energy supply. Part of the growing Probabilistic forecasting gives an estimate of where the wind
challenge is the possibility of significant levels of uncertain wind production will lie within a certain degree of confidence. In this
generation being installed on power system. This brings new review some of the representative methods are considered,
challenges involving the management of wind intermittence. including the parametric approach, non-parametric approach,
Accurate forecasting of wind power is necessary to meeting these and ensemble forecast. There are not many parametric approaches
challenges. available due to the complex stochastic nature of wind speed.
The main objectives of wind speed and power forecasting are Thus, non-parametric approaches, and distribution-free
to estimate the wind speed and power as quickly and accurately as approaches, have been more appealing. In addition, ensemble
possible. With more accurate forecasting integration of large forecasting models for probabilistic forecasting are used in order
amounts of wind power into power system can be achieved. to obtain the expected spread of weather conditions and assess the
Accurate forecasting tools reduce the financial risk and lead to probability of particular weather events.
improved scheduling and unit commitment plans. In future, major wind power developments are expected to be
This paper provides an overview of forecasting tools that are installed offshore. More frequent and powerful wind speeds
available for wind generation. The wind speed and power fore- represent the main advantages of offshore power production.
casting approaches reviewed have different characteristics, with However, offshore wind forecasting is in an early phase of
different approaches providing better results for different forecast development. All state-of-the-art forecasting models were origin-
time horizons and locations. This paper divides forecasting meth- ally designed for onshore. For physical forecasting approaches
ods into five categories: wind speed and power forecasting, spatial special modeling considerations have to be considered for off-
correlation forecasting, regional forecasting, probabilistic forecast- shore. However, with statistical approaches precise knowledge of
ing, and offshore forecasting. An overview of existing research in offshore conditions is not needed.
each category is presented. Using advanced wind speed and power forecasting approaches
Several wind speed and power forecasting approaches, includ- introduced in Section 2, it is possible to provide significantly better
ing physical, statistical, and combination approaches, have been forecasts than the simplistic or reference forecasting approaches.
implemented and are now commercially available. The statistical However, it is difficult to say which model is the best because wind
approaches provide good results in the majority of cases, including speed and power forecasting is not “plug-in-play” due to the site-
short-term, medium-term, and long-term forecasting. However, in dependence. Therefore, an evaluation of accuracy relative to the
the very short-term and short-term horizon, the influence of forecasting model performance is discussed. The general forecast-
atmospheric dynamics becomes more important, so that the use ing accuracy is then considered along with the level of forecasting
of the physical approaches becomes necessary. errors that can be expected depending on the forecast time
The statistical forecasting approaches, such as conventional horizon, location complexity, seasonal variability, and weather
statistical approaches, ANN, and ANN–Fuzzy, use large amounts of conditions. The forecasting accuracy goes down with increasing
historical data as input. The conventional statistical approaches are forecast time horizon, terrain complexity, and unstable weather
mostly aimed at very short-term and short-term forecasting. ANN regimes.
is able to model complex non-linear relationships and extract the Finally, several potential techniques for improving the accuracy
dependence between the inputs and outputs through learning of the models are discussed: (1) Kalman filtering of systematic
and training. Fuzzy logic outperforms others when dealing errors coming from NWPs. (2) The combination of different NWPs
with reasoning problems, while learning and training abilities or forecasting models. (3) Combining power outputs of distributed
are mediocre. The merging of the ANN approach with the Fuzzy wind farms providing spatial smoothing effects. (4) Estimating
logic, creating the ANN–Fuzzy approach, provides excellent perfor- the wind speed at wind farm locations by statistical down
mance. scaling. (5) Appropriate selection of input parameters. (6)
It is difficult to say which model is the best because different Improved power curve modeling. (7) Choosing the appropriate
models perform differently in different situations. The combina- forecasting parameters.
tion approach attempts to leverage the strengths of different
modeling approaches. This approach is able to not only improve 5.2. Future work
the forecast accuracy, but also reduces the risk from extreme
events. The reviewed wind speed and power forecasting approaches in
In order to select a suitable location of wind generation, it is this paper provide a foundation for future research. Future work to
necessary to assess the wind power sufficiency of the neighboring be considered includes the following:
regions. In this case, spatial correlation forecasting may be applied
to estimate the characteristics of the wind resource at sites where Further improvements in both physical and statistical forecast-
insufficient information is not available. Current models use ing approaches and improved approaches for the combination
functions that describe the wind flow depending on the distance of different forecasting approaches can be expected to further
between two sites by selectively using the topography of the area, reduce forecasting errors.
Extending the forecasts further into the future, enabling long- [10] Zack J. Overview of wind energy generation forecasting. Albany, NY:
term planning. TrueWind Solutions, LLC. & AWS Scientific, Inc.; 2003.
Regional forecasting is useful in feasibility studies of prospec-
[11] Soman SS, Zareipour H, Malik O, Mandal P. A review of wind power
and wind speed forecasting methods with different time horizons. In:
tive installations and assessment of the potential impact on Proceedings of the 2010 north american power symposium (NAPS); 2010.
system reliability. The choice of reference sites and their p. 1–8.
[12] Colak I, Sagiroglu S, Yesilbudak M. Data mining and wind power prediction: a
combinations should be considered in regional forecasting. literature review. Renew Energy 2012;46:241–7.
Forecasting the aggregated wind power on a regional basis [13] Bhaskar M, Jain A, Venkata Srinath N. Wind speed forecasting: present
would be useful for management of total wind power produc- status. In: Proceedings of the 2010 international conference on power system
technology (POWERCON); 2010. p. 1–6.
tion. It would also be useful to detect extreme events so that
[14] Tsikalakis AG, Hatziargyriou ND, Katsigiannis YA, Georgilakis PS. Impact of
market operators can manage trading as wind power ramps up wind power forecasting error bias on the economic operation of autonomous
and down. power systems. Wind Energy 2009;12:315–31.
Probabilistic forecasting could provide probability distributions [15] Kennedy J, Fox B, Morrow DJ. Distributed generation as a balancing resource
for wind generation. Renew Power Gener, IET 2007;1:167–74.
of the forecasting errors, providing the opportunity to reduce [16] Marti I, San Isidro MJ, Gaston M, Loureiro Y, Sanz J, Perez I. Economic impact
the reserve capacity for balancing wind power forecast errors. of wind power forecast. In: Proceedings of the 2004 european wind energy
The development of operational ensemble forecasts using the conference and exhibition; 2004.
[17] Milligan MR, Miller AH, Chapman F. Estimating the economic value of
data from several NWPs would help improve the forecast wind forecasting to utilities. Golden, Colorado: National Renewable Energy
accuracy. Laboratory; 1995.
Improved offshore forecasting models are needed to improve [18] Lowery C, O0 Malley M. Impact of wind forecast error statistics upon unit
commitment. IEEE Trans Sustain Energy 2012;3:760–8.
the offshore forecasts that are currently being done with [19] Mahoney WP, Parks K, Wiener G, Yubao L, Myers WL, Juanzhen S, et al. A
models developed for onshore. wind power forecasting system to optimize grid integration. IEEE Trans
New approaches for complex terrains are required. These Sustain Energy 2012;3:670–82.
[20] Usaola J, Ravelo O, González G, Soto F, Dávila MC, Díaz-Guerra B. Benefits for
approaches could incorporate higher resolution maps.
wind energy in electricity markets from using short term wind power
Improving and tuning the NWP model to predict the hub- prediction tools; a simulation study. Wind Eng. 2004;28:119–27.
height wind would help improve the forecasting accuracy. It is [21] Wang J, Botterud A, Bessa R, Keko H, Carvalho L, Issicaba D, et al. Wind
sometimes required to forecast accurate pressure, temperature, power forecasting uncertainty and unit commitment. Appl Energy
2011;88:4014–23.
and other related weather parameters. [22] Botterud A, Wang J, Miranda V, Bessa RJ. Wind power forecasting in U.S.
Additional use of online wind measurement data has the electricity markets. Electr J 2010;23:71–82.
potential for improved forecasts, especially for very short- [23] Smith JC, Ahlstrom ML, Zavadil RM, Sadjadpour A, Philbrick CR. The role of
wind forecasting in utility system operation. In: Proceedings of the 2009
term and short-term wind power forecasting. IEEE power and energy society general meeting; 2009. p. 1–5.
Further research on automatic adaptive parameter estimation [24] Ernst B, Oakleaf B, Ahlstrom ML, Lange M, Moehrlen C, Lange B, et al.
is required since wind speed and power forecasting is very site- Predicting the wind. Power Energy Mag, IEEE 2007;5:78–89.
[25] Dica C, Dica CI, Vasiliu D, Comanescu G, Ungureanu M. Wind power short-
dependent. term forecasting system. In: Proceedings of the 2009 IEEE Bucharest Power-
Increasing the spatial and time resolution of NWPs, better Tech; 2009. p. 1–7.
taking into account local phenomena, would help to improve [26] Gubina AF, Keane A, Meibom P, O0 Sullivan J, Goulding O, McCartan T, et al.
New tool for integration of wind power forecasting into power system
the forecast.
operation. In: Proceedings of the 2009 IEEE Bucharest PowerTech; 2009. p.
Forecasting extreme events, as opposed to common behavior, 1–6.
would be very valuable to protecting the power system. [27] Yutong Z, Ka Wing C. The impact of wind forecasting in power system
reliability. In: Proceedings of the 2008 third international conference on
electric utility deregulation and restructuring and power technologies
(DRPT); 2008. p. 2781–5.
[28] Ahlstrom Jones L, Zavadil R, Grant W. The future of wind forecasting and
Acknowledgment utility operations. Power Energy Mag, IEEE 2005;3:57–64.
[29] Ahlstrom ML, Zavadil RM. The role of wind forecasting in grid operations &
reliability. In: Proceedings of the 2005 IEEE/PES transmission and distribu-
The authors would like to thank Electrical Distribution Design, tion conference and exhibition: Asia and Pacific; 2005. p. 1–5.
Inc., for funding and providing technical assistance for this study. [30] Lange M, Focken U. Physical approach to short-term wind power prediction.
New York, NY: Springer; 2006.
[31] Landberg L, Giebel G, Nielsen HA, Nielsen T, Madsen H. Short-term predic-
tion – an overview. Wind Energy 2003;6:273–80.
Reference
[32] Nils S. Development of methods for regional wind power forecasting [Ph.D.].
MINES ParisTech; 2008.
[1] United States Department of Energy. 20% Wind energy by 2030; 2008. [33] Landberg L. Short-term prediction of local wind conditions. J Wind Eng Ind
[2] Lei M, Shiyan L, Chuanwen J, Hongling L, Yan Z. A review on the forecasting of Aerodyn 2001;89:235–45.
wind speed and generated power. Renew Sustain Energy Rev [34] Landberg L. Short-term prediction of the power production from wind farms.
2009;13:915–20. J Wind Eng Ind Aerodyn 1999;80:207–20.
[3] Foley AM, Leahy PG, Marvuglia A, McKeogh EJ. Current methods and [35] Landberg L. A mathematical look at a physical power prediction model. Wind
advances in forecasting of wind power generation. Renew Energy Energy 1998;1:23–8.
2012;37:1–8. [36] Focken U, Lange M, Waldl H-P. Previento – a wind power prediction system
[4] Cutler N. Wind energy forecasting issues paper. Centre for Energy and with an innovative upscaling algorithm. In: Proceedings of the 2001
Environmental Markets, The University of New South Wales; 2006. european wind energy conference & exhibition; 2001.
[5] Dongmei Z, Yuchen Z, Xu Z. Research on wind power forecasting in wind [37] Marti I, Cabezon D, Villanueva J, Sanisidro MJ, Loureiro Y, Cantero E, et al.
farms. In: Proceedings of the 2011 IEEE power engineering and automation LocalPred and RegioPred. Advanced tools for wind energy prediction in
conference (PEAM); 2011. p. 175–8. complex terrain. In: Proceedings of the 2003 european wind energy
[6] Foley AM, Leahy PG, McKeogh EJ. Wind power forecasting & prediction conference & exhibition; 2003.
methods. In: Proceedings of the 9th international conference on environ- [38] McGowin C. California wind energy forecasting system development and
ment and electrical engineering (EEEIC); 2010. p. 61–4. testing. Phase 1: Initial testing. Palo Alto, California: Electric Power Research
[7] Kariniotakis G, Pinson P, Siebert N, Giebel G, Barthelmie R. The state of the art Institute (EPRI); 2003.
in short-term prediction of wind power – from an offshore perspective. 2004 [39] Milligan M, Schwartz M, Wan Y. Statistical wind power forecasting models:
Sea Tech Week; 2004. results for U.S. wind farms. WINDPOWER 2003; 2003.
[8] Costa A, Crespo A, Navarro J, Lizcano G, Madsen H, Feitosa E. A review on the [40] Milligan M, Schwartz MN, Wan Y. Statistical wind power forecasting for U.S.
young history of the wind power short-term prediction. Renew Sustain wind farms. In: Proceedings of the 17th conference on probability and
Energy Rev 2008;12:1725–44. statistics in the atmospheric sciences/2004 american meteorological society
[9] Negnevitsky M, Mandal P, Srivastava AK. An overview of forecasting annual meeting; 2004.
problems and techniques in power systems. In: Proceedings of the 2009 [41] Liu H, Tian H-Q, Chen C, Li Y-f. A hybrid statistical method to predict wind
IEEE power and energy society general meeting; 2009. p. 1–4. speed and wind power. Renew Energy 2010;35:1857–61.
[42] Peng L, Lili Y. Short-term wind speed forecasting based on non-stationary [73] Kusiak A, Haiyang Z, Zhe S. Short-term prediction of wind farm power: a data
time series analysis and ARCH model. In: Proceedings of the 2011 mining approach. IEEE Trans Energy Convers 2009;24:125–36.
International Conference on Multimedia Technology (ICMT); 2011. p. [74] Zheng H, Kusiak A. Prediction of wind farm power ramp rates: a data-mining
2549–53. approach. J Sol Energy Eng 2009;131:031011–8.
[43] Meng-Di W, Qi-Rong Q, Bing-Wei C. Short-term wind speed forecasting [75] Kusiak A, Zheng H, Song Z. Wind farm power prediction: a data-mining
combined time series method and arch model. In: Proceedings of the 2012 approach. Wind Energy 2009;12:275–93.
international conference on machine learning and cybernetics (ICMLC); [76] Jung J, Tam K-S. A frequency domain approach to characterize and analyze
2012. p. 924–7. wind speed patterns. Appl Energy 2013;103:435–43.
[44] Xiaodong Z, Jianwen Z, Yinping L, Rongbao Z. Short-term forecasting of wind [77] Hibon M, Evgeniou T. To combine or not to combine: selecting among
speed based on recursive least squares. In: Proceedings of the 2011 international forecasts and their combinations. Int J Forecast. 2005;21:15–24.
conference on electrical and control engineering (ICECE); 2011. p. 367–70. [78] Sánchez I. Short-term prediction of wind energy production. Int J Forecast
[45] Erdem E, Shi J. ARMA based approaches for forecasting the tuple of wind 2006;22:43–56.
speed and direction. Appl Energy 2011;88:1405–14. [79] Marti I, Kariniotakis G, Pinson P, Sanchez I, Nielsen TS, Madsen H., et al.
[46] Miranda MS, Dunn RW. One-hour-ahead wind speed prediction using a Evaluation of advanced wind power forecasting models – results of the
Bayesian methodology. In: Proceedings of the 2006 IEEE power engineering Anemos Project. In: Proceedings of the 2006 european wind energy
society general meeting; 2006. p. 1–6. conference and exhibition; 2006.
[47] Kani SAP, Riahy GH. A new ANN-based methodology for very short-term [80] Bouzgou H, Benoudjit N. Multiple architecture system for wind speed
wind speed prediction using Markov chain approach. In: Proceedings of the prediction. Appl Energy 2011;88:2463–71.
2008 IEEE Canada electric power conference (EPEC); 2008. p. 1–6. [81] Qian Z, Kin Keung L, Dongxiao N, Qiang W. Wind park power forecasting
[48] Jursa R. Variable selection for wind power prediction using particle swarm models and comparison. In: Proceedings of the 2012 fifth international joint
optimization. In: Proceedings of the 9th annual conference on genetic and conference on computational sciences and optimization (CSO); 2012. p. 27–
evolutionary computation; 2007. p. 2059–65. 31.
[49] Amjady N, Keynia F, Zareipour H. Wind power prediction by a new forecast [82] Xiao Q, Cong J, Jun W. Online clustering for wind speed forecasting based on
engine composed of modified hybrid neural network and enhanced particle combination of RBF neural network and persistence method. In: Proceedings
swarm optimization. IEEE Trans Sustain Energy 2011;2:265–76. of the 2011 chinese control and decision conference (CCDC); 2011. p. 2798–
[50] Amjady N, Keynia F, Zareipour H. Short-term wind power forecasting using 802.
ridgelet neural network. Electr Power Syst Res 2011;81:2099–107. [83] Shi J, Guo J, Zheng S. Evaluation of hybrid forecasting approaches for wind
[51] Li G, Shi J. On comparing three artificial neural networks for wind speed speed and power generation time series. Renew Sustain Energy Rev
forecasting. Appl Energy 2010;87:2313–20. 2012;16:3471–80.
[52] Kitajima T, Yasuno T. Output prediction of wind power generation system [84] Cadenas E, Rivera W. Wind speed forecasting in three different regions of
using complex-valued neural network. In: Proceedings of the 2010 SICE Mexico, using a hybrid ARIMA–ANN model. Renew Energy 2010;35:2732–8.
annual conference; 2010. p. 3610–3. [85] Liu H, Tian H-q, Li Y-f. Comparison of two new ARIMA–ANN and ARIMA–
[53] Blonbou R. Very short-term wind power forecasting with neural networks Kalman hybrid methods for wind speed prediction. Appl Energy
and adaptive Bayesian learning. Renew Energy 2011;36:1118–24. 2012;98:415–24.
[54] Catalão JPS, Pousinho HMI, Mendes VMF. Short-term wind power forecasting [86] Li G, Shi J, Zhou J. Bayesian adaptive combination of short-term wind speed
in Portugal by neural networks and wavelet transform. Renew Energy forecasts from neural network models. Renew Energy 2011;36:352–9.
2011;36:1245–51. [87] Haque AU, Mandal P, Kaye ME, Meng J, Chang L, Senjyu T. A new strategy for
[55] Bhaskar K, Singh SN. AWNN-assisted wind power forecasting using feed- predicting short-term wind speed using soft computing models. Renew
forward neural network. IEEE Trans Sustain Energy 2012;3:306–15. Sustain Energy Rev 2012;16:4563–73.
[56] Guo Z, Zhao W, Lu H, Wang J. Multi-step forecasting for wind speed using a [88] Guo Z, Zhao J, Zhang W, Wang J. A corrected hybrid approach for wind speed
modified EMD-based artificial neural network model. Renew Energy prediction in Hexi Corridor of China. Energy 2011;36:1668–79.
2012;37:241–9. [89] Thogersen ML, Motta M, Sorensen T, Nielsen P. Measure-correlate-predict
[57] Liu H, Chen C, Tian H-q, Li Y-f. A hybrid model for wind speed prediction methods: case studies and software implementation. In: Proceedings of the
using empirical mode decomposition and artificial neural networks. Renew 2007 EWEA conference; 2007.
Energy 2012;48:545–56. [90] Kwon S-D. Uncertainty analysis of wind energy potential assessment. Appl
[58] Celik AN, Kolhe M. Generalized feed-forward based method for wind energy Energy 2010;87:856–65.
prediction. Appl Energy 2013;101:582–8. [91] Miranda MS, Dunn RW, Li F, Shaddick G, Bell KRW. Bayesian inferencing for
[59] Zhao P, Wang J, Xia J, Dai Y, Sheng Y, Yue J. Performance evaluation and wind resource characterisation. In: Proceedings of the 2006 international
accuracy enhancement of a day-ahead wind power forecasting system in conference on probabilistic methods applied to power systems (PMAPS);
China. Renew Energy 2012;43:234–41. 2006. p. 1–6.
[60] Douak F, Melgani F, Benoudjit N. Kernel ridge regression with active learning [92] Alexiadis MC, Dokopoulos PS, Sahsamanoglou HS. Wind speed and power
for wind speed prediction. Appl Energy 2013;103:328–40. forecasting based on spatial correlation models. IEEE Trans Energy Convers
[61] Sideratos G, Hatziargyriou ND. An advanced statistical method for wind 1999;14:836–42.
power forecasting. IEEE Trans Power Syst 2007;22:258–65. [93] Bechrakis DA, Sparis PD. Correlation of wind speed between neighboring
[62] Catala JPS, Pousinho HMI, Mendes VMF. Hybrid wavelet-PSO-ANFIS measuring stations. IEEE Trans Energy Convers 2004;19:400–6.
approach for short-term wind power forecasting in Portugal. IEEE Trans [94] Damousis IG, Alexiadis MC, Theocharis JB, Dokopoulos PS. A fuzzy model for
Sustain Energy 2011;2:50–9. wind speed prediction and power generation in wind parks using spatial
[63] Hong Y-Y, Chang H-L, Chiu C-S. Hour-ahead wind power and speed correlation. IEEE Trans Energy Convers 2004;19:352–61.
forecasting using simultaneous perturbation stochastic approximation [95] Barbounis TG, Theocharis JB, Alexiadis MC, Dokopoulos PS. Long-term wind
(SPSA) algorithm and neural network with fuzzy inputs. Energy speed and power forecasting using local recurrent neural network models.
2010;35:3870–6. IEEE Trans Energy Convers 2006;21:273–84.
[64] Mori H, Kurata E. Application of Gaussian Process to wind speed forecasting [96] Velázquez S, Carta JA, Matías JM. Influence of the input layer signals of ANNs
for wind power generation. In: Proceedings of the 2008 IEEE international on wind power estimation for a target site: a case study. Renew Sustain
conference on sustainable energy technologies (ICSET); 2008. p. 956–9. Energy Rev 2011;15:1556–66.
[65] Peng K, Feng G. Sparse heteroscedastic Gaussian process for shortterm wind [97] Lobo MG, Sanchez I. Aggregated wind power prediction methods based on
speed forecasting. In: Proceedings of the 2012 international joint conference distances between weather forecasting vectors. In: Proceedings of the 2009
on neural networks (IJCNN); 2012. p. 1–8. international conference on clean electrical power; 2009. p. 242–7.
[66] Liu H, Shi J, Erdem E. Prediction of wind speed time series using modified [98] Pinson P, Siebert N, Kariniotakis G. Forecasting of regional wind generation
Taylor Kriging method. Energy 2010;35:4870–9. by a dynamic fuzzy-neural networks based upscaling approach. In:
[67] Chi-Yo H, Yu-Wei L, Wei-Chang T, Po-Yen W. Short term wind speed Proceedings of the 2003 european wind energy conference and exhibition;
predictions by using the grey prediction model based forecast method. In: 2003.
Proceedings of the 2011 IEEE green technologies conference (IEEE-Green); [99] Murugesan S, Junshan Z, Vittal V. Finite state Markov chain model for
2011. p. 1–5. wind generation forecast: A data-driven spatiotemporal approach. In:
[68] Shu F, Liao JR, Yokoyama R, Luonan C, Wei-Jen L. Forecasting the wind Proceedings of the 2012 IEEE PES innovative smart grid technologies
generation using a two-stage network based on meteorological information. (ISGT); 2012. p. 1–8.
IEEE Trans Energy Convers 2009;24:474–82. [100] Bludszuweit H, Dominguez-Navarro JA, Llombart A. Statistical analysis of
[69] Took CC, Strbac G, Aihara K, Mandic DP. Quaternion-valued short-term joint wind power forecast error. IEEE Trans Power Syst 2008;23:983–91.
forecasting of three-dimensional wind and atmospheric parameters. Renew [101] Hodge B, Milligan M. Wind power forecasting error distributions over
Energy 2011;36:1754–60. multiple timescales. In: Proceedings of the 2011 IEEE power and energy
[70] Li G, Shi J. Application of Bayesian model averaging in modeling long-term society general meeting; 2011. p. 1–8.
wind speed distributions. Renew Energy 2010;35:1192–202. [102] Juban J, Siebert N, Kariniotakis GN. Probabilistic short-term wind power
[71] Jiang Y, Song Z, Kusiak A. Very short-term wind speed forecasting with forecasting for the optimal management of wind generation. In: Proceedings
Bayesian structural break model. Renew Energy 2013;50:637–47. of the 2007 IEEE Lausanne PowerTech; 2007. p. 683–8.
[72] Zhang W, Wu J, Wang J, Zhao W, Shen L. Performance analysis of four [103] Bessa RJ, Miranda V, Botterud A, Jianhui W, Constantinescu EM. Time
modified approaches for wind speed forecasting. Appl Energy 2012;99: adaptive conditional kernel density estimation for wind power forecasting.
324–33. IEEE Trans Sustain Energy 2012;3:660–9.
[104] Pinson P, Nielsen HA, Møller JK, Madsen H, Kariniotakis GN. Non-parametric [127] Nielsen TS, Joensen A, Madsen H, Landberg L, Giebel G. A new reference for
probabilistic forecasts of wind power: required properties and evaluation. wind power forecasting. Wind Energy 1998;1:29–34.
Wind Energy 2007;10:497–516. [128] Madsen H, Kariniotakis G, Nielsen HA, Nielsen TS, Pinson PA. Protocol for
[105] Bessa RJ, Miranda V, Botterud A, Zhou Z, Wang J. Time-adaptive quantile- standardizing the performance evaluation of short-term wind power pre-
copula for wind power probabilistic forecasting. Renew Energy 2012;40: diction models. In: Proceedings of the global windpower 2004 conference;
29–39. 2004.
[106] Pinson P, Kariniotakis G. On-line assessment of prediction risk for wind [129] Madsen H, Pinson P, Kariniotakis G, Nielsen HA, Nielsen T. Standardizing the
power production forecasts. Wind Energy 2004;7:119–32. performance evaluation of shortterm wind power prediction models. Wind
[107] Nielsen HA, Madsen H, Nielsen TS. Using quantile regression to extend an Eng 2005;29:475–89.
existing wind power forecasting system with probabilistic forecasts. Wind [130] Lange M, Focken U. State-of-the-art in wind power prediction in Germany
Energy 2006;9:95–108. and international developments. In: Proceedings of the second workshop of
[108] Bremnes JB. Probabilistic wind power forecasts using local quantile regres- international feed-in cooperation; 2005.
sion. Wind Energy 2004;7:47–54. [131] Louka P, Galanis G, Siebert N, Kariniotakis G, Katsafados P, Pytharoulis I, et al.
[109] Carpinone A, Langella R, Testa A, Giorgio M. Very short-term probabilistic Improvements in wind speed forecasts for wind power prediction purposes
wind power forecasting based on Markov chain models. In: Proceedings of using Kalman filtering. J Wind Eng Ind Aerodyn 2008;96:2348–62.
the 2010 IEEE 11th international conference on probabilistic methods applied [132] Cassola F, Burlando M. Wind speed and wind energy forecast through
to power systems (PMAPS); 2010. p. 107–12. Kalman filtering of numerical weather prediction model output. Appl Energy
[110] Sideratos G, Hatziargyriou ND. Probabilistic wind power forecasting using 2012;99:154–66.
radial basis function neural networks. IEEE Trans Power Syst [133] Stathopoulos C, Kaperoni A, Galanis G, Kallos G. Wind power prediction
2012;27:1788–96. based on numerical and statistical models. J Wind Eng Ind Aerodyn
[111] Peng K, Feng G, Xiaohong G, Jiang W. Prediction intervals for wind power 2013;112:25–38.
forecasting: using sparse warped Gaussian process. In: Proceedings of the [134] Lange M, Focken U. New developments in wind energy forecasting. In:
2012 IEEE power and energy society general meeting; 2012. p. 1–8. Proceedings of the 2008 IEEE power and energy society general meeting –
[112] Giebel G, Landberg L, Badger J, Sattler K, Feddersen H, Nielsen TS, et al. Using conversion and delivery of electrical energy in the 21st century; 2008. p. 1–8.
ensemble forecasting for wind power. In: Proceedings of the 2003 european [135] Lange M, Focken U, Meyer R, Denhardt M, Ernst B, Berster F. Optimal
wind energy conference and exhibition; 2003. combination of different numerical weather models for improved wind
[113] Lang S, Mohrlen C, Jorgensen J, Gallachoir BO, McKeogh E. Application of a power predictions. In: Proceedings of the 6th international workshop
multi-scheme ensemble prediction system for wind power forecasting in large-scale integration of wind power and transmission networks for off-
ireland and comparison with validation results from Denmark and Germany. shore wind farms; 2006.
In: Proceedings of the 2006 european wind energy conference and exhibi- [136] Lueder von B. Combination of deterministic and probabilistic meteorological
tion; 2006. models to enhance wind farm power forecasts. J Phys: Conf Ser.
[114] Pinson P, Nielsen HA, Madsen H, Kariniotakis G. Skill forecasting from 2007;75:012050.
ensemble predictions of wind power. Appl Energy 2009;86:1326–34. [137] Vaccaro A, Mercogliano P, Schiano P, Villacci D. An adaptive framework based
[115] Taylor JW, McSharry PE, Buizza R. Wind power density forecasting using on multi-model data fusion for one-day-ahead wind power forecasting.
ensemble predictions and time series models. IEEE Trans Energy Convers Electr Power Syst Res 2011;81:775–82.
2009;24:775–82. [138] Focken U, Lange M, Mönnich K, Waldl H-P, Beyer HG, Luig A. Short-term
[116] Lange M, Heinemann D. Relating the uncertainty of short-term wind speed prediction of the aggregated power output of wind farms – a statistical
predictions to meteorological situations. In: Proceedings of the 2003 eur- analysis of the reduction of the prediction error by spatial smoothing effects.
opean wind energy conference and exhibition; 2003. J Wind Eng Ind Aerodyn 2002;90:231–46.
[117] Papaefthymiou G, Pinson P. Modeling of spatial dependence in wind power [139] Yu H, Liuchen C. A study of the reduction of the regional aggregated wind
forecast uncertainty. In: Proceedings of the 2008 10th international con- power forecast error by spatial smoothing effects in the Maritimes Canada.
ference on probabilistic methods applied to power systems (PMAPS); 2008. p. In: Proceedings of the 2010 IEEE electric power and energy conference
1–9. (EPEC); 2010. p. 1–6.
[118] Tastu J, Pinson P, Kotwa E, Madsen H, Nielsen HA. Spatio-temporal analysis [140] Kariniotakis G, Halliday J, Brownsword R, Marti I, Palomares A, Cruz I, et al.
and modeling of short-term wind power forecast errors. Wind Energy Next generation short-term forecasting of wind power – overview of the
2011;14:43–60. ANEMOS Project. In: Proceedings of the 2006 european wind energy
[119] Lange M, Heinemann D. Accuracy of short term wind power predictions conference and exhibition; 2006.
depending on meteorological conditions. In: Proceedings of the 2002 global [141] Galanis G, Badger J, Marti Perez I, Louka P, Kallos G, Palomares AM, et al.
windpower conference; 2002. Short-term forecasting using advanced physical modelling – the results of
[120] Cutler N, Kepert J, Outhred H, MacGill I. Characterizing wind power forecast the Anemos Project. In: Proceedings of the 2006 european wind energy
uncertainty with numerical weather prediction spatial fields. Wind Eng conference and exhibition; 2006.
2008;32:509–24. [142] Nielsen TS, Madsen H, Nielsen HA, Pinson P, Kariniotakis G, Siebert N, et al.
[121] De Giorgi MG, Ficarella A, Tarantino M. Error analysis of short term wind Short-term wind power forecasting using advanced statistical methods. In:
power prediction models. Appl Energy 2011;88:1298–311. Proceedings of the 2006 european wind energy conference and exhibition;
[122] Kariniotakis G, Mayer D, Moussafir J, Chevallaz-Perrier R, Usaola J, Sanchez I, 2006.
et al. ANEMOS: development of a next generation wind power forecasting [143] Al-Yahyai S, Charabi Y, Al-Badi A, Gastli A. Nested ensemble NWP approach
system for the large-scale integration of onshore & offshore wind farms. In: for wind energy assessment. Renew Energy 2012;37:150–60.
Proceedings of the 2003 european wind energy conference and exhibition; [144] Lazić L, Pejanović G, Živković M. Wind forecasts for wind power generation
2003. using the Eta model. Renew Energy 2010;35:1236–43.
[123] Tambke J, von Bremen L, Barthelmie R, Palomares AM, Ranchin T, Juban J, [145] Cheng WYY, Liu Y, Liu Y, Zhang Y, Mahoney WP, Warner TT. The impact of
et al. Short-term forecasting of offshore wind farm production – develop- model physics on numerical wind forecasts. Renew Energy 2013;55:347–56.
ments of the Anemos Project. In: Proceedings of the 2006 european wind [146] Larson KA, Westrick K. Short-term wind forecasting using off-site observa-
energy conference and exhibition; 2006. tions. Wind Energy 2006;9:55–62.
[124] Tambke J, Lange M, Focken U, Wolff J-O, Bye JAT. Forecasting offshore wind [147] De Giorgi MG, Ficarella A, Tarantino M. Assessment of the benefits of
speeds above the North Sea. Wind Energy 2005;8:3–16. numerical weather predictions in wind power forecasting based on statistical
[125] von Bremen L. Optimal Linkage of NWP models with neural networks for methods. Energy 2011;36:3968–78.
offshore wind power predictions. In: Proceedings of the 6th international [148] Lange B, Rohrig K, Ernst B, Schlogl F, Cali U, Jursa R, et al. Wind power
workshop large-scale integration of wind power and transmission networks prediction in Germany – recent advances and future challenges. In: Proceed-
for offshore wind farms; 2006. ings of the 2006 european wind energy conference and exhibition; 2006.
[126] Gallego C, Pinson P, Madsen H, Costa A, Cuerva A. Influence of local wind [149] Negnevitsky M, Potter CW.Innovative short-term wind generation prediction
speed and direction on wind power dynamics – application to offshore very techniques. In: Proceedings of the 2006 IEEE PES power systems conference
short-term forecasting. Appl Energy 2011;88:4087–96. and exposition (PSCE); 2006. p. 60–5.

Renewable and Sustainable Energy Reviews: Jaesung Jung, Robert P. Broadwater

Uploaded by

Copyright:

Available Formats

Renewable and Sustainable Energy Reviews: Jaesung Jung, Robert P. Broadwater

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Renewable and Sustainable Energy Reviews: Jaesung Jung, Robert P. Broadwater

Uploaded by

Copyright:

Available Formats

Renewable and Sustainable Energy Reviews 31 (2014) 762–777

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews

Current status and future advances for wind speed

3.3.3. The accuracy by seasonal variability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 772

1. Introduction presented in Section 4. Finally, conclusions and future works are

represents the relation between wind power or speed forecasting

2.1.2.1. Conventional statistical approach. In the conventional

where X t represents the forecasting parameter at time t, φi is the

Fig. 2. The statistical approach to forecasting wind speed and power.

Wind Speed Wind Speed or Power Forecast

Fig. 3. ANN approach for wind speed and power forecast.

RMSE2k ¼ Bias2k þSDE2k ð15Þ

You might also like