**4. Variability and predictability of wind energy**

Low variability and great predictability are required for a reliable energy source. While the modest variance is acceptable, poor predictability is not, and can result in significant revenue loss. Wind energy fluctuation is caused by a heavy reliance on weather, which varies during the day and annually. As a result, precise weather forecasting is required to produce a useful wind power forecast, **Figure 2** shows Dualstep wind power prediction approach based on a hybrid wavelet transform (WT)-ant colony optimization algorithm (ACO)-feedforward artificial neural network (FFANN).

It is commonly known that the accuracy of weather forecasts improves as the forecast horizon shortens. Combining forecasts from multiple numerical techniques can also be advantageous. As a result, wind farms rely on a variety of weather forecasts given by different models at different times of day or week. Although the weather cannot be controlled, the wind sector may take advantage of advances in artificial intelligence to increase the predictability of the energy supply.

In the chapter, there are several ways for forecasting wind power are categorized as Physical models, statistical models, and hybrid.

### **5. Current forecasting and prediction methods**

Models for wind energy forecasting can be classified into two categories. The first is based on historical wind time series analysis, while the second is based on anticipated values from a numerical weather prediction (NWP) model. However, physical methods, classic statistical or 'black box' methods, and more recently, so-called learning approaches, artificial intelligence, or "gray box" methods are used to characterize wind power forecasts. All of these can be included into hybrid approaches.

#### **Figure 2.**

*Dual-step wind power prediction approach based on hybrid wavelet transform (WT)-ant colony optimization algorithm (ACO)-feedforward artificial neural network (FFANN).*

The first category of models utilizes a statistical approach to anticipate mean hourly wind speed or directly forecast electric power production. To anticipate wind power N-steps ahead, the models in the second category use explanatory variables (often hourly mean wind speed and direction) generated from a meteorological model of wind dynamics. In the majority of cases, the models in the first group produce good results in the estimation of mean monthly or even higher temporal scale (quarterly, annual) wind speed.

However, the influence of atmospheric dynamics becomes more important in the short term (mean daily or hourly wind speed predictions), making the adoption of the second group's models necessary [11].

In wind power forecasting, there are three steps: first, determining wind speed from a model; second, calculating the wind power output forecast or prediction; and finally, regional forecasting or upscaling or downscaling, which can be implemented over various time horizons. Statistical models are typically used in very short-term forecasting. Ensemble forecasting is utilized to overcome these statistical and learning method conditions [12].

Nielsen et al. [12] demonstrated that if several NWP forecasts are used the forecast error decreases. Louka et al. [13] showed that the Kalman filter can remove systematic forecast errors in NWP wind speed forecasts. Wind forecasting can be separated based on the prediction horizon into three categories:


For short-term forecasting, several tools have been created, including WPPT, Predictor, Zephyr, Ewind, WPFS Ver1.0, and AWPPS. A number of case studies in Spain, Germany, Denmark, Ireland, Greece, and France have used these models [15, 16].

The Wind Power Prediction Tool is a well-known model with a wide range of applications for this time frame (WPPT). It can be used to generate short-term (say, up to 120 hours, or 36 hours) wind power output projections. Because the system can provide prediction values as a total including not only a single wind farm, but also a region, it is extremely flexible. The system also gives accurate estimates of the tools' uncertainty, which is critical for efficient scheduling or trading. WPPT uses advanced nonlinear techniques.

Because it may produce prediction values as a total spanning not just a particular wind farm, but also a region, the technique is extremely versatile. The system also provides accurate estimations of the tools' uncertainty, which is critical for optimum trading or scheduling. Advanced nonlinear statistical models underpin WPPT. A semi-parametric power curve model for wind farms that take both wind speed and direction into account, as well as dynamical forecasting models that describe the dynamics of wind power and any diurnal variations, are among the models included in the package. Self-calibrating and self-adaptive models have been developed.

As a result, they update parameters automatically in response to changes in the number of turbines and their features, the environment, the NWP models, and non-explicit model attributes like roughness and filthy blades. WPPT can automatically calibrate to the observed circumstances using artificial intelligence [15]. The system requires online wind power measurements in its simplest configuration. However, the following data is taken into account depending on the configuration: Wind power measurements are now available online. Energy readings from all (or almost all) turbines in a region aggregated (for regional forecasting). Wind speed and direction forecasts by meteorologists for wind farms and regions.

Other measurements or predictions, such as local wind speed, stability, and the number of active turbines, are available. Prediktor, a tool developed by the meteorology research program, is another useful tool (MET). Unlike WPPT, however, Prediktor's main goal is to represent as much as possible using physical models. Every 6 hours, the system provides the predicted production of wind farms for up to 48 hours. All it requires is online access to NWP model output.

The basic processes are as follows: a NWP model predicts overall weather patterns. Only the entire wind can be predicted by such a model, and only correct forecasts can be made at a given site. Then, if needed, these projections are tailored. The WAsP model tailors the wind turbines to each other by modeling local characteristics such as roughness, horography (ridges and hills), and obstructions, as well as the influence of the wind turbines on each other.

Since no model can simulate nature perfectly, two MOS (model output statistics) filters are used in Prediktor to correct shortcomings. The wind power observed is used to adjust the parameters of these filters. The final output of the model is the expected production of the wind farm every 3 hours over the next 48 hours. Furthermore, Prediktor forecasts or will forecast in the near future for up to 50 wind farms in Ireland, Denmark, Germany, France, and Spain in 2025 [15].

The AWPPS is the only instrument available that estimates confidence intervals for wind power predictions at a predetermined level of certainty (i.e. 85 percent, 90 percent, and 95 percent). The intervals are generated using an important international dedicated to the problem of wind prediction. The Prediction Risk Module allows to forecast uncertainty for the next 24 hours based on projected weather stability. Furthermore, the online use of this module allows for the development of appropriate techniques for optimizing the value of power forecasts [17, 18].

A general overview of wind forecasting models is presented in **Table 2**. This section is divided into three parts based on the time-scales, and for each of them and its applications.

• Immediate short-term forecasting Models

Medium-term forecasts (from 6 hours up to a day) are used to make decisions for switching the turbine on or off for safety or conditions on the market.

WPMS has been adapted for performance in the ICT settings of various grid operators and carriers of major wind parks, as one prominent example of immediateshort-term wind forecasting [20].

WPMS deployed artificial neural networks (ANN) in wind farms that were trained using a large amount of historical data. A preprocessor translated input data, output data measured in wind farms, and forecasted meteorological parameters into XMLformat before being sent to the program core, which consists of prediction and transformation modules.

• Long-term forecasting

Long-term wind forecasting methods have been studied in a few researches. And there aren't many prediction tools available for this timeframe. Simple models can no


**Table 2.**

*Time-scale classification for wind forecasting [19].*

#### *Wind Power Forecasting Models DOI: http://dx.doi.org/10.5772/intechopen.103034*

longer match the criteria because to the extended ahead-forecasting time, hence NWP or hybrid NWP models are being investigated. Modern wind power forecasting methods, which are typically based on NWP, provide forecasts over a time range of up to several days. To put it another way, the NWP is the source of all information about the future of wind forecasting.

The national weather service or private weather data provider supplies a collection of NWP data that can be used to predict wind speed and power. In the future, it is becoming more common to use NWP for long-term forecasting [21]. Previento is comparable to Prediktor, but it utilizes more severe physical downscaling and specific upscaling techniques. It provides a reliable forecast of projected wind power for any locations and regions in Germany, Europe, and the rest of the world up to 10 days ahead of time, with a temporal resolution of up to 15 minutes. The wind power forecast is based on the best possible mix of meteorological models, as well as the local conditions of the wind farm's surrounds and the NWP [22].

The Previento system involves a physical approach with data from a large-scale weather prediction model, such as the German Weather Service's Lokalmodell. It simulates roughness, horography, and wake effects in the boundary layer. The daily variation of the thermal stratification of the atmosphere, which is employed to adjust the logarithmic profile, is critical for calculating wind speed at hub height. The expected power output for single sites is derived using the turbine's particular power characteristic. The total amount of power generated by wind in a certain region is computed using data from chosen wind farms.

For long-term planning, long-term forecasts (from a day to a week or even a year) are utilized (to schedule the maintenance or unit commitment, optimize the cost of operation). Maintenance of offshore wind farms can be extremely costly, thus proper planning of maintenance activities is essential. Wind power predictions have a temporal resolution of 10 minutes to a few hours (depending on the forecast length). Wind power forecasting improvements are concentrating on using additional data as input to the models involved, as well as offering uncertainty estimates alongside the standard predictions.

Wind forecasting schemes as **Figure 3** can also be classified based on their methodology into many categories:

#### **Figure 3.**

*The conceptual mind on wind energy prediction.*

#### **5.1 Physical approach to wind power forecasting**

Approaching the situation physically (deterministic approach), the physical approach, also known as the deterministic technique, is based on weather forecast data such as temperature, pressure, surface roughness, and obstructions in the lower atmosphere, or numerical weather prediction (NWP).

Established several physical models based on weather data to predict wind speed and estimated wind power [23]. Physical models often rely on global databases of meteorological information or atmospheric mesoscale models, but to provide accurate results, they require massive computer systems [24].

To estimate wind power production, the physical method uses a thorough description of the lower atmosphere. Cellura et al. [23] provide an overview of some of the neural, geostatistical, and hybrid models that have been applied in space-temporal wind forecasting. Dynamic models (also known as prognostic) and kinematic models (also known as diagnostic) are the two main forms of numerical codes for wind field modeling across rugged terrain [25, 26]. The momentum and energy equations are not explicitly solved in these models; instead, parametric relations and/or wind data are used to examine them implicitly [27].

To account for the local circumstances of the physical topography, computational fluid dynamics (CFD) is utilized as an alternative to the power law [28]. Model output statistics (MOS) are frequently employed to reduce systematic forecasting mistakes and to compensate for unknowns in the expected power output [29].

Forecasts are provided at specified nodes on a grid that covers a certain area. Due to the fact that wind farms are not located on these nodes, these estimates must be extrapolated to the required location and turbine hub height. Physical-based forecasting methods are comprised of multiple sub-models that work together to translate wind forecasts at various grid points and model levels to power forecasts at the actual site.

Converting wind speed to power at the level of the wind farm and at hub height depending on the using theoretical power curves supplied by the wind turbine manufacturer. However, since multiple studies have demonstrated a preference for empirically obtained power curves over theoretical ones; theoretical power curves are becoming less and less important. When using a physical methodology, the function that calculates wind generation from NWPs at various locations around the wind farm is modeled once and for all. The calculated transfer function is then applied to the current weather predictions. Physical simulations frequently integrate Model Output Statistics (MOS) for post-processing power forecasts to account for systematic forecasting errors that may be due to the NWP model or modeling approach, **Figure 4** shows steps forecasting wind farm with NPW.

#### **5.2 Statistical approach to wind power forecasting**

Statistical approach statistical method is based on the vast amount of historical data without considering meteorological conditions. It usually involved artificial intelligence (neural networks, neuron-fuzzy networks) and time series analysis approaches [30, 31]. Statistical models, the set of models includes a semi-parametric power curve model for wind farms taking into account both wind speed and direction, and dynamical forecasting models describing the dynamics of the wind power and any weather variation, etc.

Statistical forecasting approaches are based on one or more models that establish the relationship between historical power values, historical and future values of

#### **Figure 4.** *Steps forecasting wind farm with NPW.*

meteorological variables, and wind power measurements. The physical events are not deconstructed and accounted for, despite the fact that problem expertise is required for selecting the appropriate meteorological variables and developing appropriate models.

Model parameters are calculated using a collection of previously known data, and they are updated on a frequent basis during online operation to account for any new information that becomes available (i.e. meteorological forecasts and power measurements).

Linear and nonlinear statistical models, as well as structural and black-box models, are all examples of statistical models. Structural models rely on the analyst's knowledge of the phenomenon of interest, whereas black-box models are built from data in a fairly mechanical manner and require little subject-matter knowledge.

Structural models for wind power forecasting would include diurnal wind speed changes modeling or an explicit function of meteorological variable predictions. Neural-Networks (NNs) and Support Vector Machines are examples of black-box models (SVMs). Some models, on the other hand, are 'in-between' the extremes of being entirely structural or completely black-box. Expert systems, for example, learn from experience (from a dataset) and can be programmed with prior information. The subject of gray-box modeling is then discussed.

Statistical models are often made up of two parts: an autoregressive portion for capturing the wind's persistent behavior, and a "meteorological" part for nonlinear transformation of meteorological variable projections. The autoregressive component provides for considerable gains in forecast accuracy across horizons up to 6–10 hours

ahead, when the use of meteorological forecast information alone may not be adequate to exceed persistence.

Statistical approaches to wind power prediction are currently focusing on the use of multiple meteorological forecasts as input and forecast combination, as well as the best use of spatially distributed measurement data for prediction error correction or issuing warnings on potentially large uncertainty.

Calculate a statistical relationship between the essential input data and the generation of wind energy. They entail utilizing a statistical model to directly turn the input factors into wind generation. With these models, a one-step direct calculation of wind power from input parameters is achievable. Most data mining-based models (e.g., ANN, SVM, fuzzy model, model trees), as well as time series analysis methods, can be used as output models (e.g. ARIMA, fractional ARIMA).

A massive quantity of data is processed in the statistical technique, and meteorological processes are not clearly represented. The relationship between historical power output and weather is established, and this information is then used to anticipate future power output. Statistical methods, unlike physical methods, simply require one step to convert input variables to power output. As a result, the procedures used are referred to as "black box." In most cases, a statistical relationship is established between the weather forecast or projection and the wind farm's prospective power output. Other statistical approaches employed include the Box-Jenkins methodology, the use of the Kalman filter, and the use of autoregressive (AR), moving average (MA), autoregressive moving average model (ARMA), and autoregressive integrated moving average model (ARIMA).

Torres et al. [30] discovered that compared to persistence, it was possible to get a 20% error reduction when forecasting average hourly wind speed for a 10 h forecast horizon at a number of locations using nine variables.

Classical time series analysis is not the only approach to model a statistical relationship between data points. Artificial neural networks (ANN) and fuzzy systems are the most common soft computing (or machine learning) techniques utilized, however other models such as gray predictors and support vector machines (SVM) have also been used. Artificial intelligence (AI) approaches are another term for learning approaches. They're known as learning techniques since they take historical time series to learn about the relationship between projected wind and predicted power production. They've been dubbed "gray box" approaches in recent years.

#### *5.2.1 Parametric methods*

The presentation of parametric statistical methods directly inspired from the physical equation. Parametric modeling according to the wind speed only, the investigated the simplest parametric models, namely linear regression and logistic regression, with the wind speed as the unique explanatory variable. If the predicted power at time t is denoted by Yˆt, these models are given by

$$\mathbf{Y}^{\star}\mathbf{t} = \mathbf{a}\_0 + \mathbf{a}\_1 \mathbf{W}\_{\mathbf{t}}, \text{ and } \tag{1}$$

$$\mathbf{Y}^{\star}\mathbf{t} = \mathbf{C}\_{1} + \exp\left(\mathbf{a}\_{0} + \mathbf{a}\_{1}\mathbf{W}\_{t}\right),\tag{2}$$

where the parameters a0, a1, C are estimated using the associated methodology.

#### *5.2.2 Logistic regression*

Logistic regression has also been considered to mimic more closely Eq. (1). More precisely, the model is then defined by:

$$\mathbf{Y^\uparrow t} = \mathbf{C} / \left(\mathbf{1} + \exp\left(\mathbf{a\_0} + \mathbf{a\_1}\mathbf{W\_t} + \mathbf{a\_2}\mathbf{W\_t}^2 + \mathbf{a\_3}\mathbf{W\_t}^3\right)\right),\tag{3}$$

where ai, i = 0, … , 3 and C are estimated parameters.

This model is using not only wind speed as a predictor, but also wind direction, (coded by its cosine and sine: Dcos and Dsin), temperature T, and the variances of the wind speed WS and direction, DS, Re and DS, Im.

#### *5.2.3 Lasso model*

The Lasso method, which simultaneously performs variable selection and regularization through the least squares criterion penalized by the ` 1 norm of the regression coefficients has been investigated as well (see for instance [21]). The model is defined by.

$$\mathbf{Y}^{\star}\mathbf{t} = \mathbf{a}\mathbf{0} + \mathbf{a}\_1\mathbf{W}^{\mathrm{t}} + \mathbf{a}\_2\mathbf{D}\_{\mathrm{t}}^{\mathrm{cos}} + \mathbf{a}\_3\mathbf{D}\_{\mathrm{t}}^{\mathrm{sin}} + \mathbf{a}\_4\mathbf{T}\mathbf{t} + \mathbf{a}\_5\mathbf{W}\mathbf{S}\_{\mathrm{t}} + \mathbf{a}\_6\mathbf{D}\_{\mathrm{t}}^{\mathrm{S,Re}} + \mathbf{a}\_7\mathbf{D}\_{\mathrm{t}}^{\mathrm{S,Im}}, \quad \text{(4)}$$

with a0, … , a7 minimizing.

$$\begin{aligned} \mathbf{1} & \text{1/n} \sum\_{\mathbf{n}}^{i=1} \left( \mathbf{Y}\_{\mathbf{i}} - \mathbf{a}\_{0} - \mathbf{a}\_{1} \mathbf{W}\_{\mathbf{i}} - \mathbf{a}\_{2} \mathbf{D}\_{\mathbf{i}}^{\text{cos}} - \mathbf{a}\_{3} \mathbf{D}\_{\mathbf{i}}^{\text{sin}} - \mathbf{a}\_{4} \mathbf{T}\_{\mathbf{i}} - \mathbf{a}\_{5} \mathbf{W}\_{\mathbf{i}}^{\text{S}} - \mathbf{a}\_{6} \mathbf{D}\_{\mathbf{i}}^{\text{S,Re}} - \mathbf{a}\_{7} \mathbf{D}\_{\mathbf{i}}^{\text{S,Im}} \right)^{2} \\ & + \lambda \sum\_{\mathbf{j}=1}^{7} |\mathbf{aj}|. \end{aligned} \tag{5}$$

#### **5.3 Hybrid approach energy power forecasting**

Hybrid method, which combines physical methods and statistical methods particularly uses weather forecasts and time series analysis.

ANEMOS is a hybrid wind forecast tool that takes into account a variety of time horizons. The development of combining high-resolution meteorological predictions and appropriate prediction models for the offshore is emphasized [19, 31].

Hybrid models aim to combine the advantages of each model in order to produce the best predicting results possible. Because the information provided in individual forecasting techniques is restricted, a hybrid approach can take use of the available data, integrate individual model data, and maximize the benefits of many forecasting methods, improving prediction accuracy [32].

Many techniques, such as mixing physical and statistical procedures or short-term and medium-term models, are included in hybrid methods. A number of hybrid models were utilized to anticipate wind power. Here are some examples of potential combinations:


Global Forecasting System (GFS) with the Weather Research and Forecasting (WRF) system. **Figure 5** shows an example of an ANN structure with 4 inputs and 2 hidden layer.

Shi et al. [34] proposed two hybrid models for wind speed and power forecasting: ARIMA-ANN and ARIMA-SVM. Based on two case studies on wind speed and wind power generation, this research analyses the application of the suggested hybrid models in a systematic and thorough manner. The findings imply that hybrid approaches are feasible alternatives for predicting both wind speed and wind power generation time series, but that they do not always provide better forecasting performance for all forecasting time horizons investigated.

Guo et al. [35] proposed a novel hybrid wind speed forecasting method based on a back propagation neural network and the notion of seasonal exponential adjustment to exclude seasonal effects from real wind speed datasets. A proposed technique outperformed the single back propagation neural network in the tests.

For short-term wind power forecasting in Portugal, Catalo et al. [19] presented a hybrid approach based on the combination of ANN and wavelet transform. To deconstruct the wind power series into a set of better-behaved constituent series, the wavelet transform is applied. The test findings show that the proposed hybrid technique for forecasting wind output has a lot of potential.

Finally, hybrid models (e.g. [19, 36]) are based on the combination of the physical and statistical models, the combination of models with several time horizons, and the combination of alternative statistical models

### **5.4 Spatial correlation models**

The spatial correlation models take into account the spatial link between wind speeds at different sites. The wind speed time-series of the projected point and its neighbors is used to predict the wind speed in spatial correlation models [36]. When predicting wind speed at one location based on observations taken at another, a spatial correlation model is used. Data obtained over a seven-year period [37] was used to test its behavior and provide adequate verification.

Based on cross-correlation at surrounding sites, Alexiadis et al. [38] demonstrated a technique for forecasting wind speed and power output up to several hours ahead.

**Figure 5.** *ANN structure with 4 inputs and 2 hidden layer.*

This research established an ANN technique based on spatial correlation models that outperform the persistence forecasting model in terms of forecasting accuracy [39].

Barbounis and Theocharis [40] proposed the use of local feedback dynamic fuzzy neural network (LF-DFNN) to forecast wind speed using spatial correlation. Remote meteorological stations are installed at two reference sites in accordance with the location of the base site so that the three sites are aligned along the prevailing wind direction. Using spatial information from remote meteorological stations, the LF-DFNN is used in this paper to predict multi-step forward wind speed in the base site. The LF-DFNN outperforms other network models tested in this application, according to simulation data.

#### **5.5 Artificial intelligence methods**

Various novel AI algorithms for wind speed and power prediction have recently been developed as a result of the advancement of artificial intelligence (AI). Artificial neural networks (ANN), adaptive neuro-fuzzy inference system (ANFIS), fuzzy logic approaches, support vector machine (SVM), neuro-fuzzy network, and evolutionary optimization algorithms are among the newly developed methodologies.

Through the training process, ANN models can represent a complex nonlinear relationship and extract the dependency between variables [40]. Back propagation neural networks, recurrent neural networks, radial basis function (RBF) neural networks, ridgelet neural networks, and adaptive linear element neural networks are examples of ANN-based methods. The application of an ANN-based method to the problem of wind power forecasting is appropriate.

ANN might handle nonlinear and complex scenarios in terms of categorization or forecasting. ANN models can depict a complex nonlinear relationship and extract the link between variables through the training phase [40]. Examples of ANN-based techniques include back propagation neural networks, recurrent neural networks, radial basis function (RBF) neural networks, ridgelet neural networks, and adaptive linear element neural networks. It appears that applying an ANN-based technique to the problem of wind power forecasting is a good idea.

Using time series analysis, Sfetsos [41] proposed an ANN technique for forecasting mean hourly wind speed data. The proposed methodology also has a benefit for utilities that have a high level of wind penetration and utilize hourly intervals for power system operational procedures like economic dispatch and unit commitment.

Chang [42] discussed back propagation neural network-based wind power forecasting algorithms. The created model for short-term wind forecasting demonstrated excellent accuracy when utilized to supply energy to a 2400 kW (WECS) on the Taichung coast. Back propagation neural networks and recurrent neural networks were used in More and Deo's [43] wind forecasting methodology. Traditional statistical time series analysis has been found to be less accurate than neural network forecasting [44].

Chang [45] described a method for forecasting wind power generation time series using an RBF neural network. The numerical results show that the suggested forecasting method is accurate and dependable, with good matches between realistic values and predicting values.

Guo et al. [46] studied a feed-forward neural network (FNN) wind forecasting approach based on modified empirical mode decomposition (EMD). Through multistep forecasting of mean monthly and daily wind speeds in Zhangye, China, the proposed technique outperforms basic FNN and unmodified EMD-based FNN [47].

Li and Shi [48] used three types of conventional ANNs to anticipate wind speed: adaptive linear element, back propagation, and radial basis function.

The outcomes of comparing three types of ANN reveal that no single ANN model outperforms another universally in terms of all evaluation measures, even for the same wind dataset. Furthermore, the type of ANN to use for the best results is determined by the data sources.

Yang et al. [49] proposed an ANFIS approach for interpolating missing and incorrect wind data. Twelve measured wind data sets from a wind farm in North China are interpolated and examined for performance testing. The ANFIS method's effectiveness was demonstrated by the test results. A SVM-based technique for wind power forecasting was described by Zeng and Qiao [32]. Real wind speed and wind power data obtained from the National Renewable Energy Laboratory are used in simulation research.

The suggested SVM method outperforms the persistence model and the RBF neural network-based model, according to the results. For one-step ahead, wind speed forecasting, Zhou et al. [50] described a systematic investigation on fine-tuning least-squares support vector machines (LSSVM) model parameters. Three SVM kernels are implemented: linear, Gaussian, and polynomial kernels. LSSVM approaches are proven to outperform the persistence model in the vast majority of scenarios. For short-term wind power forecasting, Xia et al. [51] introduced a neuro-fuzzy network technique.

For the wind power forecasting of a practical wind farm in China, the forecasting approach is used. The results of the tests revealed that the trained neuro-fuzzy networks are capable of predicting and forecasting wind power.

Jursa and Rohrig [52] proposed a new short-term prediction technique based on the automated specification of neural networks and the nearest **neighbor** search using evolutionary optimization algorithms. The test results demonstrated that by employing the proposed automated specification method, the wind power forecast error can be decreased.

### **6. Conclusion and future advances for wind power prediction**

Wind Forecasting in the Future The forecast accuracy of wind power prediction systems is becoming increasingly significant due to the high penetration of wind power in the energy grid. Many academics have been working on wind power forecasting in recent years. Forecast accuracy has steadily increased and intensive research and development efforts are projected to be underway soon. In order to improve wind power projections even more, various literature [33, 53] suggest that future studies should focus on the following areas:

