The accurate assessment of wind power potential requires not only the detailed knowledge of the local wind resource but also an equivalent power curve with good effect for a local wind farm. Although the probability distribution functions (pdfs) of the wind speed are commonly used, their seemingly good performance for distribution may not always translate into an accurate assessment of power generation. This paper contributes to the development of wind power assessment based on the wind speed simulation of weather research and forecasting (WRF) and two improved power curve modeling methods. These approaches are improvements on the power curve modeling that is originally fitted by the single layer feedforward neural network (SLFN) in this paper; in addition, a data quality check and outlier detection technique and the directional curve modeling method are adopted to effectively enhance the original model performance. The proposed two methods, named WRFSLFNOD and WRFSLFNWD, are able to avoid the interference from abnormal output and the directional effect of local wind speed during the power curve modeling process. The data examined are from three stations in northern China; the simulation indicates that the two developed methods have strong abilities to provide a more accurate assessment of the wind power potential compared with the original methods.
Currently, for both developed and developing countries, the heavy dependence on fossil fuels has caused serious environmental problems, such as atmospheric pollution and soil and water contaminations. The threat of global warming and the rapid depletion of nonrenewable energy resources are now driving one of the greatest transitions in the energy field in the history of human civilization. Consequently, renewable energy is considered to be the most promising alternative energy resource because it plays a significant role in securing the longterm sustainable energy supply and reducing global greenhouse gases emissions [
Wind energy is characterized by strong instability and intermittency, mainly as a result of the atmospheric circulation, making it difficult to estimate the generated power that is needed to be injected into the grid and also causing difficulties for energy transportation [
Researchers have exerted great effort toward the wind power assessment of a local site, which can be categorized into two major groups when considering the application of the stageconstruction wind farms [
Although the power output from wind turbines (WTs) varies with the cube of the wind speed at the hub height, the wind speed is not constant with time, which makes it difficult to evaluate the accurate power output. To effectively evaluate the wind power available for a particular site, the statistical analysis, which concerns the use of various probability density functions (pdfs) to describe wind speed frequency distributions, is generally used. The pdf of wind speed distribution indicates how often wind of different speeds will occur at a local site with a certain average wind speed; there are different distribution functions for determining the wind energy potential, including Weibull [
As is well known, the WT converts the energy of wind into electric power output at the grid connection interface, which depends on the integrated view of the complex aerodynamic, mechanic, electromagnetic, and control aspects; there is a direct connection between the wind resource and the power output. To accurately describe the characteristics of windrelated power generation and to accurately evaluate the wind energy potential, the link between the wind speed and the power produced by a wind generator is given by the socalled power curve, typically provided by the WT manufacturer. In general, a power curve describes the power delivered by a WT by representing the turbine power output as a function of the wind speed at hub height. With such a curve, the power output from a WT can be estimated without the detailed knowledge of turbine operations and its control schemes [
The power curve of a WT is obtained by the manufacturers from the field measurements of wind speed and power and partly from the environmental values (temperature, pressure, and relative humidity) [
The major highlight of this paper is its development of a wind power assessment based on the wind speed outputs of the numerical weather models and the developed power curve modeling methods that consider the artificial intelligence (AI) algorithm, the data quality check, and the outlier detection technique for the turbine outputs and the enhanced power curve modeling according to wind directions. Specifically, as the current generation “community” physicsbased atmospheric model, the weather research and forecasting (WRF) model is now the widely used mesoscale system serving both the operational forecasting and atmospheric research needs. In this paper, the wind speed outputs from the WRF simulation are used for future power assessment; calculations indicate that the WRFbased power assessment is far more accurate than the evaluations from the basic wind speed distributions when considering the total power production of a local farm. At the first step, the single layer feedforward neural network (SLFN) is employed for power curve modeling by learning the input/output transition through an AIbased nonlinear mapping. The SLFN is chosen mainly because of its strengths in capturing the complex nonlinear input/output relations identical to the transition from wind speed to the windgenerated power output; it enables the WRFSLFN method to be far more effective and accurate than the original methods.
Next, two improvements based on the original WRFSLFN are proposed to provide better assessment results. The first, named WRFSLFNOD, is a mixture of not only the WRF wind speed simulation and SLFN structure but also an additional data quality check and outlier detection technique. It enables the assessment process to first check and eliminate the abnormal records within the raw data set; then, the cleaned data set is regarded as the input of the original WRFSLFN process. The second, named WRFSLFNWD, contains a preanalysis of the wind direction distribution and its influence on the power output, before the curve modeling process. Wind speeds are grouped into twelve directional sectors, and the most frequent wind speeds always carry the primary and most effective information of the local wind resource, considering the power potential assessment. The simulation indicates that both the WRFSLFNOD and WRFSLFNWD methods can be significant enhancements for the assessment problem.
To validate the effectiveness and accuracy of the proposed methods according to the wind power potential assessment problem in this paper, the available wind speed and corresponding power output records are collected from three sites in northern China (the site description is displayed in Figure
The rest of this paper is organized as follows. Section
This section reviews the existing works in the literature that are related to the wind power potential assessment based on the measured records of wind speed (see Figure
The commonly used wind speed distributions.
Distribution fitting according to the three stations.
The analysis of the wind speed distribution at hub heights is commonly used when assessing the wind potentials at a proposed site. In the bibliography, the wind speed frequency distribution was represented by various probability density functions; the most common are Weibull, Rayleigh, Gamma, and Lognormal distributions; this paper also considers these commonly used functions.
The Weibull distribution is one of the most important and widely used frequency functions in the study of wind climate and wind energy [
The Rayleigh distribution is another case of the Weibull distribution when setting the shape parameter as
The pdf of Gamma distribution is as follows:
The pdf of Lognormal distribution is defined by
Figure
The wind turbine is the direct transducer that delivers wind into the electric power output. The power curve is the most common representation of the relation between the wind speed and the power production, which is defined as
There are three key points on the power curve: (i) the cutin speed, below which the turbine will not produce power, (ii) the rated speed, at which the rated power of the turbine is produced, and (iii) the cutoff speed, beyond which the turbine is not allowed to deliver power. Then, the power produced in a given period can be calculated by
For the future potential assessment of windgenerated power output, this paper considers a combination of the atmospheric numerical simulation and the WT power curve. Specifically, the WRF model is chosen to provide a physical prediction of the future wind speed; then, the assessment of power potential can be obtained through the power curve modeling.
The WRF model is now the current generation “community” physicsbased atmospheric model, serving both the operational forecasting and atmospheric research needs; the WRF model has now become one of the most popular and widely used tools for numeric weather prediction. In the WRF model, a grid is a set of threedimensional points in space containing weather data, such as wind speed and atmospheric pressure. Each grid has a current time and an associated stop time. The WRF simulates the atmosphere using physics calculations based on the grid data and a specific physics model; then, the current time of the grid is advanced by a unit of time called a timestep [
In this paper, the wind speed from the WRF outputs will be used for future power assessment with the proposed power curve modeling strategies. The physical options of the WRF model are described in Table
Model configuration of the WRF simulation.
Physical options  

Cumulus parameterization  Grell 3D ensemble cumulus scheme 
Shortwave radiation  RRTM scheme 
Longwave radiation  Dudhia scheme 
Surface layer physics  Eta similarity 
Land surface processes  Fractional seaice 
Planetary boundary layer  MellorYamadaJanjic scheme 
After acquiring the wind speed outputs from the WRF simulation process, the WT power curve can be employed to transform the wind speed into power generation; thus, the future wind power potential of the chosen promising site can be obtained and can be used for further construction and operation of the wind farm.
Different from the power assessment based on the historical measurements discussed in Section
Power curve modeling is the most straightforward approach to accurately match the wind speed and the actual power output according to a specific condition of power generation. A single turbine power curve is determined by measuring the turbine output and inflow wind speed at the hub height. While considering the power potential assessment for a wind farm, the power curves are adversely affected by the wind farm layout and the changing environmental and topographical conditions. In fact, mainly due to the wakes created by the wind turbines upstream, the wind speed reduces the efficiency of the turbine array. Therefore, the equivalent power curve modeling is of great significance, incorporating the effect of the array efficiency, the high wind speed cutout, the topographic effect spatial averaging, the availability, and the electrical losses [
The neural network (NN) technique is an informationprocessing model simulating the operation of the biological nervous system, which is widely used to model complex functions for various applications. An NN model consists of interconnected groups of artificial neurons that emulate the function of neuron cells; therefore, it is able to identify the complicated pattern within a certain data set [
The chosen neural network in this paper is the SLFN model, which is considered as a powerful and effective network frame to address complex problems, such as the power curve modeling problem.
Consider a set of
Then, the mean square error (MSE) of the SLFN simulation according to the given data can be defined as
In this paper, the relative error (RE) is selected for accuracy evaluation according to the equivalent power curve modeling; the definition of RE is as follows:
This section employs the SLFN to model the complex relationship between wind speed and actual power output, which is usually mismatched according to the WT power curve provided by the manufacturer. Figures
Original curve fitting result of station 1.
Original curve fitting result of station 2.
Original curve fitting result of station 3.
Therefore, the SLFN is regarded as a powerful tool to learn the transition relation. It is clear that the SLFNfitted power curve is complicated and is unlike the form of the power curve provided by the manufacturer. When the wind speed value is relatively small, the SLFNfitted power curve exhibits a steady upward trend as the wind speed increases. For the larger wind speeds, the fitting result is complex and is obviously somewhat lower than the normal power outputs. This may result from the abnormal outputs that existed in the data sets, which have a large wind speed but quite low power output; a detailed discussion about abnormal outputs will be provided in the next section.
Traditional power assessment is usually based on the frequency distribution of the wind speed, and it then transforms the wind speed into the power output. A significant undertaking is to compare the performance among the developed WRFSLFN method with the methods according to a certain pdf of the wind speed. Figure
Performance comparison between the WRFSLFN and the four pdfSLFN methods.
RE of wind power assessment  

0–10 MW  10–20 MW  20–30 MW  30–40 MW  40–50 MW  Total  
Site 1  WeibullSLFN  0.22  0.35  0.73  1.06  −0.41  −0.28 
RayleighSLFN  0.38  0.71  0.75  0.61  −0.66  −0.23  
GammaSLFN  0.54  0.49  0.46  0.32  −0.53  −0.15  
LognormalSLFN  0.04  0.66  1.51  1.36  −0.19  −0.56  
WRFSLFN  0.40  0.99  0.56  0.05  −0.86  −0.09  


0–20 MW  20–40 MW  40–60 MW  60–80 MW  80–100 MW  Total  


Site 2  WeibullSLFN  0.40  0.50  0.32  1.27  −1.00  −0.19 
RayleighSLFN  0.40  0.81  0.59  0.71  −1.00  −0.19  
GammaSLFN  0.71  0.65  0.23  0.41  −0.70  −0.14  
LognormalSLFN  0.13  0.70  0.86  1.30  −0.81  −0.37  
WRFSLFN  0.44  1.07  0.46  −0.03  −0.85  −0.10  


0–10 MW  10–20 MW  20–30 MW  30–40 MW  40–50 MW  Total  


Site 3  WeibullSLFN  0.62  0.34  1.15  0.99  −1.00  −0.20 
RayleighSLFN  0.72  0.63  1.40  0.42  −1.00  −0.18  
GammaSLFN  0.77  0.54  1.29  0.37  −1.00  −0.14  
LognormalSLFN  0.29  0.56  2.10  1.07  −0.89  −0.40  
WRFSLFN  0.76  1.02  0.60  0.57  −0.85  −0.20 
Comparison between the WRFSLFN and the assessments based on wind speed distributions.
The WRFSLFN approach performs much better compared with the other four pdfSLFN methods, which was introduced in the above text; the developed method also has its weakness, mainly due to the use of unfiltered raw data sets.
It is clear that the wind plant power curve may have a general shape that is similar to the power curve of a single turbine. Figures
Abnormal power output within the raw data sets.
The data quality check and outlier detection technique adopted in this paper was introduced in [
The key concept of this method is to find an optimal
During the power curve modeling process addressed by the WRFSLFN method, it is demonstrated that using the entire data sets is not necessary and will not improve the model performance. As a consequence, this section displays the power curve modeling result by using an improved methodology that combines the SLFN with the outlier detection technique; this improved power curve modeling method is named the SLFNOD. During the SLFNOD process, the data quality check and outlier detection technique introduced in Section
Curve fitting using the cleaned data from station 1.
Curve fitting using the cleaned data from station 2.
Curve fitting using the cleaned data from station 3.
Table
Performance comparison between the WRFSLFN and WRFSLFNOD methods.
RE of wind power assessment  

0–10 MW  10–20 MW  20–30 MW  30–40 MW  40–50 MW  Total  
Site 1  WRFSLFN  0.40  0.99  0.56  0.05  −0.86  −0.09 
WRFSLFNOD  −0.48  0.24  0.41  0.01  −0.01  −0.07  


0–20 MW  20–40 MW  40–60 MW  60–80 MW  80–100 MW  Total  


Site 2  WRFSLFN  0.44  1.07  0.46  −0.03  −0.85  −0.10 
WRFSLFNOD  −1.00  −0.83  −0.21  0.31  0.11  −0.07  


0–10 MW  10–20 MW  20–30 MW  30–40 MW  40–50 MW  Total  


Site 3  WRFSLFN  0.76  1.02  0.60  0.57  −0.85  −0.20 
WRFSLFNOD  −0.61  0.28  0.45  0.12  −0.15  −0.06 
During the abovementioned modeling processes, discussions on wind speed only focused on the numerical value; however, there is another important characteristic, wind direction, which will be quite helpful in windrelated modeling such as the wind power assessment discussed in this paper.
Considering a specific local site, wind direction has its distinctive distribution and characteristic. Exemplified by the data collected from station 1, Figure
Wind rose of station 1 and the division of the wind direction.
The layout of turbines at a wind farm is always designed to take advantage of the prevailing wind directions because the wind from other directions may not produce as much power as the wind from the prevailing direction as a result of the wake effect. Thus, what is the relationship between the directional wind speed and the actual power output? Figure
Power output performance according to the different wind direction sectors, exemplified by station 1.
Wind rose of stations 2 and 3.
The above discussion indicates that the wind farm production varies under the different wind directions; the wind direction is also a significant factor for an accurate estimation of the wind power output. This section develops an improved FLFNhandled power curve modeling with the wind direction factor; it is designated the SLFNWD. Figures
Curve fitting using the data in the prevailing direction of station 1.
Curve fitting using the data in the prevailing direction of station 2.
Curve fitting using the data in the prevailing direction of station 3.
The performance of the WRFSLFNWD is represented in Figure
Performance of power curve modeling considering the wind direction.
To build up wind farms, it is essential to perform an accurate assessment of the wind energy potential at the promising sites because it is a necessary and crucial step both in the preevaluation for the site selection and in the further planning of the project. As one of the most important factors during the assessment process, the knowledge of wind power transformation, from the wind speed to the windgenerated power output, is required, but it is not an easy task. The statistical analysis, concerning the use of the various pdfs to describe the wind speed frequency distributions, is generally used. However, their performance is expected to be enhanced in the actual applications.
This original method of wind power assessment is a combination of the WRF wind speed simulation and the SLFN algorithm, due to the recognized strengths of both methods. In most cases, the WRFSLFN method demonstrates good performance compared with the pdfbased power assessment. Next, two improvements are proposed through the aid of the abovementioned data quality check and the outlier detection technique and the directional power curve modeling method, named the WRFSLFNOD and the WRFSLFNWD. Specifically, the first improved method contains a data filtering process that eliminates the abnormal outputs within the raw data set and uses the cleaned set as the input of the original SLFNfitting structure. The WRFSLFNOD method leads to an RE reduction of 40.74%, which is the mean value from the simulation of all three stations compared with the original RE. Then, the WRFSLFNWD method includes a data analysis according to the twelve directional sectors of the local wind speeds; the data with the prevailing directional wind speeds is chosen for further power curve modeling and power potential assessment. The result indicates that power curve modeling with unselected raw data may introduce additional interferences instead of leading to an error reduction; thus, both of the proposed methods are of great significance to filter the effective information of the power transformation and to provide an accurate assessment of the local wind power potential.
The authors declare that there is no conflict of interests regarding the publication of this paper.
This work was supported by CAS Strategic Priority Research Program Grant no. XDA05110305.