Renewable Energy 115 (2018) 448e461

Contents lists available at ScienceDirect

Renewable Energy
journal homepage: www.elsevier.com/locate/renene

Statistical analysis of wind energy characteristics in Santiago island,

Cape Verde
Xiangyun Qing
Key Laboratory of Advanced Control and Optimization for Chemical Processes (East China University of Science and Technology), Ministry of Education, East
China University of Science and Technology, Shanghai, China

a r t i c l e i n f o a b s t r a c t

Article history: As a volcanic archipelago, the Republic of Cape Verde relies dominantly on diesel to power its electricity
Received 7 November 2016 supply. Recognizing the financial and environmental burden of diesel generation and risk of energy
Received in revised form security, the government of Cape Verde has launched an ambitious goal of 50% electricity from re-
17 August 2017
newables by 2020, since the country is endowed with high potential of renewable energy resources such
Accepted 26 August 2017
Available online 30 August 2017
as wind and solar. Although the annual average penetration rate of wind power has reached 24% of total
electricity production generated in Cape Verde, raising the wind energy penetration level in future will
pose numerous challenges for the operation and control of the power system because of wind's inherent
Keywords:
Wind characteristics
intermittency and unpredictability. In this study, a statistical analysis of the wind characteristics in
Wind energy potential assessment Santiago island, is presented by using historical wind speed and power data of the Santiago wind farm in
Statistical analysis 2014. A two-parameter Weibull distribution is first applied to model the wind speeds on various
Wind power curve timescales and to determine wind energy potential in Santiago island, Cape Verde. The annual average
Wind ramp wind speed was 8.57 m/s with a standard deviation close to 3.29 m/s. The monthly Weibull scale pa-
Uncertainty set rameters varied from 5.64 m/s to 13.7 m/s, while the monthly Weibull shape parameters varied from 1.97
to 9.13. Although the monthly mean power density of the rainy season from August to September was
low, the annual mean power density shows that Santiago has good wind potential. Then, an approach to
modeling the equivalent power curve based on available wind speed and power output data from the
wind farm is proposed. By utilizing the estimated power curve, the uncertainty set of wind power
generation, resulted from the uncertainty of wind speed forecast, can be obtained to quantify the power
system reserve requirements. A statistical analysis of wind power ramp is also given for estimating the
power capacity requirement of the energy storage system that can be considered as a reasonable way to
mitigate the wind intermittency and minimize curtailment of wind. Results of this study contribute to
assess the wind energy potential of Cape Verde for investors, and can be used to quantify the un-
certainties of wind power generation for the power system operator.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction
In recent years, electricity generation by photovoltaic or wind
power has captured considerable attention worldwide. In partic-
ular, from the point of view of security of power supply, for a
country like the Republic of Cape Verde, which does not have
known fossil fuel resources or reserves, renewable energy sources
play an essential role in reinforcing levels of energy security,
improving energy sustainability and mitigating risks from a fluc-
tuant international energy market. The Republic of Cape Verde is an
archipelago consisting of 10 islands and 13 islets approximately
E-mail address: xytsing@ecust.edu.cn.
400 km off the coast of Senegal in the Atlantic Ocean. The country
has a total land area of 4033 square kilometers and a total popu-
lation of 542,000 inhabitants [1]. It has limited natural resources
and poor rainfall. In past, Cape Verde's energy was characterized by
a high dependency on imported oil products [2]. Although 10
islands had been electrified, electricity supply was based on diesel
and gasoline generators. Furthermore, since Cape Verde is not on a
major shipping route and has to transport fuel to its islands, diesel
is quite expensive. Therefore, a National Energy Plan 2003e2012
was published in 2003 laying out a pathway for the consolidation of
the Energy Sector and a guarantee of national energy security [3].
Consequently, renewable energy has been rapidly invested in Cape
Verde. The Cape Verde Wind Power project with the 32 turbines
installed on four onshore wind farms on four islands (Santiago, Sao

http://dx.doi.org/10.1016/j.renene.2017.08.077
0960-1481/© 2017 Elsevier Ltd. All rights reserved.
 X. Qing / Renewable Energy 115 (2018) 448e461
Vicente, Sal, and Boa Vista), was also awarded the ’Best Renewable
Project in Africa’ prize at the Africa Energy Awards 2011 in Johan-
nesburg. Currently, more than 25% of the electricity produced in
Cape Verde is based on the use of renewable energy resources.
According to annual report 2014 of Cabeo  lica, a private-public
company, the four wind farms together, produced a total of
80.9 GWh (Santiago 42%, Sao Vicente 27%, Sal 21%, and Boa Vista
10%) [4], as shown is Table 1. The annual average penetration rate of
the four wind farms was 24% compared to total production
generated in the country in 2014.
Recently, Cape Verde has set an ambitious goal of meeting 50%
of its energy needs through renewable energy by 2020. The country
has also planned to have at least one of ten islands 100% energy
self-sufficient through renewable means by 2020. Since Cape Verde
has excellent conditions for wind and solar power generation,
renewable resources integration holds an important position in
relieving dependence on foreign oil products, cutting operating
cost and reducing the emission of pollutants. Although the goal is
lofty and the corresponding government incentives have been
established, Cape Verde will facing with a great deal of challenging
issues. One of the most important issues in renewable integration is
intermittency, which creates difficulties in meeting load demand.
The power system in Cape Verde is a hybrid plant which is an
integration of diesel generators with renewable energy resources,
such as solar and wind. The hybrid plant must continuously
manage the fluctuations of the load and power output of the
renewable resources to maintain the nominal frequency of the
power system, which is required for the satisfactory operation of
the power system. That the diesel generator in the hybrid plant
must operate above a minimum loading constraint will further
exacerbate the operation and management of the power grid.
Moreover, the wind farms in Cape Verde with low geographic di-
versity exhibit frequent occurrences of extremely high and low
power outputs. Wind ramps caused by surge of wind speed will
lead to higher ancillary service requirements, or curtailing or
limiting the ramp rate of wind generation, to maintain the power
system balance. Therefore, in 2014, the total wind power wasted on
the four islands in Cape Verde approached 25% of total available
wind energy. Most wind impact studies have been restricted to
penetrations below 20%. As penetrations increase above 20% their
impacts on power system operations including frequency control
and transient stability, become more severe. To mitigate the wind
intermittency and minimize curtailment of wind, the energy stor-
age system (ESS) can be considered as a reasonable way to cater to
such high wind penetrations in future. Recently, there are a number
of innovations in technological development and a huge leap in
business mode about the ESS. The ESS can decrease the total diesel
consumption, carbon emissions, and fuel costs while increasing the
value of renewable energy investment and improving the renew-
able generation output controllability. However, since there are
always some errors of wind speed and wind power forecast, the
relationship between the sizing power capacity of the ESS and the
uncertainties of wind speed and power forecast, should be inves-
tigated thoroughly.
Table 1
Wind energy in Cape Verde [4].
Island Installed Energy produced Wind speed Availability
capacity (MW) (MWh) (m/s) (%)
Santiago 9.35 33,898 8.6 99.5
S. Vicente 5.95 22,211 10.0 99.0
Sal 7.65 16,640 8.9 98.7
Boa Vista 2.55 8130 8.9 99.1
Total 25.50 80,878 9.1 99.1
449
The main objective of this paper is to gain insight into the wind
resource of Cape Verde by analyzing the historical wind data ob-
tained in Santiago island in 2014. Statistical analysis of wind
resource data is critical to have an idea of how the future power
system will operate with a higher wind penetration. Simple sta-
tistics extracted from the data can answer questions, such as, which
month has the largest mean wind speed? What are the distribu-
tions of the monthly wind speeds? What is the diurnal variation of
the wind speed? What is the relationship between the wind speeds
and the active power outputs? How should ramps be defined
especially in light of increasing penetration? Do severe ramps occur
rarely or are they relatively frequent? For the planning of a wind
farm or the preparation of expanding the capacity of a wind farm,
one can assess the potential of the wind energy by using the his-
torical wind speed and wind power data. The historical data can
always be a promising source to find the uncertainties of wind
speeds for any specific regions installing wind turbines. Moreover,
the most suitable capacity of an ESS smoothing the wind farm
power output, can be effectively determined by analyzing the his-
torical wind data and considering the uncertainties of wind speed
and power forecasting. Of course, long-term data sets of wind
speed and responding active power measurements should be
desirable to evaluate the impact of integrating wind energy into the
isolated hybrid electrical network. As pointed out in Ref. [5], one
year of records of wind speed and power behavior is sufficient to
predict the long-term wind characteristics within an accuracy of
10% and a confidence level of 90%.
There are relatively few studies published for statistical analysis
of wind characteristic, wind energy potential and economic evalu-
ation of investments of wind energy for Cape Verde. In the pio-
neering work published in 2000 [2], Alves et al. had studied the
renewable energy technologies and the corresponding social, eco-
nomic and environmental benefits for Cape Verde. They alleged that
renewable energy can compete with conventional energies for is-
land and remote regions, such as Cape Verde. Therefore, their ap-
peal that energy and environment policies should maximize the use
of alternative sources of energy by introducing appropriated tech-
nologies, is insightful. In Ref. [6], five scenarios of development of
electricity production for Santiago Island were presented to analyze
the potential influence of clean development mechanism, as well as
the influence of declining prices of renewable energy technologies.
Due to the fact the technology prices of wind in 2002 were very
high, the study in Ref. [6] concluded that wind was not economically
viable on the island of Santiago. Nørgård and Fonseca simulated the
power systems at three of the Cape Verdean islands with varying
expected demand developments and varying wind power capac-
ities by using the IPSYS time series simulation tool [7]. Freire et al.
used the computer simulation program HOMER to evaluate the
economic feasibility of adding a large wing turbine to an existing
electric system in Cape Verde [8]. The simulation result is encour-
aging that the implementation of the turbines is viable for Cape
Verde. Segurado et al. analyzed the way to increase the penetration
of renewable sources by using pumped hydro storage in S. Vicente
Island, Cape Verde [9]. However, their study was based on existing
load data and meteorological data. Teixeira et al. analyzed the sta-
bility of power systems in Santiago Island with wind farm by using
MATLAB/Simulink [10]. In Ref. [11], Ranaboldo et al. analyzed the
designs of off-grid electrification projects based on hybrid wind-PV
energies in 3 rural communities in Cape Verde. They assessed the
wind resource by using the numerical wind atlas. Recently, a policy
analysis tool was proposed to quantify the benefits and costs of
implementing a grid-connected onshore wind project and applied
to the appraisal of a wind farm in Santiago Island [12].
Other studies focusing on statistical analysis of wind energy
characteristics around the world besides Cape Verde can also be
450 X. Qing / Renewable Energy 115 (2018) 448e461
learned. In Ref. [13], the hourly measured time series wind speed
data of Keban-Elazig in Turkey were utilized to derive the proba-
bility density function (PDF) and identify the distribution param-
eters. Two probabilistic models, namely the Weibull and the
Rayleigh models, were applied to study the wind energy potential
of the location. One of the most important outcomes of the study
was that the Weibull distribution provided better power density
estimations in all 12 months than the Rayleigh distribution. Carta
et al. determined the maximum wind sourced energy potential in
the Canary Islands (Spain) by estimating the PDF of wind speed
[14]. Their studies show that the Weibull distribution of two pa-
rameters presents a series of advantages, such as its flexibility, the
dependence on only two parameters, the simplicity of the esti-
mation procedure. Ayodele el al. presented a comparative assess-
ment of statistical distribution model for the coastal regions of
South Africa by using historical data consisting of 10-min average
time series wind speeds, temperatures, and standard deviation
measured at anemometer height of 20 and 60 m for a period of 1
year [15]. The results of fitting wind speed distribution also indicate
that the Weibull model gives better fittings. Boudia and Guerri has
used ten years of wind data from Oran meteorological station in
Algeria to evaluate the potential of wind power on the Oranie re-
gion [16]. Based on their statistical analysis results, appropriate
wind turbine model has been recommended for wind farms con-
structing. Shu, Li, and Chan presented a statistical assessment of the
wind characteristics and wind energy potential at typical sites in
Hong Kong by using 6-year wind data [17]. That the Weibull dis-
tribution function can give a good representation of the measured
wind data was also held in this study. Moreover, they provided the
most promising locations for wind energy exploitation in Hong
Kong according to the statistical analysis. Instead of using the wind
speed data, Stoutenburg et al. used simulated historic wind power
data containing the 10-min resolution power output of a simulated
30 MW wind farm operating from 2004 to 2006 located at 2881
onshore sites in California [18]. In Ref. [19], the wind power
ramping behavior from 2004 to 2009 was statistically analyzed by
using 10-min and hourly average wind power data from Electricity
Reliability Council of Texas (ERCOT). Because ERCOT system is large,
the statistical analysis results indicate that the level of wind power
at that time is not likely to cause severe damage events for ERCOT
grid operators. In Ref. [20], dimension reduction and classification
techniques were utilized to analyze historical data from the wind
farms in the Tehachapi Pass and mid-Columbia Basin regions to
support control room operators for making well-informed sched-
uling decisions.
The paper is organized as follows. Section 2 describes the site,
the data collection and the preprocessing algorithm on the Santiago
wind farm of Cape Verde. Section 3 provides empirical statistics of
wind speed data. Section 4 presents the statistical analysis of wind
power outputs. Section 5 presents the statistics of the wind ramp
events. Finally, conclusions close the work.
2. Data collection and preprocessing
2.1. Description of site and wind speed measurement
Cape Verde has wind energy resources from the trade winds
providing a strong northeasterly flow for most of the year. The
Santiago wind farm is located in the south of the Santiago Island, on
Monte de Sao Filipe, near the city of Praia, as shown in Fig. 1. It was
officially unveiled on October 21, 2011 and became the first wind
farm to begin operation in Cape Verde. It occupies about 30 ha of
land and contains of 11 Vesta V52 wind turbines, using a special
asynchronous 4-pole generator with wound rotor and slip rings.
The rated power of the V52 turbines is 850 kW, and the rotor
diameter is 52 m. Therefore, all rated installed capacity of the wind
farm in Santiago Island is 9.35 MW.
Following the International Electrotechnical Commission (IEC)
wind turbine standard 61400-12-1 regarding power performance
measurements of electricity producing wind turbines [21], all
measured 10-min data for each wind turbine are acquired by a
supervisory control and data acquisition (SCADA) system, including
time stamps (year, month, day, hour and minute), wind speeds
recorded at a height of about 60 m, active power outputs, and
frequencies of turbines, averaged over a 10-min intervals. In this
study, the data were collected from the period 1/1/2014 0:00 to 12/
31/2014 23:50.
2.2. Wind data preprocessing
Similar to [22] and [23], the raw wind data set (52560 points)
contains some invalid data points which should be filtered for
further statistical analysis. An initial preliminary data preprocess-
ing was performed on the 10-min average data streams to eliminate
obvious invalid data points such as no data, negative values of
active power output readings. Let W and P denote the matrices of
wind speed data and active power data, respectively. The elements
Wi,j and Pi,j of matrices W and P denote the wind speed and active
power output of the jth wind turbine at the ith time stamp,
respectively. The dimension of both the two matrices is T
N,
where T ¼ 52560 and N ¼ 11. The flowchart in Fig. 2 shows the
preprocessing algorithm for the data matrix A equivalently stand-
ing for W or P: the internal cycle for each time instant t, with N
iterations, missing data and negative active power data of N wind
turbines are first labeled. At the same time, the invalid entries are
counted and their number of these entries is stored in k. Then, these
invalid entries of wind speed and power matrices are filled with
appropriate values if the number at the time instant t is less than 4.
Otherwise, all data at the time instant t are labeled as ‘NaN's and
discarded because too many invalid data can severely affect the
following statistical analysis. Consequently, the mean values of
wind speed and the aggregate power outputs for the wind farm are
denoted as vectors W and P, respectively.
This preprocessing algorithm identified about 2% of the 10-min
wind speed data points and 12.8% of the 10-min wind power data
points as the invalid data points.
3. Statistical analysis of wind speeds
3.1. Wind analysis model
The Weibull distribution model is applied to represent the wind
speed distribution because of its simplicity to estimate the mean
and standard deviation of the total wind power density and its
ability to assume the characteristic of many different types of dis-
tributions. Moreover, numerous studies, as reviewed in the intro-
duction section, have shown that the two-parameter Weibull
distribution provides a good fit to wind data for evaluating the wind
energy potential. Although a two-component mixture Weibull
distribution has the ability to represent heterogenous wind re-
gimes [24,25], there are not significant performance improvements
by adopting the mixture mode in this study, compared to the
typical two-parameter Weibull model. The PDF of the two-
parameter Weibull distribution f(v) of measured wind speed v is
defined by the following equation:

 k1   k 
k v v
f ðvÞ ¼ exp  (1)
c c c
where c represents the Weibull scale parameter (in unit of m/s) and
 X. Qing / Renewable Energy 115 (2018) 448e461 451

Fig. 1. Geographical location of the Santiago wind farm (from Wiki).

k denotes the dimensionless Weibull shape parameter. Both are

estimated by using the maximum likelihood method in the present 


study. The Weibull shape parameter k, as known as the Weibull 2 1 1∕2
s ¼ c, G 1 þ  G2 1 þ (4)
slope, indicates how skewed the distribution is. A large shape k k
parameter gives a left tail, whereas a smaller shape parameter gives
The most probable wind speed vmp of the region, and maximum
a right tail. Therefore, the characteristics of the different distribu-
energy carrier vmaxE of wind speed for wind energy converters are
tions can be modeled by changing the shape parameters. The
determined as.
Weibull scale parameter c of the wind speed model determines
how windy the region is, like the mean defines a normal curve's
position. The larger the value of c, the more the location will have

1 1∕k
wind energy potential. vmp ¼ c, 1  (5)
k
The Weibull cumulative distribution function (CDF) F(v) is
defined as:
  k 

v 2 1∕k
FðvÞ ¼ 1  exp  (2) vmaxE ¼ c, 1 þ (6)
c k

The average wind speed vm is calculated by using the estimated The two wind speeds are very useful to the wind energy in-
Weibull parameters as below. vestors. Essentially, the vmp simply tells the most frequently wind
speed for a given wind probability distribution. The vmaxE is utilized


1 to estimate the wind turbine design of rated wind speed because
vm ¼ c,G 1 þ (3)
k the speed carrying the maximum amount of energy and the design
wind speed of the turbine should be close as possible [26].
where G is the Gamma function. The standard deviation of wind The wind power density PD using the Weibull model is
speed can also be calculated as: expressed as:
452 X. Qing / Renewable Energy 115 (2018) 448e461
Fig. 2. Flowchart of the proposed wind speed and wind power data preprocessing.


1 3 3
PD ¼ rc G 1 þ
2 k
where r is the air density (assumed to be 1.225 kg ∕ m3 in this
study).
3.2. Estimation and evaluation of distribution parameters
Both the monthly and annual values of these parameters are
summarized in Table 2. In Santiago island, during the year 2014, the
annual average wind speed was 8.57 m/s with a standard deviation
close to 3.29 m/s. The monthly mean wind speeds varied between
4.95 m/s in August and 12.48 m/s in February. August and
September, composing the rainy season of Cape Verde, had little
wind with the monthly vm values of 4.95 m/s and 5.90 m/s,
respectively. The monthly average wind speeds obtained from the
actual data, instead of the Weibull models, are also presented in
Fig. 3, which are quite consistent with the estimated parameter
values vm obtained from the Weibull models. The monthly values of
(7) the scale parameter c varied between 5.64 and 13.17, while the
monthly values of the shape parameter k varied between 1.97 and
9.13. It is clear that a wide variation in the shape and scale pa-
rameters occurs over the months. It can be seen from Table 2 than
for all months except September the values of shape parameter k
are greater than 2. Generally, the value of k remains close to 2,
which means that the wind distribution is regular and uniform. For
k < 2.6, the Weibull PDF has a right tail. For 2.6 < k < 3.7, the Weibull
PDF may approximate the normal PDF. If k > 3.7, it has a left tail.
Therefore, the wind speed in September is approximately uniform,
as shown in Fig. 4. The PDF of speed data in February has the
maximum values of scale and shape parameters, as shown in Fig. 5.
The annual probability density and CDFs of wind speeds are
displayed in Fig. 6. The measured PDF represented as a histogram, is
obtained by binning the 10-min wind speed data. This binning is in
the range from 0 m/s to 20 m/s and has the width 0.2 m/s. It is seen
that the estimated Weibull PDF matches well with the measured
histogram because only one hump is found on the histogram
 X. Qing / Renewable Energy 115 (2018) 448e461 453

Table 2
Wind speed characteristics and wind energy potential.

c (m/s) k vm (m/s) s (m/s) vmp (m/s) vmaxE (m/s) PD (W∕m2)

Jan 11.45 5.78 10.60 2.13 11.08 12.05 815.17

Feb 13.17 9.13 12.48 1.64 13.01 13.46 1251.00
Mar 11.24 4.95 10.32 2.38 10.74 12.04 778.60
Apr 10.67 3.07 9.54 3.39 9.39 12.56 736.95
May 10.58 3.61 9.54 2.93 9.67 11.95 681.98
Jun 10.65 3.33 9.56 3.17 9.56 12.27 712.12
Jul 7.34 2.29 6.50 3.01 5.71 9.66 284.89
Aug 5.56 2.68 4.95 1.99 4.68 6.85 111.54
Sep 6.64 1.97 5.90 3.13 4.63 9.49 243.21
Oct 7.68 3.91 6.95 1.99 7.12 8.54 256.20
Nov 8.87 4.72 8.12 1.96 8.44 9.56 384.15
Dec 9.48 3.79 8.57 2.52 8.74 10.59 484.59

Annual 9.62 2.82 8.57 3.29 8.24 11.63 560.94

14

12.51

12

10.60 10.32
10 9.55 9.53 9.58
Mean wind speed (m/s)

8.58
8.12
8
6.96
6.50
5.88
6
4.96

0
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Month

Fig. 3. Monthly average wind speed at Santiago island.

outline. The PDF illustrates the fraction of time for which given
wind speed probably prevails at this site. Moreover, the peak of the
PDF represents the most frequent wind speed occurred at this site.
It can be observed from Fig. 6 that the most frequent wind speed is
about 8.2 m/s which is consistent with the value vmp of estimated
Weibull distribution using the annual data in Table 2.
Fig. 7 shows the q-q plot between the real wind speed data and
the data sampled from the estimated Weibull distribution. The q-q
plot shows the linear relationship between the measured and
sampled data when the wind speeds are less than 14 m/s. When the
wind speeds are greater than 14 m/s, the q-q plot is not a straight
line, suggesting that the wind speeds fail to follow the estimated
Weibull distribution. However, the frequency of the wind speeds
being greater than 14 m/s is relatively low. Thus, the estimated
Weibull distribution is still feasible. From Fig. 6, we observe that the
estimated Weibull CDF is also consistent with that of measured one.
The CDF represents the fraction of time that the wind speed is
below a particular speed. As an example, wind speeds being greater
than or equal to cut-in wind speed 3.0 m/s have frequencies of
about 96.3% for the Santiago wind farm with the V52 turbines. Fig. 4. Wind speed frequency distribution for September data.
The diurnal wind speeds are also modeled by using the Weibull
454 X. Qing / Renewable Energy 115 (2018) 448e461
Fig. 5. Wind speed frequency distribution for February data.
Fig. 6. Histogram and cumulative distribution of the annual wind speeds.
distribution. The hourly mean wind speeds using the estimated
Weibull parameters are shown in Fig. 8. It is found that the hourly
wind speed increases at around 3 a.m. and reaches the peak at
round 11 a.m. then it decreases. The wind speeds in the day usually
are greater than that in the night.
3.3. Estimation of wind power potential
The monthly mean power density had the minimum value in
August and the maximum in the February, ranging from
111.54 W∕m2 to 1251.00 W∕m2 with annual average of
560.94 W∕m2. According to PNL wind power classification scheme
[27], in August, the monthly mean power density falls into class 1
(PD  200¡/b¿ W∕m2 and Vm  5:6¡/b¿ m/s). Therefore, the site has
non-ideal wind power generation in August. In September, the
wind power is marginal acceptable. These results coincide with the
local climate that the Cape Verde may have heavy tropical rainfall at
the rainy season from August to September. In June and October,
the wind power is acceptable and can be improved by using taller
wind turbine towers. From November to July of next year, the site is
extremely suitable for wind power development. The annual mean
power density falls into between class 5 and class 6, which means
Fig. 7. q-q plot between real wind speed data and the data sampled from the Weibull
distribution.
Fig. 8. Mean annual diurnal profile of mean wind speed.
that the site has good wind energy potential.
4. Statistical analysis of wind power outputs
Because the wind farm consists of 11 turbines connected to the
power system, the focus of this study is the entire farm's perfor-
mance. Accurate estimation of wind farm power curve graphically
representing of the turbine electric power output as a function of
the wind speed, is essential to operational management of wind
energy as well as predicting wind farm power output for a given
wind forecast. A theoretical power curve supplied by a turbine
manufacturer is utilized to characterize performance of a single
wind turbine under ideal meteorological and topographical con-
ditions. Therefore, there are at lease two important problems to
handle. The first is how to obtain an equivalent power curve for the
entire farm. The second is how to deal with the substantial differ-
ence between the theoretical and real working conditions of the
wind turbine due to the location of the turbine, air density, wind
 X. Qing / Renewable Energy 115 (2018) 448e461 455

velocity distribution, wind direction, mechanical and control issues

[28]. Consequently, a simple sum of the manufacture power curves
for the wind farm will typically yield overestimated power outputs.
In this paper, we develop an approach to modeling the equivalent
power curve based on available wind and power output data from
the wind farm. We present a ninth degree polynomial regression
model for power curve fitting.
In addition, the power outputs from adjacent identical turbines
even under the same wind speed are not be necessarily equal, and
so aggregating across the entire wind farm can yield the spread of
potential values of aggregate wind power, in which case the stan-
dard deviation is an appropriate measure for uncertainty of the
power curve [29]. Moreover, the accurate estimation of the spread
of power outputs throughout the wind speeds is also required to
size the storage capacity when uncertainty of the wind speed
forecasting is considered. For long-term economic dispatch, the
long-term forecasting of wind speed may be more accurate and
easily available than the long-term forecast of power output. We
need to predict the resulted variation range of power output from Fig. 9. Normalized histogram and fitted Beta distribution of the power output.
the possible forecast errors of wind speeds. Based on the variation
range, the power size of storage capacity can be appropriately
determined. rated wind power capacity 9.35 MW in this study.

4.1. Statistical model of wind power output 4.2. Filtering the unnatural data

Before we generated an equivalent power curve and performed Fig. 10 is a scatter plot of the wind farm power output and wind
the statistical analysis of wind power outputs, a data quality check speed data after performing the preprocessing procedure. The red
was carried out again. For each time stamp t, if WS [t] or P [t] is power curve is the sum of 11 idealized power curves of a 850 kW
denoted as ’NaN0 , then both WS [t] and P [t] are denoted as ’NaN0 , wind turbine with cut-in speed of 3 m/s, rated wind speed of 16 m/
namely invalid data points. Consequently, about 86% of data points s, and cut-out wind speed of 25 m/s. The power curve provides pair
are saved to the statistical analysis. Fig. 8 shows a histogram of of points for wind speed-power every 1 m/s from the manufac-
aggregate wind power from the wind farm. A Beta distribution, turer's datasheet.
with parameters a and b, is chosen for approximating the It is clear from Fig. 10 that the idealized power curve over-
normalized histogram of wind power, as shown as a red curve in estimates the true in situ relationship between wind speed and
Fig. 8. The Beta distribution of normalized aggregate wind power wind power, even though the general trend shown in the measured
Pagg is expressed as: data tends to identify with the manufacture's rated curve. The
 b1 phenomenon can be boiled down to some non-ideal working
 P aagg
1

1  Pagg conditions of turbines including dependency of power output on
f Pagg ¼ (8)
Bða; bÞ wind direction, efficiency degradation over time. Moreover, there
also exist some unnatural data with low power outputs during high
where B (a,b) is the Beta function: wind speed periods because the wind farm may have one or more
wind turbines out of service at any given time. We note that
Z1 analyzing and utilizing these unnatural data for turbine monitoring
Bða; bÞ ¼ xa1 ð1  xÞb1 dx (9)
0

and a and b are shape parameters and x is the variable of inte-

gration. The a and b parameters of the Beta distribution are derived
by the maximum likelihood estimation method, whose values are
also shown in Fig. 9. Because the values of a and b parameters are
close to 1, the estimated Beta distribution approximates to a Uni-
form distribution, which means that the distributions from near-
zero to near-full power outputs are almost uniform. The expected
value E(Pagg) and the variance V (Pagg) of the Beta distribution, are
given as.
 a
E Pagg ¼
aþb
(10)
 ab
V Pagg ¼ 2
ða þ bÞ ða þ b þ 1Þ
Thus, the expected value of aggregate power outputs is 0.507
p.u, which is meaningfully interpreted as the wind farm capacity
factor of 50.7% in the wind power generation system, as discussed
in Ref. [30]. Note that 1 p.u power in per-unit system denotes the Fig. 10. Scatter plot of wind farm power output and wind speed.
456 X. Qing / Renewable Energy 115 (2018) 448e461

is not the focus of this study. Therefore, these unnatural data should
be discard for accurately learning the wind farm power curve.
Given the aggregate manufacturer's rating curve
Pr ½i; i ¼ 1; 2; …; 20¡/b¿, wind speeds W [t] and power outputs
P½t; t ¼ 1; 2; …; T¡/b¿, the flowchart of the algorithm filtering the
unnatural data is shown in Fig. 11. In Fig. 11, the floor function
returns the largest integer j that is smaller than or equal to
measured wind speed W [t]. Therefore, the integer j equals to the
bin value of wind speed in the manufacturer's power curve. The
ideal power Pi related to the wind speed W [t] is calculated using
linear interpolation. If the power output P [t] at the wind speed W
[t] is lower than the threshold value a,Pi ¡/b¿, both the power
output P [t] and wind speed W [t] are assigned to ’NaN's. The value
of parameter a is empirically set according the magnitude of wind
speed. Consequently, the unnatural wind power data indicating
significantly low power outputs related to specific wind speeds are
filtered.
Fig. 12 shows the obtained scatter plot of 43397 data points after
performed the proposed filtering algorithm. Obviously, those un-
natural data points have been erased.

4.3. Generating an equivalent power curve for the wind farm

The ”methods of bins” is employed to obtain an equivalent po-

wer curve for the wind farm based on these filtered data points, as
given in Ref. [31]. Wind speed is divided into bins of 0.5 m/s. For
each bin i, the average wind speed vi ¡/b¿ and average wind power
P i ¡/b¿ are calculated as follows:

P
Ni
vi;l
vi ¼ l¼1 (11)
Ni

P
Ni
Pi;l
P i ¼ l¼1 (12)
Ni

where Ni is the total number of points in bin i,vi,l and Pi,l are the
values of the l-th point of wind speed and wind power in bin i,
respectively. The standard deviation sPi ¡/b¿ of the power in each bin
i is calculated:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
uPNi  2
u Pi;l  P i
u
t
sPi ¼ l¼1 (13)
Ni

Similarly, the standard deviation sws i ¡/b¿ of the wind speed in

each bin i can be obtained. In Fig. 12, the red curve represents the
produced power curve based on fitting a curve to ðvi ; P i Þ¡/b¿ pairs by
using the method of bins. In each bin, the uncertainty range of
power is chosen as ½P i  3sPi ; P i þ 3sPi ¡/b¿, which corresponds to
the lower and upper bounds as shown in Fig. 12.
A ninth degree polynomial is used to fit the ðvi ; P i Þ¡/b¿ pairs as:

b
PðvÞ ¼ c1 v9 þ c2 v8 þ c3 v7 þ c4 v6 þ c5 v5 þ c6 v4 þ c7 v3 þ c8 v2
þ c9 v þ c10
(14)

where c1 ; …; c10 ¡/b¿ are coefficients that are to be found out, and
b
PðvÞ¡/b¿ is a p.u power. The coefficients are estimated by using
curve fitting toolbox of MATLAB, as listed in Table 3.
Coefficient ri, is defined as:

Fig. 11. Flowchart of the proposed algorithm for filtering the unnatural wind power
data.
ri ¼ sPi ∕P i (15)
to quantify the relative varying level of standard variation of
power output. Fig. 13 shows the ðP i ; ri Þ¡/b¿ pairs. It can be seen from
Fig. 11 that the coefficients linearly decrease as normalized P i ¡/b¿
increase when they are greater than 0.15 p.u. When P i ¡/b¿ is small,
the variation values of wind farm power output can be tolerated by
 X. Qing / Renewable Energy 115 (2018) 448e461 457


ri P i ¼ 0:0789P i þ 0:0936 (16)

The fitted curve denoted as a blue line, is also shown in Fig. 13.
The corresponding fitted standard deviation b s ¡/b¿ can be
calculated as:
 
b
sb P ¼ r PðvÞ b
PðvÞ (17)

for a wind speed v falling into the range between the cut-in and
rated wind speeds. Fig. 14 shows a fitted power curve Pðv b Þ¡/b¿ to
i
the wind speed vi ¡/b¿. The fitted lower and upper bounds are
b Þ  3s
calculated as Pðv b Þ þ 3b
b P ¡/b¿ and Pðv s P ¡/b¿, respectively.
i i i i
Coefficient bi, as the relative standard deviation of the wind
speed, is also expressed as:

bi ¼ sws
i ∕vi (18)
Then, factor ti defined as:

Fig. 12. Scatter plot of filtered wind farm power output and wind speed. ti ¼ ri ∕bi (19)
is introduced to quantify the scaling effect of the power curve
Table 3 that describes the variation of power output in the wind speed due
Coefficients of ninth degree polynomial fit.
to the local slope of the power curve, as proposed in Ref. [32]. The
Coefficients Values ðvi ; ti Þ¡/b¿ pairs are shown in Fig. 15. The factor ti, as the ratio be-
c1 1.4088
108 tween the relative standard deviations of power output and wind
c2 1.1844
106 speed, decreases from about 8 to 0.6 as the wind speed increases
c3 4.3230
105 from 3 m/s to 16 m/s. Smaller the wind speed is, the greater the
c4 8.9902
104 scaling effect of the power curve will be. Note, however, that the
c5 0.0117
c6 0.0988
absolute values of the variations of power outputs resulting from
c7 0.5339 the nonlinearity of the power curve should be small.
c8 1.7803
c9 3.2953
4.4. Estimation of uncertainty set of wind power
c10 2.5761

It is because of the nonlinearity of the power curve that the

the power grid, although the relative varying levels of standard
variations are large. Thus, a linear model is fitted for modeling the
linear dependencies between the coefficients obtained by dividing
the average power outputs by the corresponding standard varia-
tions and the average power outputs when normalized power
outputs are greater than 0.15 p.u, as given by:
probability distribution of wind power forecast error is not
Gaussian [33]. Alternatively, consider that error of the wind speed
forecast is described to be within ±r%¡/b¿ variation on the forecast
value e
v¡/b¿ m/s, generally meaning that the actual wind speed value
v falls into a range ½e
v  r%,e
v; e
v þ r%,e
v¡/b¿ with a confidence level of
95% or a greater value. As the power curve of the wind farm is
estimated, uncertainty of the corresponding wind power forecast
can be readily evaluated with a step-by-step illustration.

Fig. 13. Coefficients corresponding to variation of power output for the wind farm. Fig. 14. Fitted power curve.
458 X. Qing / Renewable Energy 115 (2018) 448e461

Fig. 16. Predicted power curve, the lower and upper bounds of power forecasts based
Fig. 15. Factors that quantify the scaling effect of the power curve from the variations
on the wind speed forecast values.
of wind speeds to the variations of power outputs.

r r It is observed from Figs. 16 and 17 that the absolute variation

Step 1: Calculate the lower bound e
vl ¡/b¿ and upper bound evu ¡/b¿ values of power forecast due to the errors of wind speed forecasts,
r r
e
of wind speed forecast as: v ¼ maxðv ; v  r%,vÞ¡/b¿;e
e e v reach to their maximum values when the predicted wind speeds
l cutin u
¼ minðvrate ; e
v þ r%,evÞ¡/b¿; are within the range from 8m∕s to 13m∕s. It can also be seen that
Step 2: Calculate the corresponding point estimation values even small errors of wind speed forecasts can strongly amplify the
e er ¡/b¿ p.u, and Per ¡/b¿ p.u of power forecast with e variation range of wind power forecasts in this range which is the
P¡/b¿ p.u, P l u v¡/b¿,
r r
steep part of the estimated power curve. However, the wind speeds
e
vl ¡/b¿ and e
vu ¡/b¿ by using Eq. (14). falling into this range account for about 46% in the all year.
r r
Step 3: Calculate the corresponding coefficients e rl ¡/b¿ and eru ¡/b¿ Therefore, for quantify the system reserve requirements, the wind
er ¡/b¿ p.u, and P
er ¡/b¿ p.u by using Eq. (16), respectively. power uncertainty due to the uncertainty of wind speed forecast,
with P l u
r r r r r r should be considered in the system planning and operation to
sl ¼ erl ,Pel ¡/b¿ and seu ¼ eru ,Peu ¡/b¿ according to Eq.
Step 4: Set e guarantee reliability.
(15).
Step 5: Obtain the lower bound P b r ¡/b¿ and the upper bound
l 4.5. Modeling the relationship between the maximum deviation
b r ¡/b¿ of power forecast with the 95% confidence level as:
P values of power forecasts and the uncertain levels of wind speed
u

! forecasts
b r ¼ max 0; P
P er  1:96s
er (20)
l l l Define the maximum deviation value dPr ¡/b¿ of power forecast
using the error of wind speed forecast with the uncertain level r%
!
b r ¼ min 1; P
P er þ 1:96s
er (21)
u u u

Step 6: Calculate the lower bound err rl ¡/b¿ and upper bound
err ru ¡/b¿ of deviation of power forecast using the wind speed
forecast value e v¡/b¿ with ±r%¡/b¿ forecast error, as:

r
eP
errrl ¼ P b (22)
l

br  P
errru ¼ P e (23)
u

Consequently, the uncertainty of wind power generation is

characterized by an uncertainty set, which is particularly attractive
to robust unit commitment and robust power system expansion
planning [34,35]. Fig. 16 shows the predicted power curve, the
lower and upper bounds of power forecasts using the wind speed
forecast values with varying r from 5 to 20 with step size 5. The
corresponding lower and upper bounds of deviation of power
forecasts are shown in Fig. 17. Fig. 17. Lower and upper bounds of deviation of power forecasts.
 X. Qing / Renewable Energy 115 (2018) 448e461 459

as:




dPr ¼ max errrl e
v ; err ru e
v ; ce
v2½vcutin ; vrate  (24)

Fig. 18 shows the ðr%; dPr Þ¡/b¿ pairs as blue points. As Fig. 18
makes clear, the maximum deviation values dPr ¡/b¿ of power fore-
casts linearly increase as the uncertain levels r% increase. Therefore,
a linear model is fitted the ðr%; dPr Þ¡/b¿ pairs as:

dPr ¼ 1:5312*ðr%Þ þ 0:0623 (25)

For instance, if the wind speed forecast has 20% of the uncertain
level for the medium-term economic dispatch, the maximum de-
viation of the power forecast can reach 37% of the rated wind power
capacity with the 95% confidence level. If the uncertain level can be
controlled to be lower than 10% by advanced forecasting methods,
the maximum deviation of the power forecast will be lower than
21% of the rated wind power capacity.
Fig. 19. Daily wind power profile and power ramps in the hour on January 1, 2014.

5. Statistical analysis of wind ramp events

Several definitions of wind ramp event, and their advantages
and disadvantages are illustrated in Ref. [36]. In this study, a simple
definition of wind ramp event, as given in Ref. [37], is adopted,
although it has a problem of counting large ramps multiple times
[38]. A ramp event is considered to occur at the start of an interval
[t,t þ Dt] if the fluctuation magnitude of wind power power at the
time Dt is greater than a predefined power ramp threshold, PRval:
jP½t þ Dt  P½tj > PRval
Furthermore, the ramp can be categorized into positive and
negative ramps according to the sign of incremental generation
over time. Generally, the negative ramps pose more challenges for
the operation of the power system because other means should be
found to increase generating capacity, while the positive ramps can
be efficiently limited by using the power rate limiter in the wind
farm main controller. As an example, Fig. 19 shows large positive
and negative ramps of the wind farm on January 1, 2014. From 8:30
a.m. to 9:30 a.m., a large positive ramp with the time interval of 1 h
occurred that the generation increased by about 3 MW. From 14:40
p.m. to 15:40 p.m., the generation decreased by about 1.7 MW,
which can be considered as a large negative ramp event.
Fig. 20 shows the magnitudes of the increase or decrease in the
wind power over 30-min intervals. Negative value represents a
drop of generation from the start of the interval to 30 min later. The
red dashed lines dictate the threshold values of positive and
negative ramps, which is equal to 15% of the rated capacity of the
wind farm. The average value and standard deviation of such step
power changes are 7.43
104 p.u and 0.082 p.u, respectively.
(26) The average value tends to be zero because the wind capacity is
limited and the positive ramp events are always balanced by the
negative ramp events [19]. If the value of three times standard
deviation (3s) is chosen to determine the power size of the ESS, 25%
of the total wind capacity should be reasonable for power capacity
requirement of the ESS.
Table 4 summarizes how many data points (each 10 min interval
is considered on data point) are the start of a change of a given
magnitude for a given time interval Dt. We can observe from Table 4
that small intervals will lead to small number of changes in gen-
eration. According to the suggestion that an important negative
ramp occurs only if the power change in 60 min is at least 15% of
total capacity [36,39], provided by a renewable energy consulting

Fig. 18. Linear relationship between the maximum deviations of the power forecast
and the uncertain levels of wind speed forecast. Fig. 20. Change in wind energy generated for 30 min intervals.
460
Table 4
Statistics of changes in wind generation over different time interval values.
jP½t þ Dt  P½tj 10 min 30 min
pos neg pos neg
 0:15ðp:uÞ; < 0:25ðp:uÞ¡/b¿ 172 183 1191 1140
 0:25ðp:uÞ; < 0:35ðp:uÞ¡/b¿ 36 23 244 259
 0:35ðp:uÞ; < 0:45ðp:uÞ¡/b¿ 5 8 73 58
 0:45ðp:uÞ; < 0:55ðp:uÞ¡/b¿ 6 4 24 22
 0:55ðp:uÞ; < 0:65ðp:uÞ¡/b¿ 0 0 12 11
 0:65ðp:uÞ; < 0:75ðp:uÞ¡/b¿ 0 0 1 9
 0:75ðp:uÞ; < 0:85ðp:uÞ¡/b¿ 0 0 0 1
 0:85ðp:uÞ¡/b¿ 0 0 0 0
company named as AWS Truepower, the number of the important
negative ramps is 3277, which makes up about 6.2% of total data
points using 10-min measured data. Note that the definition of the
wind power ramp does not check if the ramp will continue in the
same direction. Consequently, the statistics overestimate the
number of the wind power ramp events. The average value and
standard deviation of the step power changes in 60 min
are 0.0013 p.u and 0.113 p.u, respectively. Therefore, in case of 3s,
a conservative power capacity requirement of the ESS for satisfying
the power ramp constraint in 60-min time interval is about 34% of
the installed wind power capacity.
6. Conclusions
In the present study, the 10-min wind speed and power output
data of the Santiago wind farm in 2014, have been statistically
analyzed. The PDFs of the annual, monthly and diurnal wind speed
data have been been derived from the historical data, and the
distribution parameters were identified by the maximum likeli-
hood method. The annual and monthly mean power densities have
been calculated and were used to evaluate the wind energy po-
tential of this island. The characteristics of the aggregate power
output of the wind farm have also been analyzed. An approach to
modeling the equivalent power curve based on available wind
speed and power output data from the wind farm has been pro-
posed. By utilizing the estimated power curve, the uncertainty set
of wind power generation, resulted from the uncertainty of wind
speed forecast, can be obtained to quantify the power system
reserve requirements. A statistical analysis of wind power ramps
has also been presented for estimating the power capacity
requirement of the energy storage system. The important outcomes
of the present study are summarized as follows:
(1) The annual average wind speed was 8.57 m/s with a standard
deviation close to 3.29 m/s. The monthly Weibull scale
parameter varied from 5.64 m/s to 13.7 m/s, while the
monthly Weibull shape parameter varied from 1.97 to 9.13.
(2) The diurnal wind speed had small values during the night
and large values during the day.
(3) The annual mean wind power density value was
560.94 W∕m2. Although the monthly mean power density of
the rainy season from August to September was low, the
annual mean power density value shows that Santiago has
good wind potential. Using taller wind turbine towers is
recommended for the future wind energy development.
(4) The distributions of wind power outputs from near-zero to
near-full power outputs were almost uniform. The wind farm
capacity factor in the wind power generation system was
50.7%.
(5) If the uncertain levels of the wind speed forecasting increase
from 10% to 20%, the maximum deviations of the power
X. Qing / Renewable Energy 115 (2018) 448e461
forecast will linearly increase, ranging from 21% to 37% of the
rated wind power capacity with the 95% confidence level.
60 min (6) A conservative power capacity requirements of the energy
pos neg
storage system for satisfying the power ramp constraint in
60-min and 30-min time interval, are about 34% and 21% of
2303 2344
the installed wind power capacity, respectively.
693 631
204 187
69 70 All taken into account, we can safely recommend that 30% of the
32 22 installed wind power capacity can be the power capacity require-
7 22
ment of the energy storage system for mitigating the wind inter-
2 1
0 0
mittency and minimizing curtailment of wind in the future hybrid
power plant with higher levels of wind penetration in Santiago
island, Cape Verde.
Acknowledgments
This work was supported by the Fundamental Research Funds
for the Central Universities under Grant, No.222201717006. The
author would like to thank Truewin Renewables Technology
(Shanghai) Co., Ltd., for financial backing in field investigating the
Santiago wind farm and Electra Company for providing wind data.
The author wishes also to thank Dr. Huanji Xie, vice general man-
ager of the Truewin Renewables Technology (Shanghai) Co., Ltd.
and O.C.Nogueira of Electra Company for their valuable help.
References
[1] A.S. Antonia, Summary of Cape Verde Renewable Energy Plan, 2010. www.
ecreee.org.
[2] L.M.M. Alves, A.L. Costa, M. da G. Carvalho, Analysis of potential for market
penetration of renewable energy technologies in peripheral islands, Renew.
Energy 19 (2000) 311e317.
[3] International Renewable Energy Agency, Renewable Islands: Setting for Suc-
cess, 2014. http://www.irena.org/DocumentDownloads/Publications/GREIN_
Settings_for_Success.pdf.
[4] Cabeo  lica, Annual Report, 2014, 2014. http://www.cabeolica.com/site1/about-
us/annual-reports.
[5] M. Jamil, S. Parsa, M. Majidi, Wind power statistics and an evaluation of wind
energy density, Renew. Energy 6 (5e6) (1995) 623e628.
[6] N. Duic, L.M. Alves, F. Chen, M. da G. Carvalha, Potential of Kyoto protocol
clean development mechanism in transfer of clean energy technologies to
small island developing states: case study of Cape Verde, Renew. Suistainable
Energy Rev. 7 (2003) 83e98.
[7] P. Nørgård, J. Fonseca, Ultra High Wind Penetration in Simple Wind-diesel
Power Systems in Cape Verde, in: 2009 European Wind Energy Conference
and Exhibition, Marseille, France, 2009, pp. 9e12.
[8] F. Freire, J.P.S. Catalao, J.C.O. Matias, C.M.P. Cabrita, Contribution to investment
decisions on wind energy in Cape Verde, in: Proceedings of the 12th
Portuguese-Spanish Conference on Electrical Engineering - XIICLEEE, at Ponta
Delgada, Portugal, CD-R, 30 June - 2 July, 2011.
[9] R. Seguardo, G. Krajacic, N. Duic, L. Alves, Increasing the penetration of
renewable energy resources in S. Vicente, Cape Verde, Appl. Energy 88 (2011)
466e472.
[10] L. Teixeira, Y. Zhang, Y. Kang, Stability analysis of power system in Cape Verde
with wind power under wind gust and short circuit fault, in: 2011 4th In-
ternational Conference on Electric Utility Deregulation and Restructuring and
Power Technologies (DRPT), Weihai, Shandong, China, 6-9 July, 2011,
pp. 1072e1075.
[11] M. Ranaboldo, B.D. Lega, D.V. Ferrenbach, Moreno R.P. Ferrer-Martí, A. García-
Villoria, Rewable energy projects to electrify rural communties in Cape Verde,
Appl. Energy 118 (2014) 280e291.
[12] S. Salci, G.P. Jenkins, An Economic and Stakeholder Analysis for the Design of
IPP Contracts for Wind Farms, April 6, 2016. Available at SSRN: http://ssrn.
com/abstract¼2760137.
[13] E.K. Akpinar, S. Akpinar, A statistical analysis of wind speed data used in
installation of wind energy conversion systems, Energy Convers. Manag. 46
(2005) 515e532.
[14] J.A. Carta, P. Ramírez, S. Vel azquez, A review of wind speed probability dis-
tributions used in wind energy analysis Case studies in the Canary Islands,
Renew. Sustain. Energy Rev. 13 (2009) 933e955.
[15] T.R. Ayodele, A.A. Jimoh, J.L. Munda, J.T. Agee, Wind power potential in the
coastal region of South Africa, Int. J. Green Energy 10 (2013) 814e834.
[16] S.M. Boudia, O. Guerri, Investigation of wind power potential at Oran,
northwest of Algeria, Energy Convers. Mangement 105 (2015) 81e92.
[17] Z.R. Shu, Q.S. Li, P.W. Chan, Statistical analysis of wind characteristics and
wind energy potential in Hong Kong, Energy Convers. Manag. 101 (2015)
 X. Qing / Renewable Energy 115 (2018) 448e461
644e657.
[18] E.D. Stoutenburg, N. Jenkins, M.Z. Jacobson, Variability and uncertainty of
wind power in the California electric power system, Wind Energy 17 (2014)
1411e1424.
[19] Y. Wan, Analysis of Wind Power Ramping Behavior in ERCOT, 2011. Technical
Report NREL/TP-5500e49218.
[20] C. Kamath, Y.J. Fan, Using data mining to enable integration of wind resources
on the power grid, Stat. Anal. Data Ming 5 (2012) 410e427.
[21] IEC. 61400-12-1, Wind Turbines - Part 12-1: Power Performance Measure-
ment of Electricity Producing Wind Turbines. s.L, International Electro-
technical Commission, 2005. IEC 61400-12-1.
[22] Z.Q. Liu, W.Z. Gao, Y.H. Wan, E. Muljadi, Wind power plant prediction by using
neural networks, in: Proc. IEEE Energy Convers. Congr. Expo, 2012,
pp. 3154e3160.
[23] L. Zheng, W. Hu, Y. Min, Raw wind data preprocessing: a data-ming approach,
IEEE Trans. Sustain. Energy 6 (1) (2015) 11e19.
[24] O.A. Jaramillo, M.A. Borja, Wind speed analysis in La Ventosa, MexicoL a
bimodal probability distribution case, Renew. Energy 29 (2004) 1613e1630.
[25] J.A. Carta, Ramírez, Analysis of two-component mixture Weibull statstics for
estimation of wind speed distributions, Renew. Energy 32 (2007) 518e531.
[26] S. Mathew, K.P. Pandey, V.A. Kumar, Analysis of wind regimes for energy
estimation, Renew. Energy 25 (2002) 381e399.
tiveau, Wind potential assessment of
[27] A. Ilinca, E. McCarthy, J.J. Chaumel, J.L. Re
Quebec province, Renew. Energy 28 (2003) 1881e1897.
[28] S. Shokrzadeh, M.J. Jozani, E. Bibeau, Wind turbine power curve modeling
using advanced parameteric and nonparametric methods, IEEE Trans. Sustain.
Energy 5 (4) (2014) 1262e1269.
461
[29] H. Louie, J.M. Sloughter, Probabilistic Modeling and Statistical Characteristics
of Aggregate Wind Power. Large Scale Renewable Power Generation, Springer,
2014, pp. 19e51.
[30] H. Louie, Characterizing and modeling aggregate wind plant power output in
large systems, in: Proc. 2010 IEEE Power and Energy Society General Meeting,
2010, pp. 1e8.
[31] Wind Turbine Power Performance Testing, 2006. European Standard
EN61400-12-1.
[32] M. Lange, On the uncertainty of wind power predictions - analysis of the
forecast accuracy and statistical distribution of errors, J. Sol. Energy Eng. 127
(2005) 177e184.
[33] G. Liu, K. Tomsovic, Quantifying spinning reserve in systems with significant
wind power penetration, IEEE Trans. Power Syst. 27 (4) (2012) 2385e2393.
[34] Y. Guan, J. Wang, Uncertainty sets for robust unit commitment, IEEE Trans.
Power Syst. 29 (3) (2014) 1439e1440.
[35] S. Dehghan, N. Amjady, A.J. Conejo, Reliability-constrained robust power
system expansion planning, IEEE Trans. Power Syst. 31 (3) (2016) 2383e2392.
[36] Y. Gong, Q. Jiang, R. Baldick, Ramp events forecast based wind power ramp
control with energy storage system, IEEE Trans. Power Syst. 31 (3) (2016)
1831e1844.
[37] C. Kamath, Understanding wind ramp events through analysis of historical
data, in: Proc. 2010 IEEE PES Transmission and Distribution Conf. Expo, 2010,
pp. 1e6.
[38] R. Sevlian, R. Rajagopal, Detection and satistics of wind power ramps, IEEE
Trans. power Syst. 28 (4) (2013) 3610e3620.
[39] Truewind a. AWS Truewind's Final Report for the Alberta Forecasting Pilot
Project, 2008. Wind Power Forecasting PILOT Project.