Wind farm electrical power production model for load ow analysis

Isidoro Segura-Heras, Guillermo Escriv-Escriv

*
, Manuel Alczar-Ortega
Institute for Energy Engineering, Universidad Politcnica de Valencia, Camino de Vera, s/n, edicio 8E, escalera F, 2
a
planta, 46022 Valencia, Spain
a r t i c l e i n f o
Article history:
Received 26 January 2010
Accepted 8 September 2010
Keywords:
Wind energy
Wind farms
Weibull distribution
Correlated wind speeds
Distributed generation
Reliability
a b s t r a c t
The importance of renewable energy increases in activities relating to new forms of managing and
operating electrical power: especially wind power. Wind generation is increasing its share in the elec-
tricity generation portfolios of many countries. Wind power production in Spain has doubled over the
past four years and has reached 20 GW. One of the greatest problems facing wind farms is that the
electrical power generated depends on the variable characteristics of the wind. To become competitive in
a liberalized market, the reliability of wind energy must be guaranteed. Good local wind forecasts are
therefore essential for the accurate prediction of generation levels for each moment of the day.
This paper proposes an electrical power production model for wind farms based on a new method that
produces correlated wind speeds for various wind farms. This method enables a reliable evaluation of the
impact of new wind farms on the high-voltage distribution grid.
2010 Elsevier Ltd. All rights reserved.
1. Introduction
Wind farms are the most common and fastest growing appli-
cation of wind energy [1]. Most installed wind power in Spain is
used for this purpose and a total of nearly 20 GW were operating in
2009. Further massive installation of wind farms is planned in
Spain and Europe in the near future [2].
The typical conguration of a wind farm connected to the
transmission grid is formed of a set of wind generators connected
througha mediumvoltage (MV) network withsharedinfrastructure
for access and control [3]. Transformers connect the wind farms
to the transmission grid are sized according to the rated power
of the plants. Electricity from wind farms differs from electricity
producedbytraditional generators becausethe power owbetween
wind farm and the transmission grid depends on an uncontrollable
source ethe wind. Accurate windforecasts are therefore essential to
forecast generation for each moment of the day.
A wind farm is a complex system with many elements required
for modelling purposes. For the implementation of an electrical
power production model of a wind farm, it is rst necessary to
create a model for the wind speed at each moment. Secondly,
a probabilistic outage model based on the failure probabilities of
the wind farm elements is needed. Finally, the probability of supply
is calculated from the current values of the wind speed; the failure
probabilities; and the wind farmconguration. Acompromise must
also be made in the selected model between the simplication level
and the time needed for calculation.
2. Proposed wind farm production model
A model to calculate production in a wind farm is proposed in
order to evaluate the impact of a wind farm on the power trans-
missiongrid. This model converts windspeeddata intothe electrical
production data injected into the PQ-bus of the grid. The following
factors have been taken into account when designing the model.
- Wind farm generators are asynchronous.
- Wind speeds in the various locations of the wind farms are
obtained using Monte Carlo techniques, considering correla-
tion among wind speeds in the various wind farms, and that
the probability distribution of wind speed is a Weibull distri-
bution [3].
- Power generated by a wind turbine is a function of the wind
speed, according to a power curve given by the manufacturer.
The total electrical power of a wind farm is the sum of the
power generated by all the modelled generators while taking
into account the failure probabilities of each.
- Reactive power consumed by asynchronous generators is
obtained as a function of the generated active power, Qf(P)
[4,5].
The process to develop the wind farm model was implemented
in the following steps.
* Corresponding author. Tel.: 34 963 879 240; fax: 34 963 877 272.
E-mail address: guieses@die.upv.es (G. Escriv-Escriv).
Contents lists available at ScienceDirect
Renewable Energy
j ournal homepage: www. el sevi er. com/ l ocat e/ renene
0960-1481/$ e see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.renene.2010.09.007
Renewable Energy 36 (2011) 1008e1013
- Modelling of the wind speed in the wind farm.
- Correlationamongthewindspeeds inthe consideredwindfarms.
- Modelling the number of wind turbines running in each wind
farm.
In the rst step it is well known that wind speed probability
distribution is generally considered to followa Weibull distribution.
This function depends on three parameters (x
0
, a, and c) as given by
the formula:
Fx 1 exp
_

_
x x
0
a
_
c
_
; x > x
0
(1)
where:
x wind speed.
a scale parameter. The value of this parameter is related with the
average wind speed at the location. If the wind speed probability
distribution is given in m/s then
a0:01; 20
- c shape parameter, also known as the slope, because the
value of c is equal to the slope of the regressed line in
a probability plot. Different values of the shape parameter can
have marked effects on the behaviour of the distribution. In the
case of wind speed probability distribution:c0:25; 2:5.
- x location parameter. In this particular case, as the minimum
wind speed is 0, x
0
0.
It is therefore easy to generate pseudorandom observations
using approaches such as:
U FX; x
0
; a; c 1 exp
_

_
X x
0
a
_
c
_
(2)
while inverting the transformation produces:
X x
0
aln1 U
1=c
(3)
After the generation of pseudorandom observations U of
a uniform distribution (0, 1), the required observations X of a Wei-
bull distribution can be obtained from the previous expression for
specic values of x
0
, a, and c. This simple process is easily
programmable in a computer.
3. Wind speed correlation
There are many situations in which the generation of correlated
random variables is desirable. It is possible to nd in the literature
many studies that develop algorithms that generate multivariate
normal distributions; or multivariate lognormal distributions when
the variables are positive. An alternative to these methodologies is
proposed in Section 3.2.
3.1. Current methods for obtaining correlated wind speeds
Current methods do not produce predictions on a future horizon
and therefore produce a series of wind speeds that are converted
into electrical power outputs by the models which evaluate the
impact of the wind farms on the electrical power system (SEP).
These methods can be classied in two groups.
(a) Methods based on wind speed distribution simulations [4,5].
These methods start from standardised random variables that
are difcult to adapt to a two-parameter Weibull distribution. A
non-standard method based on the Cholesky decomposition of the
covariance matrix is used. This method is only valid for normal
distributions and also presents the following drawbacks.
- Weibull distributions with a xed parameter are used (Ray-
leigh distributions).
- Negative wind speeds appear in some series.
(b) Methods based on chronological wind speed series [6,7].
These methods use temporary wind speed series and this
implies that the speed observed in period t inuences the following
period. In this case, it is not a random sample.
3.2. Proposed method to obtain correlated wind speeds
The method proposed in this paper belongs to type (a) and
a new algorithm has been developed to avoid the described
inconveniences. This method enables the wind speed to be
considered as a randomvariable that follows a Weibull distribution.
This idea is broadly discussed and accepted [3].
The generation method proposed here requires a series of
restrictions on the space of the generated distribution parameters;
in particular, the covariance will not be negative (meaning that
correlation among the wind speeds of the various wind farms must
be positive) [8]. To generate these random values, the developed
algorithm starts with the simulation (using Monte Carlo tech-
niques) of the n m independent variables that follow a Weibull
distribution; with n being the number of wind farms, and m being
a parameter used only for simulation purposes. These variables are
used in a non-lineal programming problem to obtain the Weibull
distribution multivariate with the desired covariance matrix. To
obtain the parameter c the estimator proposed by Menon (1963)
[9] is used. The steps in the proposed method are as follows.
Step 1: Simulate k values of n m independent uniform random
variables (following the scheme mentioned previously)
and being n <m:
U
i
0; 1; i 1; .; n
U
j
0; 1; j 1; .; m
Step 2: Calculate n m independent Weibull randomvariables:
X
i
a
i
$ ln1 U
i

1=c
i
; i 1; .; n
Y
j
a
0
j
$
_
ln1 U
j

_
1=c
0
j
; j 1; .; m (4)
Step 3: Dene n m T matrix of the repeated value 1 (adjacent
matrix).
Step 4: Calculate n Weibull random variables: Z X T * Y
Z
1
X
1
t
11
Y
1
. t
1m
Y
m

Z
n
X
n
t
n1
Y
1
. t
nm
Y
m
(5)
Steps 1, 2, 3, and 4 enable us to obtain the n correlated
random variables that follow a multivariate Weibull distribution.
Each of these variables is a lineal combination of the m1
independent random variables and follows a univariate Weibull
distribution [8].
Step 5: Calculate n scale parameters:
a
1
a
1
t
11
a
0
1
. t
1m
a
0
m

a
n
a
n
t
n1
a
0
1
. t
nm
a
0
m
(6)
Step 6: Calculate n variables:
I. Segura-Heras et al. / Renewable Energy 36 (2011) 1008e1013 1009
H
1

6
p
p
$

1
k1

k
t 1
_
lnZ
1t
m
lnZ1
_
2

H
n

6
p
p
$

1
k1

k
t 1
_
lnZ
nt
m
lnZn
_
2

(7)
In steps 5 and 6, we obtain the multivariate Weibull distribution
parameters: the scale factor a from the proposed lineal combi-
nation (step 5); and the shape parameter H from the estimator
proposed by Menon (step 6) [9].
Step 7: Fix the value of the multivariate Weibull distribution
parameters required to generate:
a
00
1
; .; a
00
n
; c
00
1
; .; c
00
n
Step 8: Add restrictions:
c
00
1

1
H
1
; .; c
00
n

1
H
n
a
00
1
a
1
; .; a
00
n
a
n
(8)
In step 7, the multivariate Weibull distribution parameters that
must be generated are introduced. In the optimal solution, these
parameters must be matched with those calculated in steps 5 and 6.
Step 9: Calculate the covariance matrices S(X), S(Y).
Step 10: Calculate the covariance matrix
SZ SX
_
T*SY*T
0
_
Step 11: Calculate fromthe above data the correlation matrix R(Z)
D
1/2
*S(Z)*D
1/2
, where D is a diagonal matrix and its
elements are the variances of the Z
i
(i 1,.,n) variables.
Steps 8, 9, and10enable us to calculate the correlationmatrix R(Z)
associated with the multivariate Weibull distribution, using the
n m independent randomvariables.
Step 12: Introduce the correlation values of the multivariate Wei-
bull distribution that must be generated: P(Z).
Step 13: Solve the optimization problem:
Min trSZ
s:a: : PZ RZ
Step 7
a
1
; .; a
n
; a
0
1
; .; a
0
m
;
c
1
; .; c
n
; c
0
1
; .; c
0
m
> 0;
t
ij
< i 1; .; n; j 1; .; m
Finally, in this step we solve the non-linear programming
problem that must be optimized to obtain the desired multivariate
Weibull distribution.
The selected objective function, tr(S(Z)), enables us to obtain the
minimal generalized variance of the Weibull multivariate distri-
bution. This distribution has a multivariate grade since we have
only one variability measurement for each variable, and the
combined variability comes fromthe variance-covariance matrix. It
is usual to work with the tr(S(Z)) or with jS(Z)j as variability
measurements of the multivariate distribution.
The proposed non-linear programming problem has been
solved using the solver.dll library, and the algorithm has been
programmed in Visual Basic 6.0. The solver.dll library has imple-
mented the GRG2 code (generalized reduced gradient, Lasdon and
Warren). In this prime method the feasibility stays active during
the optimization process. The portability of this GRG2 library
provides an easy and user-friendly interface for the non-technical
user [10].
This paper presents an optimization algorithm for obtaining
the Weibull multivariate distribution. This procedure uses an
invertible distribution function; as well as Monte Carlo tech-
niques together with articial random variables. It is important
to stress that these random variables are independent and follow
a Weibull distribution, although not necessarily with the same
parameters.
4. Modelling the number of turbines running in wind farms
The number of wind turbines running in wind farms is obtained
by a statistical probability calculus that takes into account the
failure probabilities of the wind farm components. Let us consider
the conguration shown in Fig. 1: groups of wind generators
(G
1
,.,G
N
), with their corresponding protection elements (S
1
,.,S
N
),.
These are connected in parallel to the internal distribution system
(20e30 kV) in groups of two, with their respective transformers
(T
p1
, T
p2
) and protection elements (S
p1
,.,S
pN
). Connection to the
high-voltage distribution grid (132 kV) is made by the substation
transformer. This is composed of transformers with adequate
measurement and protection elements. In Fig. 1, this is simplied
with switches S
d1
and S
d2
and the transformer T
d1
.
For the calculation of the effective probability of power being
generated in the farm, a logical diagram construction method is
applied with a sequential grouping of the model elements in series
and/or in parallel [11].
Using the probabilities of some of the elements not being
operational e namely, the generation units (p
g1
, p
g2
,.,p
gN
); the
protection elements (p
s1
, p
s2
,.,p
sN
); the transformers (p
tp1
,
p
tp2
,.,p
tpN
); and the transformer protection elements (p
p11
,
p
p12
,.,p
pN1
,.,p
pN2
), the probability of some of the generators (p
i
)
not supplying power as consequence of a failure in the branches, is
given by:
Fig. 1. Wind farm model with asynchronous generators.
I. Segura-Heras et al. / Renewable Energy 36 (2011) 1008e1013 1010
p
1
1
_
1 p
g1
__
1 p
tp1
_
1 p
s1

_
1 p
p11
__
1 p
p12
_
p
2
1
_
1 p
g2
__
1 p
tp2
_
1 p
s2

_
1 p
p21
__
1 p
p22
_

p
N
1
_
1 p
gN
__
1 p
tpN
_
1 p
sN

_
1 p
pN1
__
1 p
pN2
_
(9)
Similarly, by considering the elements whose failure would
cause all of the generators to stop; namely, the elements in the
connection branch to the high-voltage distribution grid, S
d1
, S
d2
and
T
d1
(with failure probabilities p
d1
, p
d2
and p
td1
, respectively) a failure
probability (p
r
), is obtained:
p
r
1 1 p
d1
1 p
d2
1 p
td1
(10)
It is also necessary to account for the probability of the wind
generator being out of service, due to maintenance tasks, and this is
dened by the probability p
a
.
With the failure probabilities previously calculated, the proba-
bility of having a wind generator or turbine in a non-operational
status is given by the expression:
p
hi
p
i
$p
r
$p
a
(11)
where p
hi
represents the probability of generator i in wind farm h
being in a non-operational status.
Once p
hi
, is known, the number of turbines T
hi
running in wind
farm h for wind speed z
hi
can be obtained using Monte Carlo
techniques by generating m
h
Bernoulli variables (Y
hj
) for each wind
farm. The sum of these variables will be the binomial variable (T
hi
)
that we require. The process steps are:
1. Generate for every speed Z
hi
, U
hj
yU(0,1), where h 1,.,n;
i 1,.,k and j 1,.,m
h
.
2. If Z
hi
v
h1
then T
hi
0
3. If v
h1
Z
hi
v
h3
then:
If U
hj
1 p
h1
then Y
hj
1, in another case Y
hj
0.
4. If v
h3
Z
hi
then T
hi
0.
5. Calculate T
hi
Y
h1
Y
h2
.Y
h,mh
where:
- k is the number of wind speed values of the generated
distribution.
- n is the number of wind farms.
- m
h
is the number of turbines in wind farm h, being h 1,.,n.
- v
hs
is the limiting speed, differentiating between normal and
extreme situations. In which s 1 is considered the minimum
speed, s 2 is considered high speed and, nally, s 3 is
considered extreme speed.
- p
hs
is the non-operational probability of a turbine in normal or
extreme situations. In which s 1 is considered a normal
situation, and s 2 is considered an extreme situation. This
failure rate is calculated on the basis of:
average number of faults innormal situations (incidents/year).
average number of faults in extreme situations (incidents/
year).
average period of awaiting repair unavailability in normal
situations (h/h year).
average period of awaiting repair unavailability in extreme
situations (h/h year).
average period of unavailability due to maintenance in
normal situations (h/h year).
average period for unavailability due to maintenance in
extreme situations (h/h year).
- h indicates the farms 1, 2, or n.
- T
hi
is the number of available turbines in wind farm h for wind
speed z
hi
.
5. Wind farm electrical power production
Once the wind speeds and the number of available turbines are
obtained with the proposed model, the next step is to obtain the
electrical power generatedineachwindfarmunder those conditions.
Fig. 2. Power curve of a wind turbine.
Table 1
Model simulation results.
Date/h z
1
z
2
z
3
z
4
m
1
m
2
m
3
m
4
P
1
P
2
P
3
P
4
13/09/2009 14:42 14.01 16.82 13.72 14.72 26 28 29 29 21.77 28.00 23.04 27.62
13/09/2009 14:52 7.73 8.94 8.18 9.73 28 28 30 28 5.34 7.72 6.62 9.52
13/09/2009 15:02 6.68 8.20 12.23 9.33 30 29 28 30 3.81 6.43 16.65 9.22
13/09/2009 15:12 9.14 7.69 7.23 7.83 28 29 30 29 8.17 5.45 4.77 5.71
13/09/2009 15:22 4.40 7.09 7.37 6.47 29 30 30 30 0.47 4.52 5.02 3.45
13/09/2009 15:32 10.16 11.03 12.73 14.58 29 30 30 30 10.97 13.86 19.73 27.86
13/09/2009 15:42 4.82 6.83 5.96 6.09 30 29 28 28 1.02 3.93 2.46 2.66
13/09/2009 15:52 6.47 6.23 6.40 6.19 28 30 29 29 3.23 3.07 3.22 2.90
13/09/2009 16:02 3.52 3.19 2.82 3.18 0 0 0 0 0.00 0.00 0.00 0.00
13/09/2009 16:12 7.03 6.04 6.52 8.32 30 29 28 29 4.42 2.67 3.30 6.68
13/09/2009 16:22 10.30 10.38 9.65 9.62 30 30 29 27 11.72 11.94 9.69 8.93
13/09/2009 16:32 5.47 9.92 9.64 9.69 29 30 30 26 1.85 10.69 9.97 8.75
13/09/2009 16:42 9.28 7.69 7.34 9.41 28 28 30 30 8.40 5.27 4.97 9.40
13/09/2009 16:52 1.88 1.83 2.37 2.46 0 0 0 0 0.00 0.00 0.00 0.00
z
h
wind speed in wind farm h, in m/s.
m
h
number of turbines in operational status in wind farm h.
P
h
electric power generated in wind farm h, in MW.
I. Segura-Heras et al. / Renewable Energy 36 (2011) 1008e1013 1011
The capacity C(z) associated with each turbine for a wind speed
z can be evaluated by using various functions: linear, exponential,
or others. The authors have chosen an exponential function with
the following form:
Cz
_

_
0 0 z z
min
R
_

_
1
1exp
_

3
2
,
z
max
z
z
max
z
min
_
1expf
3
2
g
_

_
z
min
z z
max
R z
max
z
(13)
where R is the maximumelectrical power generated by the turbine;
Z
min
is the minimum speed (in m/s) for power generation; Z
max
is
the maximum speed (in m/s) for power generation; and Z
ext
is the
extreme speed, above which the turbine is stopped to avoid failure.
Graphically, the form of this function is shown in Fig. 2.
Finally, Table 1 shows an example of the results obtained with
the proposed model for four wind farms. To match the real data
wind speed measurements, a difference of 10 min has been allo-
cated between each group of generated wind speeds. The wind
farms are located in various parts of the region of Valencia in Spain.
The scale and shape parameters of the wind speed probability
distribution are known for each wind farm. These values have been
used as inputs for the model simulation; and combinations of
simultaneous wind speeds for all the wind farms have been
obtained with very small deviations compared with the simulta-
neous real wind speed data (95% of data compared, deviation is less
than 3%).
6. Application of the model to load ow analysis
This method of obtaining the power generated by several wind
farms using the correlated wind speeds is applied to study the
inuence of wind farms on a load ow analysis of the electrical
transmission grid.
Fig. 3 shows an electrical power network with wind farms
connected to nodes 2 and 3. The main focus is to establish the
power generated by the wind farms, which mainly depends on
local winds. These winds are correlated to reliably evaluate the
working conditions of the interconnection lines. The electrical
power generated by the wind farms connected to nodes 2 and 3 for
each situation is obtained using the proposed model. For example,
on the 13/09/09 at 14:42 when the wind speeds inwind farmsites 2
and 3 were 16.82 m/s and 13.72 m/s, the power generated in those
wind farms was 28 MW and 23.04 MW, respectively (Table 1).
The reactive power of the wind farms is considered as a function
of the real power generated, Qf(P) [4,5].
7. Conclusions
The method proposed in this paper enables the correlated wind
speed of various wind farms to be obtained using a new algorithm.
This algorithm considers wind speed as a random variable
(following a Weibull distribution) and it offers advantages when
compared with the methods currently used to correlate wind
speeds. This method uses the scale and shape parameters of the
wind speed probability distribution for each wind farm as inputs,
and produces combinations of simultaneous wind speeds for all the
wind farms. Deviations are very small when compared with the
simultaneous real wind speed data.
The electrical power produced by the wind farms is calculated
using a probabilistic method and the following data is obtained: the
number of turbines running at each wind farm; the power curve of
the turbines; and the correlated wind speeds.
Electrical power is obtained from the wind speed and the result
can be used to evaluate the power production of the farm. This
information can then be used in load ow analysis to evaluate the
impact of the massive introduction of wind farms on the electric
power transmission grid. With this information we can analyse
from a technical point of view: power ows, power losses, variation
of voltage levels, as well as stability and its contribution to the fault
level. Moreover, we can analyse available transfer capability (ATC)
and similar factors from a planning point of view.
References
[1] Joselin Herbert GM, Iniyan S, Sreevalsan E, Rajapandian S. A review of wind
energy technologies. Renew Sustain Energy Rev 2007;11(6):1117e45.
[2] Garca-Garca M, Paz Comech M, Salln J, Llombart A. Modelling wind farms
for grid disturbance studies. Renewable Energy 2008;33:2109e21.
[3] Freris LL. Wind energy conversion systems. Prentice Hall; 1990.
Fig. 3. Electrical power system with wind farms connected in nodes 2 and 3.
I. Segura-Heras et al. / Renewable Energy 36 (2011) 1008e1013 1012
[4] Feijo A, Cidrs J, Dornelas JLG. Wind speed simulation in wind farms for
steady-state security assessment of electrical power systems. IEEE Trans
Energy Convers December 1999;14(4):1582e8.
[5] Karaki SH, Salim BA, Chedid RB. Probabilistic model of a two site wind
energy conversion system. IEEE Trans Energy Convers December 2002;
17(4):530e6.
[6] Huang Z, Chalabi ZS. Use of time-series analysis to model and forecast wind
speed. J Wind Eng Ind Aerodynam 1995;56:311e22.
[7] Torre MC, Poggi P, Louche A. Markovian model for studying wind speed time
series in Corsica. Int J Renew Energy Eng 2001;3:311e9.
[8] Ronning G. A simple scheme for generating multivariate gamma distributions
with non negative covariance matrix. Technometrics 1977;2(19):179e83.
[9] Menon MV. Estimation of the shape and scale parameters of the Weibull
distribution. Technometrics 1963;5:175e82.
[10] Lasdon LS, Warren AD. Generalised reduced gradient software for linearly and
nonlinearly constrained problems. In: Greenbaerg HJ, editor. Design and imple-
mentation of optimisation software. Sijthoff & Noordhof: Alphen van de Ryn;
1978. p. 335e62.
[11] Billinton R, Allan RN. Reliability evaluation of power systems. New York:
Plenum Press; 1995.
I. Segura-Heras et al. / Renewable Energy 36 (2011) 1008e1013 1013