Skip to main content

Full text of "Statistical analysis of wind speed fluctuation and increments of non-stationary atmospheric boundary layer turbulence"

See other formats


Statistical analysis of wind speed fluctuation and 
increments of non-stationary atmospheric 
boundary layer turbulence 

T. Laubrich^, F. Ghasemi^, J. Peinke^, H. Kantz^ 

^ Max-Planck-Institut fiir Physik komplexer Systeme, 
D-01187 Dresden, Germany 
laubrichSpks . mpg . de 

^ Institut fiir Theoretische Physik der 
Westfiilischen Wilhelms-Universitat Miinster, 
D-48149 Miinster, Germany 

Institut fiir Physik der Universitat Oldenburg, 
D-26111 Oldenburg, Germany 

Abstract 

We study the statistics of the horizontal component of atmospheric 
boundary layer wind speed. Motivated by its non-stationarity, we in- 
vestigate which parameters remain constant or can be regarded as being 
piece-wise constant and explain how to estimate them. We will verify the 
picture of natural atmospheric boundary layer turbulence to be composed 
of successively occurring close to ideal turbulence with different parame- 
ters. 

The first focus is put on the fluctuation of wind speed around its mean 
behaviour. We describe a method estimating the proportionality factor 
between the standard deviation of the fluctuation and the mean wind 
speed and analyse its time dependence. The second focus is put on the 
wind speed increments. We investigate the increment distribution and use 
an algorithm based on superstatistics to quantify the time dependence of 
the parameters describing the distribution. 

Applying the introduced tools yields a comprehensive description of 
the wind speed in the atmospheric boundary layer. 

1 Introduction 

The statistical analysis of turbulent flow has a long tradition and has revealed a 
lot of insight int o the pr op erties of turbulence, starting with the pioneering works 
of iKolmogorovl (|l941al lbl. |l962| ). However, in the transition regimes between 
isotropic turbulence as one idealisation and laminar flow as another idealisation, 
our knowledge is still incomplete. This is even more the case when turbulence 
outside the laboratory is studied. The air flow in the atmosp heric boundary layer 
(ABL ), i.e. in the lowest few 1-2 km of the atmosphere (see Wallace and Hobb^ . 



l2006l) . is strongly influenced by surface roughness and hence orography and land 
use, but even more by geothermal effects through heating from the ground. Both 



1 



effects do not only introduce additional structures into the turbulent flow, but 
also cause non-stationarities because these effects depend, e.g. on the intensity 
of solar radiation and on the direction of the surface wind, which both change 
much faster than large scale pressure differences which generate the overall wind 
conditions. 

In several applications, a better understanding of the statistical properties of 
boundary layer turbulence under realistic conditions is essential, in particular in 
view of the cost efficient use of wind power. One example is that more realistic 
input wind fields than just laminar flows are desired for numerical simulations 
of the flow around an obstacle. Another example is the need of good statistical 
evidence of extreme wind gusts, their relative frequency, their spatial extension, 
and also their temporal correlations for the estimation of loads on structures 
and their expected lifetimes. 

Motivated by these considerations, we will here discuss the detailed analysis 
of bou ndary layer win d fields. Moreover, since theoretical concepts such as 
that of ICastaing et al. (Il990l) were developed for an idealised turbulence, we 



investigate in how far these results hold true for ABL turbulence. We therefore 
consider time series recordings obtained by a single anemometer at fixed height 
above ground. The wind field at position r and time t is denoted by u{f,t). 
The time series is given by the horizontal component 



r, -) + ul f r, -) (1) 



for n = 0, 1, 2, . . . a nd v standing for the measurement frequency. Making use 
of the lTavloi](ll938h hypothesis, i e. temporal correlations can be translated into 



spatial longitudinal correlations, our analysis aims at a quantitative characteri- 
sation of the statistics of horizontal wind speed data. 

Throughout our analysis the non-stationarity of the data plays a major role 
and we intend to answer the question which parameters remain constant or can 
be regarded as being piece-wise constant and how to estimate them. We draw 
the conclusion that natural ABL turbulence is a composition of successively 
occurring close to ideal turbulence with different parameters. 

It is worth ment i oning that the same conclusion was drawn in a recent work 
bv lBoettcher et ah ( 2007 ). However, the authors applied a different statistical 



method and concentrated on the wind speed increments. 

As for experimental data we study win d speed recordings acquired at 10 m 
altitude with a frequency of 8 Hz at the LammefiordI ( 19871 ) site. It clearly 



exemplifies ABL turbulence. The results of data gathered at 20 m and 30 m 
above ground do not differ qualitatively. 

The plan of the paper is as follows. In the first part we study the fluctu- 
ation of the horizontal wind speed around its mean behaviour. Empirically, 
the standard deviation of the fluctuation grows linearly with the mean wind 
speed. We explain a method how to estimate the proportionality factor. The 
second part deals with the statistics of wind speed increments. The empirical 
results are compared to theoretical works which assume the correctness of the 



2 



intermittency hypothesis of turbulence (see Kolmogorov , 1962 : Obukhovlll962 ). 
These works state that the distribution of short time increments is strongly lep- 
tokurtic. The parameters describing this distribution are in good approximation 
piece- wise constant. The superstatistical approach, which the third part of this 
paper deals with, is sensitive enough to actually quantify the dynamics of the 
distribution parameters. Finally, the last section contains the conclusions. 



2 Conditioned Fluctuation Distributions 

The first method which we want to give an introduction to analyses the fluctua- 
tion of the wind speed around its window mean over m = 2m -I- 1 sample points 
with TO = 0, 1, 2, ... . In other words, we consider the fluctuation series 

/(™) = x„ - X (™) (2) 

where 

^ rh 

— / , ^n+k- (3) 
TO '^-^ 



Figure [T] (top row) shows three days of measurement at the iLammefiordl (jlQSTn 



site illustrating the non-stationarity of ABL wind speed. The same kind of 
non-stationarity is inherited in xl™'' so that the mean of the fluctuation //i™^ 
is (at least nearly) constant with n, namely zero. The second row of the figure 
displays the fluctuation series for m chosen exemplarily to be 101. Thus, 
the fluctuation corresponds to the wind speed deviation at time n from the 
12.5 s window mean around n. It can be seen that the fluctuation series is 
centered around zero and that its volatility becomes larger as the wind speed 
Xn, and hence the mean wind speed increases. We are interested in the 

functional dependence of the volatility of the fluctuation from the mean wind 
speed. Hence, collecting the events 

n e AA(™) (V) ^ {n : x^^^ = V} (4) 

allows us to estimate numerically the variance of the set : n G A/''-™-'(V^)| 

for each 24 h sample individually. This variance can be identified with the con- 
ditioned volatility a^^\v)'^ under the assumption that the conditioned variance 

Var fn"'''\V does not depend on n G Af^"^\V). Figure [2] shows the empirical 

result: The proportionality a''^\v) (x V can be verified for each 24 h data 
individually. Figure [2] (right) reveals that the proportionality becomes better 
as V becomes larger. It is shown in figure [T] that the wind speed can be low 
during the night hours where the wind is less turbulent causing a^J^\v)/V to 
be comparably small. 

The proportionality factor differs slightly from day to day. Consequently, 
splitting the set A/"*-™-* (V) into two disjoint subsets A/^' (V) and A/'2™-' (V) whose 



3 



day 1 86 



day 191 



day 192 




4 8 12 16 20 24 
time n [h] 




12 16 20 24 
time n [h] 



1 1 










. 1 






1 I- 






12 16 20 24 
time n [h] 



Fig ure 1: The top row shows the wind speed data for three different days of 
the lLammefiord (1987) measurement. The second row displays the fluctuation 
according to ([2]) for m = 101 which corresponds to a time window of 12.5 s 



day 


181 


-e- 187 


-e- 


182 


-t^ 188 


-a- 


183 


-e- 189 


-9- 


184 


190 


fi 


185 


-e- 191 


-9- 


186 


H5- 192 






> 

6" 




day 


181 


-e- 187 


-e- 


182 


-B- 188 


-s- 


183 


-o- 189 


-9- 


184 


^^- 190 


-9- 


185 


-e- 191 


-9- 


186 


-e- 192 


-O- 



mean wind speed V for m=1 01 [m s 



2 4 6 8 10 12 
mean wind speed V for m=1 01 [ms"^] 



Figure 2: The fluctuation standard deviation a 
(right) as a function of the m 



(left) and a''P\v)/V 
101 window mean wind speed V estimated from 



24 h recordings at the iLammefiordl (jl987f ) site 



4 



union is Af^™\V), ABL data show that the variance of the set {/n™^ : S 
Afi"^\v)} might not coincide with the variance of {/I'"'' : n e A/'j™' (t/)}. We 
therefore need to refine the method. 

From a meteorological point of view, the considered variances are mainly de- 
termined by the stratification and thus the Richardson number. The ABL wind 
field is strongly influenced by geothermal effects changing with time (day/night 
cycle, clouds etc.). We should therefore expect 



Var 



Jn 



a„(™)4"^ (5) 



with a time dependent a„(rn). We assume that the proportionality factor 
changes on a larger time scale than the fluctuation so that it can be approxi- 
mated by being piece-wise constant over Az9 time steps. We therefore assume 

a„(m) « a('i?, m) (6) 

for ■& — ^ < n < ■& + ^ where ?9 represents the time of the day. Defining 



T 

and assuming that the volatility of the fluctuation at time n e A/"^™"* (i?, V) does 
not depend on n, the normalised fluctuation 

r{rn) 

(m) ^ Jn_ .g-) 

with ill™'' 7^ has a time-independent volatility for n e A/"^™' F). We de- 
note the mentioned volatilities by (t|™'' (i?, F)^ and ag"^\'d,V)^ , respectively, 

Tlr\^,V) ^ ^a^p\^,V) (9) 



and write 



so that 



9 V-W y 

a'f\^,V) =ai^,m)V ^ (t^J^\§,V) ^ a{d,m). (10) 

In other words, the standard deviation of the fluctuation grows linearly with 
V if the standard deviation of the normalised fluctuation does not depend on 
V . The latter is equivalent to saying that the volatility of .gi™'' remains con- 
stant for — A'0/2 < n < + A'0/2. Under this assumption we can estimate 
the proportionality factor a{'d,m) by computing the standard deviation of the 

set l^i™^ : n e A/'^™^(i?, F)|. Figure [3] depicts the so obtained proportionahty 
factors over a period of eleven days of measurement. The top panel contains 
a(-(?, m = 101) as a step function of time -d with A?? = 8 and 24 h. The lower 
three panels depict the proportionality factor approximated by a step function 
with 1/2, 2, and 8 h plateaus exemplarily for the days 186, 191, and 192 of 



5 



E 



E 




day 1 86 



120 144 168 
time* [h] 

day 191 



day 192 




4 8 12 16 20 24 
timeiJ [h] 



0.11 
0.1 
0.09 
0.08 
0.07 



Ai» [h] 

1/2 

2 



_l I I L. 



0.1 



II 0.08 
E 



E 



4 8 12 16 20 24 
time* [h] 



0.06 



0.04 



"T r — I r 



A{> [hi 
1/2 
2 



_i I I i_ 



4 8 12 16 20 24 
time* [h] 



Figure 3: Standard deviation of the normalised fluctuation with m — 101 at the 
Lammefiordl ( 1987 ) site estimated in successive periods of lengths Ad. The top 



panel shows the "long-term" behaviour over eleven days of measurement (A-i? = 
8,24h) whereas the lower three panels depict three 24 h samples individually 
(Ai9 = 1/2, 2, 8h) 



6 




Figure 4: Estimated distribution of the fluctuation conditioned o n mean wind 
speed V over m = 101 time steps for the wind data acquired at the lLammefiord 
(|l987l ) site. The crosses symbolise the estimation whereas the dashed hues 
represent Gaussian distributions given in (|12p with a{d,m) being estimated by 
the 24 h standard deviation of the normahsed fluctuation, see (IT51) and figure [3] 
(top panel). The standard deviation of the fluctuation is again plotted on top 
of the boxes. The black dots represent the numerical estimation whereas the 
dotted line corresponds to a line through the origin with slope a{'&, m) 



7 



the measurement. It can be seen that a{i!), m) changes with time leading to the 
question whether its approximation of being constant over a period is suitable. 



We check if treating a{'d,m) as remaining constant over 24 h is a good ap- 
proximation by estimating the histograms of the fluctuation conditioned on 
a window mean wind speed V and comparing it with symmetric Gaussian 
distributions whose standard deviation is proportional to V. The propor- 
tionality factor is set equal to the estimated standard deviation of the set 
{ffi™^ : n e A/l^^li?, V)Y That is, computing 



q^"^Hf\v) = (sif - f(r^)) (11) 

and comparing it with 

^(™)(/|y) = ^ e"2''(''.")^^^ (12) 



2T:a{^,m)V 



where 



a(^?,m) = ^Var j^i™^ :^„^<„<^+^| (13) 

for Ai? ~ 24 h. Figure [4] shows the estimated histograms for m — 101 and i? 
representing day 181, 191, and 192 of the Lammefjord measurement. It can 
be seen that they are in good agreement with the above mentioned Gaussian 
distributions verifying 

a^f'\v)ocV. (14) 

On top of the boxes the conditioned standard deviation of the fluctuation is 
drawn vs. V. The dotted line represents a proportional dependence with an 
a{-d, m) estimated via p^ . It can be concluded that the 24 h estimation yields 
reasonable results. 

As a side remark, the estimated co nstant a(??, m ) is als o referred to as the 



turbulence intensity (TI) as defined in iBurton et al.l (|2004[ ) examined over the 
period T = m/v around the time It is customary in turbulence research 
to decompose the wind speed, in our case the horizontal component, during a 
period of time T as 

u{r,t) = U{r) + u'{r,t) (15) 

where U{r) is the average value of u{f,t) over the period T which is typically 
chosen to be 10 min or 1 h. The TI is defined as the root mean square of 
u'{r,t) over the period T in units of ?7(r) and it states the percentage of the 
mean flow which are represented by the velocity fluctuation. The assumed linear 
relation between the standard deviation of the fluctuation and the mean wind 
speed is in accordance with this interpretation. The TI describes the strength 



^ Under the assumption that E[rE„|y] = V, the mean of the fluctuation vanishes, 

E\fir^\v] =0. 



8 



of the instantaneous turbulence at time d and does for instance depend on the 
weather situation. Hence, it can vary from one measurement period to another 
measurement period. 

The proportionaHty cj™'' (V) oc 1^ is not a generic property of a random 
process. For instance, the variance of the fluctuation around ^ in a white noise 
(wn) process f„ does not depend on V, i.e. a^J!2n(^) ~ const. The independence 
between ^„ and f„+s for s ^ allows to conclude that E [£,n\V] — E [^n+s|V^] = 
V and E[f2|^j ^ E[^2^JF] for |s| < m. The conditioned variance of the 
fluctuation is equivalent to the expectation of the biased variance estimation of 
the set {^ri-Tji, • ■ • , S.n+m} with sample mean V and therefore independent from 
V: 

^ rh 

E [{in -Vf\V]^-Y.¥. [(e„+. - Vf\V] 

s——rh 
~ ^ rh 
1 Y: iin+s - Vf 



= E 

m — 1 



m 

s— — m 



V 



m 



(16) 



In general, any stationary process y„ with mean E [y„] does not show a pro- 
portionality between the standard deviation of the fluctuation and the window 
mean. This is because the variance of the numerically estimated series yi™'' is 
nearly zero for sufficiently large m. The mean of j/i'"'' coincides with the global 
mean E [?/„]. In other words, yi™'' ^ const so that the variance of the fluctua- 
tion conditioned on yi™^ — V can only be evaluated if y = E [?/„] making the 
condition in fact redundant. The distribution of the fluctuation coincides with 
th e distribution of ?/„, s hifted by — E [y„]. 



Renner et al.l (|200l[ ) measured the air speed in an air into air round free jet 
experiment. The acquired data series is an example of stationary laboratory 
turbulence. Figure [5] shows the distribution of the air speed in the free jet 
experiment which is in good agreement with a Gaussian shaped curve. In other 
words, the shown histogram corresponds to one slice of the three-dimensional 
plots in figure m 

As a conclusion, analysing a period — At?/2 < n < + Ai9/2 of ABL 
wind speed data reveals that the distribution of the fluctuation fn""^ around the 
mean speed Xn™^ = is well described by a symmetric Gaussian distribution 
with a standard deviation being proportional to V. The proportionality factor 
0(19, m) can be estimated by evaluating the standard deviation of the normalised 
fluctuation fn"^^ /xll^^ over the period of interest. 

In a further study, we propose a stochastic process which can be used to 
generate a time series having the same fluctuation behaviour as ABL wind 
speed data. The work is still in progress. 

Additionally, a conclusion about wind speed increments Xs-n — Xn+s — Xn 
for sufficient large s can be drawn. If Xn+s and be treated as being 



9 




1.5 2 2.5 3 

air speed [m s'^] 



Fig ure 5: The distribut ion of the air speed measured in the free jet experiment 
(see lRenner et all 120011) . The dashed hne corresponds to a Gaussian curve with 
mean 2.25 ms~^ and standard deviation 0.341 ms~^ 



independent from each other and choosing m such that x„ « 2;„™{,, the in- 
crement Xs-n is of a symmetric normal distribution with a standard deviation 
being proportional to xi™^. Taking the statistics of Xs;„ over a much larger 
time period than the time scale on which Xn™^ changes, e.g. 24 h, the increment 
distribution ps {xs ) corresponds to a superposition of symmetric Gaussians with 
different volatilities. The variance of the volatilities is determined by the vari- 
ance of xl^K It (nearly) vanishes if xl™' ~ const in time making ps(xs) coincide 
with a Gaussian distribution. Otherwise, Ps{xs) is fat tailed. 

The following two sections intend to analyse the increment series, its statis- 
tics, and its limit for large increment lengths. 



3 Increment Statistics 

This section is dedicated to the statistics of wind speed increments. The incre- 
ment series is defined as 

with s denoting the time over which the increment is measured — the increment 
len gth. 

Castaing et al. I (|l990h deduced an analytical expression for the marginal in- 



crement distribution ps{xs) from the assumption of a log-normally distributed 
energy transfer rate in turbulence. Castaing's hypothesis assumes that the in- 
crement series in a small time window is of a Gaussian distribution 

pixM^^e-^^'^. (18) 



10 



In each window the parameter /3 can be regarded as being constant. However, 
it varies between the windows according to the log-normal distribution 

/.(/3) = -=i— e-K^i"^r. (19) 



As a result, the increment distribution of ABL wind speed is given by 

/•OC 

Ps{Xs)= / d(3fs{P)p{Xs\f3) 



It is symmetric, leptokurtic, and described by two positive parameters (3s and 
As which are called position and shape parameter, respectively. The latter is 
directly related to the kurtosis of the increment series by 



ks = = 3c-^= > 3 (21) 

{x^f 

so that the shape parameter can be estimated using 

A?-ln^. (22) 

The position parameter of the increment process can be estimated via the vari- 
ance and kurtosis by 

A = ^# (23) 

In the limit of — > the volatility distribution /s(/3) in (flQl) turns into 

/,(/3)^-^''<5(/3-A) (24) 

making the increment distribution coincide with a Gaussian distribution with 
variance = l/f3s and kurtosis fc^ = 3: 

pM'^^^e-'f^'. (25) 

Figure [6] depicts the increment distribution Psix^ ) for a variety of s a nd for 
three different days (day 186, 191, and 192) of the iLammefiord (1987) mea- 
surement. The fitted Castaing distributions according to and are 
drawn with solid lines. It can be seen that the empirical histograms are in good 
agreement with Castaing distributions. Furthermore, the histograms for the 
data acquired at the days 191 and 192 approach a Gaussian distribution as s 
becomes larger. 



11 



day 191 



day 192 




s 


1 


+ 


2 


X 


4 




8 


□ 


16 





32 


A 


64 


V 


128 


o 


256 


■ 


512 


• 


1024 




2048 


T 



Figure 6: Increment hist ograms Psjxs ) io'r d ay 186 (left), day 191 (centre), 
and day 192 (right) of the Lammefiord ( 19871) measurement. The normalised 
histograms are drawn with points whereas the solid lines represent a Castaing 
distribution with the same variance and kurtosis as the increment series. The 
histograms are shifted and drawn in a semi-logarithmic plot for better visibility 



12 




increment length s [1] 



Figure 7: The position parameter (3s an d shape parameter as a function of 
s for the days 186, 191, and 192 of the iLammefiordl (|l987| ) measurement are 
plotted in the upper and lower panel, respectively 



The intermittency hypothesis and the conclusion drawn in Sec.[2]imply that 
the shape parameter decreases as s gets larger. In case of (nearly) stationary 
turbulence it converges to zero, i.e. 

\l 0. (26) 



Boettcher et al. I (|2003h obtained results from laboratory and ABL turbulence 



showing (I26p . Consequently, with increasing s both the kurtosis kg and in- 
verse variance 1/(7^0 decrease causing the position parameter to decrease, too. 
Figure [7] depicts j3s vs. s and A^ vs. s for the increment distributions of the 
Lammefjord ABL wind speed data. It can be seen that the shape parameter A^ 
of the days 191 and 192 goes to zero as s becomes larger whereas it does not 
reach zero for the day 186 data. In other words, the data acquired at day 186 
have leptokurtic increment distributions at large increment lengths. 

It is worth mentioning that for laboratory turbulence the empirical increment 
histograms are in perfec t agre ement with the hypothetical Castaing shaped 
curves fsee iRenner et al" . 2001 . figure 8). Additionally, laboratory turbulence 



data yield a shape parameter which decreases with s according to a power law 
and approaches zero for sufficiently large s (see Renner et al.l . l200ll . figure 9). 

As a summary, this simple approach showed an agreement between the em- 
pirical histograms and the hypothetical distributions. The effect of increasing 
Gaussianity and hence decreasing shape parameter for increasing s can be seen 
and is well supported by the intermittency hypothesis of idealised turbulence. 



^The variance becomes larger due to = 2a'^(l — 7(s)) where cr^ and 7(-) denote the 
variance and auto correlation function of the series x„, respectively. 



13 




Figure 8: Schematic sketch of the superstatistical approach. Tlic increment 
series is considered to consist of successive normally distributed segments of 
length Ts- Within such a period the mean is considered to be zero and the 
standard deviation (std dev) approximated by the estimation over the interval 



Some measurements show a fat tailed increment distrib ution even for large 
increment lengths s. According to iBoettcher et alJ (|2007l ) and the conclusion 
drawn in Sec. [21 a fluctuating mean wind speed might be the reason for this 
behaviour. It is however the drawback of this technique not being suitable 
to verify a connection between this kind of non-stationarity and the shape of 
increment distributions. 



4 Superstatistics 

We scrutinise the increment series a little further in order to understand the 
different increment statistics behaviour of ABL wind. The method which this 
section has its focus on aims to validate the Castaing hypothesis ([20)1 in another 
way: by computing the, a priori unknown, distribution /s(/3) and comparing it 
with a log-normal distribution. 

Th e concept of superposing two statistics was generalised bv lBeck and Cohen 
and called "superstatistics". It describes a driven non-equilibrium sys- 
tem of an intensive parameter /?, which in our case is the inverse volatility, 
fluctuating on a large spatio-temporal scale T whereas the system itself changes 
on a short spatio-temporal scale t <^T. 



14 



mean local kurtosis k; 



1 / 

day 186 / 4 


y ' "1 y ' ' / 





1 



1 10 100 

window length m 



16 
32 
64 



1000 



K=3 




10 100 
window length m 



1000 



Figure 9: Mean window kurtosis as function of the window size m for different 



increment length s of the ABL wind speed measurement at the iLammefiord 
(1987) site. The dashed line corresponds to k = 3. The intersection kI™^ = 3 is 
an estimation for the large time scale shown in figure [TU] 



Based on this idea Beck et al] ( 2005f) and Queirod ( 2007 ) proposed an algo- 
rithm to approximate the volatility by being piece- wise constant, namely 1//3, 
and treat the increment series as being of a Gaussian distribution in each seg- 
ment. Figure [5] sketches schematically the approximation. 

We use this algorithm to estimate the statistics of (3 from the increment 
series {xs;n)n=a '^^ ABL wind speed data. The essential step is to find the large 
time scale Tg and estimate /3 in each segment. 

The large time scale can be identified with the scale on which the increment 
series is of a normal distribution. Hence, being a measure for Gaussianity, the 
sample kurtosis of a population A is defined as 



KwtA = \A\ X 



(Eae^(«-a)'^ 



(27) 



with a = ■pjX^ae^'* denoting the sample average and |^| denoting the sam- 
ple size. If the population stems from a Gaussian distributed population, the 
kurtosis equals to three if the population size is reasonably large. The above 
defined sample kurtosis is biased with respect to the sample size. If A consists 



15 



of only one or two elements, i.e. A = {a} or A = {ai,a2}, the sample kurtosis 
Kurt^ = 1. In order to find the large time scale, the increment series {xs-n)n=o 
is split into sub-series: int [N/m] sub-series of size m and if necessary one 
remaining sub-series of size less than m. For each sub-series the sample kurtosis 
is computed. The average of the Nm sub-series kurtosis is taken as measure of 
sub-series Gaussianity. Changing m means changing the length of the sub-series 
so that we arrive at the mean sub-series kurtosis written as 

4™^ = {^^rt {Xs-km, ■ ■ ■ ,Xs-(k+l)m-l}) (28) 

being a function of the window length m. As mentioned above, kI™^ = 1 for 
m < 2 . As m gets larger tends to one. This results in a Ki™"* being equal to 
the sample kurtosis of the whole increment series whose kurtosis is larger than 
three because we know from the increment statistics that the increment series 
is of a leptokurtic distribution. Hence, somewhere between 1 < m < N there is 
a value m where ^1™""^°-' is closest to three. 



which, according to Beck et al. ( 2005 ) and lQueiro i (l2007l) . is taken as an esti 



mation for the large time scale Tg. Figure [9] displays the mean window kurtosis 
as function of the window size m for different incre ment lengths s for the ABL 
wind speed measurement at the LammefiordI ( 19871 ) site. It can be seen that it 



increases with increasing m and has an intersection with three. 

As Ts is supposed to be large there should be no problem with using the 
biased kurtosis estimator defined above. Using an unbiased estimator would give 
curves like kI™-* w 3 for to < Ts and ni"^^ > 3 for m > Tg. From a statistical 
point of view it is more involved to find the point Tg that way making the 
algorithm unnecessarily more complicated. 

Knowing the length Ts of the sub-series such that they show in average 
Gaussian behaviour, leads to the question whether the relaxation time of the 
process is small compared to Ts- It is reflected by the short time scale being 
estimated by the decay of the auto correlation function js;t- 

Ts = min {t : js-t < e~^}. (30) 

t=l,...,N-l 

If Ts is much smaller than Ts the existence of two separated time scales is justified 
so that it is suitable to estimate the the variance in each sub-series of length Ts 
by its (unbiased) sample variance and identify its inverse as 

Pk = -r—f (31) 

Var |Xs;fcT,, ■ • ■ ,Xs-^{k+i)T,-ii 

Its estimated distribution reads 

fsW^{6{P-Pk))k (32) 



16 



which, hypothetically, has the shape of a log-normal distribution. Thus, it is 
more convenient to consider the variables 



Afe-ln/3fe (33) 

and their estimated distribution 

h^{A) = {S(A-A,)),. (34) 



The hypothesis reads 



Ho: /i,(A) = ^=L-e-K^r (35) 



with As = In (3s turning (I20p into 

Ps{xs) = 7r^ dAe >-s ) ^ " J. (36) 

This approach provides an alternative way to fit the Castaing parameters by 
ln/3s = As =Mean{Ao,...,AAr^^,_i} (37) 

and 

=Var{Ao,...,AjvA^^_i}. (38) 

Figure [10] shows the estim ated large and sho rt time scale of wind speed 
increment data acquired at the iLammefiord (1987) site as a function of s in its 



top and central panel, respectively. T^, whose estimation is bounded from above 
by the number of points which the time series consists of, generally increases 
with increasing s. It describes the scale on which the increment process is 
of a normal distribution and is hence a measure for Gaussianity. Therefore, an 
increasing large time scale is in full agreement with the approach to Gaussianity 
as the increment length gets larger. Additionally, the comparably slow approach 
to Gaussianity of the data gathered at day 186 can be recovered, cf. figure [71 
The plot in the lowest panel of figure [TOj verifies the existence of two separated 
time scales. It visualises the ratio Ts/ts being of the order of magnitude of 10 
or larger. 

The series (Afc)^^ ^ for the Lammefjord day 191 data is computed ac- 



cording to pTj) and ([33|l . Figure [TT] shows exemplarily that the estimated dis- 
tribution /is (A) for s = 8 is close to Gaussian. But it has a systematic and 
statistically significant deviation. In fact, it has a positive skewness. The non- 
Gaussianity is underlined by the quantile-quantile plot in the top right panel 
of this figure. However, the increment distribution is in good agreement with a 
Castaing shaped distribution, cf. figure [6} This allows the statement that the 
superstatistical approach is sensitive enough to discover small deviations from 
Ca staing's hypothe sis. 



Boettcher et al.l ((20071) showed that the distribution of ABL wind speed in- 



crements can be understood as a superposition of different subsets of isotropic 



17 



large and short time scale 




increment length s [1] 



Figure 10: The top and central panel show the estimated large and short time 
scales, Ts arid Ts in time steps, respectively, for three 24 h data acquired at the 
Lammefiord (|l987t ) site (day 186, 191, and 192). A day consists of 691200 data 
points being the upper limit for the estimation of Tg. The central panel displays 
the short time scale Ts which is estimated by the decay of the auto correlation 
function for small lags. The ratio Ts/t^ is depicted in the lowest panel 



turbulence. Indeed, the depicted A-series in the lowest panel of figure [TT] indi- 
cates that the hypothesis "A is of a normal distribution" might be fulfilled on 
a smaller period than 24 h. Hence, the 24 h increment series {xs-n)n=o 
vided into twelve 2 h sub-samples where each of which represents a time d of the 
day. Each sub-sample was analysed with respect to superstatistics individually. 
After computing and comparing the time scales Ts{d) and Ts{'d), the A-series for 
each sub-sample around i) was extracted and tested for the hypothesis of being 
nor mally distribu ted. This test was done using a Kolmogorov-Smirnov test (see 
e.g. Daniell . [19901 chapter 8). Its test value corresponds to the largest devia- 



tion of the empirical cumulative distribution function (c.d.f.) S's(A,'i9) from the 
corresponding Gaussian c.d.f. <f>o (^^^T7i^) ^^^^ fitted As(??) and Ag(^9). It is 
denoted by 

dKS(^) =SUp 



A 



(39) 



The hypothesis Hq in ([55]) is rejected on a significance level a if 

jKS/ 



The critical value is 1.63 for a = 1 % and N^-si'^) > 40 where the latter denotes 
the number of da ta points in the A-series for increment length s and time (see 
e.g. Daniell . 1990L table A. 18 for tabulated values). 



18 




Figure 11: The t op left panel display s the estimated 24 h A-distribution hs{A) 
for s = 8 of the iLammefjordl ( 1987f ) measurement day 191. hl{A) denotes a 
Gaussian distribution with mean (A) and variance ((A - (A))2). The quantile- 
quantilc plot is shown in the upper right panel. The bottom panel displays the 
A-series for that day. The numbers 1 to 12 illustrate the twelve 2 h sub-samples 
which are individually investigated with respect to superstatistics 



19 



day 191 



I 12 16 

time d of the day [h] 



20 



s 


1 


+ 


64 


V 


2 


X 


128 


O 


4 




256 


■ 


8 


□ 


512 


• 


16 





1024 


A 


32 


A 


2048 


T 




: 1 1 1 1 1 : 


- ^ ^ 

: i 8 1 





24 



a = 1 % 



Figure 12: The top (central) panel shows the mean (variance) of A for each sub- 
sample of the data acquired at day 191 at the LammefiordI (1987) site and for 
several s. The mean and variance can be identified with the position and shape 
parameter, respectively. The solid line corresponds to s = 8 and is therefore re- 
lated to the graphs in figure [TT] The test value of the Kolmogorov-Smirnov test, 
d^^{'&) X •\/ A'a;s(i^), is plotted in the lowest panel. The dashed line corresponds 
to the critical value for the significance level a = 1 % 



20 



day 186 (s=8, V59) 



quantile-quantile plot 




Figure 13: The t op left panel displays the estimated 24 h A-distribution /is (A) 
for s = 8 of the iLammefiordl (|1987) measurement day 186. /i*(A) denotes a 
Gaussian distribution with mean (A) and variance ((A - (A))2). The quantile- 
quantile plot is shown in the upper right panel. The bottom panel displays the 
A-series for that day. The numbers 1 to 12 illustrate the twelve 2 h sub-samples 
which are individually investigated with respect to superstatistics 



Figure \n\ shows the results for each 2 h sub-sample of the Lammefjord day 
191 data plotted as a function of time ^ for a variety of increment lengths s. 
The top and central panel depict the mean and variance of A for each sub- 
sample, respectively. According to (|37p and ([55)1 they can be identified with 
As = In /3s and A^, respectively. The bottom panel shows the test value of 
the Kolmogorov-Smirnov test and the critical value for the significance level 
a = 1 %. The graph allows the conclusion that the hypothesis "the A-series of 
each sub-sample around ■& and for different s is of a normal distribution" can not 
be rejected on a significance level a — 1 %. Moreover, the A^-plots in the central 
panel reveal that for large s the shape parameter of each sub-sample approaches 
zero as it is expected from the intermittency hypothesis. Therefore, it can be 
concluded that Castaing's hypothesis is fulfilled during time intervals of 2 h. 
But it is violated on larger time scales, such as 24 h, due to non-stationarity 
which are reflected by the time dependence of As(??) and A^(??). In other words, 
the distribution shown in figure fTTl is a superposition of Gaussians with different 
means As(i?) and variances A^(i^) for s — 8 and is thus not exactly a normal 
shaped distribution. 

The same analysis was done with the Lammefjord day 186 data which did 
not show a clear cross-over behaviour in figure [6l The superstatistical algorithm 
was used to extract the A-series from the 24 h time series {xs-n)n=o ^'^^ tested 



21 



dayie 



X rxv:; X 

■ ■ . I 



I 12 16 

time d of the day [h] 



I • 



20 





' X + ' ' ' + ' 


X : 
1 '■ 


- ^ g 


1 ^ i i i t M 

1 1 1 1 


! S - 

' 1 : 



s 


1 


+ 


64 


V 


2 


X 


128 





4 




256 


■ 


8 


□ 


512 


• 


16 





1024 


A 


32 


A 


2048 


T 



24 



a = 1 % 



Figure 14: The top (central) panel shows the mean (variance) of A for each sub- 
sample of the data acquired at day 186 at the LammefiordI (1987) site and for 
several s. The mean and variance can be identified with the position and shape 
parameter, respectively. The solid line corresponds to s = 8 and is therefore re- 
lated to the graphs in figure [151 The test value of the Kolmogorov-Smirnov test, 
d^^{'&) X •\/ A'a;s(i^), is plotted in the lowest panel. The dashed line corresponds 
to the critical value for the significance level a = 1 % 



22 



for Gaussianity. The upper left panel of figure [1^1 shows the histogram /is (A) for 
s = 8 which is clearly non-Gaussian shaped. This explains the small deviation 
of Ps{xs) in figure [5] from the fitted Castaing distribution for s = 8. The 
24 h increment series {xs-n)n=o ^^^^ divided into twelve 2 h sub-samples 
where each of which was analysed with respect to superstatistics individually. 
Figure [HI shows the Castaing parameters As{'d) and Xgid) as a function of 
time i). The time span 8 h < < 18 h is the only region where Castaing's 
hypothesis can not be rejected on a 2 h scale with a ~ 1 %: the Kolmogorov- 
Smirnov test value df^ {d) x ■y/iVA;s('i?) is below the critical value and the shape 
parameter X^i^) goes to zero as s gets larger. That means that both Castaing's 
hypothesis and the cross-over behaviour can be recovered in the mentioned time 
span on windows of 2 h period. Outside this region, the length of 2 h for the sub- 
samples is still too large or in other words the resolution is too low for recovering 
a normally distributed A-series. However, going to smaller sub-samples gives 
worse statistics due to the reduced number of data points in each sub-sample. 
Nevertheless, the increments of this day are much more non-stationary than 
the data gathered at day 191. It exemplifies that high non-stationarity of wind 
speed data cause a fat tailed increment distribution for large increment length 
s. 

The data acquired at day 192 at the Lammefjord site show similar behaviour 
to the day 191 data when analysed with respect to superstatistics. 

As a summary, the superstatistical approach is sensitive enough to detect 
sub-regions of the increment series where Castaing's hypothesis is fulfilled. In 
such a region the increment distribution takes the shape given in ([36|l . The 
position and shape parameter change with time as they differ between different 
sub-regions. This makes this approach be capable of determining the dynamics 
of As{'d) = ln/3s(t?) and A^(z9). Moreover, it can be used to verify that wind 
speed data with large fluctuating Castaing parameters have a non-Gaussian 
increment distribution at large increment lengths. 

5 Conclusions 

Our study verified the picture of natural ABL turbulence to be composed of 
successively occurring close to ideal turbulence with different parameters. 

We showed that in good approximation the fluctuation of the the wind speed 
around its window average is of a symmetric normal distribution with a 
standard deviation being proportional to V. The proportionality factor can 
be estimated by the standard deviation of the normalised fluctuation. The 
investigation of the time dependent volatility of the normalised fluctuation leads 
to the time resolution of the proportionality factor. Our analysis showed that 
approximating it by being constant over 24 h is reasonable. However, within 
24 h the mean wind speed changed frequently leading to two separate time 
scales: the time scale on which the mean wind speed changes and the time 
scale on which the proportionality factor between the mean wind speed and the 
standard deviation of the fluctuation can be approximated as being constant. 



23 



The mean speed of stationary laboratory turbulence docs not alter with time so 
that the V dependence of the fluctuation distribution can not be investigated 
by means of such experiments. Nevertheless, the fluctuation is of a normal 
distribution, too, leading to the conclusion that ABL turbulence is a sequence 
of stationary turbulence with varying mean. 

The intermittency behaviour of ABL wind speed was tested using the in- 
crement statistics approach. The coincidence between the empirical histograms 
and the hypothetical distributions makes the increment series in this represen- 
tation look stationary. However, Castaing's intermittency hypothesis involves 
a time scale separation. On a small scale the wind speed increments arc of a 
symmetric normal distribution whose variance alters on a larger scale according 
to a log-normal distribution. 

We checked this time scale separation using a supcrstatistical approach. We 
found that there is a "critical" time scale below which the increment series 
behaves normally distributed and above which the increment series is of a lep- 
tokurtic distribution. In the terminology of supcrstatistics this time scale is 
referred to as the large time scale in contrast to the small time scale reflecting 
the relaxation time of the increment process. If the latter is small compared to 
the large time scale it is possible to estimate the variance of each Gaussian seg- 
ment and analyse their statistics. We found that their logarithm is not exactly 
of a stationary normal distribution but rather of a sequence of normal distri- 
butions with varying mean and variance. This incorporates a third time scale 
into the ABL wind speed increment series on which the log-normal distribution 
of the variances in the sense of Castaing's hypothesis can be approximated as 
being stationary. 

We additionally verified that highly non-stationary turbulence might show 
a fat tailed increment distribution even at large increment length s. 

Acknowledgements. The study was supported by Germany's Federal Min- 
istry for Education and Research (BMBF) under grant number 03SF0314. It is 
part of the joint project "Statistical analysis and stochastic modelling of turbu- 
lent gusts in surface wind" . 

References 

Beck, C., Gohen, E., and Swinney, H. (2005). From time series to superstatistics. 
Phys. Rev. E, 72:056133. 

Beck, C. and Cohen, E. G. D. (2003). Superstatistics. Physica A, 322:267 275. 

Boettcher, F., Barth, S., and Peinke, J. (2007). Small and large scale fluctuations 
in atmospheric wind speeds. Stoch. Environ. Res. Risk Assess., 21:299-308. 

Boettcher, F., Renner, C, Waldl, H. P., and Peinke, J. (2003). On the statistics 
of wind gusts. Bound. -Layer Meteor., 108:163-173. 



24 



Burton, T., Sharpe, D., Jenkins, N., and Bossanyi, E. (2004). Wind Energy 
Handbook. John Wiley. 

Castaing, B., Gagne, Y., and Hopfinger, E. J. (1990). Velocity probability 
density-functions of high Reynolds number turbulence. Physica D, 46:177- 
200. 

Daniel, W. (1990). Applied Nonparametric Statistics. PWS-Kent, second edi- 
tion. 

Kolmogorov, A. N. (1941a). Dissipation of energy in locally isotropic turbulence. 
Dokl. Akad. Nauk SSSR, 32:16-18. 

Kolmogorov, A. N. (1941b). The local structure of turbulence in incompressible 
viscous fluid for very large Reynolds number. Dokl. Akad. Nauk SSSR, 30:299- 
303. 

Kolmogorov, A. N. (1962). A refinement of previous hypotheses concerning the 
local structure of turbulence in a viscous incompressible fluid at high Reynolds 
number. J. Fluid Mech., 13:82-85. 

Lammefjord (1987). Lammefjord data obtained from the Ris0 Na- 
tional Laboratory in Denmark, ,http : //www . r isoe . dk/vea through 
|http : / / www ■ winddata . com 

Obukhov, A. (1962). Some specific features of atmospheric turbulence. J. Fluid 
Mech., 13:77-81. 

Queiros, S. M. D. (2007). On new conditions for evaluate long-time scales in 
superstatistical time series. Physica A, 385:191-198. 

Renner, C, Peinke, J., and Friedrich, R. (2001). Experimental indications for 
Markov properties of small-scale turbulence. J. Fluid Mech., 433:383-409. 

Taylor, G. (1938). The spectrum of turbulence. Proc. R. Soc. Lond. A, 164:476- 
490. 

Wallace, J. M. and Hobbs, P. V. (2006). Atmospheric Science. Academic Press 
Elsevier, second edition. 



25