Journal of Wind Engineering & Industrial Aerodynamics 174 (2018) 424–436

Contents lists available at ScienceDirect

Journal of Wind Engineering & Industrial Aerodynamics

journal homepage: www.elsevier.com/locate/jweia

Stochastic Fourier spectrum model and probabilistic information analysis

for wind speed process
Xu Hong a, Jie Li a, b, *
a
School of Civil Engineering, Tongji University, 1239 Siping Road, Shanghai 200092, China
b
State Key Laboratory of Disaster Reduction in Civil Engineering, Tongji University, China

A R T I C L E I N F O A B S T R A C T

Keywords: From the view point of physical mechanism, stochastic dynamic excitations, such as earthquake and strong wind,
Stochastic Fourier spectrum subjecting on structures can be represented by a physical model with elemental random variables. In this article, a
Fluctuating wind speeds physical model for wind speed process is investigated to obtain the probability density function of wind speed
Shear velocity process. Firstly, the stochastic Fourier spectrum, consisting of wave-number spectrum and phase spectrum, is
Roughness length introduced. The wave-number spectrum displays the energy distribution over frequency domain. The phase
Cut-off wave number
spectrum is viewed as the evolutionary result driven by characteristic velocity of air vortex. After that, the
Probability density evolution method
elemental variables are carefully selected in term of physical relation and the measurement data collected at a
wind observation station is analyzed to gain the statistics of the elemental random variables. With the help of
probability density evolution method (PDEM), the probability density information of wind speed process can be
obtained. The comparison with the measurement data validates the effectiveness of the stochastic Fourier spec-
trum to simulate the wind speed process.

1. Introduction of power spectral density. However, in traditional methods, the proba-

For a large number of structures, such as high buildings and long span
bridges, wind load is one of the most important dynamic excitations
which may causes the loss of the serviceability or even totally failure of
the structure. Because of highly stochastic and non-linearly properties,
the fluctuating wind simulation becomes one of the most important is-
sues in wind excitation modeling and lots of pioneer researchers have
devoted their effort into this field.
In history, Von Karman was the earlier one to model the fluctuating
wind. In his work, by introducing Gaussian process assumption, the
concept of power spectrum density (PSD) was proposed (Von Karman,
1948). Then, the well-known Davenport spectrum was proposed in 1960s
based on the records at different heights (Davenport, 1961). The
following researchers, such as Kaimal, developed this approach to satisfy
the requirements for different structures and terrain (Kaimal et al., 1972).
Virtually, the information of PSD is not enough for description of the
time-domain process since the initial phases are undefined. In general
simulation schemes, such as spectral representation method (Shinozuka,
1971; Di Paola, 1998), it is often assumed that the phases are indepen-
dent random variables uniformly distributed in the domain [0, 2π]. These
works established the frame for simulating fluctuating wind on the base
bilistic characteristics of fluctuating wind process are quite limited,
where the information higher than 2nd order is omitted. Also, researches
have revealed that for strong wind, such as typhoon, behavior of fluc-
tuating wind process may distinctly depart from the Gaussian situation
(Kareem, 1978; Gurley et al., 1996; Gong and Chen, 2014; Hui et al.,
2017). These drawbacks arise from that the PSD models are not capable
of describing the integrated probabilistic information of wind speeds. It is
then desirable to obtain the probability density function which encom-
pass the whole probabilistic information.
In fact, PSD model is a phenomenology based approach so that it is
difficult to represent the behavior of fluctuating wind comprehensively.
Therefore, physical way is needed for exquisite depiction. Along with the
basic idea of physical stochastic system, it is some elemental variables in
the physical model are uncontrollable that the process possesses the
stochastic features (Li, 2006). Based on this idea, Li and Yan (2009) and
Li et al. (2012) propose a stochastic Fourier function which stresses the
key role of physical relation in the stochastic excitations modeling. Ac-
cording to the energy spectrum equation of turbulence, a bilinear
amplitude spectrum is proposed by Li and Yan (2009) and Li et al. (2012)
which display the distribution of energy spectra over inertial sub-range
and energy containing sub-range. Then the phase spectrum is

* Corresponding author. 1239 Siping Road, Shanghai 200092, China.

E-mail address: lijie@tongji.edu.cn (J. Li).

https://doi.org/10.1016/j.jweia.2018.01.021
Received 16 October 2017; Received in revised form 13 January 2018; Accepted 13 January 2018

0167-6105/© 2018 Elsevier Ltd. All rights reserved.

X. Hong, J. Li Journal of Wind Engineering & Industrial Aerodynamics 174 (2018) 424–436

FðΘ; nÞ ¼ jFðΘ; nÞjeiφðΘ;nÞ (3)

where jFðΘ; nÞj is the amplitude spectrum and φðΘ; nÞ the phase spec-
trum.
It is commonly recognized that the amplitude describes the distri-
bution of energy over frequency domain where the phase spectrum
controls the shape of the time history. The amplitude spectrum can be
related to wave-number spectrum if Taylor's hypothesis is accepted:
rffiffiffiffiffi
U
jFðΘ; kÞj ¼ jFðΘ; nÞj
2π
(4)
2π
k¼ n
U
Fig. 1. Illustration of the homogeneous shear flow. To model the wave-number spectrum rationally, basic knowledge of
energy spectrum of turbulence is needed. The energy spectrum equation
established with the help of the concept, starting-time of phase evolution, of turbulence (See Fig. 1) reads (Hinze, 1975):
which gives a physical description of the phase evolution in wind speed
process (Li et al., 2013). The analysis using the measured data collected ∂ dU1
EðkÞ þ ζðkÞ ¼ F3 ðkÞ  2νk 2 EðkÞ (5)
in east China validated approach preliminarily (Yan, 2011). On the other ∂t dx2
hand, the newly developed probability density evolution method pro-
vides a useful tool to capture the probability density function of sto- where EðkÞ ¼ jFðkÞj2 is the energy spectrum of turbulence in wave-
chastic dynamic system (Li and Chen, 2008, 2009). Thus, This article is number domain, F3 ðkÞdenotes the Fourier transform of the third order
dedicated to propose a new approach, based on stochastic Fourier spec- velocity correlation which transport the energy from low wave-number
trum and probability density evolution method, to obtain the probability domain to high wave-number domain, ζðkÞ represents the energy trans-
density function and consequently other probabilistic information of portation caused by the shear in main flow.
wind speed process, which can be of significance in reliability assessment Integrate the equation (5) from 0 to k over wave-number domain, we
of engineering structures. Furtherly, the model is validated by the data can obtain:
collected from an strong wind observation station.
dU1 ∞
ε ¼ 2ν∫ 0 k 2 EðkÞdk 
k k
The contents are arranged as following: the stochastic Fourier spec- ∫ ζðkÞdk  ∫ 0 F3 ðkÞdk (6)
dx2 k
trum mainly deduced from the dynamic equation for the energy spectrum
is introduced and updated in section 2; the statistics of the elemental ∞
where ∫ 0 ζðkÞdk ¼ u1 u2 and ε ¼ u1 u2 dU
dx2 are utilized.
1
random variables are modeled using the wind measurements collected in
Xiamen in section 3; section 4 introduces the probability density evolu- To solve the integral equation (6), the Heisenberg's theorem of energy
tion method which will be applied to research the probability density transfer is introduced (Katul and Chu, 1998):
evolution of wind speed process; section 5 provides the probability rffiffiffiffiffiffiffiffiffi
∞ EðkÞ
∫ 0 F3 ðkÞdk ¼ 2α'∫ k
k k
density distribution of the wind history derived from the physical model dk∫ 0 k2 EðkÞdk (7)
integrated with the probability density evolution method, and the k3
contrast of the simulation and observations demonstrates the effective- where α' is a constant.
ness of this model; moreover, some important statistical characteristics In the inertial subrange, Tchen's assumption is adopted since the
deduced from stochastic Fourier spectrum are also inspected in section5. interaction between the turbulence and the main flow is relatively weak
(Katul and Chu, 1998):
2. Physical model for fluctuating wind speeds
!2 rffiffiffiffiffiffiffiffiffi
dU1 ∞ dU1 ∞ EðkÞ
2.1. Stochastic Fourier spectrum  ∫ ζðkÞdk ¼ α'' ∫k dk
dx2 k dx2 k3
From the view point of physical stochastic system, a physical model
with elemental variables can be established to represent the process where α'' is another constant. Meanwhile, a negligible shear rate can be
assumed. As a result, an approximate solution of the integral equation (6)
logically. For instance, let uðΘ; tÞdesignate a record of fluctuating wind
speeds while Θ is the elemental variables which determine the process in the inertial subrange can be derived as follows:
through the definitive physical model uðΘ; tÞ. Virtually, a Fourier trans-  2=3
8
form of uðΘ; tÞ, denoted by FðΘ; nÞ, contains the same information as the EðkÞ ¼ ε2=3 k 5=3
9α'
original process uðΘ; tÞ according to Winner-Khintchine formula:
In the energy containing subrange, an intense interaction between the
1 ∞
FðΘ; nÞ ¼ pffiffiffiffi ∫ ∞ uðΘ; tÞei2π nt dt (1) turbulence and main flow exists and dominates the integral equation (6)
T comparing to other terms, thus we have:
pffiffiffiffi ∞ rffiffiffiffiffiffiffiffiffi
uðΘ; tÞ ¼ T ∫ ∞ FðΘ; nÞei2πnt dn (2) dU1 ∞ dU1 h k 2 i1=2
∞ EðkÞ
 ∫ ζðkÞdk ¼ α'' 2∫ 0 k EðkÞdk ∫k dk
dx2 k dx2 k3
On account of this equivalence, the Fourier function FðΘ; nÞ can be
defined as the physical model of fluctuating wind speeds. It is believed which combined with (7) leads to:
the source of the randomness of fluctuating wind lies in the randomness
of Θ. Once the Fourier function is established and the distribution of 1 ε
EðkÞ ¼ k 1
elemental variables is identified, the model can be used to carry out the α'' dU1
dx2
analysis for the stochastic system.
Generally, the Fourier function FðΘ; nÞ is a complex and can be To conclude, the solution of equation (6) over energy containing sub-
rewritten as the product of two parts: range and inertia sub-range have the following form:

425
X. Hong, J. Li Journal of Wind Engineering & Industrial Aerodynamics 174 (2018) 424–436

where kc denotes the cut-off wave number, u* denotes the shear velocity
whose physical meaning will be elaborated later.
As for the phase spectrum, since the energy of air vortex depends on
its character velocity, it is reasonable to assume the energy spectrum
involves in the phase evolution, and the larger energy the vortex contains
the faster the phase changes. An illustration is given by Fig. 3, the
characteristic velocity of the vortex with the wave-number
k ¼ 2π n=Ureads:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
vðnÞ ¼ jFðnÞj2 Δn (10)

where jFðnÞj is the corresponding value of amplitude spectrum.

After a period of time, Δτ, the phase change consistent with the vortex
motion can be represented by (Li et al., 2013):

vðnÞΔτ vðnÞ
Δφ ¼ 2π ¼ Δτ ¼ vðnÞkðnÞΔτ
2π l l
Another important assumption is, from the zero-phase-state to the
moment of observation, all the vortex with different wavenumber
Fig. 2. Illustration of energy cascade and amplitude of the Fourier spectrum. experience the same time elapse of Te, called starting time of phase
evolution. Thus, the evolutionary phase spectrum can be represented by:

φðn; Te Þ ¼ vðnÞkðnÞTe (11)

Combining the wave-number spectrum (9) and the evolutionary
phase spectrum (11) together, the history record of fluctuating wind
speeds can be generated by inverse Fourier transform (2).
Whereas he function of the stochastic Fourier spectrum is to simulate
the fluctuating component of the wind speeds. A mean wind profile is
also necessary to model the mean component of the wind speeds and lots
of works have been done in this field (Simiu and Scanlan, 1996). In this
research, the logarithmic mean wind profile is adopted:
u* z
UðzÞ ¼ ln (12)
κ z0
Fig. 3. Illustration of phase evolution.
in which κ  0:4 is the von Karman constant, z0 is the roughness length.
 1
As we can see in the next section, the stochastic Fourier spectrum does
k k  kc
EðkÞ  (8) have close physical relations with the mean wind profile.
k 5=3 k > kc

in which kc define the boundary between inertia sub-range and energy 2.2. Elemental random variable
containing sub-range.
Thus, Li and Yan (2009) proposed a bilinear wave-number spectrum It is indicated by formula (9)–(12) that there four degrees of freedom
to model the amplitude of complex FðkÞ(See Fig. 2): remain undefined in the whole model. In the earlier model proposed by
Li and Yan (2009), the 10-min mean wind speed at 10 m height U10,
8 pffiffiffiffiffi
> u* roughness length z0, the starting time of phase evolution Te and the
>
> α1 k 1=2 k < kc
< ðκzkc Þ1=3 cut-off wave number kc are selected as elemental variables. Investigation
jFðkÞj ¼ pffiffiffiffiffi u* 5=6
(9) suggests that it is better to replace the mean wind speed with shear ve-
>
>
: α1
> k k  kc locity as a basic random variable. An intuitive requirement for elemental
ðκzÞ1=3
variable is that it must represent some global physical properties of the

Fig. 4. Illustrations for the SWOS. (The location of SWOS is stressed by the red line). (For interpretation of the references to colour in this figure legend, the reader
is referred to the Web version of this article.)

426
X. Hong, J. Li
Fig. 5. The layout of the SWOS.
system, which the mean wind speed seems failing to meet. Actually, the
function of U10 is to define the intensity of the wind field. And the shear
velocity u* reflect the momentum transfer, closely related to the field
intensity, between the mean wind speed at different height caused by the
vortex in the atmosphere, that is, the eddy viscosity (Simiu and Scanlan,
1996). In other word, the shear velocity u* connected the large-scale
description for mean wind and small-scale description of the instant
wind field. From this point of view, the roughness length z0 designates
the characterized scale of the vortex at the bottom of the boundary layer,
induced by the friction between the atmosphere and the ground surface
(Dyrbye and Hansen, 1996).
From the above discussion, the shear velocity u* , the roughness length
z0, the cut-off wave number kc and the starting time of phase evolution Te
are chosen as the elemental random variables of the stochastic Fourier
model. The statistics of these random variables are necessary for further
probabilistic analysis of wind process and the identification of the sta-
tistics are implemented in the next section.
Fig. 6. The relationship between the shear velocity at different heights (unit: m/s).
427
Journal of Wind Engineering & Industrial Aerodynamics 174 (2018) 424–436
3. Identification and statistics of elemental random variables
3.1. Observation station
A Strong Wind Observation Station (SWOS) was established in
Dadeng island (24 330 900 N, 118 190 4700 E) of Xiamen in 2015, which is
located at southeast part of China, to investigate the characteristics of
natural wind (Peng et al., 2018). To identify the probabilistic distribution
of above basic random variables, the data collected at the SWOS from
June 28, 2015 to August 19, 2015 is utilized in this study.
In order to better appreciate the terrain condition of the SWOS,
photos of the site are extracted from the Google Map and presented in
Fig. 4, in which the SWOS is highlighted by a red line and a ruler is
attached. Within a 250 m range, the SWOS is surrounded by agricultural
areas while there is an orchard (4–5 m height) towards south-east. Out of
this region, there are numerous low rise residential buildings (3-5storeys
in height) distributed nearby except the east. Overall, the site can be
regarded as terrain B (suburban area) according to Chinese National Load
Code.
The SWOS consists of 4 meteorological tower denoted by P1, P2, P3
and P4, whose horizontal layout is shown in Fig. 5. Two supersonic an-
emometers were mounted at 10 m, 20 m along the height of the P1 tower
and P4 tower, while four supersonic anemometers were mounted at
10 m, 20 m, 30 m, 40 m along the height of the P2 tower and P3 tower.
The sample frequency of all these anemometers is 10 Hz. The analysis
conclusions from P1, P2 and P4 are consistent with that from P3 so that
the following contents in the paper mainly focus on the data observed on
P3.
3.2. Identification of elemental random variables
The basic criterion for identifying the elemental variables is to
X. Hong, J. Li
Fig. 7. Identification process of basic parameters.
minimized the distance defined by two-norm differences of Fourier
amplitude spectrum as well as mean wind profile between observations
and models. For instance, let Fm(n) be the amplitude of Fourier spectrum
of a sample, the distance between Fm(n) and modeled one F (Θ, n)
determined by Θ ¼ ðu* ; kc Þ through formula (4) and (9) can be written as:
JðΘÞ ¼ kFðΘ; nÞ  Fm ðnÞk2 (13)
Thus, the shear velocity u* and cut-off wavenumber kc are readily to
be identified by minimizing the distance JðΘÞ. The similar approach is
implemented to identify the mean wind profile and relevant elemental
variables.
It is revealed the different values of shear velocity u* identified from
the stochastic Fourier spectrum at four heights are in good consistency at
most levels while u* is a little bit larger at 10 m height (See Fig. 6). This
conforms with the assumption of log-law profile that the shear velocity at
Fig. 8. A typical record of wave number spectrum.
428
Journal of Wind Engineering & Industrial Aerodynamics 174 (2018) 424–436
the surface layer is almost identical. Therefore, it is reasonable to take
shear velocity u* in the wind field as a constant, which is equal to the
average value of u* at different heights. Then the average one is
substituted into formula (12) to identify the log-law mean wind profile as
well as roughness length z0. Since the mean shear velocity provides a
better approximation at higher level, the identified log-law profile will fit
better to the measurement than lower heights. A diagram in Fig. 7 is used
to show the identification process. The typical identifications of wave-
number spectrum and mean wind profile are given in Fig. 8 and Fig. 9.
A preliminary study on the data analysis of the identification results
has been already carried out and revealed that the samples can be clus-
tered into two groups distinctly (Hong et al., 2017). It is shown that the
clustering phenomenon arises because of the disturbance by the
observing tower and the perturbed measurement is excluded according
to the cluster analysis in this study to avoid unreal information. There are
720 records remaining valid for every height of P3.
Out of the consideration of the validating the model, the measure-
ments are randomly divided into two groups with same size, one of which
is used to model the distribution of the variables, while another one
performs as contrast.
The histograms of kc of the modeling group for every height are
shown in Fig. 10, which can be modeled by log-normal distribution. Also,
the logarithmic mean value as well as the logarithmic standard deviation
displays a linear dependence with height in Fig. 11-Fig. 12. Therefore, it
is suitable to model the distribution of kc as a log-normal one dependent
with respect to height:
1 ðln xμÞ2
f ðxÞ ¼ pffiffiffiffiffie 2σ2 (14)
xσ 2π
The logarithmic mean value and the standard deviation are calculated
as follows:
μ ¼ 0:016z  0:81
σ ¼ 0:0036z þ 0:19
X. Hong, J. Li Journal of Wind Engineering & Industrial Aerodynamics 174 (2018) 424–436

1
xa1 eb ;
x
pu* ðxÞ ¼ a ¼ 4:23; b ¼ 0:12m=s (16)
ba ΓðaÞ
It should be pointed out that Te is usually as the same scale as 108 s,
and it can be assumed that the local terrain have little impact on Te and
herein the Gamma distribution which is proposed by previous study
(Yan, 2011), is utilized in this article:

1
xβ1 eθ
x
pTe ðxÞ ¼ (17)
θβ ΓðβÞ

with a scale parameter θ ¼ 8:2  108 s, and a location parameterβ ¼ 1:1.

4. Probability density evolution method

A newly developed probability density evolution method (PDEM) (Li

and Chen, 2008, 2009) presents a useful tool to analysis probabilistic
information of fluctuating wind speed process. Herein, the probability
density evolution method will be briefly introduced as following.
Without loss of generality, the expression for stochastic excitations is
readily generalized to any stochastic dynamic system:
Fig. 9. A typical record of mean wind profile.
X ¼ HðΘ; tÞ (18)

in which z designate the height with the unit “m”, and both σ and μ has where Θ is the elemental random variables, X is the physical quantities,
the unit of “m1”. and Hð⋅; tÞ is the physical model. It is less difficult to kwon that the
The histograms of z0 and u* of the modeling group are given in Fig. 13 Fourier spectral function is just an example of the dynamic systems.
and Fig. 14. Investigation indicates that the roughness length z0 can be According to the principal of preservation of probability (Li and Chen,
modeled by the lognormal distribution: 2008), in the probabilistic spaceðX; ΘÞ, the united probabilistic distri-
bution of X and Θ will obey the following equation:
1 ðln xμÞ2
pz0 ðxÞ ¼ pffiffiffiffiffie 2σ2 ; x  0; μ ¼ 0:32; b ¼ 0:54 (15)
xσ 2π D
∫ pXΘ ðx; θ; tÞdx dθ ¼ 0 (19)
Dt Dt DΘ
and the shear velocity u* is well fitted by Gamma distribution:

Fig. 10. The histogram of cut-off wave

number kc at different heights (unit: m1).

429
X. Hong, J. Li Journal of Wind Engineering & Industrial Aerodynamics 174 (2018) 424–436

Fig. 13. Statistics of roughness length z0.

Fig. 11. Variation of the logarithmic mean value of cut-off wave number kc
with height.

Fig. 14. Statistics of shear velocity u*.

Fig. 12. Variation of the logarithmic standard deviation of cut-off wave written as:
number kc with height.
∂pXΘ ðx; θ; tÞ _ ∂pXΘ ðx; θ; tÞ
þH ¼0 (22)
∂t ∂x
where pXΘ ðx; θ; tÞ is the joint probability density function (PDF) of the
system, Dt  DΘ is an arbitrary domain in the probabilistic space. It can And the density function of X, which is of most interest, is just the
be shown that the rate of change of the material integral in equation (19) marginal distribution density of ðΩ; XÞ:
can be rewritten as (Li and Chen, 2009) (See Appendix A):
pX ðx; tÞ ¼ ∫ pXΘ ðx; θ; tÞdx dθ (23)
! ΩΘ
D ∂pXΘ ðx;θ;tÞ X _ ∂pXΘ ðx;θ;tÞ
∫ pXΘ ðx;θ;tÞdx dθ ¼ ∫ Dt DΘ þ Hl dx dθ To solve the GDEE, the solution of physical model for every θ is
Dt Dt DΘ ∂t l
∂xl
necessary on account that the “velocity” drives the evolution of proba-
(20) bility density. Nevertheless, to a large number of problems, the analytical
solution is unavailable so that the numerical approach may be the most
with the help of the arbitrariness of Dt  DΘ , we have:
feasible way. The primary solving procedure can be concluded as follows
∂pXΘ ðx; θ; tÞ X _ ∂pXΘ ðx; θ; tÞ (Li and Chen, 2009):
þ Hl ¼0 (21)
∂t l
∂xl step1: Select the representative point sets in elemental random vari-
ables space and calculate the assigned probability for each point. In this
with the initial condition: article, number theoretical method is adopted;
step2: Calculate the deterministic process of every point to gain its
pXΘ ðx; θ; t0 Þ ¼ δ½x  Hðθ; t0 Þ pΘ ðθÞ “velocity”;
step3: Solving GDEE using the finite difference method for each
It is just the generalized density evolution equation (GDEE). In many
representative point;
cases, only one quantity is concerned, therefore the equation can be

430
X. Hong, J. Li
step4: Integrate all the results in step3 to capture the evolutionary
probability density.
Noting that the extreme value of the random process may occur at
different moment, it is hard to reach its probability information through
the evolutionary probability density and a few mathematic techniques
are required. Define the equivalent extreme value event as (Li et al.,
2007):
Yðθ; tÞ ¼ max Xðθ; τÞsin t; t 2 ½0; π =2 speeds rather than merely the fluctuating part. Take Fourier spectral
τ
It can be easily shown that the equivalent extreme value event have
the same extreme value and probability information as H (Θ,t), and all
the equivalent extreme value event gets its extreme value at the right end
of the interval, which make it possible to apply PDEM for the extreme
value.
Specifically, in the probability density evolution analysis of wind
speed process, the elemental variables Θ can be displayed as
ðu* ; kc ; z0 ; Te Þ and the stochastic dynamic system Hðθ; t0 Þ is comprised of
stochastic Fourier spectrum and mean wind profile. For every assigned
ðu* ; kc ; z0 ; Te Þ, the amplitude and the phase of the Fourier spectrum can
be obtained by equation (9) and equation (11); then, the fluctuating wind
speed can be achieved through the inverse Fourier transform of Fourier
spectrum using equation (2); and it is readily to simulate a wind speed
process by adding a mean wind speed part, generated by equation (12),
to the fluctuating wind speed; consequently, using the GDEE (equation
(22)) and the foregoing numerical method, the probability density of
wind speed process can be derived.
5. Probabilistic information of wind speeds
With the help of PDEM, it is very easily to get a PDF which gives an
431
Journal of Wind Engineering & Industrial Aerodynamics 174 (2018) 424–436
Fig. 15. PDF of wind speed at 10 m height.
intact description for wind fluctuation. Based on the PDF, it is readily to
obtain other probabilistic information.
5.1. Probability density function of wind speed process
Given that the wind load on structures is the result of the total wind
speeds, it is appropriate to analyze the distribution of the total wind
model and log-law mean wind profile as the dynamic system in equation
(18), then use the distribution in section 3.2 for each elemental variable
and carry out the analysis of PDEM, consequently the evolutionary PDF of
wind speeds with duration of 10 min can be obtained.
Figs. 15–18 shows the histograms of the measurements and the
probability density by PDEM at two moments (200s and 400s) for
different heights. These show that the PDEM solution shows a good
agreement with the measurements which demonstrate the proposed
model is capable to describe the aggregate probabilistic information of
fluctuating wind excitations. It is seen that the PDFs of wind speed at
different height are of the similar shape and the deviation increase with
along with the height which reflects the wind speeds get larger at the
higher position. Moreover, the total wind speeds exhibit the feature of
asymmetry obviously. The contour maps of PDF in Fig. 19 indicate that
the main characteristics of the wind process may be stationary while
some detail information display time-varying behavior.
In addition, to judge the match degree, generalized extreme value
distribution and Gamma distribution are fitted to the measurements at
every instant respectively. Then the relative entropies are calculated
between each result and measurement (see Fig. 20). In most cases, the
PDEM results achieve as good agreement as two specified distributions,
and it is clear the proposed model provides better simulations as the
Fig. 16. PDF of wind speed at 20 m height.
X. Hong, J. Li
Fig. 19. The contour of the PDF surface of wind speed.
height increases. It seems this behavior arises from the assumption that
the shear velocity is identical along the height. As previous analysis in
section 3.2, the average value of shear velocity at different heights is
quite close to values at higher level and underestimate the shear resis-
tance near ground. Thus the log-law mean wind profile fit better to the
432
Journal of Wind Engineering & Industrial Aerodynamics 174 (2018) 424–436
Fig. 17. PDF of wind speed at 30 m height.
Fig. 18. PDF of wind speed at 40 m height.
higher level value. Consequently, the identified distribution of elemental
random variables will result in the better fitness of higher level distri-
bution of total wind speed between model and measurement at higher
level.
Although relative entropy of PDEM results are a little bit larger than
X. Hong, J. Li Journal of Wind Engineering & Industrial Aerodynamics 174 (2018) 424–436

5.3. Power spectrum density

It is known that power spectrum density equals to the ensemble

average of Fourier spectrum:

1  
SðnÞ ¼ lim E FðΘ; tÞj2 (24)
T→∞ T

where E½⋅ means the average in the probabilistic space of elemental

variable Θ.
Then the power spectrum density derived by the proposed model can
be calculated according to equation (24). Fig. 22 shows the model results
and the measurements in contrast group at different heights. Here, we
use the Kaimal spectrum as reference:

nSðnÞ 200f
¼ (25)
u2* ð1 þ 50f Þ5=3

where SðnÞ is the power spectrum density of wind speed process, f ¼

nz=UðzÞ is the non-dimensional frequency. These results indicate the
Fig. 20. Relative entropy between the simulations and measurements.
secondary moment of the model behaves similarly as the naturally wind
and common spectrum in the frequency domain which is concerned by
the structural engineering.
the fitted distribution at 10 m height, it will not cause much influence
since the wind load at this height is not dominant one.
5.4. 3-s gust factor

3-s gust factor is important parameter to describe the stochastic

5.2. Extreme value distribution
characteristics of wind flows. It can be calculated by the following
equation:
Further, with the help of extreme value event, the probability of
extreme wind with a duration of 10 min can be simulated by the pro- maxðu3s Þ
posed model. Fig. 21 shows the comparison of the extreme wind between G¼ (26)
U
model results and observations. It is seen the multi peak performance can
be captured by the PDEM solution. However, the model results seem to where maxðu3s Þ is the largest mean wind speed over 3s within the wind
possess a larger tail which needs further investigation. speed process.

Fig. 21. PDF of extreme wind speed.

433
X. Hong, J. Li Journal of Wind Engineering & Industrial Aerodynamics 174 (2018) 424–436

Fig. 22. Comparison of power spectrum

density.

Fig. 23. PDF of 3-s gust wind factor.

Fig. 23 are the PDFs of 3-s gust factor by the PDEM for every height, furtherly show the model can reproduce the detail statistical information
and the measured ones are also presented. The overall good agreements in wind speeds. It is noted that at the higher level, the modeled quantities

434
X. Hong, J. Li Journal of Wind Engineering & Industrial Aerodynamics 174 (2018) 424–436

match the measurements better, which means the model may get a better
prediction for higher position.
6. Concluding remarks
In the present paper, the stochastic Fourier spectrum model for fluc-
tuating wind speed is investigated and a few conclusions can be drawn as
follows:
● In traditional methods of modeling the fluctuating wind speeds, only
2nd order characteristics of the probabilistic information can be re-
flected. It is because these methods are phenomenological approach.
To represent the behavior of wind speed process comprehensively, it
is desirable to achieve the probability density function of wind speed
which contain the whole probabilistic information. Thus, the physical
way need to be introduced into the model. With the help of the energy
spectrum equation of turbulence and the concept of phase evolution,
the stochastic Fourier spectrum is investigated to model the wind
speed process physically.
● It is believed the elemental variables of the stochastic function must
be of global physical significance. Thus, shear velocity u* is used to
take place of mean wind speed at 10 m height as an elemental vari-
able. To fully include all randomness, the cut-off wave number is also
viewed as elemental variable. Then, Based on data collected at SWOS,
the appropriate set of measurements is extracted according to the
clustering phenomenon and the probability density functions of the
elemental variables are statistically obtained.
● The probability density evolution analysis is carried out for the wind
history, showing that the model can depict the micro structures of the
stochastic wind process and it is hopeful to be a physical basis for
more accurate structural reliability analysis under wind loads.
● Key Statistical characteristics of simulating winds by the proposed
model are calculated based on the PDF and compared with the
measurements which validates its capacity of containing main infor-
mation in the natural wind process.
Acknowledgements
The supports of the National Natural Science Foundation of China
(Grant No. 51538010) are highly appreciated.

Appendix A

Derivation for generalized probability density evolution equation using integral transformation:

D
∫ pXΘ ðx; θ; tÞdx dθ
Dt Dt DΘ

D
¼ ∫ pXΘ ðx; θ; tÞjJjdx dθ
Dt Dt0 DΘ

D DjJj
¼ ∫ Dt DΘ
jJj pXΘ ðx; θ; tÞ þ pXΘ ðx; θ; tÞ dx dθ
0 Dt Dt

noting that the velocity of sample only depends on the elemental random variables in the augmented space, the previous equation can be rewritten as:

D DjJj
∫ Dt DΘ
jJj pXΘ ðx; θ; tÞ þ pXΘ ðx; θ; tÞ dx dθ
0 Dt Dt
"
∂ X
¼ ∫ Dt DΘ jJj pXΘ ðx; θ; tÞ þ jJjpXΘ ðx; θ; tÞ H_ l
0 ∂t l

#
X ∂H_ l
þpXΘ ðx; θ; tÞ dx dθ
l
∂xl

" #
∂ X
¼ ∫ Dt pXΘ ðx; θ; tÞ þ pXΘ ðx; θ; tÞ H_ l jJjdx dθ
0
DΘ
∂t l

" #
∂ X
¼ ∫ Dt DΘ p ðx; θ; tÞ þ pXΘ ðx; θ; tÞ H_ l dx dθ
∂t XΘ l

Appendix B. Supplementary data

Supplementary data related to this article can be found at https://doi.org/10.1016/j.jweia.2018.01.021.

References Di Paola, M., 1998. Digital simulation of wind field velocity. J. Wind Eng. Ind. Aerod. 74,
91–109.
Dyrbye, C., Hansen, S.O., 1996. Wind Loads on Structures. Wiley.
Davenport, A.G., 1961. The spectrum of horizontal gustiness near the ground in high
Gong, K., Chen, X., 2014. Influence of Non-Gaussian wind characteristics on wind turbine
winds. Q. J. R. Meteorol. Soc. 87, 194–211.
extreme response. Eng. Struct. 59, 727–744.

435
X. Hong, J. Li
Gurley, K.R., Kareem, A., Tognarelli, M.A., 1996. Simulation of a class of non-normal
random processes. Int. J. Non Lin. Mech. 31 (5), 601–617.
Hinze, J.Q., 1975. Turbulence. McGraw-Hill, New York.
Hong, X., Peng, Y., Li, J., 2017. Stochastic fourier spectrum based parameter cluster
analysis for fluctuating wind speeds. J. Tongji Univ. (in Chinese) under review.
Hui, Y., Li, B., Kawai, H., Yang, Q., 2017. Non-stationary and Non-Gaussian
characteristics of wind speeds. Wind Struct. 24, 59–78.
Kaimal, J.C., Wyngaard, J., Izumi, Y., Cote, O.R., 1972. Spectral characteristics of surface
- layer turbulence. Q. J. R. Meteorol. Soc. 98, 563–589.
Kareem, A., 1978. Wind Excited Motion of Tall Buildings. Colorado State University, Fort
Collins, Co.
Katul, G., Chu, C.R., 1998. A theoretical and experimental investigation of energy-
containing scales in the dynamic sublayer of boundary-layer flows. Boundary-Layer
Meteorol. 86, 279–312.
Li, J., 2006. A physical approach to stochastic dynamical systems. Sciencepaper Online 1
(2), 95–104.
Li, J., Chen, J.B., 2008. The principle of preservation of probability and the generalized
density evolution equation. Struct. Saf. 30, 65–77.
Li, J., Chen, J.B., 2009. Stochastic Dynamics of Structures. John Wiley & Sons.
436
Journal of Wind Engineering & Industrial Aerodynamics 174 (2018) 424–436
Li, J., Yan, Q., 2009. Physical models for the stochastic dynamic excitations of structures:
in the case of fluctuating wind speed (in Chinese). Eng. Mech. 26, 175–183.
Li, J., Chen, J.B., Fan, W.L., 2007. The equivalent extreme-value event and evaluation of
the structural system reliability. Struct. Saf. 29, 112–131.
Li, J., Yan, Q., Chen, J.B., 2012. Stochastic modeling of engineering dynamic excitations
for stochastic dynamics of structures. Probabilist. Eng. Mech. 27, 19–28.
Li, J., Peng, Y., Yan, Q., 2013. Modeling and simulation of fluctuating wind speeds using
evolutionary phase spectrum. Probabilist. Eng. Mech. 32, 48–55.
Peng, Y., Wang, S., Li, J., 2018. Field measurement and investigation of spatial coherence
for near-surface strong winds in Southeast China. J. Wind Eng. Ind. Aerod. 172,
423–440.
Shinozuka, M., 1971. Simulation of multivariate and multidimensional random processes.
J. Acoust. Soc. Am. 49 (1B), 357–368.
Simiu, E., Scanlan, R.H., 1996. Wind Effects on Structures. Wiley.
Von Karman, T., 1948. Progress in the statistical theory of turbulence. Proc. Natl. Acad.
Sci. 34, 530–539.
Yan, Q., 2011. Physical Mechanism Based Research on Stochastic Wind Field Model (in
Chinese). Tongji University, Shanghai.