Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17

Contents lists available at ScienceDirect

Journal of Wind Engineering & Industrial Aerodynamics

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

Prediction of wind loads on high-rise building using a BP neural network

combined with POD
Huang Dongmei a, b, *, He Shiqing a, He Xuhui a, b, Zhu Xue a
a
School of Civil Engineering, Central South University, Changsha, Hunan, 410075, China
b
High-speed Railway Construction Technology National Engineering Laboratory, Central South University, Changsha, Hunan, 410075, China

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

Keywords: According to the limited wind tunnel test results to obtain detailed data of surface wind pressure on buildings has
High-rise building important significance for the accurate calculation of cladding wind pressure and wind-induced response of
POD structures. In this paper, a backpropagation neural network (BPNN) combined with proper orthogonal decom-
BPNN
position (POD-BPNN) is proposed for the prediction of the mean, root-mean-square (RMS) pressure coefficients
Rigid model synchronisation pressure wind
and the time series of wind-induced pressures on a building surface, respectively. In this study, simultaneous
tunnel test
Time series of wind-induced pressures pressure measurements are made on a high-rise building model in a boundary layer wind tunnel and parts of the
model test data are used as the training input–output sets for BPNN and POD-BPNN models. Comparisons of the
prediction results by the POD-BPNN approach and those from the wind tunnel test demonstrate that the BPNN
combined with POD method can successfully and efficiently predict the time series of pressure data on all surfaces
of a high-rise building on the basis of wind tunnel pressure measurements from a certain number of pressure taps.

1. Introduction
With the development of construction technology and engineering
materials, more super tall buildings and large-span roof structures have
been built all over the world. These structures are super-flexible, light-
weight, have low damping ratios, and low fundamental frequencies,
which make them very sensitive to wind load. Under the action of strong
typhoons, failures in glass curtain walls and roofs have happened
frequently, and the fact that such structures suffer strong wind-induced
vibration is commonplace. Therefore, it is necessary to understand the
detailed characteristics of such wind effects on such structures.
In design of some slender structures such as super-tall, large-span roof
structures, and so on, usually it is necessary to obtain more detailed wind
load information: this generally comes from wind tunnel tests of rigid
model pressure measurement. In these wind tunnel tests, to obtain suf-
ficient detailed wind load characteristics to meet the requirements of
engineering design and research, enough measurement points should be
placed on the surface of the model. With the progress of synchronous
acquisition technology for pressure scanning valves, it is not a significant
problem to obtain synchronous wind pressure time histories for more
than 1000 points on a rigid model in a wind tunnel test; however, with
some of the more complex high-rise buildings (Huang et al., 2014,
2015a) and large span structures (Fu et al., 2007), the number of points is
still far from the design requirement. In addition, for some dynamic tests
with simultaneous pressure measurement, it is inappropriate to arrange
too many points on the model which would cause data distortion (Kato
and Kanda, 2014). Therefore, it is necessary to explore effective ways to
predict, or extend, the wind-induced pressures on the entire surface of a
structure according to the limited pressure data from the taps.
At present, there are many ways to build a multi-variable linear/
nonlinear forecasting model, such as inverse distance weighting (IDW)
(Lu and Wong, 2008), kriging (Franke, 1982), regression polynomials
(e.g. ARMA) (Kho et al., 2002), artificial neural networks (ANNs)
(Turkkan and Srivastava, 1995), etc. In these methods, the conventional
IDW method has relatively simple and fixed forecasting model, which
simplifies the interpolation process, however makes it not well adaptive
to the variation of wind pressure distribution of different regions and
different structures; the kriging method, based on the theory of spatial
statistics, uses the variogram to measure the spatial correlation and
weight of near sampling points, thereby it can well adapt to the variation
of wind pressure distribution. However, the variation function should be
selected by human experience, and when the variance function is a
combination, it is difficult to select and the amount of calculation is
increased; the regression polynomials is a most commonly used method
for selecting different empirical formula to accommodate to the wind
pressure distribution with different characteristics, however, when many

* Corresponding author. School of Civil Engineering, Central South University, Changsha, Hunan, 410075, China.
E-mail address: huangdongmei_tumu@163.com (H. Dongmei).

http://dx.doi.org/10.1016/j.jweia.2017.07.021
Received 3 January 2017; Received in revised form 26 July 2017; Accepted 30 July 2017

0167-6105/© 2017 Elsevier Ltd. All rights reserved.

H. Dongmei et al.
parameters are involved, it is very difficult to obtain an ideal empirical
formula for the function form between them is complicated and un-
known. Fortunately, ANNs with multilayer perceptrons are the equal of a
multi-mapping black box analysis function (Turkkan and Srivastava,
1995), which can describe the complex and non-linear functional re-
lationships of a large number of parameters by training some
input-output pairs from tests, even with noisy or incomplete information
(Haykin, 1999): thus it has the characteristics of self adjustment and
robustness and has been widely used in interpolation problems in various
fields (Chen et al., 2003; Yu and Xu, 2014; Ahmed et al., 2015). At the
same time, ANNs in some wind engineering interpolation problems are
also gradually becoming more widely used. For example, to investigate
wind interference between tall buildings (Khanduri et al., 1997; English
and Fricke, 1999), to predict the mean and root-mean-square (RMS)
pressure coefficients on gable roofs of low-rise buildings and large span
space structures (Chen et al., 2003; Fu et al., 2007; Gavalda et al., 2011),
and even to predict the wind-induced pressure time series on structures
(Chen et al., 2002; Fu et al., 2007), and so on (Chen et al., 2008, 2016;
Wu, and Kareem, 2011).
Although a neural network can commendably predict the mean and
RMS wind pressures on the structure, it is inadequate to predict the wind
pressure time series for too enormous input-output trained data set with
time parameter t. Therefore, it is necessary to seek an appropriate method
to transform the time-variant wind pressure field and then to let the
input-output trained data of the ANN be time-independent. Proper
orthogonal decomposition (POD) is just one of the approaches based on
this objective. Using POD, spatially-distributed multivariable random
loads can be reconstituted through a linear combination of a series of
orthogonal load modes (Azam and Mariani, 2013) weighted by the cor-
responding, unrelated modal coordinates (i.e. loading principal co-
ordinates), respectively. The orthogonal load modes are space-related
and time-independent, and the loading principal coordinates are
time-varying and space-independent. The POD involves a covariance
proper transformation (CPT) and uses proper vectors to form a covari-
ance matrix of random fluctuating wind loads at zero time lag as the load
modes. Armitt (1968) and Lumley (1970) firstly introduced the POD
technique to deal with turbulence and wind-related issues. Subsequently,
some researchers used POD in reconstructing wind fields (Tamura et al.,
1999; Bienkiewicz et al., 1995; Holmes et al., 1997), finding some sys-
tematic structures hidden in the random fields (Kikuchi et al., 1997;
Wang and Zhou, 2014), and calculating structural wind-induced re-
sponses (Solari and Carassale, 2000; Solari et al., 2007; Carassale et al.,
2007; Huang et al., 2015b).
In our work, a back-propagation neural network (BPNN), combined
with a proper orthogonal decomposition (POD-BPNN) approach is used
for the prediction of the mean, the root-mean-square (RMS) pressure
coefficients, and even the time series of wind-induced pressures on a
structure, respectively. A high-rise building is taken as an example to
conduct wind tunnel testing to obtain wind-induced pressure data for
training, and then it is used to show the validity of POD-BPNN for the
prediction of wind-induced pressures, not only the statistics relating to
the pressure data but also the time series, the spectra, and the coherence
functions thereof.
2. Wind tunnel experiments
2.1. Overview of experiments
Wind tunnel experiments were carried out in the boundary layer wind
tunnel at Central South University in China, which belongs to the High-
speed Railway Construction Technology National Engineering Labora-
tory. It is a double test section reflux type wind tunnel, in which, the low
speed test section is with a working section of 12 m width, 3.5 m height,
and 18 m length, with a wind speed of 2–18 m/s (continuously adjust-
able), and the high-speed test section is with a working section of 3 m
width, 3 m height, and 15 m length, with a wind speed of 2–90 m/s
2
Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17
(continuously adjustable). All the experiments in this paper are carried
out in the high-speed test section.
Fig. 1 shows a photo of the model (a square building) mounted in the
wind tunnel. The wind field and model were all made with a geometric
length scale of 1:350. As shown in Fig. 1, spires, grids, and roughness
elements were used to simulate a boundary layer wind flow of urban
terrain type stipulated in the Load Code of China (2012) as exposure C
category. This terrain type specifies a mean wind speed profile with a
power law exponent of a ¼ 0.22. The non-dimensional measured mean
wind speeds profile, longitudinal turbulence intensities profile, and
lateral turbulence intensities profile are shown in Fig. 2(a) and (b), and
(c), respectively (the reference point is located at 1.2 m and the corre-
sponding wind speed is 14 m/s). Meanwhile, the mean wind speeds
profile stipulated in the Load Code of China (2012) and the turbulence
intensity profile stipulated in the Japanese Load Code (AIJ, 1996) are
also shown in Fig. 2(a) and (b), respectively. As seen, the simulated wind
field is in good agreement with theoretical requirements. Moreover, the
spectra of longitudinal and lateral wind speeds at the reference point
(gradient wind height) are shown in Fig. 2(d) and (e), respectively, which
match the von Karman type spectrum. In addition, longitudinal and
lateral turbulence integral scales are shown in Fig. 2 (f) and (g),
respectively. According to the scale ratio, they are corresponding to
60–200 m and about 70 m of the actual wind field, respectively. That is
consistent with the actual situation. Moreover, it can be noted that the
vertical profiles of the u-turbulence intensity and integral length scale are
both greater than v-ones, and the u-and v-turbulence intensities and the
u-integral length scale decrease as height increases, the v-integral length
scale basically does not vary with altitude.
The size and pressure measuring taps arrangement of the model are
shown in Fig. 3. The model height is about 1.4 m and width is about
0.163 m (the wind tunnel obstruction is about 2.5% (less than 5%), and
the blockage effect thereof is negligible). There are 14 measurement
layers on the model, on each of which is arranged 36 measuring points
and the position of each layer is consistent (a total of 504 taps). In this
way, label M10-1 represents the measured value of the 1st measured
point in the 10th layer. To evaluate the prediction effect of the method
proposed in our paper for surface wind pressure with limited measure-
ment points, it is firstly necessary to determine which data are to be used
for training, and which data for prediction. As shown in Fig. 3, the
measurement points labeled ‘●’ were accepted as training data (a total of
120 points), and those marked as ‘O’ were untrained but can be used as
verified data for the prediction results by the proposed POD-BPNN
approach only using the limit training data. The electronic scanvalve
(ZOC-33, produced by Scanivalve Company in United States) is used to
gather wind pressures, its gathering frequency is 625 Hz, and the each
Fig. 1. Rigid model pressure measurement wind tunnel test.
H. Dongmei et al. Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17

Fig. 2. Wind field simulation.

scan length is 32s. at the reference point as its reference pressure; pij ðtÞ is the wind pressure
time-history of the jth measuring tap at the ith level; pr0 ðtÞ and pr∞ ðtÞ are
2.2. Pressure calculation the synchronously measured total pressure and static pressure time-
histories at the reference point respectively; pr0 ðtÞ  pr∞ ðtÞ expresses
In aerodynamics, the time-history of the dimensionless pressure co- the dynamic pressure time-history at the reference point.
efficient on structural surfaces is expressed as follows:
3. Proper orthogonal decomposition
pij ðtÞ  pr∞ ðtÞ
Cij ðtÞ ¼ (1)
pr0 ðtÞ  pr∞ ðtÞ Before the artificial neural networks are applied in the training and
predicting stages, the covariance matrix of the multi-point wind loads
where Cij ðtÞ is the time-history of the dimensionless pressure coefficient should be decomposed using POD. The mathematical expression of the
of the jth measuring tap at the ith level, which takes the dynamic pressure POD method is to transform some of the interdependent variables (i.e.

3
H. Dongmei et al. Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17

Fig. 3. Layout of the pressure taps on the surface of model (unit: mm) (‘●’- Measurement points for training, ‘o’- Measurement points not receiving training).

wind pressure field in this work) into the weighted superposition of a values and proper vectors are also real and satisfy the following
number of independent variables: the transformed variables are still able orthogonality conditions:
to maintain the characteristics of the original variables. An orthogonal
transformation to the basis of the eigenvectors of the sample covariance ΨT Ψ ¼ I; ΨT CFm Fm Ψ ¼ Λ (4)
matrix of the original variables is performed, and then the original var-
iables can be projected onto the subspace spanned by the eigenvectors where, Ψ ¼ ½ψ1 ⋯ψm  is the m  m matrix of load modes; Λ ¼
corresponding to the eigenvalues. The detailed process can be seen in diagfλ1 ; ⋯; λk ; ⋯; λm g is a diagonal matrix consisting of the proper values
literature by Watanabe (1965), and the key process is described below. of the load; I is the identity matrix. According to Eq. (4), CFm Fm can be
The covariance matrix of the interdependent variables, such as decomposed as:
random fluctuating wind pressures (coefficients) Fm ðtÞ, at zero time-lag
X
m
can be determined as follows: CFm Fm ¼ ΨΛΨT ¼ ψk ψTk λk (5)

 k¼1
CFm Fm ¼ E Fm ðtÞFTm ðtÞ (2)
According to the Karhunen-Loeve expansion (Watanabe, 1965), the
Let λ1 ; ⋯; λm denote the eigenvalues (i.e. proper values or load co- covariance orthogonal transformation can be defined as:
ordinates) of the covariance matrix CFm Fm , which are sorted in descending
order. The vectors ψ1 ; ⋯; ψm are the corresponding eigenvectors (i.e. X
m
Fm ðtÞ ¼ Ψf m ðtÞ ¼ ψk fk ðtÞ (6)
proper vectors or load modes), which are the nontrivial solutions of the k¼1
following characteristic equation:
where fm(t) ¼ {f1, …,fk …,fm}, fk(t) (k ¼ 1, …,m) is defined as the kth
ðCFm Fm  λk IÞψk ¼ 0 ðk ¼ 1 ; ⋯; mÞ (3) loading principal coordinate (which is also a zero-mean random process)
and can be determined from the following formula
where, λk is the kth proper value and ψk is the kth proper vector.
As CFm Fm is a real, symmetric and positive-definite matrix, its proper f m ðtÞ ¼ ΨT Fm ðtÞ (7a)

4
H. Dongmei et al.
fk ðtÞ ¼ ψTk Fm ðtÞ (7b)
According to Eq. (7a), we have



Cf m f m ¼ E f m ðtÞf Tm ðtÞ ¼ E ΨT Fm ðtÞFTm ðtÞΨ ¼ ΨT CFm Fm Ψ ¼ Λ
where Cf m f m is the covariance matrix of the load principal coordinates at
zero time lag.
In this way, each proper value of Cf m f m in Eq. (8) is actually the
variance of the load principal coordinate of the corresponding load mode.
Therefore, if the proper values are sorted in descending order, then the
load principal coordinates are sorted in descending order of energy. It can
be seen from Eq. (6) that, using POD, the time series of spatially corre-
lated fluctuating wind pressure field can be reconstituted through a
linear combination of a series of space-related and time-independent
orthogonal load modes weighted by the corresponding, time-varying
and space-independent modal coordinates (i.e. loading principal co-
ordinates), respectively.
4. BP artificial neural networks
Artificial neural networks (ANNs) are inspired from the biological
sciences by attempting to emulate the behaviour and complex func-
tioning of the human brain in recognizing patterns (Haykin, 1999). They
are an adaptive non-linear dynamic system, composed of a large number
of interconnected simple basic elements (neurons). ANNs can establish
complex non-linear relationships between the inputs and outputs
through the transfer functions of the neurons. It can be realised through
training on a representative set of input-output “data pairs” by some
optimization algorithms, even if the inputs are noisy or incomplete
(Haykin, 1999; Flood and Kartam, 1994). Therefore they can be used to
solve multivariate, non-linear modelling, interpolation problems.
There are various sorts of neural networks, such as multilayer per-
ceptron networks, Carpenter networks, Hopfield networks, and back-
propagation (BP) networks (Wu, 1992; Haykin, 1999). The BP neural
networks (BPNN) was proposed by Rumelhart and McCelland (1986),
which is a kind of multilayer feed forward network trained by error back
propagation algorithm. It is widely used in non-linear modelling and
Fig. 4. Topology of the backpropagation neural network (BPNN m-r-s-n).
Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17
interpolation problems for its self-adaptive, nonlinear mapping, gener-
alization abilities, etc. The topology of the BPNN is shown in Fig. 4. It is
generally composed of several layers of many simple elements operating
in parallel: an input layer, an output layer, and one or more hidden layers
(8) between the input and output layers. As shown in Fig. 4, the notation
BPNN m-r-s-n is used as a label for a net with m input variables, n output
variables, and two hidden layers with r and s neurons, respectively (there
may be more hidden layers). The input layer can be looked on as an
external stimulus, it can be perceived and transmitted by some hidden
layers (central nervous system-C.N.S.), and then responds in an
output layer.
The training process of the backpropagation algorithm mainly in-
volves two stages: feed-forward and error back-propagation. In the
feedforward stage, as shown in Fig. 5, the input to a neuron j in the
hidden layers is the sum of the weighted inputs and bias. This summation
is then passed through a nonlinear transfer function, f(⋅), thereby
generating an output signal for the neuron j, which can be written as,
!
X
m
yj ¼ f Wij xi þ bj (9)
i¼1
where Wji is the weight that connects the ith neuron of the previous layer
to the jth neuron of the current layer; xi is the ith input component; bj is
the bias associated with the jth neuron of the current layer; and f(⋅) is the
transfer function.
In Eq. (9), the transfer function has three types: linear function, hy-
perbolic tangent function(S-type), and a logarithmic function (S-type):
their expressions, graphs, and input-output mapping ranges are shown
in Table 1.
If there are some hidden layers, the output y j then becomes an input
that feeds the neurons in the next layer. In this way, the final output
vector y can be obtained. After the first feed-forward process, the final
output vector y may not match the target output Y. Then the network
output error (i.e., the sum of the square errors, defined as E ¼ 12ðy  YÞ2 )
is then back-propagated from the output neurons to the hidden neurons
and then the input layer by using a back-propagation learning algorithm.
The algorithm constantly modifies the weight values by minimising the
sum of the square errors (also call network training), until it reaches a
5
H. Dongmei et al. Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17

Fig. 5. The basic neuron j in the hidden layers and the feedforward phase.

Table 1 hidden layer, and the training effect is also easier to observe and
Transfer functions of the BPNN. adjust than adding to the number of layers. Certainly, the over-
Name Expression Graph Range fitting issue for adding neurons should be considered;
Linear function f ðnÞ ¼ a⋅n þ b (-∞, ii) There is no explicit rule for selecting the number of neurons in the
þ∞) , hidden layer. The more practical approach is to begin with a small
(-∞, number of neurons and then step-by-step to increase their number
þ∞) until the training effect is satisfactory;
iii) The convergence and the time of network training are closely
related to the initial weight values and the range of input values. It
is generally hoped that the weights of each neuron can be adjusted
in the biggest S-type activation function, so that the initial weight
Hyperbolic f ðnÞ ¼ tanhðnÞ (-∞, values and the input values are best within (-1, 1);
tangent S þ∞) , iv) It should select globally optimised, fast-convergence learning
type function (-1, þ1) methods. There are many back-propagation algorithms, such as
the standard steepest gradient descent algorithm, the conjugate
gradient algorithms, the Levenberg-Marquardt algorithm, and so
on. The first two are of single-order training speed, while the LM
algorithm has second-order training speed (needed to calculate
the weight coefficient of the Jacobian matrix), so the convergence
of the LM algorithm is much faster than that of the first two, but it
Logarithmic S f ðnÞ ¼ 1þe1 n (-∞, requires more memory (in the Neural Network Toolbox, MAT-
type function þ∞) , (0,
þ1)
LAB™, 2009). Moreover, the learning rate should take a smaller
value to ensure the stability of the system (such as one from 0.01
to 0.8).

5. Prediction of the wind-induced pressures on the building

5.1. Prediction of the mean and fluctuating pressure coefficients

To verify the effect of the BP neural network method in the prediction

stopping criterion (e.g., the network output error is less than a user-
defined target). After the training process, the ANN can produce a non-
linear complex relationship between input and output data patterns,
which are dependent on the weights, bias, and the transfer function.
It can be seen that the main factors influencing the success of an
ANNS training are: the input and output, the initial value of weights and
bias, and the back-propagation algorithm. Therefore, in the design of the
BP network, the layer number of network, the number of neurons in each
layer, the initial value, and the learning method should be considered:
i) It has been proved theoretically that, a network which is within
allowable deviation and having at least one S-type hidden layer
and one linear output layer can approach any rational function.
Increasing the number of layers reduces the error and improves
the accuracy; however, at the same time, it also makes the
network more complicated, which increases the training time for
the network weights. Moreover, improvement of the accuracy can
also be achieved by increasing the number of neurons in the
of the mean and fluctuating wind pressures on the building, the test data
for a wind direction of 0 are taken as an example. Coordinates, i.e. x, y, z,
values of the pressure taps, were taken as input variables in the BPNN,
and the output was a single variable either Cp-mean or Cp-rms. Moreover,
the input coordinate variables were normalised to (0–1) to improve the
performance of the BPNN (as mentioned in Section 3).
The input-output data pairs at the position marked with solid points
‘●’ in Fig. 3 can be selected as network training data, and those at hollow
point ‘o’ can be looked on as network verified data (they are not used for
training the network, but to evaluate the prediction precision of the wind
pressure field by the trained neural network model).
As mentioned before, different learning rates, learning algorithms,
transfer functions, and the numbers of neurons in the hidden layers will
affect computational efficiency. For the prediction of Cp-mean and Cp-rms
on building, two four-layer network structures (including the input layer)
were adopted in this work so that the four-layer fully connected BPNN is
capable of describing any continuous nonlinear mapping with sufficient
accuracy (Maren et al., 1990; Haykin, 1999).

6
H. Dongmei et al.
After determining the number of layers in the neural network, the key
is to determine the number of neurons in each hidden layer, which will
greatly affect the performance of the neural network (Maren et al., 1990;
Chen et al., 2003). If the number of neurons in the hidden layer is too
small, the convergence time of the neural network training is longer and
it may not converge if the network cannot establish sufficient refinement
in the relationship between the input and output data. On the other hand,
if the number of neurons in the hidden layer is too large, the time used in
neural network training is too long and the generalisation ability of
network is weakened. Therefore, a trial-and-error approach (Haykin,
1999; Chen et al., 2003; Fu et al., 2007), which starts with a small
number of hidden neurons, and then increases in size, until the training
performance is acceptable in our analysis, is adopted. It was found that
network structures with BPNN 3-15-10-1 and BPNN 3-18-12-1 have a
better ability to predict Cp-mean and Cp-rms, respectively. Such results are
similar to the findings of Chen et al. (2003) and Fu et al. (2007). More-
over, the selected network structures follow the general principle that a
ratio of 3:1 or 3:2 between the number of first and second hidden layer
nodes yields better prediction performance compared to other combi-
nations (Maier and Dandy, 1998). The transfer functions, training pa-
rameters, and training algorithm for the developed BPNN models are
summarised in Table 2. For the prediction of Cp-mean, the transfer function
is a hyperbolic tangent sigmoid function, but for the Cp-rms prediction, a
logarithmic sigmoid transfer function is used, since Cp-rms is positive
while Cp-mean can be positive or negative. And the transfer function
connected with output layer uses a linear function.
In addition, there are two ways to improve the fault tolerance of a
network: one is the training data additive noise method, and the other is
the hidden layer neuron multiplication training method. In this paper,
the latter is used to improve the fault tolerance of the network system.
For example, when using a 3-18-12-1 BPNN, first it uses a 3-9-6-1 BPNN
for training, and then on the basis of this, it turns to a 3-18-12-1 BPNN for
continuing, so as to obtain the final network parameters.
During the training process, whenever one of the following stopping
criteria is first met, training is terminated: (i) when the number of iter-
ation increases to a given number (e.g., 50,000 training epochs in our
work) (ii) when the training error is less than a training goal (e.g.,
1.0  109), and (iii) when the magnitude of the training gradient is less
than a small specified value (e.g., 1.0  1010). In this training regime, it
will be terminated generally when the conditions described in (ii) are
first met.
As the improved global Levenberg-Marquardt optimised algorithm
(Fan, 2003; Fan and Pan, 2009) can achieve a faster convergence speed,
and can avoid falling into a local minimum during training, it is selected
to carry out the network training in this work.
By training the BPNN models with the input-output patterns, the
multivariate nonlinear functional relationships between the pressure
coefficients and the corresponding coordinates of the taps on the building
Table 2
Parameters of the improved BPNN for prediction of Cp-mean and Cp-rms.
Terms All face
Inputs x, y, z
Outputs Cp-mean Cp-rms
ANN structure 3-15-10-1 3-18-12-1
Transfer functions hyperbolic tangent S type function logarithmic S type function (hidden
(hidden layers): layers):
f ðnÞ ¼ tanhðnÞ f ðnÞ ¼ 1þe1 n
linear function (output layers): linear function (output layers):
f ðnÞ ¼ a⋅n þ b f ðnÞ ¼ a⋅n þ b
Training epoch and 68 41
time 1.4 s 1.7 s
Training data ‘●’- measurement points in Fig. 3 ‘●’- measurement points in Fig. 3
Test data ‘●’ and -‘o’ measurement points in Fig. 3 ‘●’ and -‘o’ measurement points in
Fig. 3
Training algorithm improved global improved global
LM algorithm LM algorithm
7
Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17
model are established and thus, when the input variables (trained or not
trained) are given, the corresponding outputs, such as Cp-mean or Cp-rms
can be obtained.
To test the validity of the BP neural network in the prediction of wind
pressure, the 11th, 6th and 14th measurement layers were choose for
analysis. Of which, the 11th measurement layer is arranged not only with
some training measurement points, but also some untrained points, ar-
ranged alternately, as shown in Fig. 3. The points on the 6th layer are all
untrained, and are located between the 5th and 7th layers having
training measurement points. While the 14th layer is at the top of the
model, the points on it are also untrained and belong to the extension of
the training point set. Fig. 6 shows the mean and fluctuating wind
pressure coefficients on the four surfaces of the model. Due to the three-
dimensional flow around the top, the absolute values of mean wind
pressures at the top will be slightly reduced, especially on the windward
side. Moreover, the gradient of wind pressures on the trailing edge of
lateral side is steeper.
To verify the superiority of the global LM algorithm, the LM algorithm
and standard steepest gradient descent algorithm are used respectively to
carry out the network training for Cp-mean. Fig. 7 shows the error iteration
convergence curve, training epochs, and training time. As shown, by LM
algorithm, it took only 68 training epochs and 1.4s to converge (reaching
the training error goal of 1.0  109), much faster than the training speed
achieved by the steepest descent method on the same IBM workstation
(needing 102.8 s and reaching the maximum of 50,000 training epochs,
with a training error of 1.37  105). Figs. 8–10 show comparisons of Cp-
mean and Cp-rms between the experimental data and the prediction results
based on BPNN and inverse square distance weighting (ISDW) method
(Lu and Wong, 2008) for the 11th, 6th, and 14th measuring levels,
respectively. In 11th measuring level, some solid dots represent test data
used in network training, and the hollow dots are untrained points (the
numbers of trained and untrained taps are shown in Fig. 3). As shown in
Figs. 8 and 9, the BPNN prediction results of trained taps are in good
agreement with the test data (the accuracy is very high), and the pre-
dicted results of untrained ones are also satisfactory. However, as shown
in Fig. 10, the predicted mean wind pressures on the 14th level on the
windward side are significantly less than the test results, and the absolute
values of the mean wind pressures of the leeward and side faces are also
slightly less than the test results (due to the three-dimensional flow
around the top of the building, the mean wind pressures will decrease, as
shown in Fig. 6). The predicted fluctuating wind pressures are basically
close to the experimental results (except for the intermediate point-23rd
point on the windward side). Therefore, when selecting the training data,
it should, as far as possible, include the data at the boundary, so that the
results of the prediction are interpolated, rather than extrapolated, to
ensure the accuracy of the predicted results.
The wind pressure characteristics of the building four faces are
different, and so can be divided into four aspects for training and
Single face
x, z or y, z
Cp-mean Cp-rms
2-15-10-1 2-18-12-1
hyperbolic tangent S type logarithmic S type function
function f ðnÞ ¼ 1þe1 n
f ðnÞ ¼ tanhðnÞ linear function (output layers):
linear function (output layers): f ðnÞ ¼ a⋅n þ b
f ðnÞ ¼ a⋅n þ b
24 (all) 28 (all)
1.5 s (all) 1.8s (all)
‘●’- measurement points in Fig. 3 ‘●’- measurement points in Fig. 3
‘●’ and -‘o’ measurement points ‘●’ and -‘o’ measurement points
in Fig. 3 in Fig. 3
improved global improved global
LM algorithm LM algorithm
H. Dongmei et al. Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17

Fig. 6. Mean and fluctuating wind pressures on building: A. windward side; B. lateral side; C. leeward side; D. lateral side.

8
H. Dongmei et al. Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17

Fig. 7. Training epochs of LM algorithm and standard steepest gradient descent algorithm for Cp-mean.

Fig. 8. Comparisons of Cp-mean and Cp-rms between the test data and the BPNN predictions: 11th measuring level (trained level).

Fig. 9. Comparisons of Cp-mean and Cp-rms between the test data and the BPNN predictions: 6th measuring level (un-trained interpolation level).

predication. At this time, the input can be changed into two variables, i.e.,
coordinates x or y, and z. thus network structures to predict Cp-mean and
Cp-rms are BPNN 2-15-10-1 and BPNN 2-18-12-1, respectively. Their
neural network structure, transfer function and training regime are
shown in Table 2. Figs. 8–10 show the results of the prediction on the
four surfaces. As can be seen from the graphs and the table, the prediction
accuracy is slightly better than the predicted results from the training
network for all faces (untrained points), which shows that the general-
ization ability is better, and also needs little training time.
Moreover, it can be seen in Figs. 8–10 that the BPNN technique is
overall more superior to ISDW method, especially at the locations that
local wind pressures vary sharply.

9
H. Dongmei et al.
Fig. 10. Comparisons of Cp-mean and Cp-rms between the test data and the BPNN predictions: 14th measuring level (un-trained epitaxial level).
5.2. Prediction of wind-induced pressure time series, spectra, and
coherence functions
Pertinent to the present study, there are many reports concerning the
average wind pressure and fluctuating wind pressure on a building sur-
face (Chen et al., 2003; Gavalda et al., 2011), and also the time series
prediction of wind pressure (Chen et al., 2002; Fu et al., 2007). However,
the input of the aforementioned prediction of wind pressure time series
should generally involve four parameters, i.e., three coordinate param-
eters x, y, and z, and a time parameter t, constituted a four-dimensional
input vector. Therefore, the input-output data set is large (especially
when the acquisition time is longer and the data points are more
numerous). Moreover, the corresponding relationship between the input
and output data becomes very complicated. These would make the
training rate slow, and the convergence performance and generalisation
ability are also poor. In addition, the convergence rate of the improved
global LM algorithm (Fan, 2003; Fan and Pan, 2009) is very fast, but
requires significantly large memory usage: if it uses x, y, z, and t for the
Fig. 11. Process of training and predicting for POD-BPNN.
10
Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17
four parameters as its network input, the memory demand is high.
Therefore, in this work, the four-dimensional wind pressure field
sequence, which is associated with the conversion of the spatial co-
ordinates and time, should be first subjected to POD (see Eqs. (2)–(6)).
Thereby, it can be converted into the weight superposition of several load
modes (which is only related to spatial coordinates, and its number is
equivalent to the number of pressure points used for training) through
the generalised coordinates (sequences with only time variation). Since
the generalised coordinates of the load are space invariant, they do not
need to be used as input parameters for training. For the inherent load
modes of the wind pressure field, which are not related to each other and
do not change over time, they can be trained and interpolated respec-
tively in parallel, basically like that of the mean wind pressure. In this
way, it is equivalent to converting a complex large system to a series of
very simple, parallel systems for training and interpolation, which made
full use of the advantages of the LM algorithm (i.e., quick convergence
speed) while avoiding the drawback of onerous memory demands. After
obtaining the neural network system for each load mode, the load modes
H. Dongmei et al. Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17

Fig. 12. The first three orders of modes of wind pressures on model: A. windward side; B. lateral side; C. leeward side; D. lateral side.

11
H. Dongmei et al.
Fig. 12 (continued).
can be extended by interpolation, then the wind pressure time series for
the whole building surface (including the training and the non-training
points) are obtained using Eq. (6) (at this time, the dimension of the
load modes in the equation is the extended one). The detailed process is
shown in Fig. 11, including three processes: transformation, training, and
prediction. In the figure, the dimension of ψi is equal to the number of
selected training points (in this case, the number is equal to 120, which
b i can be equal to the
are labeled ‘●’ in Fig. 3), however the dimension of ψ
number of all selected prediction positions on building (in this case, 504
positions are selected, which are equal to the number of the taps labeled
Fig. 13. Spectra of first three orders of load generalised coordinates.
12
Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17
‘●’ and ‘o’ in Fig. 3).
To evaluate the effectiveness of the POD-BPNN, the surface wind
pressure field on the model is required to be first converted by POD.
Fig. 12 shows the contour plot for the first three orders of load modes
converted by POD. As shown in the figures, they have obvious regularity,
the 1st mode has anti-symmetric characteristics in the Y-axis, the 2nd
mode has positive symmetric characteristics in the Y-axis, and the 3rd
mode has polar symmetrical characteristics (these features are also
described in Kikuchi et al. (1997), but the 3rd mode in the present work
corresponds to their 5th mode). Fig. 13 shows the spectra of the first
three orders of load generalised coordinates, and Fig. 14 shows the
contribution ratio of each proper value and pre-i proper values. It can be
Fig. 14. Proper values and their contribution ratios.
H. Dongmei et al.
Fig. 15. Comparisons of the fluctuating pressure time series and spectra of tap 11–18: test data and POD-BPNN predictions (predictions results-6 means that the number of load modes
present is 6, and predictions results-120 means that the number of load modes is 120).
seen from Fig. 14 that, the first three orders of load modes include about
40% of the energy of the fluctuating wind field, and the first six reaches
about 52%. As shown in Fig. 13, the pattern of the power spectrum of the
1st generalised coordinates of the load mode is close to that of cross-wind
aerodynamic force spectrum, and there is an obvious peak at about
8.96 Hz thereby the reduced frequency is about 0.10, which is consistent
with the Strouhal number of a square building. Therefore, in combination
with the plot for the 1st loading mode (Fig. 12(a)), it can be concluded
that the 1st loading mode represents the main contribution of the vortex
shedding. The pattern of the power spectrum of the 2nd generalised
coordinates of the load mode is closer to that of the along-wind aero-
dynamic force spectrum, and the characteristics of corresponding mode
(Fig. 12(b)) are close to the fluctuating wind pressure, which shows that
the 2nd load mode represents the main contribution of the fluctuating
wind energy. The power spectrum of the 3rd generalised coordinates of
Fig. 16. Comparisons of the fluctuating pressure time series and spectra of tap 6–18: test data and POD-BPNN predictions.
13
Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17
the load mode also has an obvious peak at approximately 8.40 Hz and the
reduced frequency is about 0.098, which is close to that of the first order
(albeit a little smaller). This shows that the 3rd mode also represents a
part of the contribution from the energy of the vortex shedding. Thus, it
can be seen that, the action of vortex shedding on the square building has
features of alternating trend and spiral rotation, whose effects on the
structure are shown as a cross-wind, and a torsional, vibration (also
shown in Kikuchi et al. (1997)).
When the POD modes are interpolated by the neural network, the
parameters of the input of the neural network are first taken. Similar to
those of the mean and fluctuating wind pressures to train the input-
output data from the four faces, the positions of pressure taps (x or y,
z) were considered as input variables in the BPNN, and the output was the
corresponding values of wind pressure modes. Similarly, the input vari-
ables were normalised to (0, 1) for improving the performance of the
H. Dongmei et al. Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17

BPNN. It is worth noting that, in our work, the time, t, was not considered the RMS, maximal and minimal pressure coefficients, and RMSE of the
as an input variable in the BPNN, and the output was also not related to predicted and test time series data are compared.
time t. In Table 3,
After training the BPNN with the selected input-output patterns, and vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
giving the values of the input variables (x or y, z)-selected prediction u1 X N
RMSE ¼ t
2
positions on building, the corresponding outputs, i.e. corresponding ðyi  Yi Þ (10)
N i¼1
values of the extended load modes, can be obtained. Thereby, the whole
pressure coefficients on building at any time instant can be obtained
according to Eq. (6). Figs. 15–17 show comparisons of pressure time where N is the number of samples; Yi and yi are the fluctuating experi-
series (including mean values) between the test data and the POD-BPNN mental data and the ANN prediction for i ¼ 1, …, N, respectively. The
predictions. As shown in the figures, for the previously trained pressure ANN predictions are more precise if the value of RMSE is closer to 0.
tap 11–18 (the tap at one edge of the lateral side where the wind pressure It is clear that, for the trained case, the predicted time series match the
variation is significant), and the untrained pressure taps 6–18 (also the experimental results, e.g. tap 11–18, moreover, the time series (only
edge tap on the lateral side) and 14-5 (the top central tap on the wind- including the first six modes) lost a part of the high frequency energy; for
ward side), respectively. Some results are summarised in Table 3, where the untrained interpolation case, the predicted time series are in good

Fig. 17. Comparisons of the fluctuating pressure time series and spectra of tap 14-5: test data and POD-BPNN predictions.

Table 3
Comparison of the test data and the POD-BPNN prediction results of local wind pressures on typical positions.

Case Pressure Mean RMS Minimum Maximum RMSE

tap 11–18 (Trained, Interpolation) Measured 0.972 0.398 4.114 0.413 0.0000
POD-BPNN 0.972 0.398 4.114 0.413
tap 6–18 (Un-rained, Interpolation) Measured 0.737 0.387 3.259 0.309 0.1508
POD-BPNN 0.742 0.401 3.135 0.351
tap 14-5 (Un-rained, Extrapolation) Measured 0.65 0.224 0.304 1.591 0.1633
POD-BPNN 0.79 0.211 0.193 1.670

Fig. 18. Comparisons of the coherence functions between the test data and the POD-BPNN predictions: (a) taps 11–18 and 11–16; taps 11–18 and 11–17 (b) taps 6-1 and 6-2; taps 6-1 and
6-5.

14
H. Dongmei et al.
agreement with the test data, e.g. tap 6–18; however, for the extrapola-
tion case, the predicted results have something error with the expected
results (especially the mean values), e.g. tap 14-5 (similar to Pierre
(1994), and Agamennoni and Mandolesi (1997)). Therefore, when using
the ANN to predict wind-induced pressures on a building, it is better to
arrange as many pressure taps as possible at the edges of the model
surfaces for the best interpolation.
Moreover, the effectiveness of the predicted results can be analysed in
the frequency domain. The comparisons of the spectra of wind-induced
pressures between the test data and the POD-BPNN predictions for the
aforementioned three taps are also shown in Figs. 15–17. It can be
concluded from the figures, not only for the trained case but also in the
untrained case, that the predicted results are in good agreement with the
test results, however, the predicted results superimposing only the first
six modes would include most of the low-frequency energy but lose some
of the high-frequency energy.
Furthermore, the comparisons of the coherence functions of wind-
induced pressures between the test data and the POD-BPNN pre-
dictions are shown in Fig. 18. The comparisons includes the previously
Fig. 19. Comparisons of the spectra of global base shear forces and moments between the test data and the POD-BPNN predictions.
15
Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17
trained pressure taps 11–18 and 11–16, the untrained pressure taps
11–18 and 11–17 (the taps at edge on the lateral side where the wind
pressure variation is extreme), and the untrained pressure taps 6-1 and 6-
2, and 6-1 and 6-5 (the taps on the windward side), respectively. It shows
that the predicted results of the participating training are basically in
agreement with the experimental results, and the non-participating
training data are close to the experimental results.
Thus, both for wind loads analysis and wind-induced vibration
calculation in the time domain or the frequency domain, the POD-BPNN
method in this paper can be used to predict the wind pressures.
As shown above, some local predictions of the time series, power
spectrum, and coherence function are accurate. Of course they cannot
show such accuracy everywhere. Therefore, Fig. 19(a)-(e) show the
comparisons of the spectra of the base along-wind and across-wind shear
forces and moments, and the base torsional moment between the test
data and the POD-BPNN predictions, respectively. Moreover, Table 4
shows the mean and RMS values of them: the errors are all less than 5%.
It can be seen that the global capacity of this method to reproduce the
frequency structure and estimate the statistical parameters of the whole
H. Dongmei et al. Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17

Table 4 Ahmed, R., EI Sayed, M., Gadsden, S.A., jong, J.T., Habibi, S., 2015. Automotive internal
Comparison of the test data and the POD-BPNN prediction results of global shear forces and combustion-engine fault detection and classification using artificial neural network
moments. techniques. IEEE Trans. Veh. Tech. 64 (1), 21–33.
AIJ, 1996. Recommendations for Loads on Buildings. Architectural Institute of Japan,
Case Source Mean Error RMS Error Japan [English version].
Armitt, J., 1968. Eigenvector Analysis of Pressure Fluctuations on the West Instrumented
along-wind base shear force Measured 1.290 3.0% 0.118 4.2%
Cooling Tower [R]. Internal Rep. RD/L/N114/68. Central Electricity Research
POD- 1.251 0.113 Laboratories, U.K.
BPNN Azam, S.E., Mariani, S., 2013. Investigation of computational and accuracy issues in POD-
across-wind base shear force Measured 0.0153 / 0.332 4.8% based reduced order modeling of dynamic structural systems. Eng. Struct. 54 (3),
POD- 0.0141 0.316 150–167.
BPNN Bienkiewicz, B., Tamura, Y., Ham, H.J., Ueda, H., Hibi, K., 1995. Proper orthogonal
along-wind base bending Measured 0.607 3.6% 0.0626 4.8% decomposition and reconstruction of multichannel roof pressure. J. Wind Eng. Ind.
moment POD- 0.585 0.0596 Aerodyn. 54–55, 369–381.
BPNN Carassale, L., Solari, G., Tubino, F., 2007. Proper orthogonal decomposition in wind
across-wind base bending Measured 0.00482 / 0.173 4.6% engineering–Part 2: theoretical aspects and some applications. Wind Struct. 10,
moment POD- 0.00432 0.165 177–208.
BPNN Chen, Y., Kopp, G.A., Surry, D., 2002. Interpolation of wind-induced pressure time series
with an artificial neural network. J. Wind Eng. Ind. Aerodyn. 90, 589–615.
base torsional moment Measured 0.00951 / 0.178 4.0%
Chen, Y., Kopp, G.A., Surry, D., 2003. Prediction of pressure coefficients on roofs of low
POD- 0.00980 0.185
buildings using artificial neural networks. J. Wind Eng. Ind. Aerodyn. 91, 423–441.
BPNN
Chen, C.H., Wu, J.C., Jow-Hua, Chen, 2008. Prediction of flutter derivatives by artificial
neural networks. J. Wind Eng. Ind. Aerodyn. 96, 1925–1937.
Chen, B., Wu, T., Yang, Y., Yang, Q., Li, Q., Kareem, A., 2016. Wind effects on a cable-
pressure field is competent. suspended roof: full-scale measurements and wind tunnel based predictions. J. Wind
Eng. Ind. Aerodyn. 155, 159–173.
English, E.C., Fricke, F.R., 1999. The interference index and its prediction using a neural
6. Conclusions network analysis of wind-tunnel data. J. Wind Eng. Ind. Aerodyn. 83, 567–575.
Fan, J.Y., 2003. A modified Levenberg-Marquardt method for singular system of
nonlinear equations. J. Comput. Math. 21 (5), 625–636.
In our work, a prediction method involving proper orthogonal
Fan, J.Y., Pan, J.Y., 2009. A note on the Levenberg–Marquardt parameter. Appl. Math.
decomposition combined with a backpropagation neural network (POD- Comput. 207, 351–359.
BPNN) technique trained by the global LM algorithm are proposed for Flood, I., Kartam, N., 1994. Neural networks in civil engineering I: principles and
understanding. J. Comput. Civ. Eng. ASCE 8 (2), 131–148.
predicting wind loads on structures. Comparing with other method, the
Franke, R., 1982. Scattered data interpolation: test of some methods. Math. Comp. 38
proposed POD-BPNN has the flowing advantages: (157), 181–200.
Fu, J.Y., Liang, S.G., Li, Q.S., 2007. Prediction of wind-induced pressures on a large
1) The NN structure can describe the complex and non-linear unknown gymnasium roof using artificial neural networks. Comput. Struct. 85, 179–192.
Gavalda, X., Ferrer-Gener, J., Kopp, G.A., Giralt, F., 2011. Interpolation of pressure
relationships of positions and wind loads on structure by training coefficients for low-rise buildings of different plan dimensions and roof slopes using
some input (coordinates)-output (values of wind pressure) pairs from artificial neural networks. J. Wind Eng. Ind. Aerodyn. 99, 658–664.
tests based on BP algorithm. Thus it has the characteristics of self- GB50009-2012, 2002. Load Code for the Design of Building Structures. Architecture &
Building Press of China, Beijing [in Chinese]).
adjustment and robustness for predicting pressure coefficients, e.g., Haykin, S., 1999. Neural Networks: a Comprehensive Foundation, seconded. Macmillan
Cp-mean and Cp-rms, and its prediction accuracy is overall better than Publishing, New York.
the simple inverse square distance weighting (ISDW) method, espe- Holmes, J.D., Sankaran, R., Kwok, K.C.S., Syme, M.J., 1997. Eigenvector modes of
fluctuating pressures on low-rise building models. J. Wind Eng. Ind. Aerodyn. 69–71,
cially at the locations that the local wind pressures vary sharply. 697–707.
2) The proposed POD-BPNN can convert a complex, time related, pre- Huang, D.M., Zhu, L.D., Chen, W., 2014. Power spectra of wind forces on a high-rise
diction of pressure time series into a time-independent interpolation building with section varying along height. Wind Struct. 18 (3), 295–320.
Huang, D.M., Zhu, L.D., Chen, W., 2015a. Vertical coherence functions of wind forces and
of orthogonal load modes, thus it has a good convergence (convenient
influences on wind-induced responses of a high-rise building with section varying
for using the global LM algorithm to train the neural network) and along height. Wind Struct. 21 (2), 119–158.
parallelism (the interpolation process of orthogonal modes can be Huang, D.M., Zhu, L.D., Chen, W., 2015b. Covariance proper transformation-based
pseudo excitation algorithm and simplified SRSS method for the response of high-rise
carried out simultaneously). It is a successful attempt to the predic-
building subject to wind-induced multi-excitation. Eng. Struct. 100, 425–441.
tion of local wind pressure time series on a high-rise building, even Kato, Y., Kanda, M., 2014. Development of a modified hybrid aerodynamic vibration
the whole wind forces and moments, as well as their spectra and technique for simulating aerodynamic vibration of structures in a wind tunnel.
coherence functions. J. Wind Eng. Ind. Aerodyn. 135, 10–21.
Khanduri, A.C., Bedard, C., Stathopoulos, T., 1997. Modelling wind-induced interference
effects. J. Wind Eng. Ind. Aerodyn. 72, 71–79.
Moreover, the local pressure field has surely non-Gaussian charac- Kho, S., Baker, C., Hoxey, R., 2002. POD/ARMA reconstruction of the surface pressure
teristics at least in separation zones. Therefore, in subsequent studies, we field around a low rise structure. J. Wind Eng. Ind. Aerodyn. 90 (12–15), 1831–1842.
Kikuchi, H., Tamura, Y., Ueda, H., Hibi, K., 1997. Dynamic wind pressures acting on a tall
will show the results provided by this method also in terms in some high building model - proper orthogonal decomposition. J. Wind Eng. Ind. Aerodyn.
order moments, such as skewness and kurtosis. 69–71, 631–646.
Lu, G.Y., Wong, D.W., 2008. An adaptive inverse distance weighting spatial interpolation
technique. Comput. Geosci-uk 34, 1044–1055.
Acknowledgment Lumley, J.L., 1970. Stochastic Tools in Turbulence. Academic Press, San Diego.
Maier, H.R., Dandy, G.C., 1998. The effect of internal parameters and geometry on the
performance of back propagation neural networks: an empirical study. Environ.
The work described in this paper was supported by the National
Modell. Soft 13 (2), 193–209.
Natural Science Foundation (Grant no. 51208524), the Hunan Province Maren, A., Harston, C., Pap, R., 1990. Handbook of Neural Computing Applications.
Natural Science Foundation (Grant no. 12JJ4055, 2017JJ2318), the Academic Press, San Diego, USA.
Hunan Province University Innovation Platform Open Foundation (Grant Neural Network Toolbox for Use with Matlab, 2009. The Math Works Inc.
Pierre, C., 1994. Three algorithms for estimating the domain of validity of a feedforward
no. 14k104), and the Open-End Fund for the Valuable and Precision In- neural network. Neural Netw. 7 (1), 169–174.
struments of Central South University. Any opinions and concluding re- Rumelhart, D., McCelland, J., 1986. Parallel Distributed Processing, vol. 1. MIT Press,
marks presented here are entirely those of the authors. Cambridge.
Solari, G., Carassale, L., 2000. Modal transformation tools in structural dynamics and
wind engineering. Wind Struct. 3 (4), 221–241.
References Solari, G., Carassale, L., Tubino, F., 2007. Proper orthogonal decomposition in wind
engineering-Part 1: a state-of-the-art and some prospects. Wind Struct. 10, 153–176.
Tamura, Y., Suganuma, S., Kikuchi, H., Hibi, K., 1999. Proper orthogonal decomposition
Agamennoni, O., Mandolesi, P., 1997. Fast inner-outer point evaluation in a polytopic
of random wind pressure field. J. Fluids Struct. 13, 1069–1095.
generalization domain. In: Proceedings of the 1997 IEEE International Conference on
Neural Networks, Part 3, Houston, TX, USA, pp. 1547–1551.

16
H. Dongmei et al.
Turkkan, N., Srivastava, N.K., 1995. Prediction of wind load distribution for air-supported
structures using neural networks. Can. J. Civ. Eng. 22, 453–461.
Wang, H.F., Zhou, Y., 2014. POD analysis of a finite-length cylinder near wake. Exp.
Fluids 55 (8), 1–15.
Watanabe, S., 1965. Karhunen-Loeve expansion and factor analysis theoretical remarks
and application. In: Proceedings of the 4th Conference on Information Theory,
Prague, Czech Republic.
17
Journal of Wind Engineering & Industrial Aerodynamics 170 (2017) 1–17
Wu, B., 1992. An introduction to neural networks and their applications in
manufacturing. J. Intell. Manuf. 3 (6), 391–403.
Wu, T., Kareem, A., 2011. Modeling hysteretic nonlinear behavior of bridge aerodynamics
via cellular automata nested neural network. J. Wind Eng. Ind. Aerodyn. 99,
378–388.
Yu, F., Xu, X.Z., 2014. short-term load forecasting model of natural gas based on
optimized genetic algorithm and improved BP neural network. Appl. Energy 134,
102–113.