J. Wind Eng. Ind. Aerodyn.

141 (2015) 2738

Contents lists available at ScienceDirect

Journal of Wind Engineering

and Industrial Aerodynamics
journal homepage: www.elsevier.com/locate/jweia

An EMD-recursive ARIMA method to predict wind speed

for railway strong wind warning system
Hui Liu a,b,n, Hong-qi Tian a, Yan-fei Li a
a
Key Laboratory of Trafc Safety on Track of Ministry of Education, School of Trafc and Transportation Engineering, Central South University,
Changsha 410075, Hunan, China
b
Institute of Automation, Faculty of Computer Science and Electrical Engineering, University of Rostock, Rostock 18119, Mecklenburg-Vorpommern, Germany

art ic l e i nf o

a b s t r a c t

Article history:
Received 12 September 2014
Received in revised form
18 February 2015
Accepted 26 February 2015
Available online 19 March 2015

To protect running trains against the strong crosswind along Chinese QinghaiTibet railway, a strong
wind warning system is developed. As one of the most important technologies of the developed system,
a new short-term wind speed forecasting method is proposed by adopting the Empirical Mode
Decomposition (EMD) and the improved Recursive Autoregressive Integrated Moving Average (RARIMA)
model. The proposed forecasting method consists of three computational steps as: (a) use the EMD
method to decompose the original wind speed data into a group of wind speed sub-layers; (b) build the
forecasting models for all the decomposed sub-layers by utilizing the RARIMA algorithm; (c) employ the
built RARIMA models to predict the wind speed in the sub-layers; and (d) summarize the predicted
results of the wind speed sub-layers to get the nal forecasting results for the original wind speed. Since
the wind speed forecasting method is proposed for the real-time warning system, the forecasting
accuracy and the time performance of the forecasting computation are both considered. Two experiments show that: (a) the proposed method has better forecasting performance than the traditional
Autoregressive Integrated Moving Average (ARIMA) model, the Persistent Random Walk Model (PRWM)
and the Back Propagation (BP) neural networks; and (b) the proposed method has satisfactory
performance in both of the accuracy and the time performance.
& 2015 Elsevier Ltd. All rights reserved.

Keywords:
QinghaiTibet railway
Strong wind
Wind speed forecasting
Wind speed prediction
Warning system
Empirical Mode Decomposition
Recursive ARIMA
Neural networks

1. Introduction
In recent decade, there are several derailment accidents happened in the world caused by strong crosswind. A train consisting of
six vehicles was derailed between Akita station and Niigata station in
Uetsu railway line on December 25, 2005 in Japan (Train death toll
rises in Japan, 2005). The investigators from the Department of
Trafc Safety of East Japan Railway Company conrmed that the train
derailment was caused by the strong wind along the Uetsu line.
Eleven vehicles of a train were blown over in Xinjiang railway on
February 28, 2007 in China (Chen et al., 2010; Strong wind topples
Chinese train, 2007). The reason of the derailment proposed by the
Ministry of Railway of China was also the strong crosswind. Unfortunately, similar train derailment accidents also happened in Canada
(Freight train derails in southern Saskatchewan, 2014) and USA
(Shust et al., 2010).

n
Corresponding author at: Key Laboratory of Trafc Safety on Track of Ministry of
Education, School of Trafc and Transportation Engineering, Central South
University, Changsha 410075, Hunan, China.
Tel.: 86 73182655294; fax: 86 73182656374.
E-mail address: csuliuhui@csu.edu.cn (H. Liu).

http://dx.doi.org/10.1016/j.jweia.2015.02.004
0167-6105/& 2015 Elsevier Ltd. All rights reserved.

Aimed at the issue of the train derailment under the strong

crosswind, some important results have been studied and published.
A numerical-experimental procedure was provided for the aerodynamic
optimization of a new train named EMUV250 in terms of behaviors to
the crosswind (Cheli et al., 2010). In the numerical experiment, two
different train shapes were proposed by modifying the train's roof and
nose. A new framework was proposed for the consideration of the
effects of crosswind on trains based on correcting the current CEN
methodology (Baker, 2013). In the framework, an improved methodology was presented, which can be used for the train authorization and
the route risk analysis. A numerical calculation was conducted to
investigate the safe domain of a train in crosswind caused by different
attitudes (Cui et al., 2014). A methodology was put forward for the
safety assessment to the risk of a train overturning in strong cross-wind
(Andersson et al., 2004). The proposed methodology had been applied
in a high-speed railway named Botniabanan line with a maximum
speed of 250 km/h in the northeast coastal region of Sweden.
Besides the upper important studies, some scientist presented
that developing railway strong wind warning systems is also a feasible option to protect the running trains (Hoppmann et al., 2002;
Kobayashi and Shimamura, 2003; Liu et al., 2009; Pan et al., 2008). In
these warning systems, the short-term wind speed prediction is one of

28

H. Liu et al. / J. Wind Eng. Ind. Aerodyn. 141 (2015) 2738

the most important technologies. However, since the wind speed

signal is non-stationary and stochastic, it is difcult to predict it
accurately. To study the internal characters of stochastic wind speed
signals, a simulation algorithm was proposed to generate sample
functions of a stationary, multivariate stochastic process according to
its prescribed cross-spectral density matrix (Deodatis, 1996). Currently,
the wind speed forecasting methods for railway applications can be
classied as three types: physical methods, statistical methods and
intelligent methods. The physical models adopt the physical parameters (e.g., terrains, obstacles, pressures and temperatures, etc) to
estimate the future three-dimensional railway wind speed. Due to the
time performance of the physical methods, they are suitable for the
off-line estimation but not the real-time decision in the warning
systems. A representatively physical method was proposed by the
University of Genoa to study the wind hazard of the RomeNaples
High Speed (HS)/High Capacity (HC) railway line. In this physical
methods, the wind numerical simulation and the probabilistic assessment of the simulated wind speed results along the line were
completed (Burlando et al., 2010; Freda and Solari, 2010). The second
type is the statistical methods, which utilize the statistical models to
describe the changing law of the wind speed for the future prediction.
A wind speed statistical forecasting model was proposed for the
Nowcasting warning system developed by Deutsche Bahn AG in
Germany (Hoppmann et al., 2002). In this method a linear extrapolated algorithm was used to get the future wind speed prediction by
combining the average values of the historical wind speed data, the
estimated gust supplement and the error supplement. Although the
statistical methods are always simple and good at the real-time
performance, their accuracy can be further improved. The intelligent
methods focusing on the intelligent forecasting models. A wind speed
intelligent predicting method was presented using Kalman Filter
theory for the Windas system from Japanese East Railway Company
(Kobayashi and Shimamura, 2003). Although the intelligent methods
can have better accuracy than the physical and statistical ones,
sometimes they have the problems in the computational convergence.
From the upper references (Hoppmann et al., 2002; Kobayashi and
Shimamura, 2003), it can also be found that currently in the strong
wind warning systems only the wind speed signals but not the wind
direction signals are processed and used to present the warning
commands. The reasons of this condition can be explained as follows:

(a) compared to the wind speed prediction, the wind direction

prediction is more difcult. Because the wind direction signals are
always more non-stationary than the wind speed signals; (b) although
the wind direction data at a position can be measured conveniently,
the real angles between the running trains and the real-time wind
direction signals cannot be calculated accurately. Because the train
real-time postures are dynamic and nonlinear; and (c) it is not difcult
for these warning systems to measure both the wind speed data and
the wind direction data at a large number of wind stations at the same
time, but it can be a challenge for these systems to calculate the wind
speed predictions then to propose the critical safety velocities under huge different combinations of wind speed and wind direction
synchronously.
In this study, we focus on the wind speed high-accuracy predictions for the railway wind warning systems. By considering both the
forecasting accuracy and the real-time performance, the statistical
methods have been selected in this study. A new hybrid statistical
wind speed forecasting method is proposed by combing the Empirical
Mode Decomposition (EMD) and the improved recursive ARIMA
(RARIMA) model. In the proposed hybrid method, not only a new
RARIMA model is presented to calculate the wind speed predictions,
but also it is the rst time to adopt the EMD algorithm to process and
decompose the original non-stationary wind speed data. Additionally,
a comparison of the forecasting performance made by different
models is provided. The comparing models include the proposed
EMD-RARIMA model, the standard Autoregressive Integrated Moving
Average (ARIMA) model, the Back Propagation (BP) neural network
and the Persistent Random Walk Model (PRWM). The paper is
organized as follows: Section 2 presents the architecture of the strong
wind warning system; Section 3 explains the detailed steps of the
wind speed prediction; Section 4 provides the experimental results of
two cases; and Section 5 concludes this study.

2. Strong wind warning system for QinghaiTibet railway

2.1. System architecture
The Chinese QinghaiTibet railway is the highest railway line in
the world, which covers a big plateau area from the Geermu city in

Strong Wind Warning

System
GSM R (Global
System for Mobile
Communications
Railway)

Drivers of Running
Trains

Train and Track

Database

Wind Database

Train Information

Fibre
network

Network
Router

Track Information

Wind speed
station
Network
Hub

Train Dispatching Command

System Of Qinghai Tibet Railway
Company

Wireless
network
Client Computers

Wind speed
station
Fig. 1. Architecture of the strong wind warning system for the QinghaiTibet railway.

H. Liu et al. / J. Wind Eng. Ind. Aerodyn. 141 (2015) 2738

Qinghai province to the Lhasa city in Tibet province. The length

and altitude of the QinghaiTibet railway are 1142 km and
5072 km, respectively. Due to its high altitude, the strong crosswind phenomenon happens frequently along the QinghaiTibet
railway. To protect the safety of the running trains, an intelligent
warning system has been developed, as the system architecture
demonstrated in Fig. 1.
From Fig. 1, it can be seen that: (a) all the real-time information
of trains and tracks are sampled from the Train Dispatching
Command System (TDSC) of QinghaiTibet Company and sent into
the Train and Track Database (TTD) of the developed warning
system; (b) the wind speed data at the dangerous places along the
line are measured and stored in the Wind Database (WD) of the
warning system; (c) in the TTD, all the critical speeds of various
trains under different wind speed and track parameters are
calculated in advance based on the results from the Computational
Fluid Dynamics (CFD) and the wind tunnel experiments; and
(d) once the warning system proposes a warning message, the
message will be sent to the corresponding driver of the monitored
train automatically by the wireless Global System for Mobile
Communications-Railway (GSM-R) technology. The monitoring
GUI of the running wind warning system is demonstrated in Fig. 2.

29

results; and (d) the railway construction conditions. By considering

the upper aspects, the mileage positions of those wind speed stations
are selected as illustrated in Fig. 3. As shown in Fig. 3, the blue line
represents the QinghaiTibet railway and the red points are the
selected mileage positions of the wind speed stations. As shown in
Fig. 4, every wind speed station consists of two anemometers, two
solar power plates, two batteries, an on-board Micro-PC and some
mechanical structures. The working process of a wireless wind speed
station can be explained as follows: (a) once a wind speed station is
turned on, it will connect to the remote wind warning system
automatically by the GPRS/GSM-R dual redundant channels;
(b) once the wireless data transmitting channel is built, the wind
speed station will start to measure the wind speed data and sent
them to the remote server synchronously; (c) to insure the measured
wind speed data are credible, an error-checking strategy is proposed
based on the installed two anemometers. If the measured results of
the two anemometers are signicantly different (like 10%), the micropc of the wind speed station will terminate the measuring process
and sent out an error message; and (d) the sampling frequency of
every wind speed station can be revised remotely by the server of the
warning system.

2.2. Network of wind speed measurement

To measure the original wind speed data along the QinghaiTibet
railway, 52 wind speed stations are installed, including two types:
the wireless stations and the wired stations. The wireless stations use
the GPRS/GSM-R communicating channel and the wired stations
adopt the LAN TCP/IP one. In these wind speed stations, the wind cup
anemometers are adopted to sample the raw wind speed data. The
sampling frequency of the wind speed stations is ten times per
minute. The positions of those wind speed stations are decided by
considering four aspects as: (a) the historical derailment accidents;
(b) the historical meteorological data; (c) the CFD computational

Fig. 3. Mileage positions of the 52 wind speed stations along the QinghaiTibet
railway. (For interpretation of the references to color in this gure, the reader is
referred to the web version of this article.)

Fig. 2. Monitoring GUI of the strong wind warning system for the QinghaiTibet railway.

30

H. Liu et al. / J. Wind Eng. Ind. Aerodyn. 141 (2015) 2738

2.3. Application of wind speed predictions

The concept of the application of the proposed wind speed
predictions in the warning system is given in Fig. 5.
As shown in Fig. 5, the working process can be explained as
follows:
(a) The warning system reads all the train dispatching data from
the QinghaiTibet Railway Company (see Fig. 1), so it knows
the detailed dispatching information of all trains (e.g., the
arriving time and the leaving time to a train station, the
mileage distances between the running trains and the
installed wind speed stations, the train real-time running
velocities, etc).
(b) When a train reaches the rst train station 1, the warning
system will start to calculate the wind speed predictions. In
the warning system, three time parameters t1, t2 and t3 are
dened to manage the predicted wind speed results. The
parameter t1 is proposed to validate the built forecasting
model. In the t1 time, the wind speed predicted results will
not be used to dispatch the trains but only to estimate the

Fig. 4. Wireless wind speed station at Fenghuoshan Mountain.

L1

t1

L2
t2

3. Wind speed forecasting

3.1. Framework of EMD-RARIMA forecasting method

t4

t3

Fig. 5. Concept of the application of the wind speed predictions in the warning
system.

Start:
Original Railway
Wind Speed Data

forecasting performance. If the results of a forecasting model

cannot meet the requirements of the accuracy and the computational time performance in this period, this forecasting
model will be terminated. The parameter t2 is presented for
dispatching the trains. Once a model enters the t2 time
period, it means this model is qualied and the forecasting
results of this model will be adopted to generate various
warning commands. The parameter t3 is provided for the
trains' drivers to process a coming warning command.
(c) When the train enters the areas managed by the time period
t2, the warning system will start to output the multi-step
wind speed predictions. When the trains enters areas managed by the time period t3, the warning system will use the
latest multi-step wind speed forecasting value to search for
the critical safety velocity of the train. If the current train
running velocity is bigger than the expected critical safety
velocity, a warning message will be sent by the system to the
driver of the train.
(d) In this study, the parameters t2 and t3 are equal to ve
minutes and two minutes, respectively. They are both proposed by the QinghaiTibet Railway Company. They can be
revised referring to the specic parameters of a running train.
In the period t2, there are two standards adopted for the
validation of the wind speed forecasting models. The rst
validating standard is the forecasting accuracy. The current
accuracy in the warning system is required as: the MAE (Mean
Absolute Error) is less than 2 m/s and the MAPE (Mean
Absolute Percentage Error) is less than 10%. The second
validating standard is the real-time forecasting performance.
It estimates how fast the models can output the multi-step
wind speed predictions. The current requirement of the
forecasting elapsed time is less than 3 s.
(e) There is a protecting strategy in the developed warning
system. If the wind speed forecasted results cannot meet the
accuracy or the time requirements, the real-time wind speed
measured data will be used to replace the multi-step forecasting results to propose the critical safety velocities of the
monitored trains.

In this study, a new hybrid EMD-RARIMA method is proposed

to predict the wind speed. The computational framework of the
proposed method is given in Fig. 6. As shown in Fig. 6, the contents
of the method include: (a) use the EMD to decompose the original

Empirical Mode
Decomposition
(EMD)

Recursive ARIMA Modeling

Summarizing
Calculation

Decomposed Wind Speed

Sub layer

Predicted Results in the

EMD Decomposed Series

End:
Final Wind Speed
Predictions

Fig. 6. Framework of the proposed EMD-RARIMA forecasting method.

H. Liu et al. / J. Wind Eng. Ind. Aerodyn. 141 (2015) 2738

31

manually, the computational process of the EMD decomposition is

adaptive and automatic (Liu et al., 2015).
The EMD considers any signal Xt can be described using the
equation as follows:
Xt

n
X

C i t Rt

i1



where C i t
 ; i 1; 2; ; n is the IMFs in different decompositions, Rt is the residue and n is the number of the IMFs.
The computational steps of the EMD are given as follows
(Huang et al., 1998):
Fig. 7. Original wind speed time series fX 1t g.



Step 1: Identify all the local extrema of series Xt , including
local maxima and local minima.
Step 2: Connect all the local maxima
by a cubic spline line to


generate its upper
envelop X up t . Similarly the lower

envelop X low t is made with all the
 local
 minima.
Step 3: Compute the mean envelop Mt from the upper and
lower envelops as follows:


Mt X up t X low t =2
2
Step 4: Extract the details as follows:
Zt Xt Mt

Fig. 8. Original wind speed time series fX 2t g.

wind speed data into a number of wind speed sub-layers; (b) build
the recursive ARIMA models for all the sub-layers and adopt
the built models to forecast the multi-step predictions; and
(c) summarize the wind speed multi-step predictions of the sublayers to get the nal predictions for the original wind speed.
3.2. Original wind speed data
Two groups of real wind speed series measured by two
different wind speed stations (i.e., the 30th station and the 32nd
station) along the QinghaiTibet railway are adopted to demonstrate the forecasting performance of the proposed EMD-RARIMA
method, as shown in Figs. 7 and 8. The time interval of two
consecutive data points is one minute. The mileage distance
between the two wind speed stations is 41 km. They are named
fX 1t g series and fX 2t g series, respectively. The 1st300th ones of
them are used to establish the forecasting models and the 301st
400th ones are utilized to check the built models.
3.3. Wind speed decomposition by EMD
As explained in Section 3.1, the Empirical Mode Decomposition
(EMD) is adopted to decompose the originally non-stationary
wind speed samples. The EMD is a mathematical time domain
decomposing method presented by scientists N. E. Huang in 1998
(Huang et al., 1998), which can convert a group of time series into
locally narrow band components, named Intrinsic Mode Functions
(IMFs). The purpose of executing the EMD decomposition in the
study is to convert the original non-stationary wind speed series
into a number of relatively stable wind speed sub-layers, which
will decrease the difculty to realize the high-precision predictions for the original wind speed. Different to the wavelet decomposition and the wavelet packet decomposition (Liu et al., 2013a,
2013b) which both need to select the decomposing parameters





Step 5: Check whether Zt is an IMF: (a) if Zt is an IMF


then set Ct Zt and meantime
Xt with the
 replace

residual Rt Xt  Ct ; (b) if Zt is not an IMF, replace




Xt with Zt then repeat Steps 24 until the termination
criterion is satised. The following equation can be regarded as
the termination condition of this iterative calculation:

2
m
X
Z j  1 t  Z j t
r j 1; 2; ; t 1; 2; ; m
4

2
Z j  1 t
t1
where m is the number of the wind speed data points, is the
terminated parameter, and j denotes the times of iterative
calculation. In this study, the is equal to 0.2.
Step 6: The procedure of Steps 15 is repeated until all the IMFs
are found.
The EMD decomposed results of the original wind speed series
fX 1t g are given in Fig. 9. From Fig. 9, it can be seen that the fX 1t g
series have been converted into a group of wind speed sub-layers
successfully. A number of RARIMA models will be built in these
decomposed wind speed sub-layers to complete the multi-step
forecasting computation. It is obvious that it is easier for the built
RARIMA models to obtain the accurate forecasting results by using
the decomposed sub-layers than adopting the original wind
speed data.
In this study, all of the decomposed wind speed sub-layers will
be checked their stability by the Run Sequence Method (RSM)
(Hentati-Kaffel and de Peretti, 2015). The RSM is a non-parametric
statistical test to check the stability of a group of time series data.
The contents of the RSM algorithm are given as follows:
Step 1: Calculate the average value X of an original series fX t g.
Step 2: Convert the original series fX t g to a symbol series fSt g
using the criterion as below: set X t i is the ith wind speed
sample in the series fX t g; if X t i Z X, the symbol will be
adopted to replace the position of this sample; Oppositely, if
X t i o X, the symbol  will be used to replace the position of
this sample.
Step 3: Calculate the number of the positive runs N and the
number of the negative runs N  using the standard as below:

32

H. Liu et al. / J. Wind Eng. Ind. Aerodyn. 141 (2015) 2738

Fig. 9. Decomposed results of the original wind speed series fX 1t g. (a) fX D1t g series, (b) fX D2t g series, (c) fX D3t g series, (d) fX D4t g series, (e) fX D5t g series, (f) fX D6t g series,
(g) fX D7t g series, (h) fX D8t g series, (i) fX Rt g series.

H. Liu et al. / J. Wind Eng. Ind. Aerodyn. 141 (2015) 2738

in the converted symbol series fSt g, a sequence of the consecutive symbols is named a positive run, and a sequence of
the consecutive  symbols is named a negative run. For
example, the N is equal to three and the N  is equal to two
in a symbol series like    .
Step 4: The RSM considers that if the time series fX t g is stable,
its corresponding symbol series fSt g should meet the normal
distribution. So the following statistical equations can be used
to check whether the series fX t g is stable as:
Z  =

where:
2N 1 N 2 2N 1 N 2  N 1=2

N2 N  1

2N1 N 2
1
N

N N1 N2

Then, if jZ jr 1:96, the series fX t g is stable under the signicant level parameter 0:05. Otherwise, the series fX t g is
unstable.
3.4. Wind speed forecasting by RARIMA
Time series method is a statistical algorithm which can build an
explicit equation to describe the potential changing law of a
section of time series data (Wei, 1994). It has ve standard models,
including the Auto-Regressive (AR), the Moving Average (MA), the
Auto-Regressive Moving Average (ARMA) and the Auto-Regressive
Integrated Moving Average (ARIMA). The ARIMA models can
describe all the other ones.
3.4.1. Autoregressive Integrated Moving Average (ARIMA) model
The equations of the ARIMA (p, d, q) are dened as follows:

B UK d UXt B Uat

where:
8
B Xt  1=Xt
>
>
>
>
< K 1 B

B 1  1 B  2 B2  3 B3   p  1 Bp  1  p Bp
>
>
>
>
: B 1  B  B2  B3  
Bq  1  Bq
1

q1

10





where Xt is the wind speed time series, at is the random error
series assumed to be a white noise with a mean of zero and equal
variance, B is the backward shift operator, K is the difference

33

operator, p is the order of the auto-regressive part in the ARIMA

model, d is the order of the difference computation operator,
q is the
n o
order of the moving average part in the ARIMA
model,
is the

p
 
parameter of the auto-regressive part, and q is the parameter of
the moving average part.
As shown in Eq. (9), the ARIMA (p, d, q) model is actually a
combination of the AR(p) model, the MA(q) model and the
component of d times Difference Computation (DC). To demonstrate the derivation of the ARIMA (p, d, q) model, an ARIMA(2, 2,
2) model is taken as an example as follows:
1  1 B  2 B2 1  B2
 2 B2 U at
U DC U Xt 1  1 BMA2
AR2

) 1  1 B  2 B2 U 1 B2 2B U Xt 1  1 B  2 B2 U at
) 1  2 1 B 1 21  2 B2  1  22 B3  2 B4  UXt
1  1 B  2 B2 U at
) Xt  2 1 Xt  1 1 21  2 Xt  2  1  22
Xt  3  2 Xt  4 at  1 at 1  2 at  2
) Xt 2 1 Xt 1 1 21  2 Xt  2 1 22
11
Xt  3 2 Xt  4 at  1 at  1  2 at  2
3.4.2. BoxJenkins methodology
The BoxJenkins methodology, named after statisticians George
Box and Gwilym Jenkins, is presented to establish the suitable
time series models. The reason to choose the BoxJenkins methodology in the study is that it has good real-time performance due
to its simple computational equations.
The BoxJenkins methodology consists of three modeling contents:
the model identication, the selection of model order and the
parameter estimation. The details of the BoxJenkins methodology
can be explained as: (a) in the step of model identication, the
Autocorrelation function (ACF) and partial autocorrelation function
(PACF) are always utilized to decide whether an Auto-Regressive (AR)
component or a Moving Average (MA) one should be included in the
ARIMA models. If the ACF coefcients show smearing effect, the AR
component will be included. If the PACF coefcients indicate censoring
effect, the MA component will be included. Additionally, the Run
Sequence Method (RSM) is employed to identify the stability of the
EMD decomposed sub-layers. If a sub-layer has been found that it is
unstable, the Difference Computation (DC) will be executed on this
sub-layer; (b) in the step of selection of model order, some criterions
(e.g., Final Prediction Error, Akaike Information Criterion, Bayes Information Criterion, etc) will be adopted to select the best orders of the
identied ARIMA models; and (c) in the step of parameter estimation,
some methods (e.g., Least Square Estimation, YuleWalker Estimation,
Greatly Relieved Estimate, etc) will be utilized to calculate the optimal

Fig. 10. Calculated results of the ACF/PACF coefcients for the fX D1t g series. (a) ACF results, (b) PACF results.

34

H. Liu et al. / J. Wind Eng. Ind. Aerodyn. 141 (2015) 2738

parameters of the identied ARIMA models. In this study, the Final

Prediction Error (FPE) criterion is used to choose the orders of the
ARIMA models and the YuleWalker algorithm is adopted to calculate
the parameters of the ARIMA models. The detailed procedure of the
BoxJenkins methodology can be found in Pankratz (1983).
The results of the ACF and PACF coefcients for the decomposed fX D1t g series are shown in Fig. 10. From Fig. 10, it can be seen
that the ACF results have smearing phenomenon while the PACF
results have censoring effect, which can be concluded that the
fX D1t g series meet the ARIMA (p, 0, 0) model.
To select the best order p for the recognized ARIMA (p, 0, 0),
the FPE coefcients of the fX D1t g series are calculated using the
following equations as follows:
8
p
X
>
>
>
FPEp 1 p=N1 p=NR^ 0 
^ j R^ j
>
>
<
j1
12
N
k
>
X
>
>
^k 1
>
XtXt

k
R
>
N
:
t1

where p is the order of then ARIMA(p,

0, 0) model, N is the
o


^
number of the
Xt ,
is the estimated AR operating
n series
o
j
series, and R^ k is the estimated autocorrelation coefcients.
Based on the computational results completed by Eq. (12), the
best model for the fX D1t g series is selected as ARIMA(7,0,0).
To estimate the parameters of the identied ARIMA(7,0,0), the
YuleWalker computation is executed using the following equation as follows:
2
3 2
3  12 3
^
R^ p  1
R^ 1
R^ 0
R^
6 17 6
7 6 17
6
7 6 R^ 2 7
^ 7 6 R^
^
^

R
R
6 27 6 1
0
p2 7
7
13
6
76
7 6
6 7 6
7 6
7

4
5 4
5 4 5
R^ p
R^
R^
R^
^
p

p1

p2

Based on Eq. (13), the nal equation of the ARIMA(7,0,0) model

for the fX D1t g series can be obtained as follows:

3.4.3. Recursive estimation algorithm for model parameters

To strength the tracking ability of the built time series models
in their multi-step recursive computation, a new recursive algorithm is proposed as in Fig. 11.
Based on Fig. 11, the recursive ARIMA computational process of
the decomposed fX D1t g series can be demonstrated as follows:
(a) In the non-recursive ARIMA forecasting strategy, the threestep predictions can be made by using Eq. (14) as below. It can be
found that the parameters of Eqs. (15)(17) will not be updated.

 One-step forecasting:
X^ D1t 301 0:1879X D1t 300  0:1910X D1t 299
0:1988X D1t 298 0:2706X D1t 297
0:1377X D1t 296 0:3088X D1t 295
0:0873X D1t 294

15

 Two-step forecasting:
X^ D1t 302 0:1879X^ D1t 301  0:1910X D1t 300 0:1988X D1t 299

0:2706X D1t 298 0:1377X D1t 297

0:3088X D1t 296 0:0873X D1t 295

16

 Three-step forecasting:
X^ D1t 303 0:1879X^ D1t 302  0:1910X^ D1t 301 0:1988X D1t 300
0:2706X D1t 299 0:1377X D1t 298 0:3088X D1t 297

0:0873X D1t 296

17

(b) Different to the non-recursive ARIMA computational process,

in the proposed recursive ARIMA strategy the parameters of the ARIMA models will be updated in real-time by
using the previous forecasted data. For instance, the series
fX^ D1t 301; X D1t 300; X D1t 299:; X D1t 298; X D1t 297; X D1t 296;
X D1t 295gare adopted to update the parameters of Eq. (19) before
calculating the forecasting value X^ D1t 302.

X D1t t 0:1879X D1t t  1  0:1910X D1t t  2

0:1988X D1t t 3 0:2706X D1t t  4

 One-step forecasting:

0:1377X D1t t 5 0:3088X D1t t  6

0:0873X D1t t 7 aD1t

14

where t is the sampling time, and faD1t g is a white noise series.

X^ D1t 301 0:1879X D1t 300  0:1910X D1t 299 0:1988X D1t 298
0:2706X D1t 297 0:1377X D1t 296 0:3088X D1t 295

0:0873X D1t 294

Fig. 11. Calculation steps of the proposed recursive ARIMA model.

18

H. Liu et al. / J. Wind Eng. Ind. Aerodyn. 141 (2015) 2738

 Two-step forecasting:
X^ D1t 302 0:1327X^ D1t 301  0:1253X D1t 300 0:1680X D1t 299
0:2837X D1t 298 0:1102X D1t 297 0:3235X D1t 296
0:1071X D1t 295

19

35

(c) Since the RARIMA models can adopt the latest forecasted results
to update the parameters of the forecasting equations before starting a
new computational process so that the RARIMA models can have
better forecasting accuracy than the traditional ARIMA models (i.e.,
non-recursive ARIMA models).

 Three-step forecasting:
X^ D1t 303 0:1993X^ D1t 302  0:0945X^ D1t 301
0:1895X D1t 300 0:2710X D1t 299
0:0771X D1t 298 0:2953X D1t 297
0:0623X D1t 296

3.5. Estimation standard for forecasting results

20

Three error indexes are employed to estimate the accuracy

performance of all the involved forecasting models (Liu et al., 2012),

Fig. 12. Results of the one-step predictions for the original wind speed series fX 1t g by the EMD-RARIMA model, the BP neural networks, the ARIMA and the PRWM.

Fig. 13. Results of the three-step predictions for the original wind speed series fX 1t g by the EMD-RARIMA model, the BP neural networks, the ARIMA and the PRWM.

Fig. 14. Results of the ve-step predictions for the original wind speed series fX 1t g by the EMD-RARIMA model, the BP neural networks, the ARIMA and the PRWM.

36

H. Liu et al. / J. Wind Eng. Ind. Aerodyn. 141 (2015) 2738

including the Mean Absolute Error (MAE), the Mean Absolute Percentage Error (MAPE) and the Standard Deviation (SD).

 Mean Absolute Error:

MAE

M
1 X
^
j Xi  Xij
Mi1

21

 Mean Absolute Percentage Error:

MAPE

M
^
1 X
Xi  Xi
j
j
^
Mi1
Xi

22

 Standard Deviation:

v
u
M
u 1 X
^ 2
Xi  Xi
SD t
M 1 i 1

23
n

o


^
where Xi is the measured wind speed time series, Xi
is
the forecasted wind
time series and M is the number of
 speed

the terms in the Xi series.

Table 2
Analysis of the time performance of the predictions given in Figs. 1214.

4. Experimental results and analysis

In this study, two different cases are selected to demonstrate
the effectiveness of the proposed wind speed forecasting methods
to protect the safety of the running trains. In the Case One, there is
no warning message proposed by the warning system because all
the wind speed do not exceed the safety threshold. In the Case
Two, there is a warning message presented by the warning system
because some wind speed date exceed the safety threshold.
Additionally, the wind speed data in the Case One and the Case
Two show the decreasing and increasing trend, respectively.

Figs. 1214 show the prediction results of the 301st400th

original wind speed series fX 1t g by the different forecasting methods. The estimated results of the accuracy and the time performance (with a normal PC with Intel i5 CPU) for these predictions are
given in Tables 1 and 2, respectively. Based on the original results
shown in Table 1, the improvement percentages of the accuracies of
the ARIMA, the BP and the PRWM by the hybrid EMD-RARIMA
model can also be calculated as the results shown in Tables 35.
From Tables 15, it can be analyzed that:
(a) When comparing the proposed hybrid EMD-RARIMA model
with the BP neural network, the former has improved the
accuracy performance of the latter obviously. The promoting

EMD-RARIMA
1-step

Time (s)
Indexes
Time (s)

EMD-RARIMA

Indexes

MAE (%)
MAPE (%)
SD (%)

BP neural network

3-step

5-step

1-step

3-step

5-step

1.2243
1.8413
ARIMA model
1-step
3-step
0.0631
0.0752

1.9743

2.8940
3.1223
PRWM model
1-step
3-step
0.0165

3.4475

5-step
0.0937

5-step

EMD-RARIMA vs. BP neural network

1-step

3-step

5-step

63.52
63.11
62.14

75.03
75.24
70.01

82.75
83.24
76.05

Table 4
Improved accuracy percentages of the ARIMA model by the proposed hybrid EMDRARIMA model.
Indexes

MAE (%)
MAPE (%)
SD (%)

Table 1
Analysis of the accuracies of the predictions given in Figs. 1214.

EMD-RARIMA vs. ARIMA

1-step

3-step

5-step

50.40
50.00
52.22

64.15
63.81
63.32

75.41
75.76
69.30

BP neural network

1-step

3-step

5-step

1-step

3-step

5-step

MAE (m/s)
MAPE (%)
SD (m/s)

0.1607
0.76
0.2343

0.2763
1.31
0.4166

0.3248
1.51
0.5522

0.4405
2.06
0.6188

1.1067
5.29
1.3889

1.8826
9.01
2.3054

Indexes

ARIMA model

MAE (m/s)
MAPE (%)
SD (m/s)

Indexes

Table 3
Improved accuracy percentages of the BP neural network by the proposed hybrid
EMD-RARIMA model.

4.1. Case One

Indexes

accuracy percentages of the MAE indexes from one-step to

ve-step are 63.52%, 75.03% and 82.75%, respectively. The
promoting accuracy percentages of the MAPE indexes from
one-step to ve-step are 63.11%, 75.24% and 83.24%, respectively. The promoting accuracy percentages of the SD indexes
from one-step to ve-step are 62.14%, 70.01% and 76.05%,
respectively. The reason of this improving phenomenon is that
the EMD converts the original jumping wind speed series into
a series of relatively stationary wind speed sub-layers.
(b) When comparing the proposed hybrid EMD-RARIMA model
with the ARIMA model, the former has also improved the
performance of the latter considerably. The promoting accuracy percentages of the MAE indexes from one-step to vestep are 50.40%, 64.15% and 75.41%, respectively. The promoting accuracy percentages of the MAPE indexes from one-step
to ve-step are 50.00%, 63.81% and 75.76%, respectively. The
promoting accuracy percentages of the SD indexes from onestep to ve-step are 52.22%, 63.32% and 69.30%, respectively.
The reasons of this promoting phenomenon are given as: the
adopted EMD decreases the forecasting difculty of the ARIMA

Table 5
Improved accuracy percentages of the PRWM model by the proposed hybrid EMDRARIMA model.
Indexes

PRWM model

1-step

3-step

5-step

1-step

3-step

5-step

0.3240
1.52
0.4904

0.7708
3.62
1.1357

1.3210
6.23
1.7988

0.5610
2.64
0.7886

MAE (%)
MAPE (%)
SD (%)

EMD-RARIMA vs. PRWM

1-step

3-step

5-step

71.35
71.21
70.29

H. Liu et al. / J. Wind Eng. Ind. Aerodyn. 141 (2015) 2738

models by decomposing the original wind speed series; and

The RARIMA component in the hybrid EMD- RARIMA method
improves the forecasting performance of the ARIMA model by
updating the equation parameters in every step.
(c) When comparing the proposed hybrid EMD-RARIMA model
with the PRWM model, the former has still improved the
performance of the latter considerably. The promoting accuracy percentages of the MAE, MAPE and SD indexes for the
one-step predictions are 71.35%, 71.21% and 70.29%.

37

(d) All the built forecasting models can meet the elapsed time
requirements of the warning system.
4.2. Case Two
Figs. 1517 show the forecasting results of the 301st400th original
wind speed series fX 2t g by the different forecasting methods. The
results of the accuracy and the time performance estimation for these
predictions are given in Tables 6 and 7, respectively.

Fig. 15. Results of the one-step predictions for the original wind speed series fX 2t g by the EMD-RARIMA model, the BP neural networks, the ARIMA and the PRWM.

Fig. 16. Results of the three-step predictions for the original wind speed series fX 2t g by the EMD-RARIMA model, the BP neural networks, the ARIMA and the PRWM.

Fig. 17. Results of the ve-step predictions for the original wind speed series fX 2t g by the EMD-RARIMA model, the BP neural networks, the ARIMA and the PRWM.

38

H. Liu et al. / J. Wind Eng. Ind. Aerodyn. 141 (2015) 2738

Program of the Central South University (Principle Investigator: Dr.

Hui Liu). This study is partly supported by the National Natural
Science Foundation of China (Grant nos. U1134203, U1334205).

Table 6
Analysis of the accuracies of the predictions given in Figs. 1517.
Indexes

EMD-RARIMA

BP neural network

1-step

3-step

5-step

1-step

3-step

5-step

MAE
MAPE
SD

0.2103
1.53
0.2747

0.3408
2.54
0.4924

0.6620
4.84
1.0625

0.7582
5.03
1.1185

1.6859
11.12
2.1842

2.4703
16.13
3.2208

Indexes

ARIMA model

MAE
MAPE
SD

PRWM model

1-step

3-step

5-step

1-step

3-step

5-step

0.4884
3.51
0.7016

1.2543
9.32
1.8286

1.9998
14.88
2.7480

0.7606
5.75
1.0161

Table 7
Analysis of the time performance of the predictions given in Figs. 1517.
Indexes

EMD-RARIMA

BP neural network

1-step

3-step

5-step

1-step

3-step

5-step

Time (s)

1.3289

1.9023

2.0189

2.8528

3.2109

3.5723

Indexes

ARIMA model

Time (s)

PRWM model

1-step

3-step

5-step

1-step

3-step

5-step

0.0708

0.0832

0.0998

0.0185

From Tables 6 and 7, the same conclusions can be made to the

ones proposed in the Case One in Section 4.1 as follows: (a) the
proposed hybrid EMD-RARIMA model has the best forecasting
accuracy in all step predictions among all the built forecasting models;
and (b) all of the built models can meet the time performance
requirements proposed by the warning system (i.e.,o3 s).
5. Conclusions
In this study a wind warning system is developed to protect the
running trains under strong crosswind along the China Qinghai
Tibet railway. In the warning system, a hybrid EMD-RARIMA
method is proposed to forecast the multi-step wind speed. Two
experimental cases validate that: (a) the proposed EMD-RARIMA
method has satisfactory performance, which meets all the requirements of the warning system either in the forecasting accuracy or
in the real-time computation; (b) compared to the ARIMA model,
the BP neural network and the PRWM model, the proposed EMDRARIMA method improves all their forecasting accuracy considerably; and (c) in the current warning system, the forecasted wind
speed data are used to select the trains' critical safety velocities to
propose warning messages. In the future work, the wind direction
signals will also be included in the system decision to optimize the
proposal of the warning messages.
Acknowledgments
The authors would like to thank the referees for their precious
reviewing. This study is fully supported by the National Natural
Science Foundation of China (Grant no. 51308553), the Scientic
Research Fund of Hunan Provincial Education Department (Principle Investigator: Dr. Hui Liu) and the Shenghua Yu-ying Talents

References
Andersson, E., Hggstrm, J., Sima, M., Stichel, S., 2004. Assessment of trainoverturning risk due to strong cross-winds. Proc. Inst. Mech. Eng. F: J. Rail Rapid
Transit 218, 213223. http://dx.doi.org/10.1243/0954409042389382.
Baker, C., 2013. A framework for the consideration of the effects of crosswinds on
trains. J. Wind Eng. Ind. Aerodyn. A 123, 130142. http://dx.doi.org/10.1016/j.
jweia.2013.09.015.
Burlando, M., Freda, A., Ratto, C.F., Solari, G., 2010. A pilot study of the wind speed
along the RomeNaples HS/HC railway line. Part 1numerical modelling and
wind simulations. J. Wind Eng. Ind. Aerodyn. 98, 392403. http://dx.doi.org/
10.1016/j.jweia.2009.12.006.
Cheli, F., Ripamonti, F., Rocchi, D., Tomasini, G., 2010. Aerodynamic behaviour
investigation of the new EMUV250 train to cross wind. J. Wind Eng. Ind.
Aerodyn. 98, 189201. http://dx.doi.org/10.1016/j.jweia.2009.10.015.
Chen, R., Zeng, Q., Huang, Y., Xiang, J., Wen, Y., Guo, X., Yin, C., Dong, H., Zhao, G.,
2010. Analysis theory of random energy of train derailment in wind. Sci. China
Phys. Mech. Astron. 53, 751757. http://dx.doi.org/10.1007/s11433-010-0158-2.
Cui, T., Zhang, W., Sun, B., 2014. Investigation of train safety domain in cross wind in
respect of attitude change. J. Wind Eng. Ind. Aerodyn. 130, 7587. http://dx.doi.
org/10.1016/j.jweia.2014.04.006.
Deodatis, G., 1996. Simulation of ergodic multivariate stochastic processes. J. Eng. Mech.
122, 778787. http://dx.doi.org/10.1061/(ASCE)0733-9399(1996)122:8(778).
Freda, A., Solari, G., 2010. A pilot study of the wind speed along the RomeNaples
HS/HC railway line.: Part 2probabilistic analyses and methodology assessment. J. Wind Eng. Ind. Aerodyn. 98, 404416. http://dx.doi.org/10.1016/j.
jweia.2009.12.005.
Freight train derails in southern Saskatchewan [WWW Document], 2014. Prince
George Citiz. URL http://www.princegeorgecitizen.com/news/national/freighttrain-derails-in-southern-saskatchewan-cpr-says-plow-winds-were-cause-1.
1334141 (accessed 9.03.14).
Hentati-Kaffel, R., de Peretti, P., 2015. Generalized runs tests to detect randomness
in hedge funds returns. J. Bank. Financ. 50, 608615. http://dx.doi.org/10.1016/j.
jbankn.2014.07.011.
Hoppmann, U., Koenig, S., Tielkes, T., Matschke, G., 2002. A short-term strong wind
prediction model for railway application: design and verication. In: Proceedings
of the 3rd European-African Conference on Wind Engineering, J. Wind Eng. Ind.
Aerodyn. 90, 11271134. http://dx.doi.org/10.1016/S0167-6105(02)00226-X.
Huang, N.E., Shen, Z., Long, S.R., Wu, M.C., Shih, H.H., Zheng, Q., Yen, N.-C., Tung, C.
C., Liu, H.H., 1998. The empirical mode decomposition and the Hilbert spectrum
for nonlinear and non-stationary time series analysis. Proc. R. Soc. Lond. Ser.
Math. Phys. Eng. Sci. 454, 903995. http://dx.doi.org/10.1098/rspa.1998.0193.
Kobayashi, N., Shimamura, M., 2003. Study of a strong warning system. JR East Tech.
Rev., 6165.
Liu, H., Tian, H., Chen, C., Li, Y., 2013a. An experimental investigation of two
Wavelet-MLP hybrid frameworks for wind speed prediction using GA and PSO
optimization. Int. J. Electr. Power Energy Syst. 52, 161173. http://dx.doi.org/
10.1016/j.ijepes.2013.03.034.
Liu, H., Tian, H., Li, Y., 2012. Comparison of two new ARIMA-ANN and ARIMAKalman hybrid methods for wind speed prediction. Appl. Energy 98, 415424.
http://dx.doi.org/10.1016/j.apenergy.2012.04.001.
Liu, H., Tian, H., Li, Y., 2015. Comparison of new hybrid FEEMD-MLP, FEEMD-ANFIS,
Wavelet Packet-MLP and Wavelet Packet-ANFIS for wind speed predictions. Energy
Convers. Manag. 89, 111. http://dx.doi.org/10.1016/j.enconman.2014.09.060.
Liu, H., Tian, H., Pan, D., Li, Y., 2013b. Forecasting models for wind speed using
wavelet, wavelet packet, time series and articial neural networks. Appl.
Energy 107, 191208. http://dx.doi.org/10.1016/j.apenergy.2013.02.002.
Liu, H., Tian, H.Q., Li, Y.-F., 2009. Short-term forecasting optimization algorithms for
wind speed along QinghaiTibet railway based on different intelligent modeling theories. J. Cent. South Univ. Technol. Engl. 16, 690696. http://dx.doi.org/
10.1007/s11771-009-0114-3.
Pan, D., Liu, H., Li, Y., Liang, H., 2008. Short-term forecast method for wind speed
along GolmudLhasa section of QinghaiTibet railway. China Railw. Sci. 29,
129133.
Pankratz, A., 1983. Forecasting with Univariate BoxJenkins Models: Concepts and
Cases. John Wiley & Sons.
Shust, W., Pasta, C., Milhon, L., Rosenberger, W., Dumont, B.J., Miller, J.L., 2010. An
approach to wind blow-over risk reduction at BNSF Railway. In: Proceedings of
the ASME Joint Rail Conference 2010, JRC2010, 5768. http://dx.doi.org/10.1115/
JRC2010-36019.
Strong wind topples Chinese train, 2007. BBC.
Train death toll rises in Japan, 2005. BBC.
Wei, W.W.-S., 1994. Time Series Analysis. Addison-Wesley publ.