Vous êtes sur la page 1sur 14

Journal of Hydrology 470471 (2012) 158171

Contents lists available at SciVerse ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

IDF relationships using bivariate copula for storm events in Peninsular Malaysia
N.M. Ariff , A.A. Jemain 1, K. Ibrahim 2, W.Z. Wan Zin 3
School of Mathematical Sciences, Faculty of Science and Technology, Universiti Kebangsaan Malaysia, 43600 Bangi, Selangor, Malaysia

a r t i c l e

i n f o

Article history:
Received 2 April 2012
Received in revised form 6 July 2012
Accepted 22 August 2012
Available online 31 August 2012
This manuscript was handled by Andras
Bardossy, Editor-in-Chief, with the
assistance of Sheng Yue, Associate Editor
Keywords:
Intensitydurationfrequency (IDF)
Copula
Storm event
Bivariate frequency analysis

s u m m a r y
Intensitydurationfrequency (IDF) curves are used in many hydrologic designs for the purpose of water
managements and ood preventions. The IDF curves available in Malaysia are those obtained from univariate analysis approach which only considers the intensity of rainfalls at xed time intervals. As several
rainfall variables are correlated with each other such as intensity and duration, this paper aims to derive
IDF points for storm events in Peninsular Malaysia by means of bivariate frequency analysis. This is
achieved through utilizing the relationship between storm intensities and durations using the copula
method. Four types of copulas; namely the AliMikhailHaq (AMH), Frank, Gaussian and FarlieGumbelMorgenstern (FGM) copulas are considered because the correlation between storm intensity, I, and
duration, D, are negative and these copulas are appropriate when the relationship between the variables
are negative. The correlations are attained by means of Kendalls s estimation. The analysis was performed on twenty rainfall stations with hourly data across Peninsular Malaysia. Using Akaikes Information Criteria (AIC) for testing goodness-of-t, both Frank and Gaussian copulas are found to be suitable to
represent the relationship between I and D. The IDF points found by the copula method are compared to
the IDF curves yielded based on the typical IDF empirical formula of the univariate approach. This study
indicates that storm intensities obtained from both methods are in agreement with each other for any
given storm duration and for various return periods.
2012 Elsevier B.V. All rights reserved.

1. Introduction
Statistical analysis of extreme data is important in various disciplines including hydrology, engineering and environmental science (Reiss and Thomas, 2007). By performing extreme analysis
on rainfall data, damages to human lives and properties as the consequences from extreme rainfall events such as ood or landslides
may be reduced or prevented. In Malaysia, extreme analysis on
rainfall data has been explored for all sorts of purposes such as
tracing patterns and trends of daily rainfall during monsoon seasons (Suhaila et al., 2010a,b), detecting recent changes in extreme
rainfall events (Wan Zin et al., 2010) and tting probability distributions to annual maximum rainfalls by implementing various
methods (Shabri et al., 2011; Wan Zin et al., 2009a,b).
The rainfall intensitydurationfrequency (IDF) curves are
essential tools in designing hydraulic structures such as dams,
spillways and drainage systems. These hydraulic structures help
to lessen the loss caused by extreme rainfall events. Hence, for
Corresponding author. Tel: +60 603 89215784; fax: +60 603 89254519.
E-mail addresses: tqah@ukm.my (N.M. Ariff), azizj@ukm.my (A.A. Jemain),
kamarulz@ukm.my (K. Ibrahim), w_zawiah@ukm.my (W.Z. Wan Zin).
1
Tel.: +60 603 89215724; fax: +60 603 89254519.
2
Tel.: +60 603 89213702; fax: +60 603 89254519.
3
Tel.: +60 603 89215790; fax: +60 603 89254519.
0022-1694/$ - see front matter 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.jhydrol.2012.08.045

countries such as Malaysia where ood is considered as the most


signicant natural hazard (Sulaiman, 2007), correct rainfall estimation as points on the IDF curves is important. According to Koutsoyiannis et al. (1998), IDF curves are able to show the
mathematical relationship between rainfall intensity, i, duration,
d, and return period, T (the annual frequency of exceedance).
The construction of IDF curves is mostly done using univariate
rainfall frequency analysis approach because of its mathematical
simplicity (Singh and Zhang, 2007). The univariate approach of
rainfall frequency analysis is explained in detail by Chow et al.
(1988). In addition, most of the IDF curves are constructed using
the window-based analysis approach where the durations are predetermined time intervals. Thus, the durations do not represent
the actual durations of rainfall events. In other words, a smaller
window may be a subset of a longer extreme rainfall event while
a bigger sized window could contain several short duration rainfalls and dry periods.
Several attempts have been made to derive joint distributions of
rainfalls characteristics, namely; intensity, depth and duration, as
their random variables. This is to accommodate the complexity
of hydrological events. It has come to light that events such as
storms and ood always appear to be multivariate and thus, single-variable frequency analysis can only provide limited assessments of these events (Yue et al., 2001). In earlier studies, several
assumptions have been made due to the mathematical difculties

159

N.M. Ariff et al. / Journal of Hydrology 470471 (2012) 158171

in obtaining these joint distributions using standard statistical


methods. One of the assumptions is the independence assumption
between the random variables. However, Cordova and RodriguezIturbe (1985) have proven that this assumption is inappropriate
and unrealistic. Their study showed that the correlation between
rainfall duration and its average intensity gives non-negligible effect on storm surface runoff. Hence, following this discovery, most
rainfall models incorporated the dependence between variables
with some limitations imposed on the marginal distributions of
the random variables. For instance, the marginals are assumed to
be either normal (Yue, 2000) or possess the same type of probability distribution such as bivariate exponential (Favre et al., 2002;
Goel et al., 2000), bivariate Gamma (Yue et al., 2001), bivariate lognormal (Yue, 2002) and bivariate extreme value distribution (Shiau, 2003). In reality, extreme rainfall is a complicated
phenomenon and its marginal distributions are not necessarily
similar or distributed as normal. Other distributions should be considered which may produce better rainfall estimates.
The copula approach is a exible method that allow more
choices of marginal distributions and dependence structures to
be used in multivariate problems (Kao and Govindaraju, 2008).
Copulas are based on Sklars theorem (1959) which states that
for joint distribution, the analysis of the marginals and the dependence structure can be done separately. A thorough explanation on
the theory and description of copula is given by Nelsen (2006). Various types of copulas have been used in hydrology over the last
decade. Among them are the FarlieGumbelMorgenstern (FGM)
copula (Favre et al., 2004), the elliptic copulas such as the Gaussian
copula (Renard and Lang, 2007) and the Archimedean copulas (De
Michele and Salvadori, 2003; De Michele et al., 2005; Salvadori and
De Michele, 2004; Zhang and Singh, 2006).
In Malaysia, although the IDF curves play important roles, the
only ones utilized are those obtained through the univariate frequency analysis. This paper aims to derive the bivariate or the joint
distribution between intensity and duration of extreme storm
events in Peninsular Malaysia using copulas. Then, the points for
IDF curves are derived from the copula obtained and compared
with the IDF points found based on the empirical formula used
in univariate analysis.
The rst section of this paper contains the introduction of this
research, followed by a section describing the data used for analysis and the denition of storm events. The third section comprises
of the explanation on copula method and the construction of IDF
points using the copula method. Section four provides the various
copulas under consideration. The different probability distribution
functions used in this study are shown in section ve. In section
six, the construction of IDF curves using the typical empirical
method is discussed. The seventh section contains the computation
of IDF points based on the copula and empirical methods on the
hourly rainfalls data in Peninsular Malaysia and section eight provides the conclusion of the research.

2. Data and denition of storm


Hourly rainfall data obtained from stations in Peninsular Malaysia are acquired from the Department of Irrigation and Drainage
Malaysia. The stations are located in four different regions scattered across Peninsular Malaysia. These regions are based on their
geographical locations as suggested by many previous researches
(Deni et al., 2010; Suhaila and Jemain, 2009; Suhaila et al., 2011).
The regions are known as the Northwest, West, East and Southwest
region of Peninsular Malaysia. Most rainfall stations are located at
the edge since at the centre of Peninsular Malaysia lies the Titiwangsa Mountain range which is mostly unoccupied. In this study,
twenty stations are selected based on their locations and com-

Table 1
The locations and completeness of rainfall data for stations in Peninsular Malaysia.
Region

Code

Station

Latitude

Longitude

Completeness of
data (%)

Northwest

N01
N02

Alor Setar
Bukit
Bendera
Jeniang
Sungai
Pinang

6.116
5.425

100.356
100.269

97.7
99.2

5.817
5.404

100.633
100.217

96.7
99.0

Bertam
Genting
Klang
Gua
Musang
Kalong
Tengah
Kampar
Teluk
Intan

5.144
3.204

102.048
101.722

90.9
99.0

4.860

101.963

95.5

3.438

101.658

98.4

4.346
4.026

101.157
101.021

97.2
95.6

Dungun
Endau
Kampung
Dura
Kemaman
Kepasing
Paya
Kangsar

4.756
2.641
5.060

103.400
103.661
102.934

97.5
97.7
93.0

4.233
2.967
3.900

103.420
102.867
102.430

96.2
93.0
96.8

Chinchin
Johor
Bahru
Kota
Tinggi
Labis

2.290
1.463

102.474
103.755

99.7
96.9

1.733

103.720

95.1

2.395

103.017

94.4

N03
N04
West

W01
W02
W03
W04
W05
W06

East

E01
E02
E03
E04
E05
E06

Southwest

S01
S02
S03
S04

pleteness. These stations are located both at the edge and the middle of the Peninsular. Data from these stations are more than 90%
complete for the year 19752008 as presented in Table 1. The locations of these 20 stations are shown in Fig. 1.
The denition of storm-event depends greatly on the interevent time denition. The inter-event time denition (IETD) is dened as the minimum duration of dry period between two consecutive storm events. Hence, the dry duration between two
individual storm events must at least be equal to the IETD value.
If not, they would not be considered as two different events but
parts of the same storm. The IETD value is chosen such that the serial correlation between the two different storms is minimized
(Restrepo-Posada and Eagleson, 1982). For small urban catchments, the IETD is usually taken as 6 h because the time concentration of rainfall which is less than 6 h would make the runoff
response of successive storms to appear independent (Palynchuk
and Guo, 2008). Storm depth is dened as the accumulated rainfall
which begins and ends with at least one wet hour and either contains dry periods with less than 6 h or none at all. Storm duration is
dened as the time interval for a storm event and storm intensity is
the ratio of storm depth to storm duration.
The information extracted from the rainfall data is the annual
maximum storm intensity for several storm durations. This paper
focuses on the convective storms for the purpose of IDF curves construction. Convective storm is dened as short duration storm
which has great impacts on small or urban catchments (Palynchuk
and Guo, 2008).
3. A brief introduction to copula
Before we proceed further, we will introduce the notations used
in this paper for simplicity purposes. Uppercase letters (example: X
and Y) are dened as random variables and lowercase letters

160

N.M. Ariff et al. / Journal of Hydrology 470471 (2012) 158171

NORTHWEST

EAST

WEST

SOUTHWEST

Fig. 1. Map of Peninsular Malaysia and the locations of the twenty hourly rainfall stations under consideration.

(example: x and y) are realizations or specic values of random


variables. For a bivariate distribution, the joint distribution function of a pair of random variables (X, Y) is denoted as HX,Y with marginals FX and FY respectively. If we let U and V be the random
variables of these marginals, then U and V are uniformly distributed; U = FX(X)  U[0, 1] and V = FY(Y)  U[0,1]. Thus, u = FX(x) and
v = FY(y) are realizations of U and V respectively.
A copula is usually described as either a function which links a
multivariate distribution function to its marginals (cumulative distribution functions) or a function composed of uniformly distributed
marginals in [0, 1] (Nelsen, 2006). Copula is dened according to
Sklars theorem (1959) which states that for continuous random
variables X and Y with joint distribution HX,Y and marginals FX and
FY, there exists a unique copula, CU,V, such that for all x, y in R,

The construction of one-parameter copula can be summarized


into four simple steps. For this study, the copula parameter is denoted as h. Once the copula family and function to be used is identied, the copula can be obtained by rst nding the marginal
distribution functions U and V. This is done by tting probability
distribution functions to random variables X and Y. Thus, all the
parameters for the two probability distributions are approximated
and their respective distribution functions, U and V, can be obtained. Next, the dependence between the random variables is calculated. In this paper, the correlation between random variables x
and y is yielded using Kendalls s. For N paired (x, y) observations,
Kendalls s is estimated as (Kao and Govindaraju, 2007)

TN

1
C U;V u; v PU 6 u; V 6 m PX 6 F 1
X u; Y 6 F Y m
1
HX;Y F 1
X u; F Y m H X;Y x; y:

In a way copula helps to map random observations (x, y) from R2


to a bounded domain [0, 1]2. Copula reduces the complexity of
deriving the joint distribution of random variables X and Y. Unlike
the traditional bivariate joint distribution, the marginals of the random variables and the dependence between them can be determined separately one at a time.

N
2

1 X


sign xj  xk yj  yk
j<k

with

sign

8
1;
>
>
>
<

xj < xk and yj < yk


xj > xk and yj > yk

>
0;
xj  xk yj  yk 0
>
>
:
1; otherwise:

161

N.M. Ariff et al. / Journal of Hydrology 470471 (2012) 158171

Similar to the Pearson correlation coefcient q, Kendalls s is


dened in the closed interval [1, 1] with 1 implying total concordance, 1 representing total discordance and 0 showing zero or no
concordance between the random variables (Kao and Govindaraju,
2008).The third step is calculating the copula parameter h. This is
done by using the relationship between h and Kendalls s, i.e.
s = g(h). This relationship differs for different copula. Finally, the
approximated h and the marginals of the random variables can
be inserted into the candidate copula function to obtain the respective copula.
The conditional copula can be derived from the copula function
for realizations of U given the value of V = v. This conditional distribution is needed in order to attain the values for the intensity
durationfrequency (IDF) curves which involve conditional distribution of storm intensity given storm duration (Singh and Zhang,
2007). In this study, X is regarded as the storm intensity and Y is
the storm duration. For simplicity purposes, we denote them as I
and D respectively with lowercase letters, i and d, as their realizations. The conditional distribution of I given D = d is represented in
conditional copula form as the realizations of marginal U for
known value of V = v, CU|V=v. The value of v is easily obtained since
v = FD(d). The conditional copula is written as (Zhang and Singh,
2007)

C UjVm ujV m

@
C U;V u; v jVv :
@m

Similar to the return period of any conditional bivariate distributions, the relationship between the conditional copula and the selected return period is

1
C UjVm ujV m 1  :
T

Hence, by solving Eqs. (3) and (4) simultaneously for any given
values of T and v, we can solve for the corresponding u. From the
value of u, the respective i can be yielded since i F 1
I (u). This i value is then taken as one of the point on the IDF curves indicating
the intensity of storm for return period T and storm duration d.
Repeating this process for various return periods and selected values of storm durations, other points of the IDF curves can be determined and the IDF curves can be constructed.

2h
;
9

s2 


22
:
99

The conditional copula is

C UjVm ujV m u hu1  u1  v  huv 1  u:

There are various copula families and functions available for


bivariate frequency analysis. In fact, copula consists of families
with many different copula functions. For example, the Archimedean copula, which is one of the most commonly used copula family due to its mathematical tractability and simplicity, has at least
22 copula functions as its member. The choice of copula family and
its function relies on the correlation between the random variables
under consideration. The correlation between storm intensity, I,
and storm duration, D, is known to be negative (Kao and Govindaraju, 2007; Singh and Zhang, 2007; Zhang and Singh, 2007).
Among the copulas which have been applied in hydrology that
are appropriate for negatively correlated random variables are
the FarlieGumbelMorgenstern (FGM), Gaussian, AliMikhail
Haq (AMH) and Frank copula.
4.1. FarlieGumbelMorgenstern (FGM) copula
FGM copula belongs to the family of copulas with quadratic section. The FGM copula is used for modelling purposes due to their
simple analytical form with its copula function written as (Favre
et al., 2004)

h 2 1; 1:

4.2. Gaussian copula


Gaussian copula is from the elliptic copula family. Gaussian copula is good for practical applications since it possess several properties of the multivariate normal distribution (Favre et al., 2004). The
copula function for the Gaussian copula is (Schmidt, 2006)

C U;V u; v UR U1 u; U1 v


Z U1 u Z U1 v
1
p

1
1
2p 1  h2
!
2hsx  s2  x2
 exp
dsdx;
21  h2
h 2 1; 1:

where U is the cumulative distribution function of a standard normal distribution and UR is the bivariate normal distribution with
mean 0 and covariance matrix R. Thus, the conditional Gaussian
copula is

C UjVu ujV v

@
UR U1 u; U1 v jVv :
@t

h in Eq. (8) is actually the Pearson q. A one-to-one relationship between the Pearson q and Kendalls s under normality is provided by
Kruskal (1958) as

arcsin h;

s 2 1; 1:

10

4.3. Archimedean copula


According to Nelson (2006), the Archimedean copula can be
generalized as

C U;V u; v u1 uu uv :

4. Selection of copula

C U;V u; v uv huv 1  u1  v ;

The relationship between h and s for FGM copula is (Huard et al.,


2006)

11

with u is a copula generator and u1 is appropriately dened. He


stated that u(1) = 0 and u1(x) = 0 for x P u(0). The Archimedean
copula family is preferable in hydrologic analysis because it possess
desirable properties such as symmetric and associative (Favre et al.,
2004). Copulas from the Archimedean family are also easily constructed (Zhang and Singh, 2006). Below are two copulas from the
Archimedean copula family which are considered in this study.
4.3.1. AliMikhailHaq (AHM) copula
AMHs copula function is (Huard et al., 2006)

C U;V u; v

uv
;
1  h1  u1  v

h 2 1; 1

12

and



2 ln1  h 2
h
1 for
h  2h
2
3
ln1  h
h


1
2 0:181726; :
3

s1

s
13

The conditional copula for AMH is

C UjVt ujV m



u
hm1  u
1
:
1  h1  u1  m
1  h1  u1  m

14

162

N.M. Ariff et al. / Journal of Hydrology 470471 (2012) 158171

COPULA

MORE THAN ONEPARAMETER COPULA

ONE-PARAMETER COPULA

Parameters and distributions required to


construct a bivariate copula function
copula parameter
i)
ii)
U, V marginal distributions of
random variables
iii)
correlation of the two random
variables (to get )

Copulas for only positive


correlations

Copulas suitable for negative correlations

(used in this paper due to common practise in hydrological analysis)

Copulas with
quadraric section

Archimedean

Elliptic

Advantage: symmetric, associate,


easily constructed, preferable in
hydrological analysis

Gauss
Farlie-GumbelMorgenstern (FGM)

Advantage: simple
analytical form

Advantage: good for


practical applications,
possess multivariate
normal properties

Range for :

Range for :

Ali-MikhailHaq (AMH)

Frank

Range for :
Range for :

Fig. 2. A short summary of the four copulas used in this study.

4.3.2. Frank copula


Frank copula is written as (Huard et al., 2006)

C U;V u; v 

1
hh
ln
;
h hh  hhuhhm

hx 1  expx

5. Probability distribution functions for storm intensity and


duration

15

for h h 2 R{0} and

4
h

s 1  D1 h  1

16

with

Dk h

k
h

Z
0

tk
kh
:
dt
k1
expt  1

The domain for s is 1; 1f0g. From Eq. (3), the conditional
Frank copula is

C UjVm ujV m

hhu1  hhm
;
hh  hhuhhm

hx 1  expx:

17

A short summary of the four copulas considered in this study is


shown in Fig. 2.

The probability distribution of storm intensity is chosen from


either the exponential, gamma, weibull or lognormal distribution.
The exponential distribution is used to represent the probability
distribution for storm duration since storms are usually assumed
to follow the Poisson process such as in the NeymanScott and
BartlettLewis model of storms. Thus, the inter arrival time of
storm cells is taken as exponentially distributed. The probability
distribution functions for exponential, gamma and weibull probability distribution can be generalized as

f x

  c 
 ca1
x
x
exp 
b
bCa b

18

with a and c as the shape parameters and b as the scale parameter.


When both a and c have the value of one, Eq. (18) becomes an exponential probability distribution function. If a = 1, then it becomes a
gamma probability distribution function and if c = 1, it is a weibull

N.M. Ariff et al. / Journal of Hydrology 470471 (2012) 158171

163

ie gd; T id sd K T

21

probability distribution function. The probability distribution function of lognormal distribution is

1
ln x  l
f x p exp 
2b
x 2pb

!
19

where l is the location parameter and, similar to Eq. (18), b is the


scale parameter.
6. IDF curves using empirical method
In this paper, the IDF points produced by the copula method are
compared to the IDF curves found using one of the IDF empirical
formula. The formula used is known as the Sherman equation
(Nhat et al., 2006) and is written as

aT k
d bc

20

with ie as the intensity of storm in millimetres per hour (mm/h), d


as storm duration in hours (h) and T as the return period of storm in
years. The parameters a, j, b and c are estimated using the leastsquare method on ie for a set of given T and d. For the purpose of
performing least-square, ie is approximated using the design rainfall
intensity formula. The design rainfall intensity for the annual maximums of storms with duration d and return period T is a function of
d and T which can be represented as (Kottegoda and Rosso, 2008)

KT 

p
 


T
6
:
0:5772 ln ln
T 1
p

22

The IDF curves constructed from values of storm intensities ie for


various T and d are compared to the IDF points found based on
the copula method. The percentage difference, denoted as D, between storm intensity ie found using Eq. (20) and storm intensity
i derived by the copula method through Eq. (3) is written as

jie  ij
100%:
ie

23

7. Application on storm events in Peninsular Malaysia


Data from twenty stations representing four regions in Peninsular Malaysia are used for application purposes. The scatter plots of

W01

40
30
10
0

10

15

20

10

15

storm duration, h

storm duration, h

E01

S01

20

30
0

10

10

20

30

storm intensity, mm/h

40

40

storm intensity, mm/h

20

storm intensity, mm/h

30
20
10

storm intensity, mm/h

40

50

N01

20

ie

where id and sd are the mean and standard deviation of the storm
intensity for a given d. In Eq. (21), KT is dened as the frequency factor
for the return period T which depends on the probability distribution
function of ie. Eq. (21) is applicable to many probability distributions
of storm intensity that are employed in hydrologic frequency analysis (Chow et al., 1988). For rainfall data in Malaysia, the Gumbel distribution is commonly used and is deemed to be suitable (Amin et al.,
2008). Hence, in this paper, the Gumbel distribution will be utilized.
The frequency factor for Gumbel distribution is (Guo, 2006)

10

15

storm duration, h

20

10

15

20

storm duration, h

Fig. 3. Scatter plots of storm intensities against storm durations for one station from each region of Peninsular Malaysia.

164

N.M. Ariff et al. / Journal of Hydrology 470471 (2012) 158171

storm intensity versus duration for a selected station in each region is displayed in Fig. 3. From this gure, it can be seen that there
is a negative correlation between the two variables. This is because
storm intensity is the rate of storm depth which is yielded by taking the ratio of storm depth to storm duration. Thus, it is clear that
the larger the storm duration, the smaller will the storm intensity
be. Fig. 3 also shows that for short storms, for instance, storms with
not more than 12 h, there is a higher correlation between storm
intensities and durations. It can also be observed that there is a
slight variation on the characteristics of storms when different regions are compared especially for short duration storms.
The exponential, gamma, weibull and lognormal distributions
were tted to storm intensities of the twenty stations. Fig. 4 illustrates the L-moments ratio diagram for the four distributions with
the sample L-skewness and L-kurtosis of storm intensities for each
of the twenty stations. The most appropriate probability distributions tted to the storm intensities and durations of these stations
are shown in Table 2. The L-moment ratio diagram of Fig. 4 and Table 2 imply that there may be dissimilarities among the best tted
storms distribution according to different regions. From the probability distributions selected, we may obtain the marginal distribution functions, U = FI(I) and V = FD(D).
The estimated correlations between I and D, that is the Kendalls
s, are also given in Table 2. It can be seen that the values of Kendalls s for all twenty stations considered are negative which is
in accordance with Fig. 3. The values of s for all the stations lie
within the range of 0.64 6 s 6 0.42. The values are neither close
to 1 nor 0. In other words, the dependence level between the
storm intensities and durations are moderate. Since the dependence level is not high, it is unsuitable to use regression method
to represent their relationship. On the other hand, it is also inaccurate to reduce the joint distribution of random variables I and D to
the product of the marginal distributions since the dependence level is not low and s is not close to 0. The s values are found not to
be in the domain of s for the AliMikhailHaq (AMH) and Farlie
GumbelMorgenstern (FGM) copulas. Thus, only the Frank and
Gaussian copulas will be further considered for the construction
of bivariate distributions of storm intensities and durations. The
parameter, h, for both copulas are estimated using their respective
relationship to Kendalls s, as displayed in Eqs. (10) and (16).
Hence, using the estimated h and marginal distributions U and V,
both copulas are tted to the paired I and D by using their copula
functions which are Eqs. (8) and (15) respectively. The comparison
of both copulas to determine which type of copula is better to rep-

Table 2
Distribution for storm intensities and durations as well as the Kendalls s for the
correlations between the two random variables.
Station

N01
N02
N03
N04
W01
W02
W03
W04
W05
W06
E01
E02
E03
E04
E05
E06
S01
S02
S03
S04

Intensity (mm/h)

Duration (h)

Dist

a, c or l

Dist

Gamma
Gamma
Gamma
Gamma
Weibull
Weibull
Lognormal
Lognormal
Gamma
Weibull
Lognormal
Gamma
Gamma
Lognormal
Weibull
Gamma
Weibull
Weibull
Lognormal
Lognormal

2.35
2.76
2.31
3.13
1.58
1.70
2.48
2.53
4.16
1.85
2.25
3.07
1.90
2.31
1.56
2.44
1.55
1.44
2.66
2.24

5.50
5.34
6.44
4.70
14.85
16.08
0.46
0.44
4.83
15.89
0.43
4.86
7.60
0.50
13.71
4.84
15.48
16.75
0.56
0.82

Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential
Exponential

5.49
5.50
5.53
5.48
5.47
5.50
5.48
5.51
4.11
5.46
5.48
4.09
5.37
5.45
5.29
5.46
5.37
5.37
4.11
5.48

0.55
0.55
0.61
0.48
0.54
0.62
0.62
0.57
0.45
0.64
0.42
0.42
0.61
0.45
0.53
0.50
0.61
0.60
0.54
0.61

a, c = Shape parameter, b = scale parameter and l = location parameter.

Table 3
Values of h and Akaikes Information Criteria (AIC) for Gaussian and Frank copula.
Station

N01
N02
N03
N04
W01
W02
W03
W04
W05
W06
E01
E02
E03
E04
E05
E06
S01
S02
S03
S04
a

Gaussian

Frank

AIC

AIC

0.77
0.75
0.82
0.70
0.75
0.84
0.82
0.80
0.69
0.86
0.60
0.65
0.82
0.62
0.75
0.74
0.82
0.81
0.77
0.77

5929.20a
5963.20
5949.92
5963.98
5895.55
5922.30a
5876.50
5793.82
5501.10a
5709.19
5931.52
5506.81
5946.77
5906.08
5752.77
5908.46
5926.07
5974.52
5567.46
5979.21

6.45
7.05
8.29
5.17
6.58
8.07
8.42
7.40
4.83
9.43
4.82
4.54
8.70
5.15
6.53
5.92
8.81
8.01
6.71
8.49

5929.27
5946.37a
5940.52a
5963.53a
5887.88a
5930.54
5871.34a
5790.56a
5502.04
5708.67a
5923.04a
5504.58a
5929.45a
5897.35a
5744.00a
5904.32a
5911.56a
5966.42a
5563.55a
5942.27a

Smaller value of AIC.

resent the relationship between I and D is done with the help of


Akaikes Information Criteria (AIC). The AIC is given as (Zhang
and Singh, 2007)

AIC 2 logmaximized likelihood for the copula


2no: of fitted parameters

24

or

AIC N logMSE 2no: of fitted parameters:

Fig. 4. L-moment ratio diagram for the exponential, gamma, weibull and lognormal
distribution as well as pairs of (L-skewness, L-kurtosis) of the storm intensities.

25

The copula which shows a smaller value of AIC is chosen to represent the joint distribution of I and D. For all the 20 stations, the
values for h and AIC of both the Gaussian and Frank copula are presented in Table 3.
Table 3 indicates that the Frank copula provides a comparatively smaller value of AIC for most of the twenty stations. It can
also be observed that the difference between the AIC values of

165

N.M. Ariff et al. / Journal of Hydrology 470471 (2012) 158171


Table 4
Comparisons between the storm intensities obtained based on the typical IDF empirical formula and the copula method.
Station

T (years)

Storms intensities (mm/h)


1-h Storms

3-h Storms

6-h Storms

9-h Storms

12-h Storms

ie

ie

ie

ie

ie

N01

2
5
10
25
50
100

22.42
27.38
31.84
38.88
45.22
52.59

18.57
25.41
30.13
36.22
40.74
45.25

17.19
7.17
5.37
6.84
9.89
13.96

15.29
18.67
21.71
26.51
30.83
35.86

12.64
17.50
21.10
26.26
30.45
34.80

17.34
6.26
2.83
0.96
1.23
2.95

11.22
13.70
15.94
19.46
22.63
26.32

8.45
12.08
14.65
18.41
21.69
25.38

24.74
11.84
8.07
5.40
4.17
3.58

9.22
11.25
13.09
15.98
18.59
21.62

6.43
9.61
11.79
14.89
17.59
20.68

30.22
14.63
9.92
6.80
5.38
4.31

7.97
9.73
11.32
13.82
16.08
18.70

5.39
8.32
10.34
13.15
15.56
18.32

32.43
14.50
8.71
4.88
3.26
2.05

N02

2
5
10
25
50
100

25.53
30.95
35.79
43.39
50.18
58.05

21.08
27.95
32.69
38.81
43.37
47.82

17.44
9.68
8.68
10.54
13.58
17.62

17.19
20.83
24.09
29.20
33.78
39.07

14.63
19.40
22.90
27.95
32.09
36.44

14.88
6.87
4.97
4.30
4.99
6.75

12.48
15.13
17.50
21.22
24.54
28.38

10.04
13.65
16.13
19.69
22.79
26.30

19.55
9.77
7.83
7.18
7.14
7.32

10.18
12.34
14.28
17.31
20.02
23.15

7.77
11.00
13.13
16.08
18.59
21.46

23.67
10.88
8.01
7.06
7.11
7.30

8.77
10.63
12.29
14.90
17.23
19.93

6.57
9.60
11.59
14.29
16.54
19.08

25.06
9.65
5.66
4.07
4.02
4.27

N03

2
5
10
25
50
100

27.42
33.23
38.43
46.57
53.85
62.27

22.20
29.55
34.77
41.67
46.88
52.04

19.03
11.08
9.52
10.52
12.94
16.43

17.36
21.04
24.33
29.49
34.10
39.43

14.62
19.23
22.65
27.73
32.09
36.82

15.78
8.60
6.92
5.97
5.90
6.64

12.01
14.55
16.83
20.39
23.58
27.27

9.52
12.88
15.15
18.39
21.21
24.49

20.70
11.48
9.94
9.81
10.04
10.19

9.49
11.50
13.30
16.12
18.64
21.56

7.02
10.01
11.95
14.59
16.80
19.31

26.01
12.98
10.14
9.48
9.88
10.43

7.98
9.68
11.19
13.56
15.68
18.13

5.70
8.49
10.31
12.72
14.69
16.87

28.62
12.20
7.87
6.19
6.34
6.95

N04

2
5
10
25
50
100

21.92
26.54
30.67
37.14
42.92
49.61

19.72
26.62
31.18
36.90
41.07
45.20

10.03
0.30
1.65
0.64
4.32
8.90

16.95
20.53
23.72
28.73
33.20
38.37

14.53
20.02
23.93
29.25
33.36
37.48

14.31
2.46
0.87
1.81
0.47
2.32

13.78
16.68
19.28
23.35
26.98
31.19

10.57
14.97
18.08
22.52
26.22
30.13

23.28
10.29
6.24
3.54
2.84
3.39

12.07
14.62
16.89
20.46
23.64
27.33

8.61
12.53
15.26
19.14
22.45
26.08

28.65
14.28
9.68
6.42
5.07
4.58

10.95
13.26
15.33
18.56
21.45
24.80

7.59
11.25
13.79
17.38
20.45
23.86

30.75
15.19
10.02
6.34
4.70
3.78

W01

2
5
10
25
50
100

22.98
28.26
33.05
40.65
47.54
55.60

19.69
26.33
30.63
35.92
39.68
43.30

14.30
6.84
7.32
11.64
16.53
22.12

15.84
19.49
22.79
28.03
32.77
38.33

13.39
18.47
22.04
26.93
30.73
34.52

15.50
5.23
3.29
3.93
6.24
9.92

11.73
14.43
16.88
20.76
24.27
28.39

8.67
12.69
15.42
19.25
22.46
25.96

26.11
12.08
8.64
7.25
7.45
8.56

9.69
11.92
13.94
17.14
20.05
23.44

6.34
9.93
12.32
15.61
18.36
21.42

34.56
16.71
11.58
8.91
8.40
8.63

8.42
10.35
12.11
14.89
17.41
20.36

5.13
8.47
10.71
13.75
16.27
19.06

39.02
18.21
11.52
7.62
6.58
6.41

W02

2
5
10
25
50
100

23.60
27.95
31.76
37.62
42.75
48.58

21.48
27.50
31.46
36.37
39.88
43.21

9.00
1.62
0.95
3.30
6.70
11.06

16.62
19.69
22.37
26.50
30.11
34.22

14.70
19.03
22.06
26.29
29.70
33.20

11.55
3.33
1.42
0.80
1.38
2.99

12.53
14.84
16.86
19.97
22.69
25.79

9.65
13.12
15.37
18.44
21.00
23.84

22.97
11.58
8.87
7.67
7.48
7.55

10.46
12.39
14.08
16.67
18.95
21.54

7.03
10.24
12.26
14.91
17.05
19.40

32.77
17.32
12.91
10.58
10.03
9.94

9.16
10.85
12.33
14.60
16.60
18.86

5.62
8.68
10.61
13.10
15.06
17.19

38.65
20.01
13.96
10.31
9.23
8.89

W03

2
5
10
25
50
100

28.88
35.79
42.10
52.17
61.37
72.18

21.56
30.59
38.02
49.25
58.89
69.51

25.36
14.52
9.69
5.61
4.04
3.71

17.49
21.68
25.50
31.60
37.17
43.72

13.62
18.14
21.83
27.88
33.69
40.69

22.16
16.30
14.39
11.76
9.35
6.93

11.67
14.46
17.01
21.08
24.80
29.17

9.05
11.94
14.01
17.15
20.08
23.73

22.45
17.45
17.62
18.66
19.03
18.64

9.02
11.17
13.14
16.29
19.16
22.53

6.99
9.43
11.09
13.45
15.53
18.02

22.51
15.64
15.65
17.44
18.95
20.05

7.46
9.24
10.87
13.47
15.85
18.64

5.93
8.16
9.66
11.74
13.52
15.58

20.54
11.68
11.09
12.83
14.70
16.42

W04

2
5
10
25
50
100

29.83
36.93
43.41
53.76
63.19
74.27

22.05
31.66
39.44
50.99
60.79
71.46

26.08
14.27
9.16
5.16
3.79
3.79

18.46
22.86
26.87
33.27
39.11
45.97

14.31
19.57
23.89
30.94
37.53
45.26

22.46
14.38
11.08
7.02
4.02
1.54

12.53
15.52
18.24
22.59
26.55
31.21

9.68
13.09
15.61
19.53
23.25
27.91

22.77
15.67
14.43
13.55
12.41
10.57

9.79
12.12
14.25
17.64
20.74
24.38

7.59
10.44
12.44
15.39
18.07
21.36

22.49
13.89
12.68
12.79
12.88
12.40

8.16
10.11
11.88
14.71
17.30
20.33

6.52
9.12
10.92
13.49
15.76
18.48

20.12
9.82
8.14
8.34
8.89
9.11

W05

2
5
10
25
50
100

29.77
35.77
41.09
49.38
56.73
65.18

24.74
32.55
37.69
44.11
48.78
53.35

16.88
8.99
8.30
10.68
14.02
18.15

20.97
25.20
28.96
34.79
39.97
45.93

18.15
24.37
28.71
34.61
39.20
43.82

13.47
3.31
0.86
0.51
1.93
4.60

15.81
19.00
21.83
26.22
30.13
34.62

13.69
18.88
22.45
27.44
31.57
35.95

13.44
0.61
2.85
4.65
4.78
3.83

13.20
15.86
18.23
21.90
25.16
28.91

11.79
16.58
19.84
24.39
28.18
32.31

10.69
4.52
8.88
11.36
12.00
11.76

11.56
13.89
15.96
19.18
22.04
25.32

10.94
15.54
18.67
23.02
26.64
30.63

5.39
11.83
16.97
19.99
20.89
20.94

W06

2
5
10
25
50
100

24.48
29.00
32.98
39.08
44.43
50.52

21.01
25.94
29.23
33.35
36.30
39.10

14.18
10.55
11.36
14.67
18.31
22.6

16.33
19.35
22.00
26.07
29.64
33.71

14.62
18.02
20.35
23.62
26.32
29.18

10.46
6.90
7.53
9.41
11.20
13.41

11.78
13.96
15.87
18.81
21.38
24.31

9.83
12.64
14.39
16.69
18.57
20.64

16.52
9.45
9.35
11.23
13.14
15.1

9.56
11.33
12.89
15.27
17.36
19.74

7.24
9.95
11.57
13.62
15.22
16.92

24.34
12.25
10.22
10.80
12.36
14.32

8.21
9.72
11.06
13.10
14.90
16.94

5.78
8.44
10.03
12.01
13.51
15.07

29.57
13.23
9.27
8.36
9.34
11.04

E01

2
5
10
25
50
100

19.85
24.75
29.25
36.48
43.11
50.94

15.05
22.70
28.66
37.28
44.49
52.29

24.18
8.28
2.01
2.21
3.22
2.65

13.75
17.15
20.26
25.27
29.86
35.29

10.54
15.77
20.06
26.72
32.57
39.10

23.33
8.05
0.98
5.76
9.07
10.81

10.22
12.75
15.07
18.79
22.20
26.24

7.53
11.19
14.16
18.93
23.38
28.60

26.31
12.19
6.03
0.75
5.31
9.02

8.46
10.55
12.47
15.55
18.38
21.72

6.18
9.22
11.62
15.46
19.12
23.54

26.96
12.65
6.83
0.57
4.02
8.38

7.36
9.18
10.85
13.53
15.99
18.90

5.50
8.24
10.38
13.77
17.00
20.96

25.29
10.30
4.35
1.77
6.30
10.89

(continued on next page)

166

N.M. Ariff et al. / Journal of Hydrology 470471 (2012) 158171

Table 4 (continued)
Station

T (years)

Storms intensities (mm/h)


1-h Storms

3-h Storms

6-h Storms

9-h Storms

12-h Storms

ie

ie

ie

ie

ie

E02

2
5
10
25
50
100

23.28
28.73
33.68
41.57
48.74
57.15

18.54
25.60
30.29
36.21
40.54
44.81

20.36
10.89
10.07
12.90
16.82
21.60

16.06
19.82
23.24
28.68
33.63
39.43

13.03
18.59
22.55
27.97
32.20
36.46

18.88
6.22
2.98
2.48
4.24
7.53

11.90
14.69
17.22
21.25
24.92
29.22

9.42
13.94
17.16
21.76
25.58
29.63

20.89
5.10
0.34
2.38
2.63
1.40

9.83
12.13
14.23
17.56
20.59
24.14

7.93
12.03
14.95
19.11
22.64
26.48

19.30
0.87
5.05
8.84
9.94
9.68

8.54
10.54
12.36
15.26
17.89
20.98

7.28
11.18
13.96
17.93
21.30
25.01

14.77
5.99
12.91
17.50
19.05
19.24

E03

2
5
10
25
50
100

28.21
34.61
40.39
49.55
57.83
67.49

22.06
29.92
35.63
43.29
49.14
55.00

21.80
13.54
11.79
12.63
15.02
18.52

16.64
20.42
23.83
29.23
34.12
39.82

13.72
18.36
21.83
27.07
31.67
36.78

17.60
10.08
8.38
7.38
7.16
7.64

16.64
20.42
23.83
29.23
34.12
39.82

8.39
11.68
13.92
17.10
19.89
23.15

22.78
12.42
10.59
10.42
10.75
10.99

8.29
10.16
11.86
14.55
16.98
19.82

5.90
8.77
10.66
13.23
15.39
17.83

28.81
13.76
10.17
9.07
9.40
10.03

6.79
8.32
9.71
11.92
13.91
16.23

4.63
7.27
9.02
11.37
13.28
15.41

31.80
12.62
7.12
4.62
4.51
5.07

E04

2
5
10
25
50
100

23.55
29.55
35.08
44.02
52.27
62.06

16.84
26.02
33.38
44.25
53.45
63.70

28.50
11.94
4.87
0.52
2.25
2.64

15.01
18.83
22.36
28.06
33.32
39.56

11.33
17.25
22.25
30.22
37.41
45.59

24.52
8.41
0.49
7.70
12.28
15.25

10.43
13.09
15.55
19.51
23.16
27.50

7.81
11.78
15.05
20.41
25.56
31.76

25.19
10.00
3.20
4.65
10.38
15.49

8.28
10.38
12.33
15.47
18.37
21.81

6.26
9.50
12.08
16.27
20.33
25.37

24.39
8.52
2.00
5.15
10.67
16.34

6.98
8.76
10.40
13.05
15.49
18.39

5.49
8.38
10.66
14.30
17.81
22.22

21.33
4.28
2.53
9.60
15.00
20.82

E05

2
5
10
25
50
100

20.76
25.30
29.39
35.82
41.60
48.32

18.08
24.31
28.37
33.37
36.94
40.37

12.90
3.92
3.47
6.84
11.21
16.45

14.47
17.64
20.49
24.97
29.01
33.69

12.12
16.84
20.17
24.75
28.33
31.92

16.28
4.54
1.56
0.89
2.32
5.24

10.82
13.18
15.31
18.66
21.68
25.18

7.75
11.47
14.01
17.58
20.58
23.85

28.38
13.00
8.52
5.83
5.08
5.29

8.99
10.95
12.72
15.50
18.01
20.91

5.64
8.96
11.19
14.25
16.82
19.68

37.20
18.20
12.06
8.07
6.60
5.91

7.84
9.55
11.10
13.53
15.71
18.25

4.58
7.66
9.74
12.58
14.93
17.55

41.60
19.87
12.20
6.96
4.93
3.81

E06

2
5
10
25
50
100

20.23
24.77
28.87
35.34
41.19
48.00

16.65
22.91
27.17
32.62
36.66
40.66

17.72
7.52
5.89
7.70
11.00
15.29

13.98
17.12
19.95
24.43
28.47
33.18

11.53
16.16
19.55
24.34
28.17
32.08

17.52
5.62
2.01
0.36
1.06
3.29

10.38
12.71
14.81
18.13
21.13
24.62

7.85
11.36
13.87
17.54
20.70
24.19

24.36
10.55
6.30
3.25
2.01
1.74

8.58
10.51
12.24
14.99
17.47
20.36

6.08
9.15
11.30
14.37
17.04
20.09

29.16
12.89
7.73
4.12
2.43
1.33

7.46
9.13
10.65
13.03
15.19
17.70

5.16
8.01
9.99
12.78
15.20
17.96

30.78
12.35
6.20
1.91
0.05
1.45

S01

2
5
10
25
50
100

26.41
31.90
36.80
44.45
51.29
59.17

21.34
27.68
31.99
37.48
41.48
45.33

19.17
13.23
13.07
15.69
19.12
23.39

16.09
19.43
22.42
27.08
31.25
36.05

13.88
18.07
21.04
25.30
28.87
32.67

13.75
7.04
6.17
6.57
7.59
9.35

10.78
13.03
15.03
18.15
20.94
24.16

8.64
11.88
13.99
16.87
19.29
22.01

19.85
8.81
6.93
7.08
7.91
8.89

8.36
10.09
11.65
14.07
16.23
18.72

6.03
8.98
10.85
13.31
15.29
17.46

27.80
11.03
6.83
5.41
5.81
6.75

6.93
8.37
9.65
11.66
13.45
15.52

4.67
7.44
9.22
11.51
13.33
15.28

32.59
11.05
4.52
1.27
0.94
1.56

S02

2
5
10
25
50
100

29.64
36.15
42.01
51.24
59.55
69.20

23.40
31.34
36.75
43.64
48.67
53.51

21.06
13.31
12.51
14.83
18.27
22.68

17.70
21.59
25.08
30.60
35.56
41.32

14.87
20.21
24.06
29.61
34.21
39.05

15.95
6.39
4.09
3.22
3.78
5.49

11.67
14.23
16.54
20.18
23.45
27.25

9.00
13.00
15.69
19.48
22.72
26.41

22.85
8.68
5.14
3.46
3.10
3.09

8.95
10.92
12.69
15.47
17.98
20.90

6.20
9.72
12.04
15.20
17.82
20.76

30.77
10.98
5.06
1.77
0.90
0.65

7.36
8.98
10.44
12.73
14.79
17.19

4.78
8.02
10.19
13.09
15.45
18.06

35.14
10.71
2.40
2.80
4.44
5.06

S03

2
5
10
25
50
100

25.40
30.65
35.33
42.63
49.14
56.65

23.97
36.25
46.64
62.78
76.97
92.89

5.64
18.26
31.99
47.27
56.64
63.98

17.68
21.33
24.59
29.67
34.20
39.42

13.91
20.06
25.16
33.71
42.05
52.27

21.33
5.98
2.33
13.61
22.96
32.59

13.19
15.92
18.35
22.14
25.52
29.42

8.87
12.89
15.93
20.70
25.30
31.15

32.80
19.00
13.17
6.51
0.88
5.88

10.95
13.21
15.23
18.37
21.18
24.41

7.01
10.40
12.90
16.67
20.18
24.55

35.97
21.25
15.25
9.25
4.74
0.56

9.54
11.52
13.28
16.02
18.46
21.28

6.23
9.35
11.65
15.06
18.17
22.00

34.72
18.83
12.26
6.00
1.60
3.35

S04

2
5
10
25
50
100

31.78
41.37
50.49
65.71
80.21
97.90

20.66
32.84
43.81
61.76
78.31
97.61

34.99
20.61
13.23
6.02
2.37
0.30

17.43
22.68
27.69
36.03
43.98
53.68

11.21
16.37
20.89
28.86
37.08
47.63

35.69
27.84
24.55
19.90
15.69
11.28

10.73
13.96
17.04
22.18
27.08
33.05

6.51
9.38
11.59
15.12
18.61
23.19

39.34
32.82
32.02
31.85
31.27
29.82

7.88
10.25
12.51
16.28
19.87
24.26

4.61
6.85
8.49
10.95
13.23
16.09

41.45
33.17
32.18
32.77
33.43
33.67

6.27
8.16
9.97
12.97
15.83
19.32

3.70
5.66
7.07
9.14
11.01
13.27

40.98
30.73
29.05
29.50
30.48
31.34

Gaussian and Frank copula is relatively small compared to the AIC


values of individual copula at each station. This implies that both
copulas are equally suitable in tting the observed data. With this
reason and the fact that it is easier to construct an Archimedean
copula, which in this case is the Frank copula, compared to Gaussian copula which falls under the elliptic copula family, the Frank
copula is chosen to represent the twenty rainfall stations in Peninsular Malaysia.
The points on the IDF curves are yielded using the conditional
Frank copula shown in Eq. (17). The values for return period, T,
considered in this analysis are 2, 5, 10, 25, 50 and 100 years. Since,

this paper focuses on convective storms which is short duration


storms, the storm durations which are taken into account are 1,
3, 6, 9 and 12 h. By solving Eqs. (4) and (17) using these return periods and storm durations, the corresponding intensities, i, are obtained and tabulated in Table 4.
The estimated parameters of Eq. (20), a, j, b and c, for the storm
intensities, ie, acquired by the typical IDF empirical formula are given in Table 5. The values of ie for all storm durations, d, and return
periods, T, are shown in Table 4 along with the percentage of differences between i and ie, D, for all twenty stations. Fig. 5 provides the
visual comparisons between the IDF curves drawn based on Eq.

N.M. Ariff et al. / Journal of Hydrology 470471 (2012) 158171


Table 5
Parameters a, j, b and c of the typical IDF empirical formula.
Station

N01
N02
N03
N04
W01
W02
W03
W04
W05
W06
E01
E02
E03
E04
E05
E06
S01
S02
S03
S04

28.27
32.79
37.44
24.51
28.50
29.48
40.54
41.00
36.77
32.26
24.25
28.77
40.97
31.12
25.63
25.12
37.57
42.71
31.67
47.49

0.22
0.21
0.21
0.21
0.23
0.18
0.23
0.23
0.20
0.19
0.24
0.23
0.22
0.25
0.22
0.22
0.21
0.22
0.21
0.29

1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00

0.55
0.57
0.66
0.37
0.54
0.51
0.72
0.69
0.51
0.58
0.53
0.54
0.76
0.65
0.52
0.53
0.71
0.74
0.52
0.87

(20) and the points of IDF curves obtained using the copula method. The range of ie, i and D for each d and T are given in Table 6.
It can be observed from Table 4 that stations in Peninsular
Malaysia have values of percentage difference between i and ie,
D, less than 50 for most storm durations and return periods considered in this study. The only exception is the station Kota Tinggi
(S03), where the percentage difference of the return period of 50
and 100 years for 1-h storms are 56.64% and 63.98% respectively.
Eighty three percentage of D are less than 20 which implies that
the percentage of difference between i and ie are usually small. In
fact, half of the values for D are less than 10 with most of them
are from storm intensities for the return periods 10, 25, 50 and
100 years; more than 50% of D for these four return periods are less
than 10 with each recording 66%, 71%, 66% and 59% of D respectively. 54% of D for T = 5 years is in the range of 1020. The mean
of the percentage difference between i and ie for return period
5%, 10, 25, 50 and 100 are 11.63%, 8.85%, 8.43%, 9.26% and
10.62%. The differences in percentage are largest for the 2-year return period of most storm durations for all stations with the mean
of D is 23.51. This is believed to be due to the small value of storm
intensity. Hence, a small error in the intensity will translate into
signicant error in the differences (Singh and Zhang, 2007). The
difference usually increases as the storm duration increases from
3-h to 9-h and drops at 12-h storms. This can be seen from the
mean values of D according to storm duration 1, 3, 6, 9 and 12 h
which are 12.96, 8.73, 11.81, 13.81 and 13.23.
Table 6 shows that although the percentage of difference between i and ie are smaller for higher return periods; T = 25, 50
and 100 years; the sizes of their range are in average larger than
those for T = 2, 5 and 10 years. It can be seen that the average of
the minimum values of D are very small for T = 5, 10, 25, 50 and
100 years; 1.70, 1.24, 0.69, 0.91 and 1.05; while the average of
the minimum D for storms with 2-year return period is 9.12. However, the average of the maximum D is about the same for T = 2 and
T = 100 years which are 38.61 and 38.28 respectively. The averages
for the rest of the maximum values of D are 29.03, 29.96, 32.26 and
34.96. Thus, these results in the difference between the size range
of D for smaller (T = 2, 5 and 10) and higher (T = 25, 50 and 100)
return periods. The average sizes for the range of D are 29.49,
27.33 and 28.72 for smaller return periods and 31.56, 34.05 and
37.23 for higher return periods. Table 6 also indicates that the size
of the range for IDF points of storm intensities obtained from the
copula method, i, are slightly larger compared to those from the

167

typical IDF empirical formula, ie. The range for both i and ie increase
as the return period increases for all d hour storm durations. The
average values of the range size for i with respect to return periods
2, 5, 10, 25, 50 and 100 years are 7.24 mm, 8.78 mm, 10.51 mm,
13.68 mm, 17.05 mm and 21.21 mm while the corresponding values for ie are 7.78 mm, 10.25 mm, 12.49 mm, 16.16 mm,
20.26 mm and 26.10 mm. On the other hand, the ranges size of i
decreases as the storm duration increases. For storm duration 1,
3, 6, 9 and 12 h, the average range sizes for i are 28.24 mm,
11.99 mm, 10.52 mm, 7.44 mm and 7.21 mm. This downward
trend is not so prominent for ie with the average values for all stations and return periods are 28.90 mm, 12.53 mm, 11.36 mm,
12.15 mm and 12.60 mm for d = 1, 3, 6, 9 and 12 h respectively.
Generally, values of ie are mostly higher than i for all values of T
and d.
It can be seen from the results in Tables 4 and 6 that the differences between the storm intensities obtained from either the copula method or the typical IDF empirical formula does not follow
any particular pattern. As a consequence to that, the reason behind
the pattern for the percentage of differences, D, is also not apparent. Nevertheless, since most of the differences are small, we can
regard that the IDF points found by the copula method is somewhat in agreement with the IDF points obtained based on the typical IDF empirical formula. This is most likely due to the fact that
although the two methods handle the data differently, the storm
events which are used in both methods are the same since we consider similar denition of storms in both analyses. The difference is
that the storm intensity and duration in the copula method are
both regarded as random variables while only the storm intensity
is taken as a random variable in the IDF empirical formula. Furthermore, the storm duration is considered as a continuous random
variable in the copula method while it is used as a discrete and
xed value in the IDF empirical formula with representative values
are used to dene groups or categories of storm intensities. The
storm intensities of each category are assumed to have storm duration equal to the value which represents them. In a way, the typical
IDF empirical formula regards the storm intensity, ie, as a function
of x storm duration, d, and return period, T. For the copula approach, all the storms are combined irrespective of the storm duration to obtain the joint distribution for storm intensity and
duration. Meanwhile, the storms which are categorized with respect to the storm duration for the typical IDF empirical formula
are analyzed separately to get the design rainfall intensity for various return periods. Briey, the copula method t probability distributions to represent storm intensity and duration for each
station. Then, by calculating the correlation, s, between storm
intensity and duration as well as estimating the copula parameter,
h, the copula theory can be applied to obtain the joint distribution
for the two random variables in copula form. The IDF points are
then yielded based on the conditional distribution of storm intensity given storm duration d. As for the typical IDF empirical formula, the mean and standard deviation of each category of storm
intensities with x duration d are computed separately in order
to obtain ie for each storm duration d and various return periods
T. The IDF curves are then build by approximating the parameters
a, j, b and c through the nonlinear least squares method performed
on the Sherman equation, i.e. Eq. (20). In short, the results of both
methods which are the storm intensities i and ie are consistent because both methods are different storm models which represent
the same storm events.
The typical IDF curves using empirical derivations have not
much theoretical background; however, IDF points obtained based
on the copula method are supported by a sound copula theory. The
IDF points found from the latter method considers properties of actual storm characteristics such as storm intensities, depths and
durations. Furthermore, the copula method is able to provide the

168

N.M. Ariff et al. / Journal of Hydrology 470471 (2012) 158171

would have to be done separately. Moreover, for a large set or


range of storm duration, it is found easier to perform the copula
method since it involves the combination of storm intensities irrespective of storm duration as opposed to the IDF empirical formula
which requires storm intensities to be categorized with respect to
storm duration. No extra computation is needed for the copula
method but more categories of storm intensities, corresponding
to the x storm durations considered, have to be analyzed individually to obtain the mean, standard deviation and design rainfall
intensity of each category. Hence, the typical IDF empirical formula
will take considerably more time and is less efcient compared to

general form for the joint distribution of storm intensity and duration which is more informative and pertinent for further analysis.
Hence, other analysis on storm events such as obtaining the conditional distribution of storm duration given storm intensity, computing the moments and producing the quantile functions of
storm characteristics could be performed based on the probability
statement given by the general form of the joint distribution between storm intensity and storm duration. Thus, inadvertently
the copula method lessen the computational effort for future analysis compared to the IDF empirical method which does not provide
additional information on storm events and any further research

60

10

15

Durations,h

15

60

5
2

10

50

15

60
5
2

10

Durations,h

15

50

100
50

30 40

Intensities,mm/h

50

25
10

25

10

5
2

50

30 40

Intensities,mm/h

2510

100

10 20

50

15

E02

E01

100

10

Durations,h

60

10 20 30 40 50 60

Durations,h

W06

10
5

Durations,h

50

25

30 40

10

20

Intensities,mm/h

50

15

15

10

60

10

100

50

25

10

W05

10 20

Durations,h

100

30 40

Intensities,mm/h

10
5

50

15

W04

50

30 40

10

20

50

25

5
2

Durations,h

10

60

100

25
10

0
10

50

60

50

30 40

Intensities,mm/h

10

100

20

50
30 40

50

20

Intensities,mm/h

10

20

5
2

15

W02

100

25

10

Durations,h

10

25
10

10

60

60
50
30 40

50

Intensities,mm/h

100

W03
Intensities,mm/h

15

W01

Durations,h

Intensities,mm/h

10

Durations,h

N04

10 20

10

25

30 40

10

20

Intensities,mm/h

50
30 40
20

Intensities,mm/h

10

50
25

100
50

10

5
2

10

100

10

60

60
50
30 40
20

50

Intensities,mm/h

100

25

N03

N02

N01

10

Durations,h

15

10

15

Durations,h

Fig. 5. Comparisons between the IDF curves obtained from the typical IDF empirical formula and IDF points from the copula method. Note: The numbers at the beginning of
each curve indicates the respective values of return period, T = 2, 5, 10, 25, 50 and 100 years.

169

N.M. Ariff et al. / Journal of Hydrology 470471 (2012) 158171

60
50

10

15

E06

60

S02

10

15

50

5
2

15

10

15

Durations,h

5
2

10

50
25

5
2

10

50 60

S04
Intensities,mm/h

50
40

25
10

100

10 20 30 40

60

S03

30

10

Durations,h

100

20

100
50

30 40

30

15

20

40

25
10

10

Durations,h

Intensities,mm/h

60
50

50

Durations,h

25
10

10

20

5
2

100

20

Intensities,mm/h

25
10

50

15

10

60
50
30 40

50

Intensities,mm/h

10

S01

100

5
2

Durations,h

Durations,h

25
10

10

50

30 40

100

20

25
10

10

Intensities,mm/h

50

Intensities,mm/h

100

10 20 30 40 50 60

Intensities,mm/h

10 20 30 40 50 60

50
25
10

Intensities,mm/h

E05

E04

E03
100

10

15

Durations,h

10

15

Durations,h
Fig. 5. (continued)

the copula method. In addition to that, the nonlinear least squares


method applied under the IDF empirical formula could sometime
fail to approximate the parameters of the Sherman equation used
in the analysis if the initial values provided are unsuitable. Hence,
nding the IDF points with the copula method prove to be more
advantageous for researchers and hydrologists to extend their
understandings of storm events.

8. Conclusions
Intensitydurationfrequency (IDF) curves are important especially for a country like Malaysia where ood is regarded by the
Ministry of Natural Resources and Environment as the most significant natural hazard. The IDF curves currently available in Malaysia
are done using the univariate analysis and empirical formula. However, rainfall variables such as intensity and duration are related to
each other. This paper derives IDF points using bivariate frequency
analysis by utilizing the copula method. Based on this study, several conclusions are drawn.
The storm intensity, I, and duration, D, are negatively correlated
which implies that shorter storms have higher intensities. This
nding agrees with previous literatures done in other countries
(Kao and Govindaraju, 2007; Singh and Zhang, 2007). Based on this
relationship, four copulas which are commonly applied in hydro-

logical analysis are considered in this study. The four copulas are
the FarlieGumbelMorgenstern (FGM), Gaussian, AliMikhail
Haq (AMH) and Frank copula. The values of Kendalls s obtained
imply that the dependence level between I and D are neither very
high nor very low. Thus, only the Frank and Gaussian copula have
domains of s which are suitable for analyzing the storm events in
Peninsular Malaysia and these two copulas are further considered
in this analysis.
The Frank copula is deemed appropriate to represent the relationship between storm intensity and duration in Peninsular
Malaysia based on the AIC values which are mostly smaller than
the AIC values of Gaussian copula. It is found that both copulas results in relatively similar AIC values when used to t the relationship between storm intensities and durations. However, the Frank
copula has an additive advantage since it is a member of Archimedean copula family which can be easier constructed. Hence, the
Frank copula is preferable in hydrological analysis compared to
the Gaussian copula which falls under the elliptic copula family.
The conditional distribution of Frank copula is also less complicated compared to the conditional Gaussian copula, Eqs. (17) and
(9).
The difference between the IDF curves obtained based on the
typical IDF empirical formula and the points on IDF points acquired
using the copula method depends on the given storm duration d
and return period T. There is no exact pattern for the differences

170

N.M. Ariff et al. / Journal of Hydrology 470471 (2012) 158171

Table 6
Range of the storm intensities based on the typical IDF empirical formula and the copula method as well as their percentage differences.
d (hour)/T (years)

12

2
5
10
25
50
100

19.8531.78
24.7541.37
28.8750.49
35.3465.71
41.1980.21
48.0097.90

13.7520.97
17.1225.20
19.9528.96
24.4336.03
28.4743.98
33.1853.68

10.2216.64
12.7120.42
14.8123.83
18.1329.23
20.9434.12
24.1639.82

7.8813.20
10.0915.86
11.6518.23
14.0721.90
16.2325.16
18.7228.91

6.2711.56
8.1613.89
9.6515.96
11.6619.18
13.4522.04
15.5225.32

2
5
10
25
50
100

15.0524.74
22.7036.25
27.1746.64
32.6262.78
36.3078.31
39.1097.61

10.5418.15
15.7724.37
19.5528.71
23.6234.61
26.3242.05
29.1852.27

6.5113.69
9.3818.88
11.5922.45
15.1227.44
18.5731.57
20.6435.95

4.6111.79
6.8516.58
8.4919.84
10.9524.39
13.2328.18
16.0932.31

3.7010.94
5.6615.54
7.0718.67
9.1423.02
11.0126.64
13.2730.63

2
5
10
25
50
100

5.6434.99
0.3020.61
0.9531.99
0.5247.27
2.2556.64
0.3063.98

10.4635.69
2.4627.84
0.4924.55
0.3619.90
0.4722.96
1.5432.59

13.4439.34
0.6132.82
0.3432.02
0.7531.85
0.8831.27
1.4029.82

10.6941.45
0.8733.17
2.0032.18
0.5732.77
0.9033.43
0.5633.67

5.3941.60
4.2830.73
2.4029.05
1.2729.50
0.0530.48
1.4531.34

Range of storm intensities (mm)


ie

but they are generally small. Hence, storm intensities found based
on both methods are believed to be in agreement with each other.
This is probably because both storm intensities represent the same
storm events since similar denition of storms are considered in
both analyses. Although the two methods provide consistent results, there are great benets in using the bivariate copula method
as opposed to the univariate IDF empirical formula in practice. The
typical IDF empirical formula does not have much theoretical background while the IDF points based on the copula method are supported by a sound copula theory. The copula method is also
worthwhile to be used in storm analysis since it provides the general form for the joint distribution of storm intensity and duration
which is pertinent for further analysis. Furthermore, copula method is found to be easier to perform for a large set or range of storm
duration compared to the IDF empirical formula. This is due to the
fact that all storms are combined irrespective of the storm duration
for the copula approach and thus no extra computation is needed
for analysis while the storms are categorized with respect to storm
duration for the typical IDF empirical method and each storm category has to be analyzed individually. Thus, the typical IDF empirical approach is less efcient and more time consuming. Hence, the
copula method is more meaningful for researchers and hydrologists to extend their study on storm events.
The storm intensities yielded based on both methods indicate
that they depend on the geographical locations of the rainfall stations since the distribution of storm intensity and its pattern vary
for each region. Hence, the percentage difference, D, between them
also varies from one region to another. This may be caused by the
different characteristics of storms in each region. These differences
are very much inuenced by two monsoon seasons, the Northeast
and Southwest Monsoon, which occurred in Peninsular Malaysia
from November to March and from May to September each year.
The regionalization of storms in Peninsular Malaysia as well as
their characteristics will be included in future papers.
In conclusion, bivariate frequency analysis of storm events is
imperative for hydraulic applications in order to represent the
complexity of actual hydrologic events. It is believed that modelling the real events of storms will have a more signicant contribution on current hydrologic designs which are mostly based on the
univariate analysis approach.
Acknowledgements
The authors would like to thank the Department of Irrigation
and Drainage Malaysia for providing hourly rainfalls data for the

use of the study. Utmost appreciations to the Ministry of Higher


Education (MOHE) and Universiti Kebangsaan Malaysia (UKM)
for the allocation of research Grants, UKM-ST-06-FRGS0181-2010
and UKM-GGPM-PI-028-2011.

References
Amin, M.Z.M., Desa, M.N.M., Daud, Z.M., 2008. Malaysia. Asian Pacic FRIEND, 53
57.
Chow, V.T., Maidment, D.R., Mays, L.W., 1988. Applied Hydrology. McGraw-Hill,
New York, 572p.
Cordova, J.R., Rodriguez-Iturbe, I., 1985. On the probabilistic structure of storm
surface runoff. Water Resour. Res. 21 (5), 755763.
De Michele, C., Salvadori, G., 2003. A Generalized Pareto intensityduration model
of storm rainfall exploiting 2-Copulas. J. Geophys. Res. 108 (D2), 4067.
De Michele, C., Salvadori, G., Canossi, M., Petaccia, A., Rosso, R., 2005. Bivariate
statistical approach to check adequacy of dam spillway. J. Hydrol. Eng. 10,
50.
Deni, S.M., Suhaila, J., Wan Zin, W.Z., Jemain, A.A., 2010. Spatial trends of dry spells
over Peninsular Malaysia during monsoon seasons. Theor. Appl. Climatol. 99 (3),
357371.
Favre, A.C., Musy, A., Morgenthaler, S., 2002. Two-site modeling of rainfall based on
the NeymanScott process. Water Resour. Res. 38 (12), 1307.
Favre, A.C., El Adlouni, S., Perreault, L., Thiemonge, N., Bobe, B., 2004. Multivariate
hydrological frequency analysis using copulas. Water Resour. Res. 40 (1),
W01101.
Goel, N.K., Kurothe, R.S., Mathur, B.S., Vogel, R.M., 2000. A derived ood frequency
distribution for correlated rainfall intensity and duration. J. Hydrol. 228 (12),
5667.
Guo, J.C.Y., 2006. Urban Hydrology and Hydraulic Design. Water Resources
Publications, LLC.
Huard, D., Evin, G., Favre, A.C., 2006. Bayesian copula selection. Comput. Stat. Data
Anal. 51 (2), 809822.
Kao, S.C., Govindaraju, R.S., 2007. Probabilistic structure of storm surface runoff
considering the dependence between average intensity and storm duration of
rainfall events. Water Resour. Res. 43 (6), W06410.
Kao, S.C., Govindaraju, R.S., 2008. Trivariate statistical analysis of extreme
rainfall events via the Plackett family of copulas. Water Resour. Res. 44 (2),
W02415.
Kottegoda, N.T., Rosso, R., 2008. Applied statistics for civil and environmental
engineers. Wiley-Blackwell.
Koutsoyiannis, D., Kozonis, D., Manetas, A., 1998. A mathematical framework for
studying rainfall intensitydurationfrequency relationships. J. Hydrol. 206 (1
2), 118135.
Kruskal, W.H., 1958. Ordinal measures of association. J. Am. Statist. Assoc., 814861.
Nelsen, R.B., 2006. An Introduction to Copulas. Springer Verlag.
Nhat, L.M., Tachikawa, Y., Takara, K., 2006. Establishment of intensityduration
frequency curves for precipitation in the monsoon area of Vietnam. Ann. Dis.
Prev. Res. Inst. 49, 93102.
Palynchuk, B., Guo, Y., 2008. Threshold analysis of rainstorm depth and duration
statistics at Toronto, Canada. J. Hydrol. 348 (34), 535545.
Reiss, R.D., Thomas, M., 2007. Statistical Analysis of Extreme Values: With
Applications to Insurance, Finance, Hydrology and Other Fields. Birkhauser.
Renard, B., Lang, M., 2007. Use of a Gaussian copula for multivariate extreme
value analysis: some case studies in hydrology. Adv. Water Resour. 30 (4),
897912.

N.M. Ariff et al. / Journal of Hydrology 470471 (2012) 158171


Restrepo-Posada, P.J., Eagleson, P.S., 1982. Identication of independent rainstorms.
J. Hydrol. 55 (14), 303319.
Salvadori, G., De Michele, C., 2004. Frequency analysis via copulas: theoretical
aspects and applications to hydrological events. Water Resour. Res. 40 (12),
W12511.
Schmidt, T., 2006. Coping with Copulas. Chapter Forthcoming in Risk Books:
Copulas from Theory to Applications in Finance.
Shabri, A.B., Daud, Z.M., Ariff, N.M., 2011. Regional analysis of annual maximum
rainfall using TL-moments method. Theor. Appl. Climatol., 110.
Shiau, J.T., 2003. Return period of bivariate distributed extreme hydrological events.
Stoch. Environ. Res. Risk A 17 (1), 4257.
Singh, V.P., Zhang, L., 2007. IDF curves using the Frank Archimedean copula. J.
Hydrol. Eng. 12, 651662.
Sklar, A., 1959. Fonctions de repartition a n dimensions et leurs marges. Publ. Inst.
Stat. Univ. Paris 8 (1), 11.
Suhaila, J., Jemain, A.A., 2009. Investigating the impacts of adjoining wet days on the
distribution of daily rainfall amounts in Peninsular Malaysia. J. Hydrol. 368 (1
4), 1725.
Suhaila, J., Deni, S.M., Wan Zin, W.Z., Jemain, A.A., 2010a. Spatial patterns and trends
of daily rainfall regime in Peninsular Malaysia during the southwest and
northeast monsoons: 19752004. Meteorol. Atmos. Phys., 118.
Suhaila, J., Deni, S.M., Wan Zin, W.Z., Jemain, A.A., 2010b. Trends in Peninsular
Malaysia rainfall data during the southwest monsoon and northeast monsoon
seasons: 19752004. Sains Malaysiana 39 (4), 533542.

171

Suhaila, J., Jemain, A.A., Hamdan, M.F., Zin, W.Z.W., 2011. Comparing rainfall
patterns between regions in Peninsular Malaysia via a functional data analysis
technique. J. Hydrol. 411, 197206.
Sulaiman, A.H., 2007. Flood and Drought Management in Malaysia. Speech Text.
Ministry of Natural Resources and Environment, pp. 45.
Wan Zin, W.Z., Jemain, A.A., Ibrahim, K., 2009a. The best tting distribution of
annual maximum rainfall in Peninsular Malaysia based on methods of Lmoment and LQ-moment. Theor. Appl. Climatol. 96 (3), 337344.
Wan Zin, W.Z., Jemain, A.A., Ibrahim, K., Suhaila, J., Sayang, M.D., 2009b. A
comparative study of extreme rainfall in Peninsular Malaysia: With reference to
partial duration and annual extreme series. Sains Malaysiana 38 (5), 751760.
Wan Zin, W.Z., Jamaludin, S., Deni, S.M., Jemain, A.A., 2010. Recent changes in
extreme rainfall events in Peninsular Malaysia: 19712005. Theor. Appl.
Climatol. 99 (3), 303314.
Yue, S., 2000. Joint probability distribution of annual maximum storm peaks and
amounts as represented by daily rainfalls. Hydrol. Sci. J. 45 (2), 315326.
Yue, S., 2002. The bivariate lognormal distribution for describing joint statistical
properties of a multivariate storm event. Environmetrics 13 (8), 811819.
Yue, S., Ouarda, T., Bobee, B., 2001. A review of bivariate gamma distributions for
hydrological application. J. Hydrol. 246 (14), 118.
Zhang, L., Singh, V.P., 2006. Bivariate ood frequency analysis using the copula
method. J. Hydrol. Eng. 11, 150.
Zhang, L., Singh, V.P., 2007. Bivariate rainfall frequency distributions using
Archimedean copulas. J. Hydrol. 332 (12), 93109.

Vous aimerez peut-être aussi