Vous êtes sur la page 1sur 19

Non-Stationary Rainfall Intensity-

Duration-Frequency Relationship: a
Comparison between Annual Maximum
and Partial Duration Series

V. Agilan & N. V. Umamahesh

Water Resources Management


An International Journal - Published
for the European Water Resources
Association (EWRA)

ISSN 0920-4741
Volume 31
Number 6

Water Resour Manage (2017)


31:1825-1841
DOI 10.1007/s11269-017-1614-9

1 23
Your article is protected by copyright and all
rights are held exclusively by Springer Science
+Business Media Dordrecht. This e-offprint
is for personal use only and shall not be self-
archived in electronic repositories. If you wish
to self-archive your article, please use the
accepted manuscript version for posting on
your own website. You may further deposit
the accepted manuscript version in any
repository, provided it is only made publicly
available 12 months after official publication
or later and provided acknowledgement is
given to the original source of publication
and a link is inserted to the published article
on Springer's website. The link must be
accompanied by the following text: "The final
publication is available at link.springer.com”.

1 23
Author's personal copy
Water Resour Manage (2017) 31:1825–1841
DOI 10.1007/s11269-017-1614-9

Non-Stationary Rainfall Intensity-Duration-Frequency


Relationship: a Comparison between Annual
Maximum and Partial Duration Series

V. Agilan 1 & N. V. Umamahesh 1

Received: 2 July 2016 / Accepted: 28 February 2017 /


Published online: 13 March 2017
# Springer Science+Business Media Dordrecht 2017

Abstract The rainfall Intensity-Duration-Frequency (IDF) relationship is the primary input


for storm water management and other engineering design applications across the world and it
is developed by fitting an appropriate theoretical probability distribution to annual maximum
(AM) series or partial duration series (PDS) of rainfall. The existing IDF relationship devel-
oping methods consider the extreme rainfall series as a stationary series. There exist few
studies that compared AM and PDS datasets for developing rainfall IDF relationship in a
stationary condition. However, during the last few decades, the intensity and frequency of
extreme rainfall events are increasing due to global climate change and creating a non-
stationary component in the extreme rainfall series. Therefore, the rainfall IDF relationship
developed with the stationary assumption is no longer tenable in a changing climate. Hence, it
is inevitable to develop non-stationary rainfall IDF relationship and to understand the differ-
ences in non-stationary rainfall IDF relationships derived using AM and PDS datasets.
Consequently, the objectives of this study are: (1) to develop non-stationary rainfall IDF
relationships using both AM and PDS datasets; (2) to compare them in terms of return level
estimation. In particular, the non-linear trend in different durations’ PDS and AM datasets of
Hyderabad city (India) rainfall is modeled using Multi-objective Genetic Algorithm (MGA)
generated Time based covariate. In this study, the PDS datasets are modeled by the
Generalized Pareto Distribution (GPD) while the AM datasets are modeled by the
Generalized Extreme Value Distribution (GEVD). The time-varying component is introduced
in the scale parameter of the GPD and the location parameter of the GEVD by linking the
MGA generated covariate. In addition, the complexity of each non-stationary model is

* N. V. Umamahesh
mahesh@nitw.ac.in
V. Agilan
agilanvensiv@gmail.com

1
Department of Civil Engineering, National Institute of Technology Warangal, Warangal, Telangana
506004, India
Author's personal copy
1826 Agilan V., Umamahesh N.V.

identified using the corrected Akaike Information Criteria (AICc) and the statistical signifi-
cance of trend parameter in the non-stationary models is estimated using the Likelihood Ratio
(LR) test. Upon detecting significant superiority of non-stationary models, the return levels of
extreme rainfall event for 2-, 5-, 10- and 25-year return periods are calculated using non-
stationary models. From the results, it is observed that the non-stationary return levels
estimated with PDS datasets are higher than those estimated with AM datasets for short
durations and smaller return periods while the non-stationary return levels estimated with
AM datasets are higher than those estimated with PDS datasets for long durations and higher
return periods.

Keywords Climate change . Extreme rainfall . IDF relationship . Non-linear trend . Non-
stationary

1 Introduction

Methods based on the statistical Extreme Value Theory (EVT) have been used in
hydrology since very long time (Katz 2013) and they are mostly based on the assumption
of temporal stationarity, meaning, the extreme events’ frequency is assumed to be
constant over time (Jakob 2013). Constructing rainfall Intensity-Duration-Frequency
(IDF) relationship is one of the main applications of EVT in hydrology (Agilan and
Umamahesh 2017). The rainfall IDF relationships contain statistics of extreme rainfall. In
particular, for a given return period and duration, the rainfall IDF relationship describes
the intensity of extreme rainfall event. Therefore, it is the primary input for storm water
management and other engineering design applications across the world (Endreny and
Imbeah 2009). The IDF relationships are developed based on the historical rainfall time
series data by fitting a theoretical probability distribution to annual maximum (AM) (i.e.
block maximum with a block size of one year) series or partial duration series (PDS),
also called the peak over threshold (POT) series. In PDS approach, the rainfall events
with intensity higher than the defined high threshold value are considered as the extreme
rainfall series and these values are generally modeled by the Generalized Pareto
Distribution (GPD). On the other hand, the extreme rainfall series obtained by extracting
maximum value in each calendar year is generally modeled by the Generalized Extreme
Value Distribution (GEVD). Since AM series considers only one value per year, sec-
ondary events in any year which are having an intensity higher than the annual maximum
of other years will not be accounted and it can be considered as one of the shortcomings
of AM method. Though the PDS method alleviates this drawback, since all values above
the defined threshold are selected, the series may have dependency and it must be
removed before modeling PDS dataset (Coles 2001). Despite PDS alleviates some of
the shortcomings of AM method, because of several technical problems, the PDS method
is much less used for developing rainfall IDF relationships (Madsen et al. 2002; Ben-Zvi
2009; Yilmaz et al. 2014) and only the AM method is used by many researchers for
developing rainfall IDF relationships (Mailhot et al. 2007; Huard et al. 2010; Elsebaie
2012; Mirhosseini et al. 2013; Demarée and Vyver 2013; Cheng and AghaKouchak
2014; Panagoulia et al. 2014; Paixao et al. 2015; Rupa et al. 2015). However, Madsen
et al. (1997) compared the AM and PDS datasets for modeling hydrologic extremes with
different parameter estimation methods using Monte Carlo simulations and they
Author's personal copy
Non-Stationary Rainfall IDF Relationship: AM vs. PDS 1827

suggested PDS method for at-site quantile estimation. Specifically, Chang et al. (2016)
compared AM and PDS methods for developing rainfall IDF relationship using rainfall
observations in Peninsular Malaysia and they reported that the PDS method estimates the
higher value of intensity for short duration when compared to the AM method. Note that
these comparisons were made under the stationarity assumption.
However, during the last few decades, the intensity and frequency of extreme rainfall events
are increasing due to global climate change (Trenberth et al. 2003; Emori and Brown 2005;
Tramblay et al. 2012). The increasing trend in extreme rainfall intensity and frequency have
already been detected in many regions of the world (Asadieh and Krakauer 2015; Agilan and
Umamahesh 2015; Vasiliades et al. 2015). Therefore, the rainfall IDF relationship developed
with the stationarity assumption will underestimate the return level of an extreme event. In
other words, the storm water drainage networks will fail more frequently than its actual design,
if it is designed using rainfall IDF relationships which were developed with stationarity
assumption. Hence, it is inevitable to update the rainfall IDF relationship developing methods
for the non-stationary condition. Towards this, in recent years, researchers developed non-
stationary rainfall IDF relationships by modeling trend in the observed extreme rainfall time
series. In particular, Cheng and AghaKouchak (2014) developed a non-stationary IDF rela-
tionship using AM dataset by introducing linear trend in the location parameter of the GEVD
using Time covariate. Similarly, Yilmaz and Perera (2014) analyzed non-stationarity in the IDF
relationships of Melbourne, Australia using AM dataset and by incorporating linear trend in
the GEVD’s location and scale parameters using Time covariate. Yilmaz et al. (2014) used
PDS dataset to analyze non-stationarity in the IDF relationships of Melbourne by introducing
trend in the GPD’s scale parameter using Time covariate. Further, Agilan and Umamahesh
(2016) developed non-stationary rainfall IDF relationships of Hyderabad, India and
Wilmington, United States by modeling non-linear trend in the extreme rainfall series using
Multi-objective Genetic Algorithm (MGA) generated Time based covariate. In addition,
Agilan and Umamahesh (2016) also mentioned that directly applying Time covariate based
linear trend is sometimes increasing the bias of non-stationary model.
Note that the non-stationarity in the IDF relationships of Melbourne is analyzed
using both AM and PDS datasets by Yilmaz and Perera (2014) and Yilmaz et al.
(2014) respectively. However, both Yilmaz and Perera (2014) and Yilmaz et al. (2014)
did not find any superiority of non-stationary models over the stationary models for
Melbourne city and, therefore, used stationary models for developing IDF relationship.
Thus, the differences in non-stationary rainfall IDF relationship derived using AM and
PDS datasets are not reported till date and it is indispensable for researchers, urban
planners and designers to know the difference in non-stationary rainfall IDF relation-
ships derived using AM and PDS datasets. Therefore, in this study, the non-stationary
rainfall IDF relationships are developed using both AM and PDS datasets and
compared in terms of return level estimation.

2 Study Area and Data

Non-stationarity present in the intensity and frequency of Hyderabad city extreme rainfall is
already detected and attributed by Agilan and Umamahesh (2015) and they also reported that
the stationary statistical model is not even qualified as a considerable model when compared to
the non-stationary statistical model for modeling extreme rainfall of the city. Furthermore,
Author's personal copy
1828 Agilan V., Umamahesh N.V.

Agilan and Umamahesh (2016) developed non-stationary rainfall IDF relationships of the
Hyderabad city by modeling non-linear trend in the AM extreme rainfall series and reported
that the IDF relationships derived under the assumption of the stationary extreme value theory
are underestimating the return levels of extreme rainfall. Since the substantial evidence of non-
stationarity in the Hyderabad city extreme rainfall characteristics is already reported, the
Hyderabad city is chosen as the study area to compare the non-stationary IDF relationships
which are developed using PDS and AM datasets. Geographical and climatological details of
the Hyderabad city can be obtained either from Agilan and Umamahesh (2015) or Agilan and
Umamahesh (2016). Hourly observed rainfall data of the city is procured from the India
Meteorological Department (IMD) for the period of 1st January 1972 to 31st December 2013.
This data is a gauge observation and it is observed at the center of the city i.e. 78.46o E and
17.45o N. Further, these hourly observations are used to generate different duration rainfall
series (i.e. 1-h, 2-h, 3-h, 6-h, 12-h, 24-h and 48-h) using the moving window approach.

3 Methodology

The main aim of this study is to analyze the differences in rainfall IDF relationships developed
using AM and PDS datasets under non-stationary condition and the methodology adopted to
achieve this aim comprises of following three sections,

1. Developing non-stationary rainfall IDF relationship with AM datasets by modeling non-


linear trend in the AM dataset of all selected durations.
2. Similarly, developing non-stationary rainfall IDF relationship with PDS datasets by
modeling non-linear trend in the PDS dataset of all selected durations.
3. Comparing rainfall return levels estimated using AM and PDS datasets.

3.1 Non-stationary IDF Relationships Using AM Dataset

3.1.1 Non-stationary GEVD

Developing a rainfall IDF relationship using AM dataset is straightforward. Once the rainfall
series of different durations are created using the moving window approach, the AM series can
be directly extracted. Let M = m1, m2, …, mn be the AM rainfall series of any selected duration
and they are n independent and identically distributed (iid) random variable. Then the
cumulative distribution function (CDF), F, of these values converges to GEVD and it is given
by Eq. (1) (Coles 2001; Katz et al. 2002).
8 (  −1=k )
>
> k ðm−α Þ k ðm−αÞ
>
< exp − 1 þ ; β > 0; 1 þ > 0; k≠0
β β
F ðm; α; β; k Þ ¼    ð1Þ
>
> ðm−αÞ
>
: exp −exp − ; β > 0; k ¼ 0
β
where α is location parameter, β is scale parameter and k is shape parameter. This model is to
fit the distribution of AM dataset when the parameters are constant. The non-stationary GEVD,
where the parameters vary with respect to time, is given by Eq. (2).
Author's personal copy
Non-Stationary Rainfall IDF Relationship: AM vs. PDS 1829

8 (  −1=k t )
>
> k ð m −α Þ k t ðm−αt Þ
>
< exp − 1 þ
t t t
; β t > 0; 1 þ > 0; k t ≠0
βt βt
F ðmt ; αt ; β t ; k t Þ ¼    ð2Þ
>
> ðmt −αt Þ
>
: exp −exp − ; β t > 0; k t ¼ 0
βt

here the parameters αt, βt, and kt are linked with some covariate(s) which vary with time.
Though all three parameters can be allowed to vary with time, in this study, only the location
parameter (αt) is allowed to vary with time while the scale and shape parameters are defined as
time invariant. Because, the shape parameter k cannot be estimated precisely and assuming it
as a smooth function of time is unrealistic (Coles 2001). The reason for not allowing β to vary
with time is that, allowing temporal changes in β requires long-term observations (Cheng et al.
2014). The general form of incorporating time-varying component in the location parameter of
the GEVD using a covariate is given by Eq. (3).
 
α
αt ¼ ½1 ct  0 ð3Þ
α1

where, c is time-varying covariate; α0 is intercept; α1 is slope parameter. Most of the previous


studies directly used Time as a covariate (linear trend) to develop non-stationary rainfall IDF
relationships using both AM dataset (Cheng and AghaKouchak 2014; Yilmaz and Perera
2014) and PDS dataset (Yilmaz et al. 2014). However, recently, Agilan and Umamahesh
(2016) reported that the usage of linear trend for modeling non-stationarity in the time series
sometimes increase the bias of non-stationary model and they proposed Multi-objective
Genetic Algorithm (MGA) generated Time based covariate for modeling non-linear trend in
the extreme rainfall time series without increasing the bias. In particular, by considering both
long term trend caused due to global climate change and periodicity of physical processes
which are affecting extreme rainfall, they defined a general form of a covariate which is based
on Time and it is given by Eq. (4).
C ¼ T x1 þ sinðx2  T Þ þ cosðx3  T Þ ð4Þ
where T = 1, 2, 3… n. For the given extreme rainfall series, the values of x1, x2 and x3 are
estimated using MGA and it is discussed in Section 3.1.4. Then, the parameters of the non-
stationary GEVD is estimated using the method of maximum likelihood.

3.1.2 Significance and Quality of Non-stationary Model

Once the non-stationary model (GEVD or GPD) is constructed, the significance of non-
stationary model against the stationary model should be evaluated carefully before using it
for developing non-stationary rainfall IDF relationship. Normally, the Akaike Information
Criterion with small sample size correction (AICc) is used to select the best model among
different candidate models (Agilan and Umamahesh 2016, 2017) and it is given by Eq. (5).
Therefore, it can be used to compare the non-stationary model and the stationary model.
2pðp þ 1Þ
AICc ¼ −2logL þ 2p þ ð5Þ
n−p−1
where -log L is the minimized negative log likelihood of stationary/non-stationary
GEVD/GPD, p is the number parameters in the model, and n is sample size. Note that
Author's personal copy
1830 Agilan V., Umamahesh N.V.

the covariate (Eq. (4)) is generated using observed rainfall, therefore, the non-
stationary model will have three additional parameters and it should be considered
while calculating AICc value. For example, the number of parameters in the non-
stationary GEVD is 7 (i.e. α0, α1, β, k, x1, x2 and x3). In addition to AICc, the
statistical significance of trend parameter in the non-stationary model can be checked
using the Likelihood Ratio (LR) test. For information on LR test, the interested reader
is referred to Agilan and Umamahesh (2016).
Further, if the non-stationary model is superior to the stationary model and the trend
parameter in the non-stationary model is statistically significant, the quality of the non-
stationary model can be assessed using the probability-probability (PP) and quantile-quantile
(QQ) plots (Coles 2001; Katz et al. 2002). Creating PP and QQ plots of non-stationary GEVD
is explained here and for non-stationary GPD, it is discussed in Section 3.2.2. If a non-
stationary GEVD is fitted to n years of AM rainfall series, since the n years of AM rainfall data
is not identically distributed, the AM values should be converted to residuals (ε) for calculating
PP and QQ plots (Coles 2001).
  
1 mi −ðα0 þ α1  ci Þ
εi ¼ log 1 þ k ; i ¼ 1; 2; :::; n: ð6Þ
k β

where, α0 , α1 , k and β are the maximum likelihood method identified values of α0 , α1 , k and
β respectively. The PP plot points (Empirical, Model) and QQ plot points (Model, Empirical)
are given by Eq. (7) and (8) respectively.
 
i
; expð−expð−εi ÞÞ ; i ¼ 1; 2; :::; n: ð7Þ
nþ1

    
i
εi ; −log −log ; ; i ¼ 1; 2; :::; n: ð8Þ
nþ1
here, ε is the ordered values of ε.

3.1.3 Return Level Estimation

If the non-stationary GEVD is significantly superior to the stationary GEVD and quality of the
non-stationary GEVD is satisfactory, the estimated parameters are then used to calculate return
levels of extreme rainfall for various return periods. Unlike stationary model, the non-
stationary GEVD parameters (in this case only location parameter) will vary with time.
Consequently, the return level also varies with time. Hence, the return level is a set of values.
Estimation of a set of return levels for a Q year return period is given by Eq. (9) (Coles 2001).

n o
Z Q ¼ zQ Q
1 ; z2 ; :::; zn
Q

8 "  −k #
>
> β 1
>
< α0 þ α1  ci þ −log 1− −1 ; k≠0; i ¼ 1; 2; :::; n
k Q
¼     ð9Þ
>
> 1
>
: α0 þ α1  ci þ β −log −log 1− ; k ¼ 0; i ¼ 1; 2; :::; n
Q
Author's personal copy
Non-Stationary Rainfall IDF Relationship: AM vs. PDS 1831

3.1.4 Estimating Variables of Eq. (4)

To estimate variables x1, x2 and x3 of Eq. (4), Agilan and Umamahesh (2016) used
MGA. The same methodology has been used in this study and it is briefly explained in
this section. The aim is to estimate values of x1, x2 and x3 which produces a good
quality non-stationary model with less bias and it has been achieved using Non-
dominated Sorting Genetic Algorithm II (NSGA-II) (Deb et al. 2002). For each set of
MGA proposed values, the non-stationary model complexity is identified using AICc
(Eq. (5)) value while the quality of non-stationary GEVD is estimated using Root Mean
Square Error (RMSE) and correlation coefficient (r) based distance function and they
are given by Eq. (10) and Eq. (11) respectively. Though the quality of the model can be
assessed using both PP and QQ plots, the RMSE and r values are calculated between
QQ plot points, because QQ plot is more helpful in discerning models (Sugahara et al.
2009). For estimating variables x1, x2 and x3 of Eq. (4), the NSGA-II is set to minimize
AICc, RMSE and a distance function based on r (Objective functions). The NSGA-II
inputs used in this study are listed in Table 1. In the direction of selecting the best
solution among all solutions of the last generation, the objective function ratio method
is used to combine the objective functions’ value and the solution which has lowest
combined value is selected as final estimate. For more information on these inputs and
working procedure of NSGA-II, the interested reader is referred to Deb et al. (2002),
Reshma et al. (2015) and Agilan and Umamahesh (2016).
0 n   1
∑ QE i −QE QM i −QM
B C
F2 ¼ B i¼1
@1− rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n  2ffi C
2 rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n  A  100 ð10Þ
∑ QE i −QE ∑ QM i −QM
i¼1 i¼1

 0:5
1 n
F3 ¼ ∑ ðQEi −QM i Þ2 ð11Þ
n i¼1

here, QEi is empirical quantile (Eq. (8) or Eq. (17)), QMi is model quantile (Eq. (8) or
Eq. (17)), QE is empirical quantiles’ mean and QM is model quantiles’ mean.

Table 1 Inputs of NSGA-II

Input name Values/Method

Objective functions Eq. (5), Eq. (10) and Eq. (11)


Selection method Tournament selection
Crossover operator Simulated binary crossover operator
Mutation operator Polynomial mutation operator
No. of generations 1000
No. of populations 100
Crossover probability 0.85
Mutation probability 0.15
Lower and upper boundaries of x1 [−10 to 10]
Lower and upper boundaries of x2 [−10 to 10]
Lower and upper boundaries of x3 [0 to 10]
Author's personal copy
1832 Agilan V., Umamahesh N.V.

3.2 Non-stationary IDF Relationships Using PDS Dataset

3.2.1 Non-stationary GPD

As previously stated, the extreme rainfall series derived using PDS approach is generally
modeled by the GPD. Suppose X = x1, x2, …, xj are j iid random variable conditioned on X > u.
Now, the CDF of X is given by Eq. (12) (Coles 2001).
8 h
>
> x−u i−1=ξ x−u
< 1− 1 þ ξ ; σ > 0; 1 þ ξ>0
F ðx; σ; ξ; uÞ ¼  σ  σ ð12Þ
> −ðx−uÞ
>
: 1−exp ; σ > 0; ξ ¼ 0
σ

here, u is a chosen large threshold, ξ is shape parameter and σ is scale parameter. This model is
to fit the distribution of PDS dataset when the parameters are constant and the non-stationary
GPD is given by Eq. (13).
8  −1=ξt
>
> xt −ut xt −ut
< 1− 1 þ ξt ; σt > 0; 1 þ ξ >0
F ðxt ; σt ; ξt ; ut Þ ¼  t σ  σt t ð13Þ
>
> −ðxt −ut Þ
: 1−exp ; σt > 0; ξt ¼ 0
σt
here, σt, ξt, and ut are linked with a time-varying covariate(s). Like non-stationary GEVD, all
three parameters of GPD can be allowed to vary with time. However, time varying u lead to
numerical instabilities in the maximization of the likelihood (Furrer and Katz 2007), hence, it
is avoided in this study. Further, ξ is also kept time invariant because it is difficult to estimate it
precisely (Coles 2001). Consequently, the time varying component is added only in σ. The
general form of incorporating time-varying component in σ is given by Eq. (14).
  
σ
σt ¼ exp ½ 1 ct  0 ð14Þ
σ1

here, c is time-varying covariate; σ0 is intercept; σ1 is slope parameter. To ensure the positive


value of σ, the exponential function is introduced in Eq. (14). Like AM dataset, the covariate
(Eq. (4)) is generated using MGA for PDS datasets. However, the value of T in Eq. (4) cannot
take year value because there can be more than one exceedance in any year. Therefore, in the
case of PDS dataset, the value of T in Eq. (4) is continuous year value. For example, if a
rainfall event which exceeds u on dth day of yth year, then the value of T would be:

y−1 þ 365 ord 366 .
Though the PDS method alleviates some of the shortcomings of the AM method, it
remains underused (Beguería 2005). Because, unlike AM dataset, developing a rainfall
IDF relationship using PDS dataset is not simple. The most important obstacle that
comes with PDS modeling is the choice of the threshold value (Beguería 2005). Small
changes in the threshold value can cause significant modification on the lower part of the
distribution, and minor changes on the left tail of the distribution can lead to significant
differences in the estimated parameter (Beguería 2005). There are many methods for
selecting a threshold value. However, there is no globally accepted method. Also,
selection of threshold values is more complex in the non-stationary case than the
stationary case. Because the different percentile of rainfall may have a different
Author's personal copy
Non-Stationary Rainfall IDF Relationship: AM vs. PDS 1833

magnitude of trend. For example, very high percentile rainfall may have steep increasing
trend while high and moderate percentile rainfall may have increasing/decreasing trend.
Consequently, change in threshold value can change the magnitude of trend in the
extreme rainfall series or it can change the sign (i.e. from increasing to decreasing or
vice-versa). Therefore, it is not appropriate to compare PDS and AM datasets which have
a different magnitude of trend. In this study, to preserve the similar magnitude trend as
AM dataset, the minimum value of each duration AM dataset is chosen as a threshold
value of the corresponding duration for extracting PDS dataset. The other problem with
PDS method is the dependency in the series and it violates the iid assumption. Therefore,
the raw PDS should be declustered before modeling by the GPD (Coles 2001). The
consecutive rainfall events when rainfall is higher than u is an extreme rainfall cluster
and they are detached from each other by more than 24 h when rainfall is below u.
Towards declustering, only the maximum value in each cluster is retained and the
declustered PDS dataset is modeled by the GPD.

3.2.2 Quality of Non-stationary GPD

After constructing non-stationary GPD, the model complexity is identified using AICc and the
statistical significance of trend parameter (σ1) is evaluated using the LR test. Suppose the non-
stationary GPD is superior to the stationary GPD, then the quality of the non-stationary model
is evaluated using the PP and QQ plots. If the non-stationary GPD is fitted to a declustered
PDS dataset X = x1, x2, …, xj, these values should be transformed to residuals to obtain the PP
and QQ plots (Katz 2013).
 
1 ξðxi −uÞ
τi ¼ log 1 þ ; i ¼ 1; 2; :::; j: ð15Þ
ξ expðσ0 þ σ1  ci Þ

where, σ0 , σ1 and ξ are estimated values of σ0, σ1 and ξ. Eq. (16) gives the PP plot points
(Empirical, Model) and the QQ plot points (Empirical, Model) are given by Eq. (17).
 
i
; 1−expð−τ 1 Þ ; i ¼ 1; 2; :::; j: ð16Þ
jþ1

   
i
−log 1− ; τ 1 ; i ¼ 1; 2; :::; j: ð17Þ
jþ1
here, τ i is the ordered values of τ.

3.2.3 Return Level Estimation

Like non-stationary GEVD, if the non-stationary GPD is significantly superior to the stationary
GPD and quality of the non-stationary GPD is acceptable, the estimated parameters are used to
calculate return levels of extreme rainfall for various return periods. Unlike non-stationary
GEVD, estimation of return levels for Q year return period is not straightforward in GPD.
Because the number of exceedance in the declustered PDS dataset will not be equal to the
number of years (in most cases) and it must be considered while calculating return levels.
Author's personal copy
1834 Agilan V., Umamahesh N.V.

Estimation of a set of return levels for Q year return period from the non-stationary GPD is
given by Eq. (18) (Coles 2001).
8
n o < expðσ0 þ σ1  ci Þ h i
Q Q Q uþ ðϕ  QÞξ −1 ; ξ≠0; i ¼ 1; 2; :::; j
Z Q ¼ z1 ; z2 ; :::; z j ¼ ξ ð18Þ
:
u þ expðσ0 þ σ1  ci Þlogðϕ  QÞ; ξ ¼ 0; i ¼ 1; 2; :::; j

j
ϕ¼ ð19Þ
n
here, n is number of years.

4 Results and Discussion

As mentioned earlier, in order to preserve the similar magnitude trend as AM datasets, the
minimum value of each duration AM dataset is chosen as a threshold value of the correspond-
ing duration for extracting PDS dataset. After declustering PDS datasets of selected durations
with the method described in Section 3.2.1, the number of exceedance of PDS dataset is varied
from 132 to 170 and it is 3.14 to 4.05 times higher than AM dataset. In other words, the
average number of exceedance in a year is varied from 3.14 to 4.05.
Once the PDS and AM datasets of all selected durations are prepared, the Time based
covariate which will produce good quality and less complex non-stationary models is gener-
ated for all selected durations using the MGA based methodology discussed in Section 3.1.4.
Then the non-stationary GEVDs and GPDs are constructed using the method of maximum
likelihood for AM and PDS datasets respectively. Further, the superiority of non-stationary
model when compared to the stationary model is evaluated by calculating the difference
between AICc of the stationary and the non-stationary models. According to Burnham and
Anderson (2004), when a dataset is fitted with many candidate models, if the difference
between AICc of a candidate model and the model which has lowest AICc value is more than
two, that candidate model cannot be considered as a reasonable choice for modeling that
particular dataset and the model which has lowest AICc is the best model for that dataset. The
difference between AICc values of stationary and non-stationary models constructed with PDS
and AM datasets of all selected durations is given in Table 2. From Table 2, it is observed that
the AICc difference for all selected durations of PDS and AM datasets are greater than two and

Table 2 The AICc difference between stationary and non-stationary models and LR test p-value

Duration PDS AM

AICcS-AICcns LR test p-value AICcS-AICcns LR test p-value

1-h 2.82 6.86E-03 8.81 6.36E-04


2-h 5.24 7.89E-03 8.90 6.08E-04
3-h 9.17 1.39E-03 2.26 1.16E-02
6-h 6.62 4.39E-03 6.45 1.83E-03
12-h 5.05 8.79E-03 2.64 9.90E-03
24-h 5.96 6.10E-03 4.52 4.33E-03
48-h 11.52 5.03E-04 12.14 1.39E-04
Author's personal copy
Non-Stationary Rainfall IDF Relationship: AM vs. PDS 1835

it indicates that the stationary model cannot be considered as a reasonable choice for modeling
both PDS and AM datasets of all selected durations. Also, the p-value of LR test is lower than
0.01 for all PDS and AM datasets and it shows that the trend parameter in non-stationary
models is statistically significant at the 0.01 significance level.
Since the non-stationary models are significantly superior to the stationary models con-
structed with PDS and AM datasets of all selected durations and the trend parameter in the
non-stationary models is statistically significant with a very high significance level, the quality
of non-stationary GEVDs and GPDs are checked with PP and QQ plots. PP and QQ plots of
non-stationary GPD constructed with 2-h duration PDS dataset are plotted in Fig. 1a. From
Fig. 1a, a good match between empirical and non-stationary GPD derived probabilities and
quantiles is observed and it indicates that the quality of non-stationary GPD is satisfactory.
Likewise, PP and QQ plots of non-stationary GEVD constructed with 2-h duration AM dataset
are shown in Fig. 1b. Like non-stationary GPD, the quality of non-stationary GEVD is
satisfactory for 2-h duration AM dataset. Also, the quality of non-stationary GEVDs and
GPDs constructed with AM and PDS datasets of remaining durations are checked and their
quality is satisfactory. In the interest of brevity, the PP and QQ plots of remaining durations are
not shown in this paper.
Since the non-stationary GPDs and GEVDs constructed with PDS and AM datasets of
different durations are superior to the stationary models and the quality of non-stationary
models are satisfactory, the non-stationary return levels are calculated using the methods
discussed in Section 3.1.3 and 3.2.3. As previously stated, in a non-stationary context, return

(a)
Residual Probability Plot Residual Quantile Plot (Exptl. Scale)
5
0.8

4
Empirical
Model

3
0.4

2
1
0.0

0.0 0.2 0.4 0.6 0.8 1.0 0 1 2 3 4 5


Empirical Model
(b)
Residual Probability Plot Residual Quantile Plot (Gumbel Scale)
4
0.8

3
Empirical
Model

2
0.4

1
-1 0
0.0

0.0 0.2 0.4 0.6 0.8 1.0 -1 0 1 2 3


Empirical Model
Fig. 1 a PP and QQ plots of non-stationary GPD constructed with 2-h duration PDS dataset; b PP and QQ plots
of non-stationary GEVD constructed with 2-h duration AM dataset
Author's personal copy
1836 Agilan V., Umamahesh N.V.

level for the given return period will vary with time. Therefore, return level estimated using
non-stationary model is not a single value and it is a set of values. Also, the sample size in PDS
and AM datasets are not same for all selected durations. Therefore, a set of return levels
estimated using non-stationary GPD and non-stationary GEVD will not have the same size.
Hence, it is not possible to compare the return levels directly. To compare the return levels
estimated using non-stationary GPDs and non-stationary GEVDs, first, the probability density
function (PDF) of return levels is calculated using the kernel density estimation (KDE)
method. Then the CDF of return levels estimated using non-stationary GPDs and non-
stationary GEVDs are compared.
As the study area is an urban catchment and most of the urban infrastructure is designed for
return level of 25 years or less, the PDS and AM derived return levels are compared only for 2,
5, 10 and 25 year return periods. The CDFs of 2-, 5-, 10-, and 25-year return levels estimated
using non-stationary GPDs and non-stationary GEVDs are plotted in Fig. 2. From Fig. 2, it is
observed that the return levels’ CDFs of PDS and AM datasets are not matching with each
other for most of the durations and return periods. Further, in some cases, all percentiles of
PDS return levels are higher than AM return levels (see 1-h duration 25-year return levels in
Fig. 2) while, in some other instant, all percentiles of AM return levels are higher than PDS
return levels (refer 6-h duration 25-year return levels in Fig. 2). In addition, there are some
cases where only part of percentiles of PDS return levels are higher than AM return levels
(refer 1-h duration 2-year return levels in Fig. 2). Therefore, it is difficult to draw some
conclusions from Fig. 2 alone. Hence, the difference between AM and PDS derived return
levels’ quantiles (101 divisions between 0 and 1) of 2-, 5-, 10- and 25-year return periods are
calculated and the violin plot (i.e. rotated kernel density is plotted on both sides of a box plot)
of these differences are plotted in Fig. 3 for all durations and return periods. In Fig. 3, positive
values indicate that the return levels estimated with AM dataset is higher than those estimated
with PDS dataset and the negative values indicate its counterpart. From Fig. 3, it is seen that
the non-stationary return levels estimated with PDS datasets are higher than those estimated
with AM datasets for short durations and smaller return periods while the non-stationary return
levels estimated with AM datasets are higher than those estimated with PDS datasets for long
durations and higher return periods.
Since the differences are in mm/h and durations are not same, the differences in long
duration rainfall appear very small. Therefore, the percentage difference between AM and PDS
derived return levels’ quantiles of 2-, 5-, 10- and 25-year return periods are also calculated and
the violin plot of these percentage differences are plotted in Fig. 4 for all durations and return
periods. Similar to Figs. 3 and 4 also reveals that the non-stationary return levels estimated
with PDS datasets are higher than those estimated with AM datasets for short durations and
smaller return periods while the non-stationary return levels estimated with AM datasets are
higher than those estimated with PDS datasets for long durations and higher return periods. In
addition, from Fig. 4, it is also observed that the percentage difference is more for short
duration extreme rainfall of smaller return period and long duration extreme rainfall of higher
return period.

5 Summary and Conclusion

The differences in estimating the intensity of extreme rainfall event for a given return period
was examined using PDS and AM datasets in a non-stationary context. In particular, the
Author's personal copy
Non-Stationary Rainfall IDF Relationship: AM vs. PDS 1837

Fig. 2 CDF of non-stationary return level derived from AM and PDS datasets for different return periods and durations
Author's personal copy
1838 Agilan V., Umamahesh N.V.
Quantile difference (mm/h)
-8 -6 -4 -2 0 2

1-h 2-h 3-h 6-h 12-h 24-h 48-h


(a)
Quantile difference (mm/h)
-4 -2 0 2

1-h 2-h 3-h 6-h 12-h 24-h 48-h


(b)
Quantile difference (mm/h)
-4 -2 0 2

1-h 2-h 3-h 6-h 12-h 24-h 48-h


(c)
Quantile difference (mm/h)
-6 -4 -2 0 2 4

1-h 2-h 3-h 6-h 12-h 24-h 48-h


(d)
Fig. 3 Difference between AM and PDS derived return levels’ quantiles of a 2-year, b 5-year, c 10-year and d
25-year return period. The positive values indicate that the return levels estimated with AM dataset is higher than
those estimated with PDS dataset and the negative values indicate its counterpart
Author's personal copy
Non-Stationary Rainfall IDF Relationship: AM vs. PDS 1839
Quantile difference in %
0
-10
-20
-30

1-h 2-h 3-h 6-h 12-h 24-h 48-h


(a)
Quantile difference in %
-20 -10 0 10

1-h 2-h 3-h 6-h 12-h 24-h 48-h


(b)
Quantile difference in %
-20 -10 0 10 20

1-h 2-h 3-h 6-h 12-h 24-h 48-h


(c)
10 20 30 40
Quantile difference in %
-10 0

1-h 2-h 3-h 6-h 12-h 24-h 48-h


(d)
Fig. 4 Percentage difference between AM and PDS derived return levels’ quantiles of a 2-Year, b 5-Year, c 10-
Year and d 25-Year return period. The positive values indicate that the return levels estimated with AM dataset is
higher than those estimated with PDS dataset and the negative values indicate its counterpart
Author's personal copy
1840 Agilan V., Umamahesh N.V.

Hyderabad city extreme rainfall series of different durations is extracted by both PDS and AM
methods. Further, the Time-based covariate which will model the non-linear trend in the
extreme rainfall series is generated for all durations’ PDS and AM datasets using MGA. In
this study, the PDS datasets are modeled by the GPD while the AM datasets are modeled by
the GEVD. The time-varying component is introduced in the scale parameter of GPD and the
location parameter of GEVD by linking the MGA generated covariate. Once the non-
stationary models (GPDs and GEVDs) are constructed, the complexity of each model is
identified using the AICc value and the statistical significance of trend parameter is estimated
using LR test. Since the developed non-stationary models are superior to the stationary models
for both PDS and AM datasets of all selected durations, the return levels of extreme rainfall
event for 2-, 5-, 10- and 25-year return periods are calculated using non-stationary models. As
the parameters of the non-stationary models will vary with time, the return levels also vary
with time. Therefore, for the given duration and return period, the CDFs of return levels
estimated using PDS and AM datasets are compared. Also, the difference between AM and
PDS derived return levels’ quantiles of 2-, 5-, 10- and 25-year return periods are analyzed
using the violin plots. From the results, it is observed that the non-stationary return levels
estimated with PDS datasets are higher than those estimated with AM datasets for short
durations and smaller return periods while the non-stationary return levels estimated with
AM datasets are higher than those estimated with PDS datasets for long durations and higher
return periods. Furthermore, it is also noted that the difference (in percentage) is more for short
duration extreme rainfall of smaller return period and long duration extreme rainfall of higher
return period.

Acknowledgements This work was funded by the Information Technology Research Academy (ITRA),
Government of India Grant no. ITRA/15(68)/water/IUFM/01. We also thank the editor and two anonymous
reviewers whose constructive comments helped to improve the manuscript’s clarity and quality.

References

Agilan V, Umamahesh NV (2015) Detection and attribution of non-stationarity in intensity and frequency of
daily and 4-h extreme rainfall of Hyderabad, India. J Hydrol 530:677–697
Agilan V, Umamahesh NV (2016) Modelling nonlinear trend for developing non-stationary rainfall intensity–
duration–frequency curve. Int J Climatol. doi:10.1002/joc.4774
Agilan V, Umamahesh NV (2017) What are the best covariates for developing non-stationary rainfall intensity-
duration-frequency relationship? Adv Water Resour 101:11–22
Asadieh B, Krakauer NY (2015) Global trends in extreme precipitation: climate models versus observations.
Hydrol Earth Syst Sci 19:877–891
Beguería S (2005) Uncertainties in partial duration series modelling of extremes related to the choice of the
threshold value. J Hydrol 303(1–4):215–230
Ben-Zvi A (2009) Rainfall intensity–duration–frequency relationships derived from large partial duration series. J
Hydrol 367:104–114
Burnham KP, Anderson DR (2004) Multimodel inference: understanding AIC and BIC in model selection.
Sociol Methods Res 33:261–304
Chang KB, Lai SH, Othman F (2016) Comparison of annual maximum and partial duration series for derivation
of rainfall intensity-duration-frequency relationships in peninsular Malaysia. J Hydrol Eng 21(1):05015013
Cheng L, AghaKouchak A (2014) Nonstationary precipitation intensity-duration-frequency curves for infrastruc-
ture design in a changing climate. Nature: Scientific Reports 4:7093
Cheng L, Kouchak AA, Gilleland E, Katz RW (2014) Non-stationary extreme value analysis in a changing
climate. Clim Chang 127:353–369
Coles S (2001) An introduction to statistical modelling of extreme values. Springer, London
Author's personal copy
Non-Stationary Rainfall IDF Relationship: AM vs. PDS 1841

Deb K, Pratap A, Agarwal S, Meyarivan T (2002) A fast and elitist Multiobjective genetic algorithm: NSGA-II.
IEEE Trans Evol Comput 6(2):182–197
Demarée GR, Vyver VH (2013) Construction of intensity-duration-frequency (IDF) curves for precipitation with
annual maxima data in Rwanda, Central Africa. Adv Geosci 35:1–5
Elsebaie IH (2012) Developing rainfall intensity–duration–frequency relationship for two regions in Saudi
Arabia. Journal of King Saud University - Engineering Sciences 24(2):131–140
Emori S, Brown SJ (2005) Dynamic and thermodynamic changes in mean and extreme precipitation under
changed climate. Geophys Res Lett 35:L17706
Endreny TA, Imbeah N (2009) Generating robust rainfall intensity–duration–frequency estimates with short-
record satellite data. J Hydrol 371:182–191
Furrer EM, Katz RW (2007) Generalized linear modeling approach to stochastic weather generators. Clim Res
34:129–144
Huard D, Mailhot A, Duchesne S (2010) Bayesian estimation of intensity–duration–frequency curves and of the
return period associated to a given rainfall event. Stoch Env Res Risk A 24(3):337–347
Jakob D (2013) Nonstationarity in extremes and engineering design. In: Kouchak A et al (eds) Extremes in a
changing climate: detection, analysis and uncertainty. Springer, Dordrecht, pp 363–417
Katz RW (2013) Statistical methods for nonstationary extremes. In: Kouchak et al (eds) Extremes in a changing
climate: detection, analysis and uncertainty. Springer, Dordrecht, pp 15–37
Katz RW, Parlange MB, Naveau P (2002) Statistics of extremes in hydrology. Adv Water Resour 25(8):1287–
1304
Madsen H, Rasmussen PF, Rosbjerg D (1997) Comparison of annual maximum series and partial duration series
methods for modeling extreme hydrologic events: 1. At-site modeling. Water Resour Res 33(4):747–757
Madsen H, Mikkelsen PS, Rosbjerg D, Harremoes P (2002) Regional estimation of rainfall intensity-duration-
frequency curves using generalized least squares regression of partial duration series statistics. Water Resour
Res 38(11):21/1-21–11
Mailhot A, Duchesne S, Caya D, Talbot G (2007) Assessment of future change in intensity–duration–frequency
(IDF) curves for southern Quebec using the Canadian regional climate model (CRCM). J Hydrol 347:197–
210
Mirhosseini G, Srivastava P, Stefanova L (2013) The impact of climate change on rainfall intensity–duration–
frequency (IDF) curves in Alabama. Reg Environ Chang 13:25–33
Paixao E et al (2015) An integrated approach for identifying homogeneous regions of extreme rainfall events and
estimating IDF curves in southern Ontario, Canada: incorporating radar observations. J Hydrol 528:734–750
Panagoulia D, Economou P, Caroni C (2014) Stationary and nonstationary generalized extreme value modelling
of extreme precipitation over a mountainous area under climate change. Environmetrics 25:29–43
Reshma T et al (2015) Optimization of calibration parameters for an event based watershed model using genetic
algorithm. Water Resour Manag 29(13):4589–4606
Rupa C, Saha U, Mujumdar PP (2015) Model and parameter uncertainty in IDF relationships under climate
change. Adv Water Resour 79:127–139
Sugahara S, Rocha RP, Silveira R (2009) Non-stationary frequency analysis of extreme daily rainfall in Sao
Paulo, Brazil. Int J Climatol 29:1339–1349
Tramblay Y, Neppel L, Carreau J, Sanchez-Gomez E (2012) Extreme value modelling of daily areal rainfall over
Mediterranean catchments in a changing climate. Hydrol Process 26(25):3934–3944
Trenberth KE, Dai A, Rasmussen RM, Parsons DB (2003) The changing character of precipitation. Bull Am
Meteorol Soc 84(9):1205–1217
Vasiliades L, Galiatsatou P, Loukas A (2015) Nonstationary frequency analysis of annual maximum rainfall using
climate covariates. Water Resour Manag 29(2):339–358
Yilmaz AG, Perera JC (2014) Extreme rainfall Nonstationarity investigation and intensity–frequency–duration
relationship. J Hydrol Eng 19(6):1160–1172
Yilmaz AG, Hossain I, Perera JC (2014) Effect of climate change and variability on extreme rainfall intensity–
frequency–duration relationships: a case study of Melbourne. Hydrol Earth Syst Sci 18:4065–4076

Vous aimerez peut-être aussi