Вы находитесь на странице: 1из 16

Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324

https://doi.org/10.1007/s00477-018-1583-4
(0123456789().,-volV)(0123456789().,-volV)

ORIGINAL PAPER

Derivation of rainfall IDF relations by third-order polynomial normal


transform
Lingwan You1 • Yeou-Koung Tung1

Published online: 10 July 2018


Ó Springer-Verlag GmbH Germany, part of Springer Nature 2018

Abstract
Establishing the rainfall intensity–duration–frequency (IDF) relations by the conventional method, the use of parametric
distribution models has the advantage of automatic compliance of monotonicity condition of rainfall intensity and fre-
quency. However, fitting rainfall data to a distribution separately by individual duration may possibly produce undulation
and crossover of IDF curves which does not comply physical reality. This frequently occurs when rainfall record length is
relatively short which often is the case. To tackle this problem this study presents a methodological framework that
integrates the third-order polynomial normal transform (TPNT) with the least squares (LS) method to establish rainfall IDF
relations by simultaneously considering multi-duration rainfall data. The constraints to preserve the monotonicity and non-
crossover in the IDF relations can be incorporated easily in the LS-based TPNT framework. Hourly rainfall data at Zhongli
rain gauge station in Taiwan with 27-year record are used to establish rainfall IDF relations and to illustrate the proposed
methodology. Numerical investigation indicates that the undulation and crossover behavior of IDF curves can be effec-
tively circumvented by the proposed approach to establish reasonable IDF relations.

Keywords Rainfall intensity–duration–frequency relations  Rainfall frequency analysis  Third-order polynomial normal
transform  Least-square method

1 Introduction The basic procedure of developing rainfall IDF relations


at a given location involves at-site frequency analysis of
Rainfall intensity (depth)–duration–frequency (IDF, DDF) annual maximum rainfall intensity (or depth) of several
relations are widely used in hydrosystems planning, design, selected durations. Then, based on the developed rainfall
and management. Such relations define important refer- intensity–frequency relations for the selected durations, a
ences about the design rainfall intensity of the particular proper empirical IDF equation is adopted to fit the inten-
frequency (or return period). The magnitude of design sity–duration relations with each frequency individually or
rainstorm provides the basis of design criteria for various with all frequencies combined. By the conventional
hydrosystem infrastructures, such as urban stormwater method, rainfall frequency analysis is conducted by
management facilities including storm sewer networks and choosing a proper parametric probability distribution
detention basins (Akan and Houghtalen 2003; Sun et al. model to separately fit the observed annual maximum
2010). Rainfall IDF relations are also used for the evalu- rainfall data of different durations. The choice of a distri-
ation the effect of rainfall interception by different plant bution model for the rainfall intensity–frequency relations
covers (López-Lambraño et al. 2013), for the development is largely statistical without much physical justification
of flood inundation maps (Sahoo and Sreeja 2015; Doong (Singh and Strupczewski 2002). The use of parametric
et al. 2016), and many others. distribution models has the advantage of automatic com-
pliance of monotonicity condition requiring that the mag-
nitude of rainfall intensity (or depth) of a given duration
& Yeou-Koung Tung
yk2013tung@gmail.com will increase with the cumulative probability. However, it
does not guarantee that the resulting rainfall intensity–
1
Disaster Prevention and Water Environmental Research frequency curves of different durations would not intersect
Center, National Chiao-Tung University, Hsinchu, Taiwan

123
2310 Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324

within the probability range of practical application. The duration is larger than that of shorter duration for a given
crossovers of rainfall intensity–frequency curves of dif- frequency.
ferent durations, which would not happen physically in The methods mentioned above to treat undulation and
reality, occur often in the process of establishing rainfall crossover of rainfall IDF curves utilize parametric distri-
IDF relations (Porras and Porras 2001; Xu and Tung 2009; butions along with an externally postulated functional
Asikoglu and Benzeden 2014; Cleveland et al. 2015). relation defining the statistical moments and durations of
Porras and Porras (2001) stated that short record length and rainfall data. In this paper the third-order polynomial nor-
data of questionable quality contribute to occurrence of mal transformation (TPNT) is applied to establish rainfall
intersection of derived IDF curves. This is because the use IDF relations. Through the TPNT framework, one does not
of short rainfall record induces significant sampling errors have to choose a particular parametric distribution. It can
in estimated rainfall quantiles by frequency analysis. One handle multiple-duration rainfall data simultaneously
other plausible reason for the occurrence of crossover IDF while, at the same time, straightforwardly incorporate
curves is because frequency analysis of rainfall intensity/ constraints to preserve the monotonicity condition and non-
depth is performed separately or independently by dura- intersection nature of rainfall IDF curves. The TPNT has
tion. In doing so, the correlations intrinsically reside among been widely applied to different fields. Examples include,
rainfall data of different durations are not considered. but not limited to, wind power modeling (Yang and Zou
Haktanir (2003) pointed out that frequency analysis of 2012), load computation in power network planning (Cai
rainfall intensity/depth of different durations should not be et al. 2013), structural reliability analysis (Hong 1998;
performed independently of each other. Zhao and Lu 2007), dynamic stability assessment of a
In general, subjective modification is practiced to power system (Ciapessoni et al. 2017), and simulation of
remove crossovers of IDF curves. By analyzing rainfall massive public health data (Demirtas et al. 2012).
data at more than 500 recording gauges throughout Vene- One attractive feature of the TPNT in practical statistical
zuela, Porras and Porras (2001) found that at-site sample data analysis is that one does not have to know or make
mean and standard deviation of rainfall data can fit with strong assumption about the distribution model for the
duration nicely by a simple power function D = atb with random variable under consideration. A multivariate dis-
D being the rainfall depth, t being the duration, and a, tribution function can be easily established according to the
b being coefficients. They also found that the values of univariate normal transformation of non-normal random
coefficient of variation for rainfall data are relatively variables and their known correlation structure. Therefore,
stable across different durations for those stations where the TPNT is amenable to establishing rainfall IDF relations
crossover did not occur. By analyzing the behavior of by considering multi-duration rainfall data simultaneously.
Gumbel-based rainfall IDF with and without intersection The basic idea of the TPNT is to represent any random
they empirically suggested two conditions under which variable by the third-order polynomial function of the
Gumbel-based IDF curves would be less likely to intersect: standard normal variable. It was first presented by Fleish-
(1) record length of 30 years or longer; and (2) values of man (1978) for efficient generation of non-normal random
exponential coefficient ‘‘b’’ lie in the range of 0.4–0.6. variables with known parameters.
Later, several methods for deriving no-crossover IDF An essential step in applying TPNT is to estimate
relations were developed according to scaling theory which polynomial coefficients that can capture or preserve the
assumes a simple functional relationship between statistical known statistical features of the random variable under
moments of rainfalls with their durations. Haktanir (2003) consideration. Chen and Tung (2003) described four
proposed the inclusion of divergence criteria as a constraint polynomial coefficient estimation methods based on pro-
in deriving rainfall IDF relations to ensure that no cross- duct–moments (PM), L-moments (LM), Fisher–Cornish
over would occur. Along the similar idea of Porras and asymptotic expansion (FC), and the least-square (LS) cri-
Porras (2001), Asikoglu and Benzeden (2014) proposed the terion and examined their performances in the univariate
simple generalization procedure by fitting a smooth func- setting. The three moment-matching methods (i.e., PM,
tion between rainfall statistics and storm duration to cir- LM and FC methods) attempt to identify polynomial
cumvent undulation and intersection behavior of IDF coefficients that preserve the first four moments of the
curves under the lognormal and Gumbel distributions. Xu concerned random variable. The advantage of the moment-
and Tung (2009) addressed the crossover issue by utilizing matching methods is that the first four moments possess
the rainfall scale-invariant model, in conjunction with physical importance of the variables in engineering appli-
constrained least-square method, to estimate generalized cation and they can be calculated from sample data (Zhao
logistic distribution parameters in defining IDF relations in and Lu 2007).
Hong Kong. The constraints are explicitly set in the opti- To establish a valid probabilistic relation between the
mization model stipulating that rainfall depth of longer original variable and the standard normal variable, the

123
Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324 2311

TPNT coefficients are required to satisfy the monotonicity Stage-2 is to fit the derived rainfall intensity–frequency of
condition. However, the moment-matching methods have various durations obtained in Stage-1 by a simple empirical
no control in producing such coefficients that comply with IDF model for practical applications.
the monotonicity condition. This could largely be attrib-
uted to sampling error associated with the sample moments 2.1.1 Frequency analysis of extreme rainfall intensity data
from using finite number of data. Chen and Tung (2003)
derived the valid domains of the third and fourth moments Extreme rainfall data series of several durations are
for PM and LM within which monotonic relation between extracted from a rain gauge. Most commonly used data
the original variable and normal variable is held. Headrick series are annual maximum series. Sometimes, partial
(2010) uses the term ‘‘power method transform’’ and duration series or annual exceedance series are used. In an
extends the TPNT to a fifth-order and further derives at-site frequency analysis, the best-fit (or most plausible)
functional relations between the polynomial coefficients parametric distribution model is identified individually for
with parameters of various specific distributions. each rainfall data series of a particular duration and the
The exercise of establishing rainfall IDF relations corresponding model parameters are estimated. In the
involves statistical analysis of finite annual maximum conventional frequency analysis approach, rainfall data of
rainfall data of various durations. More often than not, varying durations are separately analyzed. Therefore, no
rainfall record length is not sufficiently long and the esti- consideration is given to preserve the cross-correlations
mated sample moments, especially higher-order ones that intrinsically exist among extreme rainfall data of dif-
(skew coefficient and kurtosis), are subject to significant ferent durations. From the adopted parametric distribution
sampling errors. Under this situation, the rainfall IDF model for the selected duration, rainfall quantile corre-
relations obtained from moment-matched TPNT are often sponding to a stipulated return period (or frequency) can be
improper due to the violation of monotonicity condition or obtained as:
crossover of some IDF curves. In this study, the LS-based  
1
TPNT is employed to determine the polynomial coeffi- it;T ¼ Fi1 1  ð1aÞ
t
T
cients for establish at-site rainfall IDF relations for two
main reasons: (1) treating multiple-duration rainfall data or
simultaneously; and (2) allowing the incorporation of it;T ¼ lt þ KT  rt ð1bÞ
constraints to preserve the monotonicity condition and non-
intersection nature of rainfall IDF curves. The resulting in which it;T denotes t-hr (or t-min), T-yr rainfall intensity;
rainfall IDF relations would be physically realistic because Fit ðÞ is the cumulative distribution function (CDF) of t-hr
there is no crossover between IDF curves. For illustration, rainfall intensity; T denotes the return period (in year) of t-
the proposed approach is applied to analyze hourly annual hr rainfall intensity with an annual exceedance probability
maximum rainfall data at Zhongli rain gauge in Taiwan of 1/T; lt ; rt are the mean and standard deviation,
with record length of 27-year. respectively, of t-hr rainfall intensity; and KT is the fre-
quency factor (a standardized quantile) corresponding to
annual exceedance probability of 1/T for the probability
2 Methods for establishing rainfall IDF distribution considered. Equation (1a) can be used to
relations determine t-hr rainfall intensity quantile when the CDF of
adopted distribution model has an analytical expression
For many hydrosystem engineering planning, analysis and (e.g., exponential, generalized extreme value, generalized
design, reliable information on rainfall IDF relation is logistic). On the other hand, Eq. (1b) using frequency
essential. Accurate IDF relations allow for cost-effective factor KT can be used for those distributions whose ana-
structural designs and for developing reliable flood man- lytical CDF are not available (e.g., log-normal, Pearson III,
agement systems. log-Pearson III). The value of frequency factor KT is a
function of return period T, skewness coefficient, and dis-
2.1 Conventional approach for establishing tribution model (Chow et al. 1988). References for
rainfall IDF relations hydrologic frequency analysis can be found elsewhere
(Kite 1988; Stedinger et al. 1993; Rao and Hamed 2000;
The conventional approach for establishing rainfall IDF Stedinger 2017).
relations are conducted in two stages. Stage-1 is to perform
frequency analysis on annual maximum rainfall intensity
(or depth) series of selected durations separately to derive
their respective rainfall intensity–frequency relations.

123
2312 Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324

2.1.2 Fitting empirical IDF equations a3 [ 0 and a22  3a1 a3 \0 ð5Þ

Empirical equations of simple form are often used to The TPNT technique only requires the knowledge of the
establish the IDF relations for practical engineering appli- polynomial coefficients and it is therefore a viable tool
cations. The rainfall intensity–duration relation for a par- considering its computation efficiency. Chen and Tung
ticular return period can be fitted by a proper empirical IDF (2003) examined the performance of several methods to
equation. Koutsoyiannis et al. (1998) presents a general- determine the TPNT coefficients and they are: product–
ized form of the empirical equation of rainfall intensity as: moment (PM) method, L-moments (LM) method, least-
aT square (LS) method and Fisher–Cornish asymptotic
it;T ¼ v ð2Þ expansion (FC) method. With the exception of LS method,
ð t þ bT Þ c T
T
all other three methods are moment-based.
where it;T denotes the average rainfall intensity of duration In the context of establishing IDF relations, this study
t-hr and return period T-yr; aT , bT , cT and vT are the adopts LS method for TPNT coefficients determination.
parameters in the empirical IDF equation which depend on This is mainly because using LS method has the capability
return period and their values are generally determined by to establish rainfall IDF relations by simultaneously con-
some curve-fitting optimization scheme. sidering annual maximum rainfalls with multiple durations.
Some widely used empirical IDF equations in literature It also permits the incorporation of suitable constraints to
and engineering design manuals are: Talbot, Bernard, achieve the prescribed objectives in the determination of
Kimijima and Sherman equations (Nhat et al. 2006). By TPNT coefficients. On the other hand, moment-based
letting vT = 1 and cT = 1 one has Talbot equation; bT = 0 methods (e.g., PM and LM methods) can only be used to
and cT = 1, Bernard equation (Bernard 1932); and cT = 1, determine TPNT coefficients for establishing rainfall IDF
Kimijima equation. In this study, the Sherman equation relations for each duration separately. Furthermore,
(Sherman 1931) is utilized to demonstrate the proposed moment-based methods have no control over the solution
method for establishing the rainfall intensity–frequency to comply monotonicity condition that is essential for
relations: establishing valid rainfall IDF relations.
aT To use TPNT technique to conduct at-site frequency
it;T ¼ ð3Þ
ð t þ bT Þ c T analysis, the connection between TPNT technique and
annual maximum rainfall intensity can be expressed as:

2.2 TPNT-based approach for establishing iobs 2 3


t;T ¼ a0;t þ a1;t zT þ a2;t zT þ a3;t zT ð6Þ
rainfall IDF relations where iobs
t;T is the observed rainfall intensity with t-hr
duration and T-yr return period; a0,t, a1,t, a2,t and a3,t denote
The TPNT technique can be used to establish the multi- the TPNT coefficients corresponding to rainfall of duration
variate distribution function based on the univariate normal t-hr; and zT is the standard normal random variate corre-
transformation of non-normal random variable. An attrac- sponding to return period T-yr having the exceedance
tive feature of the TPNT technique is that it does not probability of 1=T.
require the specification of complete marginal distribution
information to perform normal transformation. 2.2.2 Determining TPNT coefficients by unconstrained LS
(uLS) method
2.2.1 Basic TPNT technique
Consider t-hr annual maximum series (AMS) containing
The TPNT was proposed by Fleishman (1978) by which a random rainfall intensity data of record length Ny years
non-normal random variable is related to the standard  
it;1 ; it;2 ; . . .; it;Ny . The TPNT representation of t-hr rainfall
normal variable through the following transformation
data series can be expressed as:
relation:
iobs 2 3
t;py ¼ a0;t þ a1;t zpy þ a2;t zpy þ a3;t zpy ; for
X ¼ a0 þ a1 Z þ a2 Z 2 þ a3 Z 3 ð4Þ ð7Þ
y ¼ 1; 2; 3; . . .; Ny
where X is the original non-normal random variable; Z is
the standard normal random variable; and a0, a1, a2 and a3 where zp ¼ U1 ð pÞ with U1 ð pÞ being the inverse CDF of
denote the polynomial coefficients. To satisfy one-to-one standard normal with non-exceedance probability of p. For
 
monotonically increasing relations between X and Z, the each observation it;y y¼1;2;...;Ny , the non-exceedance
TPNT coefficients must comply with the following probability py can be estimated by a proper plotting posi-
conditions: tion formula (PPF). In frequency analysis of hydrologic

123
Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324 2313

extremes, a number of PPFs have been used (Cunnane ensure no crossover of resulting IDF curves within the
1978). Herein, the well-known probability-unbiased Wei- frequency range practical to a particular application. In the
bull formula is used to estimate the non-exceedance following section, an optimization model with constrained
probability py for the rainfall observation in the y-th year LS method is presented that incorporates monotonicity and
iobs
t;y as: no-crossover conditions in the determination of TPNT
h i coefficients.
Rank iobs
t;y
py ¼ ð8Þ 2.2.3 Determining TPNT coefficients by constrained LS
Ny þ 1
(cLS) method
By LS criterion, the unconstrained LS (uLS) solution for
the vector of t-hr polynomial coefficients To comply the monotonicity and no-crossover conditions
 0 0
at ¼ a0;t ; a1;t ; a2;t ; a3;t , with ‘‘ ’’ standing for the trans- of IDF curves, the multi-duration TPNT coefficients can be
pose of a vector or matrix, in Eq. (7) can be determined by determined by solving the constrained optimization model
(Montgomery et al. 2012): stated below:
 0
1 Ny
Nt X
X
at ¼ Zp Z p Zp iobs
t ð9Þ Minimize : SSEcLS ¼ e2k;y ð12Þ
 k¼1 y¼1
in which Zp ¼ 1; zp ; z2p ; z3p is a n 9 4 matrix containing
 0 Subject to:
column vector zrp ¼ zrp1 ; zrp2 ; . . .; zrpNy for r = 1, 2, 3.
(a) Fitting observed rainfall intensity data:
To extend the above univariate single-duration AMS of 
rainfall intensity data to rainfall data with Nt durations acLS cLS cLS 2 cLS 3 obs
0;k þ a1;k zpy þ a2;k zpy þ a3;k zpy þ ek;y ¼ ik;y
simultaneously, the uLS solution for polynomial coeffi- for k ¼ 1  Nt ; y ¼ 1  Ny
cients of all considered durations can be obtained by
ð13Þ
solving the following optimization problem:
Ny
Nt X
X (b) Satisfying monotonicity condition for rainfall inten-
Minimize : SSEuLS ¼ e2k;y ð10Þ sity–frequency of the k-th duration:
k¼1 y¼1
 acLS
3;k [ 0; for k ¼ 1  Nt ð14aÞ
Subject to : auLS
0;k þ auLS
1;k zpy þ auLS 2
2;k zpy þ auLS 3
3;k zpy þ ek;y ¼ iobs
k;y  2
acLS
2;k 3acLS cLS
1;k a3;k \0; for k ¼ 1  Nt ð14bÞ
for k ¼ 1  Nt ; y ¼ 1  Ny
ð11Þ (c) Intersections avoidance constraints:
  
where SSEuLS denotes the total sum of squared errors acLS  a cLS
0 þ a cLS
 a cLS
0 z T þ a cLS
 a cLS
0 z
2
0;k 0;k 1;k 1;k 2;k 2;k T
(residuals) by the uLS method; Nt ; Ny are, respectively, the 
cLS cLS 3
total number of durations and years in rainfall data set; k, y þ a3;k  a3;k0 zT [ 0;
are indicators for duration and year, respectively; ek;y is the
for two durations tk \tk0
discrepancy (or residual) between observed rainfall inten-
sity of the k-th duration and y-th year; and ð15Þ
auLS uLS uLS uLS
0;k ; a1;k ; a2;k ; a3;k are TPNT coefficients of the k-th where SSEcLS denotes the total sum of squared errors
duration estimated by the uLS approach. by constrained LS (cLS) method; T is the upper
By considering rainfall data of multiple durations in the limit of selected rainfall return period below which
above optimization model, the best set of TPNT coeffi- no crossover of IDF curves is permitted to occur; zT
cients to establish rainfall IDF relations for all considered is the standard normal quantile obtainable from

durations can be simultaneously determined. The optimal
U1 1  T1 ; and all other notations are defined in
uLS solution is obtained by trading-off errors of estimated
rainfall intensity among different durations, not just within Eqs. (10)–(11). Comparing with uLS formulation,
one single duration as the univariate case. It should be the cLS model has additional 2Nt monotonicity
pointed out that the above optimization framework for constraints and Nt - 1 no-crossover constraints. The
simultaneous determination of TPNT coefficients by min- optimization method used for solving the above
imizing SSEuLS does not guarantee that the results will constrained LS model, Eqs. (12)–(15), is the
produce rainfall intensity–frequency relations that satisfy sequential quadratic programming technique which
monotonicity condition for all durations. Also, it does not is described in the next sub-section.

123
2314 Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324

2.2.4 Solving TPNT/LS-based models for establishing IDF of the data series but may vary significantly in the two
relations ends. It has been shown that the Weibull (1939) PPF pro-
vides the best estimate for the underlying non-exceedance
The TPNT/LS-based optimization models presented above probability and the superiority of the Weibull formula gets
is solved by the sequential quadratic programming (SQP) more pronounced with decreasing number of observations
algorithm developed by Wilson (1963). The SQP solves an (Makkonen et al. 2013). Hence, the main PPF in TPNT
optimization problem having a nonlinear objective function technique with LS method used in this study is Weibull
with both linear and nonlinear constraints. The SQP tackles (1939). Several other plotting position are also used in
a nonlinear optimization problem by successively finding engineering applications (Cunnane 1978). The effects of
the approximated solution to the quadratic programming different PPFs on TPNT-based rainfall IDF determination
representation of the original problem. The approximated are examined in Sect. 3.4.
solution hopefully would improve by iteratively solving the
(b) Determine TPNT coefficients by uLS method for
quadratic programming problem. Some useful properties of
establishing IDF relations—
SQP algorithm are elaborated by Boggs and Tolle (1995).
If the TPNT coefficients by uLS passed the monotonicity
In this study, a subroutine sqp.m in Matlab is implemented
and no-intersection conditions, then an empirical IDF
to solve both TPNT/uLS model [Eqs. (10)–(11)] and
equation is selected to fit TPNT/uLS-based IDF relations.
TPNT/cLS model [Eqs. (12)–(15)]. Figure 1 shows the
Otherwise, cLS method is used to determine TPNT coef-
overall framework to establish TPNT-based IDF relations
ficients. Note that if TPNT/uLS method produces coeffi-
by the two LS methods. The contents of each block are
cients that satisfy monotonicity and no-intersection
briefly described below.
conditions, the constraints for TPNT/cLS method (i.e.,
(a) Prepare rainfall data— Eqs 13–15) will be redundant.
Based on the annual maximum rainfall observations of (c) Determine TPNT coefficients by cLS method for
different durations, arrange data of each duration in establishing IDF relations—
ascending order and assign a probability to each observa- If TPNT/uLS failed to produce IDF relations that comply
tion by a suitable PPF. The general form of PPF is: with monotonicity and/or no-intersection conditions, the
  ma cLS method is used by solving Eqs. (12)–(15).
Pr X  xðmÞ ¼ ð16Þ
nþ1b
(d) Fitting TPNT/LS IDF relations by empirical IDF
where m is the rank of the annual maximum rainfall equation—
intensity in ascending order; a C 0 and b C 0 denote the Once the TPNT-based IDF relations (by either uLS or cLS
constants which vary with different PPFs; and n is the method) pass the monotonicity and no-intersection condi-
number of observations in data series. In fact, all of the tions, they are used to fit a selected empirical IDF equation.
PPFs give similar cumulative probabilities near the center For example, using Sherman’s IDF equation [Eq. (3)], its

(b) Apply uLS to determine TPNT


(a) Prepare rainfall data
coefficients.

Pass
(c) Apply cLS to No
determine TPNT monotonicity & no-crossover
coefficients conditions?

Yes

(d) Fit selected empirical IDF equation

Fig. 1 Flowchart of TPNT-based optimization model for establishing IDF curves

123
Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324 2315

three parameters can be determined, according to the LS squared (uLS) method, i.e., Eqs (10), (11), and the results
principle, as: are listed in Table 1. The TPNT/uLS coefficients of two
NT X
X Nt shortest durations (i.e., 1- and 2-h) do not satisfy the
Minimize : e2tk ;Tj ð17Þ monotonicity conditions [Eq. (5)] over the entire proba-
j¼1 k¼1 bility range. The TPNT/uLS coefficients for 1-h rainfall can
a Tj TPNT
satisfy monotonicity for return period 1.07-yr  T 
Subject to : c þ etk ;Tj ¼ itk ;Tj ; 230-yr whereas for 2-h rainfall, the valid return period
ðtk þ bTj Þ Tj ð18Þ
k ¼ 1  Nt ; j ¼ 1  NT range is T [ 1.01-yr. It appears that IDF relations derived
by TPNT/uLS herein are applicable for estimating rainfall
where aTj , bTj and cTj denote the parameters in the Sherman intensity in frequency range of practical engineering
IDF equation which depend on the j-th return period Tj; tk problems.
is the k-th duration; and etk ;Tj is the difference between the Figure 3 shows the TPNT/uLS-based rainfall IDF curves
TPNT-based IDF values iTPNT tk ;Tj obtained by either uncon-
at the gauge site. The intersection of IDF curves occur for
strained LS method (Eqs. 10, 11) or constrained one high return periods (T [ 200-yr) when rainfall duration is
(Eqs. 12–15) and those estimated by the Sherman IDF short (t  1-h). This is mainly caused by non-mono-
equation. tonicity of 1-h intensity–frequency relation. It implies that
the crossover problem of IDF curves can be partially
resolved by constraining TPNT coefficients to satisfy the
3 Application monotonicity condition. In addition to intersection of IDF
curves, the undulation appearance of some curves (espe-
In this section, application of TPNT-based approaches is cially occur in high return periods) indicate potential
made to establish at-site rainfall IDF relations using hourly instability of estimating rainfall intensity along the curves.
annual maximum rainfall intensity of various durations at a To get some clue as to IDF curves intersect at t = 1-h,
rain gauge at Zhongli City (Station number C0C520) in values of sample coefficient of variation (Cv) are com-
Taoyuan, Taiwan. The data have record period of 27 years puted. According to the empirical observations by Porras
during 1988–2015, within which 1992 data are not used and Porras (2001) that Cv values should be relatively
due to gauge malfunctioned (i.e., Ny= 27) much of the constant over the range of duration for no-crossover IDF
time. Results of the application are used to examine the curves. At the concerned raingauge, the Cv values corre-
performance of the proposed TPNT-based approaches for sponding to 1-h and 2-h rainfalls are 0.325 which happens
establishing rainfall IDF relations. Ten storm durations to be the lowest with 16.4% lower than the averaged Cv
(i.e., Nt= 10) considered are: 1-, 2-, 5-, 12-, 18-, 24-, 48-, across the 10 durations considered. Almost all other dura-
72-, 96- and 120-h. tions have Cv values larger than averaged Cv with variation
By the conventional frequency analysis approach, the less than 10%. This is indicative that IDF curves could
first four sample L-moments for each duration were cal- intersect around storm duration of 1- and 2-h.
culated for estimating parameters of eight candidate dis- According to the TPNT/uLS-based IDF relations, the
tributions (normal, lognormal, Pearson 3, log-Pearson 3, optimal Sherman IDF model parameters in Eq. (3) are
Gumbel, generalized extreme value, generalized Pareto, sought to fit Fig. 3. The values of parameters determined
and generalized logistic). L-moment ratio diagram was for the nine return periods (NT= 9) from T = 2–1000-yr are
used to identify the best-fit distribution (Vogel and Fen- listed in Table 2 and the corresponding fitted Sherman IDF
nessey 1993). Figure 2 shows the results of rainfall IDF curves are shown in Fig. 4. Clearly, the wavy appearance
relations using the individual best-fit distribution model for of IDF curves in Fig. 3 is removed. However, some
each durations considered. It is clearly that the intensity– Sherman-based IDF curves remain intersected at the longer
duration curves start to undulate for rainfall with high durations.
return period (say, [ 100 yr) and the curves appear to
converge with possible crossing over as the duration gets 3.2 Establishing TPNT-based IDF
shorter. with constraints

3.1 Establishing TPNT-based IDF Establishing physically realistic IDF relations can be cast
without constraints into a constrained optimization framework. Constraints on
the TPNT-based IDF model are imposed to preserve
Using Weibull PPF, the TPNT coefficients for all durations monotonicity [Eq. (14a–b)] and to avoid intersection
are determined simultaneously by the unconstrained least- [Eq. (15)].

123
2316 Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324

Fig. 2 IDF relations at Zhongli


rain gauge from L-moment-
based conventional frequency
analysis approach using best-fit
distribution models

Table 1 TPNT/uLS coefficients


TPNT coeff. Duration (hr)
(with Weibull PPF) at Zhongli
raingauge 1 2 6 12 18 24 48 72 96 120

a0 52.20 35.92 18.01 10.85 8.53 7.10 4.42 3.23 2.72 2.31
a1 22.58 13.04 6.49 2.32 1.87 1.54 1.54 1.24 1.04 0.88
a2 3.25 2.96 2.55 2.26 1.40 1.00 0.78 0.49 0.28 0.21
a3 - 1.92 0.10 0.86 1.55 1.00 0.69 0.34 0.19 0.10 0.07
SSE = 293.95 (mm/hr)2; RMSE = 1.13 (mm/hr)

Fig. 3 TPNT/uLS-based IDF


relations (with Weibull PPF) at
Zhongli rain gauge

Table 2 Parameters in Sherman


IDF eq. para. Return period, T (year)
IDF Eq. fitted to TPNT/uLS-
based IDF relations 2 5 10 20 50 100 200 500 1000

aT 68.62 92.55 116.52 126.69 302.49 3298.70 5644.68 8440.76 10,342.27


bT 0.50 0.45 0.69 0.74 3.34 13.37 17.99 24.32 28.37
cT 0.70 0.72 0.68 0.63 0.80 1.33 1.37 1.36 1.35
SSE = 908.55 (mm/hr)2; RMSE = 3.80 (mm/hr)

123
Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324 2317

Fig. 4 TPNT/uLS-based IDF


curves fitted by the Sherman
IDF equation

Table 3 TPNT/cLS coefficients


TPNT coeff. Duration (hr)
(Weibull PPF) at Zhongli
raingauge 1 2 6 12 18 24 48 72 96 120

a0 52.39 35.94 18.01 10.85 8.53 7.10 4.42 3.23 2.72 2.31
a1 18.68 12.81 6.49 2.32 1.87 1.54 1.54 1.24 1.04 0.88
a2 3.02 2.94 2.55 2.26 1.40 1.00 0.78 0.49 0.28 0.21
a3 0.16 0.23 0.86 1.55 1.00 0.69 0.34 0.19 0.10 0.07
2
SSE = 399.76 (mm/hr) ; RMSE = 1.32 (mm/hr)

Using Weibull PPF, Table 3 lists the TPNT coefficients IDF equation with the best-fit parameters are listed in
for each duration obtained by the constrained least-square Table 4 and the corresponding IDF curves are shown in
(cLS) approach. Figure 5 shows the rainfall IDF curves Fig. 6. The undulation behavior of IDF curves in Fig. 5 is
obtained from the TPNT/cLS-based approach. It is clearly smoothed out. Comparing with Figs. 2 and 5, the TPNT/
that the intersection of IDF curves in high return period cLS-based approach yields more reasonable rainfall IDF/
range and short duration shown in Figs. 2 and 3 is cir- DDF relations than those from the conventional frequency
cumvented. However, the undulation appearance remains. analysis approach.
Based on the TPNT/cLS-based IDF relations, the Sherman

Fig. 5 TPNT/cLS-based IDF


relations (with Weibull PPF) at
Zhongli raingauge

123
2318 Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324

Table 4 Parameters in Sherman


IDF eq. para. Return period, T (year)
IDF Eq. fitted to TPNT/cLS-
based IDF relations 2 5 10 20 50 100 200 500 1000

aT 66.82 93.71 110.54 132.71 177.97 256.25 427.79 1322.19 4574.42


bT 0.41 0.51 0.57 0.78 1.33 2.37 4.27 9.56 16.37
cT 0.71 0.70 0.67 0.65 0.65 0.70 0.78 1.00 1.24
2
SSE = 1547.12 (mm/hr) ; RMSE = 4.96 (mm/hr)

Fig. 6 TPNT/cLS-based IDF


curves fitted by the Sherman
equation

3.3 Comparison of TPNT/uLS-based and TPNT/ that a0 is only affected by the non-intersection constraints
cLS-based IDF relations (Eq. 15); a1 and a2 are affected by both nonlinear mono-
tonicity constraints (Eq. 14b) and no-intersection constraints
Figures 7a, b show the variation of TPNT coefficients with Eq. (15); and a3 is affected by all constraints (Eqs. 14, 15).
storm duration by the cLS and uLS approaches. Comparing Hence, TPNT coefficients a3 and a0, respectively, have the
the TPNT coefficients by the two LS approaches, one notices maximum and minimum relative difference (RD) [defined by

Fig. 7 Variation of TPNT coefficients with duration by uLS and cLS approaches (with Weibull PPF) at Zhongli raingauge. a By uLS method,
b By cLS method

123
Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324 2319

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
,
Eq. (19)] and this implies that the TPNT/cLS-based u Ny  2
uX
approach has the most restriction on a3 and the least re- Duration k : RMSDk;m ¼ t ^
ik;y;m  ik;y obs
DFk ;
striction on a0. This is reflected in Fig. 8, which shows that a3 y¼1
has the largest amplitude of variation between the two LS k ¼ 1; . . .; Nt ; m ¼ uLS; cLS ð21Þ
approaches whereas a0 has the smallest variation.
acLS  auLS where iobs
k;y is the observed rainfall intensity of the k-th
r r
RD ¼  100%; for r ¼ 0; 1; 2 and 3 duration in year-y; ^ik;y;m denotes the TPNT-based estima-
auLS
r
tion of rainfall intensity of the k-th duration in year-y by
ð19Þ
method m (m = uLS or cLS); Nt, Ny are number of rainfall
where auLS
r and acLS
r denote the TPNT coefficients obtained durations and record length, respectively (Nt = 10 and
from uLS and cLS approaches, respectively. Ny = 27 in this example application); and DFo, DFk denote
Figures 9a, b compare TPNT-fitted rainfall intensity–fre- the degrees of freedom for computing the overall and
quency relations for 1- and 2-h rainfall intensity by the two LS individual duration-based performance indicator, each
approaches in which monotonicity condition is not satisfied by respectively equals to Nt (Ny - 4) and Ny - 4.
the uLS method. For 1-h rainfall, Fig. 9a shows that the esti- As shown in Table 5 that the values of RMSDo by uLS
mated rainfall intensity by the two LS methods has significant and cLS are close with the uLS method yielding smaller
difference. As for 2-h rainfall (see Fig. 9b) and other longer value as expected. Also shown in Table 5 is that the values
durations, the two LS approaches yield practically the same of duration-based RMSDk are almost identical for rainfall
estimated intensity–frequency relations. This is also evi- duration t  6-h by the LS approaches. This explains that
denced by almost identical TPNT coefficients by the two LS major discrepancy between uLS and cLS approaches is
approaches for the duration t  6-h (see Tables 1 and 3). contributed from the 1-h and 2-h TPNT-based IDF rela-
Table 5 shows the comparison of the goodness-of-fit tions, especially from the 1-h rainfall.
assessment in terms of the root-mean-squared differences To have a close examination of the difference in TPNT-
(RMSD) of TPNT-based intensity–frequency relations for based rainfall IDF, Fig. 10 reveals that the TPNT-based
the ten rainfall durations considered. The RMSD is defined intensity–duration relations obtained by the two LS
for overall performance with all rainfall durations com- approaches differ relatively small when T  50-yr. On the
bined by the LS approach m (RMSDo,m) and individual other hand, the figure shows that, when T [ 50-yr, the
duration by LS approach m (RMSDk,m), respectively, as: difference of the two TPNT-based intensity–duration
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
, relations becomes more pronounced with rainfall duration
u Nt Ny  2
uX X around 1-h. Figure 10 also reveals that the 500-yr inten-
Overall : RMSDo;m ¼ t ^ik;y;m  i obs
DFo ;
k;y sity–duration curve derived by the uLS method has a clear
k¼1 y¼1
tendency to intersect with the 50-yr curve in the region
m ¼ uLS; cLS ð20Þ

Fig. 8 Relative difference in


Weibull-based TPNT
coefficients by uLS and cLS
approaches

123
2320 Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324

Fig. 9 Comparison of fitted 1-h and 2-h rainfall intensity that do not satisfy monotonicity by TPNT using uLS and cLS approaches. a 1-h rainfall,
b 2-h rainfall

Table 5 Comparison of
m RMSDo,m (mm/hr)* RMSDk,m (mm/hr)*
goodness-of-fit assessment of
TPNT-based IDF model k=1 k=2 k=3 k=4 k=5 k=6 k=7 k=8 k=9 k = 10
1-h 2-h 6-h 12-h 18-h 24-h 48-h 72-h 96-h 120-h

uLS 1.13 2.69 1.19 1.28 1.29 0.59 0.53 0.32 0.23 0.23 0.12
cLS 1.32 3.44 1.20 1.28 1.29 0.59 0.53 0.32 0.23 0.23 0.12
*RMSDo,m Overall root-mean-squared differences for all data by LS method m; RMSDk,m Root-mean-
squared differences for rainfall data of the k-th duration by LS method m

Fig. 10 Comparison of TPNT-


based intensity–duration
relations by uLS and cLS
approaches (with Weibull PPF)
at Zhongli raingauge

where storm duration is shorter than 1-h. If this trend 3.4 Effects of plotting position formula on IDF
indeed occurred, the rainfall IDF relations would not relations
comply the physical reality.
Referring to Sect. 2, the estimation of TPNT coefficients
by the LS approaches requires the stipulation of probability
by adopting a PPF. As the result, the rainfall intensity–

123
Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324 2321

Fig. 11 Comparison of TPNT/cLS-based IDF relations by duration using different PPFs. a t = 1-h, b t = 2-h, c t = 6-h, d t = 72-h

frequency relations for a selected duration defined by   m  0:35


Hosking ð1985Þ : Pr X  xðmÞ ¼ ð25Þ
TPNT is affected by the chosen plotting position. In this n
section, the effect of probability PPF on TPNT-based IDF
Different probability PPFs produce different values of
relations is investigated. The Sherman IDF equation is used
TPNT coefficients rendering different IDF relations.
to fit the TPNT/cLS-based IDF. Among the various PPFs,
Figures 11 and 12 show the comparison of rainfall
four PPFs are selected herein to derive TPNT-based rain-
intensity–frequency relations of varying storm duration
fall IDF. They are: Weibull (1939), Hosking et al. (1985),
under the four selected PPFs. With sample size of n = 27, it
Blom (1958) and Cunnane (1978) and their equations are
can be observed that all four PPFs yield very close results
listed below:
when return period is less than 5-yr. As the return period
  m
Weibull ð1939Þ : Pr X  xðmÞ ¼ ð22Þ increases, the difference in rainfall intensity under the
nþ1 Weibull and Hosking formulae becomes more significant,
  m  0:375 whereas the Blom and Cunnane formulae yield practically
Blom ð1958Þ : Pr X  xðmÞ ¼ ð23Þ
n þ 0:25 the same IDF curves. Among the four PPFs considered, the
  m  0:4 Weibull formula yields the highest rainfall intensity of all
Cunnane ð1978Þ : Pr X  xðmÞ ¼ ð24Þ durations whereas the Hosking formula results in the
n þ 0:2
smallest values for the rainfall data.

123
2322 Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324

Fig. 12 Comparison of TPNT/cLS-based IDF relations by duration under different PPFs fitted by the Sherman IDF equation. a t = 1-h, b t = 2-h,
c t = 6-h, d t = 72-h

Table 6 Assessment of goodness-of-fit of different PPF of TPNT/cLS-based IDF model


Plotting positions formula RMSDo (mm/hr) RMSDk (mm/hr)
k=1 k=2 k=3 k=4 k=5 k=6 k=7 k=8 k=9 k = 10
1-h 2-h 6-h 12-h 18-h 24-h 48-h 72-h 96-h 120-h

Weibull 1.32 3.44 1.20 1.28 1.29 0.59 0.53 0.32 0.23 0.23 0.12
Hosking 1.54 4.27 1.55 1.25 0.91 0.46 0.48 0.22 0.16 0.20 0.10
Blom 1.42 3.85 1.32 1.24 1.14 0.53 0.51 0.27 0.20 0.22 0.11
Cunnane 1.43 3.89 1.34 1.24 1.12 0.53 0.51 0.27 0.19 0.22 0.11

Table 6 lists the values of RMSDo and RMSDk of TPNT/ Interestingly, the ranking of the four PPFs varies with the
cLS-based IDF relations according to the selected PPFs. rainfall duration. For short durations of 1-h and 2-h, the
For the overall fitting by RMSDo, the best PPF in this study four PPFs rank the same as the overall performance. On the
is the Weibull, followed by Blom, Cunnane, and Hosking. other hand, for intermediate and longer durations (t [ 6-

123
Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324 2323

h) the order of ranking reversed, i.e., the Hosking is the four PPFs yield very close results when return period is less
best, followed by Cunnane, Blom, and Weibull. However, than 5-yr. As the return period increases, the difference in
the difference in RMSDk is quite insignificant for duration t rainfall intensity under the Weibull and Hosking formulae
[ 6-h and, therefore, the four PPFs could be considered becomes more significant, whereas the Blom and Cunnane
comparable for practical purpose. If a probability-unbiased formulae yield practically the same IDF curves. Among the
PPF is preferred for the data in a study, then the Weibull four PPFs considered, the Weibull formula yields the
formula would be the best choice (Makkonen et al. 2013). highest rainfall intensity of all durations whereas the
Hosking formula results in the smallest for the rainfall data
analyzed. There have been some debates about the use of
4 Summary and conclusions plotting position formula. Weibull formula has been argued
and proved to be the best choice on the ground being
Rainfall IDF relations are commonly used in hydrosystems probability-unbiased (Makkonen et al. 2013). In the con-
analysis, planning, and design. Such relations are typically text of IDF derivation under the proposed TPNT/LS
derived through frequency analysis of annual maximum framework, a more extensive numerical experiment is
rainfall data with different durations. Conventional necessary.
approach is to adopt parametric distribution models to fit The TPNT-based framework presented in this paper
rainfall data of individual duration separately without deals with extreme rainfall data that is stationarity. In light
considering their correlations. It is not uncommon to pro- of imminence of global warming induced climate change,
duce physically unrealistic IDF relations in which rainfall many studies show that extreme precipitations around the
intensity–duration curves of different frequency may globe are getting intensified (Asadieh and Krakauer 2015).
intersect and/or have wavy shape. This feature is especially This implies that the long-term behavior of extreme rain-
likely to occur when the rainfall record is relatively short falls is becoming non-stationary. Consequently, the rainfall
and the IDF curves are extend to the range of high return IDF relations would be subject to non-stationarity. How-
period beyond the record length. ever, the present form of TPNT/LS-based approach cannot
This study addresses the issues of crossover and undu- explicitly account for non-stationarity in the rainfall data
lation of rainfall IDF curves by applying the third-order without overhauling the entire procedure. A viable means
polynomial normal transform (TPNT), in conjunction with to explicitly account for non-stationarity is to bring time-
the constrained least square (LS) method. The TPNT has dependent statistical features of non-stationary extreme
the advantages of: (1) not having to adopt parametric dis- rainfalls into the TPNT formulation, similar to what have
tributions of particular form; and (2) being able to estab- been done by Sarhadi and Soulis (2017) and Yan et al.
lishing IDF relations with simultaneous consideration of (2017) who express distributional parameters adopted in
multi-duration rainfall data when used in the constrained rainfall frequency analysis as function of time.
LS framework. However, like any statistical analysis
method, the TPNT is not immune from the sampling error Acknowledgements This study is supported by the Joint Research
under the National Research Foundation (Korea)-Ministry of Science
induced by the use of limited rainfall data. The two prob- & Technology (Taiwan) Cooperative Program (MOST 105–2923-E-
lematic features of TPNT-based IDF curves can be easily 009-004-MY2). All data used in this paper is properly cited and
taken care of in the LS framework by imposing the referred to in the reference list.
monotonicity condition to preserve probabilistic relation
and constraining no-crossover in the IDF curves within the
desired range of return period. The undulation of rainfall References
intensity–duration curves of different frequencies can be
Akan AO, Houghtalen RJ (2003) Urban hydrology, hydraulics, and
removed by using a proper empirical equation to fit the IDF stormwater quality. Wiley, Hoboken
relations obtained by the TPNT/LS-based approach. Asadieh B, Krakauer N (2015) Global trends in extreme precipitation:
For illustration, the proposed analysis framework is climate models versus observations. Hydrol Earth Syst Sci
implemented to establish rainfall IDF relations at a rain- 19:877–891
Asikoglu OL, Benzeden E (2014) Simple generalization approach for
gauge in Taiwan with record length of 27 years. The IDF intensity–duration–frequency relationships. Hydrol Process
relations obtained by using TPNT/cLS approach appear to 28(3):1114–1123
be more reasonable than those by the conventional fre- Bernard M (1932) Formulas for rainfall intensities of long duration.
quency analysis. Furthermore, to examine the effect of Trans ASCE 96(1):592–606
Blom G (1958) Statistical estimates and transformed beta-variates.
plotting position formula (PPF) on the derived IDF rela- Wiley, New York
tions, four PPFs commonly used in hydrology (i.e., Wei- Boggs PT, Tolle JW (1995) Sequential quadratic programming. Acta
bull, Hosking, Cunnane, and Blom) are used in the Numer 4:1–51
proposed TPNT/LS-based approach. It was found that all

123
2324 Stochastic Environmental Research and Risk Assessment (2018) 32:2309–2324

Cai D, Shi D, Chen J (2013) Probabilistic load flow computation with Montgomery DC, Peck EA, Vining GG (2012) Introduction to linear
polynomial normal transformation and Latin hypercube sam- regression analysis, 5th edn. Wiley, New York
pling. IET Gener Transm Distrib 7(5):474–482 Nhat LM, Tachikawa Y, Sayama T, Takara K (2006) Derivation of
Chen XY, Tung YK (2003) Investigation of polynomial normal rainfall intensity–duration–frequency relationships for short-
transform. Struct Saf 25(4):423–445 duration rainfall from daily data. In: Proceedings, international
Chow VT, Maidment DR, Mays LW (1988) Applied hydrology. symposium on managing water supply for growing demand,
McGraw-Hill Book Company, New York technical document in hydrology, vol 6, pp 89–96
Ciapessoni E, Cirio D, Massucco S, Morini A, Pitto A, Silvestro F Porras PJ Sr, Porras PJ Jr (2001) New perspective on rainfall
(2017) Risk-based dynamic security assessment for power frequency curves. J Hydrol Eng 6(1):82–85
system operation and operational planning. Energies Rao AR, Hamed KH (2000) Flood frequency analysis. CRC
10(4):475–490 Publications, New York
Cleveland, T.G., Herrmann, G.R., Tay, C.C., Neale, C.M., Schwarz, Sahoo SN, Sreeja P (2015) Development of flood inundation maps
M.R., and Asquith, W.H. (2015). New rainfall coefficients— and quantification of flood risk in an urban catchment of
including tools for estimation of intensity and hyetographs in Brahmaputra river. J Risk Uncertain Eng Syst Part A Civ Eng.
Texas, final report. Texas Department of Transportation. https://doi.org/10.1061/ajrua6.0000822
Research Project Number 0-6824 Sarhadi A, Soulis ED (2017) Time-varying extreme rainfall intensity–
Cunnane C (1978) Unbiased plotting positions—a review. J Hydrol duration–frequency curves in a changing climate. Geophys Res
37:205–222 Lett. https://doi.org/10.1002/2016gl072201
Demirtas H, Hedeker D, Mermelstein RJ (2012) Simulation of Sherman CW (1931) Frequency and intensity of excessive rainfalls at
massive public health data by power polynomials. Stat Med Boston, Massachusetts. Trans ASCE 95:951–960
31(27):3337–3346 Singh VP, Strupczewski WG (2002) On the status of flood frequency
Doong DJ, Lo WC, Vojinovic Z, Lee WL, Lee SP (2016) analysis. Hydrol Process 16:3737–3740
Development of new generation of flood inundation maps—a Stedinger JR (2017) Ch.76—flood frequency analysis. In: Singh VP
case study of the coastal city of Tainan, Taiwan. Water (ed) Handbook of applied hydrology, 2nd edn. McGraw-Hill,
8:521–540. https://doi.org/10.3390/w8110521 New York
Fleishman AL (1978) A method for simulating non-normal distribu- Stedinger JR, Vogel RM, Foufoula-Georgiou E (1993) Ch.18—
tions. Psychometrika 43(4):521–532 frequency analysis of extreme events. In: Maidment D (ed)
Haktanir T (2003) Divergence criteria in extreme rainfall series Handbook of hydrology. McGraw-Hill Inc, New York
frequency analyses. Hydrol Sci J 48(6):917–937 Sun SA, Djordjević S, Khu ST (2010) Decision making in flood risk
Headrick TC (2010) Statistical simulation: power method polynomi- based storm sewer network design. Water Sci Technol
als and other transformations. Chapman & Hall, Boca Raton 64(1):247–254
Hong HP (1998) Application of polynomial transformation to Vogel RM, Fennessey NM (1993) L-moment diagrams should replace
normality in structural reliability analysis. Can J Civ Eng product moment diagram. Water Resour Res 296:1745–1752
25(2):241–249 Weibull W (1939) A statistical theory of the strength of materials.
Hosking JRM, Wallis JR, Wood EF (1985) Estimation of the R Swed Inst Eng Res Proc 151:1–45
generalized extreme value distribution by the method of Wilson RB (1963) A simplicial method for convex programming,
probability-weighted moments. Technometrics 27(3):251–261 PhD thesis, Harvard University, Boston, MA
Kite GW (1988) Frequency and risk analyses in hydrology. Water Xu Y, Tung YK (2009) Constrained scaling approach for design
Resources Publications, Littleton CO rainfall estimation. Stoch Env Res Risk Assess 23(6):697–705
Koutsoyiannis D, Kozonis D, Manetas A (1998) A mathematical Yan L, Xiong L, Guo S, Xu CY, Xia J, Duc T (2017) Comparison of
framework for studying rainfall intensity–duration–frequency four nonstationary hydrologic design methods for changing
relationships. J Hydrol 206:118–135 environment. J Hydrol 551:132–150
Lopez-Lambrantilde A, Fuentes C, González-Sosa E, López-Ramos Yang, H., and Zou, B. (2012). The point estimate method using third-
A, Pliego-Dı́az M, Gómez-Meléndez D, Altamirano-Corro A order polynomial normal transformation technique to solve
(2013) Effect of interception by canopy in the IDF relation in a probabilistic power flow with correlated wind source and load.
semiarid zone. Afr J Agric Res 8(43):5285–5295 In: Proceedings, Asia–Pacific power and energy engineering
Makkonen L, Pajari M, Tikanmaki M (2013) Discussion on ‘‘Plotting conference, Shanghai, pp 1–4
positions for fitting distributions and extreme value analysis’’. Zhao YG, Lu ZH (2007) Fourth moment standardization for structural
Can J Civ Eng 40:130–139 reliability assessment. J Struct Eng ASCE 133(7):916–924

123

Вам также может понравиться