International Journal of Engineering Issues

Vol. 2015, no. 2, pp. 65-72

ISSN: 2458-651X
Copyright Infinity Sciences

Rainfall Frequency Analysis Using L-moments of

Probability Distributions
N. Vivekanandan

Central Water and Power Research Station, Pune, Maharashtra, India

E-mail: [email protected]
Abstract - Frequency analysis of rainfall would enhance the management of water resources applications as well as
the effective utilization of water resources. This can be carried out by fitting of probability distribution to the series
of annual 1-day maximum rainfall data. In this paper, method of L-moments (LMO) is used for determination of
parameters of Generalized Extreme Value (GEV), Gumbel (EV1) and Frechet (EV2) probability distributions. The
adequacy of fitting of probability distribution adopted in rainfall frequency analysis is evaluated by applying
Goodness-of-Fit (GoF) tests viz., Chi-square and Kolmogorov-Smirnov and diagnostic test (using D-index). Based
on GoF and diagnostic test results, the study identifies the EV1 is better suited probability distribution for estimation
of rainfall at Narwar whereas GEV for Banswara.
Keywords: Chi-square; D-index; Generalized Extreme Value; Gumbel; Kolmogorov-Smirnov; L-moments; Rainfall
I. INTRODUCTION
Rainfall is one of the most important input parameter to crop production and its occurrence and distribution is
erratic, temporal and spatial variations in nature. Most of the hydrological events occurring as natural phenomena are
observed only once. One of the important problem in hydrology deals with the interpreting past records of
hydrological event in terms of future probabilities of occurrence [1]. Under these circumstances, Rainfall Frequency
Analysis (RFA) would enable us to determine the expected rainfall for a given return period. This information can be
used to prevent floods and droughts, and applied to planning and designing of water resources related to engineering
such as reservoir design, flood control work and soil and water conservation planning [2].
Out of number of probability distributions, Extreme Value Distributions (EVDs) include Generalized Extreme
Value (GEV), Gumbel (EV1) and Frechet (EV2) distributions are generally applied for RFA. EVDs arise as limiting
distributions for the sample of independent, identically distributed random variables, as the sample size increases.
Generally, Method of Moments (MoM) is used for determination of parameters of the distributions [3]. But, the
MoM is not giving satisfactory results though the method exists for a longer period. It is sometimes difficult to assess
exactly what information about the shape of a distribution is conveyed by its moments of third and higher order; the
numerical values of sample moments particularly when the sample is small, can be very different from those of the
probability distribution from which the sample was drawn; and the estimated parameters of distributions fitted by the
MoM are often less accurate than those obtained by other estimation procedures such as maximum likelihood
method, method of least squares and probability weighted moments. To overcome this, the alternative approach,
namely L-moments (LMO) is discussed in this paper and also used in RFA [4].
In the recent past, number of studies has been carried out by different researchers on adoption of probability
distributions for RFA. Topaloglu [5] reported that the frequency analysis of the largest, or the smallest, of a sequence
of hydrologic events has long been an essential part of the design of hydraulic structures. Guevara [6] carried out
hydrologic analysis using probabilistic approach to estimate engineering design parameters of storms in Venezuela.
Kumar and Chatterjee [7] employed the LMO to define homogenous regions within 13 gauging sites of the north
Brahmaputra region of India. Di Balldassarre et al. [8] used the LMO for regionalization of annual precipitation in
65

N. Vivekanandan / International Journal of Engineering Issues

northern central Italy. Eslamian and Feizi [9] carried out EVA using monthly maximum rainfall for an arid region in
Isfahan Province (Iran) through LMO. Gonzalez and Valdes [10] applied LMO for regionalization of monthly
rainfall in the Jucar River basin. Yurekli et al. [11] found GEV and 3-parameter Log-Normal (LN3) distributions
(using LMO) as the regional distribution functions for the maximum daily rainfall of Cekerec watershed, Turkey.
Gubareva and Gartsman [12] analysed the extreme hydrometeorological characteristics adopting GEV, Generalised
Pareto, LN3 and Pearson distributions through LMO. Badreldin and Feng [13] carried out the regional RFA for the
Luanhe Basin, Hebei-China by using LMO and Cluster Techniques. But there is no general agreement in applying
particular distribution for RFA for any region or country. Moreover, when different distributional models are used
for modelling of rainfall data series, a common problem that arises is how to determine which model fits best for a
given set of data. This can be answered by formal statistical procedures involving Goodness-of-Fit (GoF) and
diagnostic tests; and the results are quantifiable and reliable. GoF tests such as Chi-square (2) and KolmogorovSmirnov (KS) are applied for checking the adequacy of fitting of probability distributions to the recorded rainfall
data. A diagnostic test of D-index is used for the selection of most suitable probability distribution for estimation of
rainfall. The procedures adopted in RFA using probability distributions and computation of GoF and diagnostic tests
are described in the following sections.
II. METHODOLOGY
LMO is analogous to the conventional moments but can be estimated linear combination of order statistics, i.e.,
by L-statistics. LMO is less subject to bias in estimation and approximate their asymptotic normal distribution more
closely in finite samples.
A) Theoretical Description of LMO
Method of LMO is a modification of the probability weighted moments method explored by Hosking and Wallis
[14]. Parameters of the distribution are estimated by equating the sample LMO (l r) with the distribution of LMO (br).
In practice, LMO must be estimated from a finite sample. Let R 1n R 2 n ... R nn be the ordered sample of size n. The
sample LMO is given by:
( 1) r k ( r k )!
bk
2
k 0 ( k! ) ( r k )!
r

l r 1

(1)

where, l r 1 is the r+1th sample moment and b k is an unbiased estimator of k with

( i 1)( i 2 ).....( i k )
R in
i k 1 ( n 1)( n 2 ).....( n k )
n

b k n 1

(2)

The first two sample LMOs are expressed by:

l1 b 0 and l 2 2 b1 b 0

(3)

Table 1 gives the details of quantile function and parameters of EV1, EV2 and GEV distributions (using LMO).
S.No.
1
2

Table1. Quantile function and parameters of probability distributions

Distribution
Quantile function (RT )
Parameters of distribution (using LMO)
EV1
R T ln( ln(1 (1 / T)))
l1 0.5772157 ; l 2 / log 2
(

ln(

ln(
1

(
1
/
T
))))
/
k
EV2
By using the logarithmic transformation of the
R T e
recorded data, parameters of EV1 are initially
obtained by LMO; and used to determine the
parameters of EV2 from e and k=1/(scale
parameter of EV1).
k
GEV
R T (1 ln(1 (1/ T )) ) / k z ( 2 /(3 t 3 ) (ln 2 / ln 3); k 7.8590 z 2.9554 z 2 ;
l 2 k /(1 2 k )(1 k ); l1 (( (1 k ) 1) / k )

66

N. Vivekanandan / International Journal of Engineering Issues

In Table 1, , , k are the location, scale and shape parameters respectively; (or R ), (or SR) and CS (or ) are
the average, standard deviation and coefficient of skewness of the recorded data; P is the probability of exceedance;
1 is the inverse of the standard normal distribution function and 1 Z P ( P 0.135 (1 P) 0.135 ) / 0.1975 ; R T is the
estimated rainfall by probability distributions corresponding to return period (T); sign(k) is 1 depending on the sign
of k [15].
B) Goodness-of-Fit Tests
GoF tests such as 2 and KS are applied for checking the adequacy of fitting of probability distributions to the
series of recorded rainfall data. Theoretical description of GoF tests are as follows:
2 test:
NC (O j (R ) E j (R )) 2

j1

(4)

E j (R )

where, O j ( R ) is the observed frequency value of jth class, E j ( R ) is the expected frequency value of jth class and
NC is the number of frequency classes [16]. The rejection region of 2 statistic at the desired significance level ()
is 2C 12 , NC m 1 . Here, m denotes the number of parameters of the distribution.
KS test:
N

KS Max ( Fe ( R i ) FD (R i ))
i 1

(5)
Here, Fe ( R i ) i /( n 1) is the empirical CDF of R i in which i is the rank assigned to the sample values
arranged in ascending order and FD ( R i ) is the computed CDF of R i [17].
Test criteria: If the computed values of GoF tests statistic given by the distribution are less than that of the
theoretical values at the desired significance level (), then the distribution is found to be acceptable for RFA.
C) Diagnostic Test
The selection of most suitable probability distribution for RFA is performed through D-index test (USWRC,
1981), which is defined as below:
D-index = 1 R R i R *i
6

(6)

i 1

Here, R is the average value of the recorded data whereas R i and R *i are the six highest recorded and
corresponding estimated values by probability distribution. The distribution having the least D-index is considered as
better suited distribution for estimation of rainfall [18].
III. APPLICATION
In this paper, a study was carried out to evaluate the probability distributions adopted for EVA of rainfall for
Narwar and Banswara. The AMR series was extracted from the daily rainfall data recorded at Narwar for the period
1988 to 2004 and Banswara for the period 1969 to 2012 and used for EVA. For Narwar, the data for the missing
years were replaced with the series maximum value of the recorded data based on AERB guidelines and the series
with imputed data used for EVA. For Banswara, there is no missing data in the rainfall series and therefore no data
was imputed. Table 2 gives the descriptive statistics of AMR recorded at Narwar and Banswara rain gauge stations.

67

N. Vivekanandan / International Journal of Engineering Issues

Rain gauge
station
Narwar
Banswara

Table 2. Descriptive statistics of AMR

Statistical parameters (SD: Standard Deviation; CV: Coefficient of Variation)
Mean (mm)
SD (mm)
CV (%)
Skewness
Kurtosis
143.9
76.0
52.8
1.243
0.419
186.3
97.1
52.1
0.572
-0.841

IV. RESULTS AND DISCUSSIONS

The procedures described above for estimating 1-day maximum rainfall have been implemented adopting
computer codes and used in RFA. The program gives the (i) parameters of GEV, EV1 and EV2 distributions (using
LMO); (ii) estimated rainfall for different return periods; and (iii) GoF tests statistic and D-index values.
A) Estimation of Rainfall by Probability Distributions
The parameters of EV1, EV2 and GEV distributions were determined by LMO and used for estimation of
rainfall at Narwar and Banswara sites; and the results are presented in Tables 3 and 4.
Table 3. Estimated 1-day maximum rainfall at Narwar
Return period
Estimated 1-day maximum rainfall (mm)
(year)
EV1
EV2
GEV
2
131.8
117.8
121.4
5
196.9
175.7
182.8
10
240.0
228.9
234.5
20
281.3
294.9
294.4
50
334.8
409.5
390.3
100
374.9
523.8
478.7
200
414.8
669.2
584.1
500
467.5
924.7
755.3
1000
507.4
1180.8
914.3
Table 4. Estimated 1-day maximum rainfall at Banswara
Return period
Estimated 1-day maximum rainfall (mm)
(year)
EV1
EV2
GEV
2
169.5
149.2
169.1
5
259.7
227.9
259.3
10
319.4
301.8
319.4
20
376.7
395.0
377.3
50
450.8
559.7
452.8
100
506.3
726.7
509.7
200
561.7
942.7
566.6
500
634.7
1328.8
642.2
1000
689.9
1722.4
699.6
From Table 3 and 4, it may be noted that the estimated rainfall using EV2 distribution is consistently higher
when compared to the corresponding values of EV1 and GEV distributions for return period 20-year and above.
B) Analysis Based on GoF Tests
The adequacy of fitting of EV1, EV2 and GEV distributions were evaluated by applying GoF tests to the
recorded rainfall data. The GoF tests statistic values were computed by the probability distributions through Eqs. (4)
and (5), and the results are presented in Table 5.

68

N. Vivekanandan / International Journal of Engineering Issues

Table 5. Computed and theoretical values of GoF tests using EV1, EV2 and GEV distributions
GoF
EV1
EV2
GEV
tests Computed Theoretical Computed Theoretical Computed Theoretical
Narwar
2
11.471
5.990
4.882
5.990
7.706
3.840
KS
0.177
0.318
0.130
0.318
0.119
0.318
Banswara
8.091
9.490
14.364
9.490
8.091
7.810
2
KS
0.133
0.205
0.182
0.205
0.132
0.205
By applying the test criteria to the GoF tests results, as given in Table 5, it may be noted that:
i) The 2 test supported the use of EV2 distribution for RFA for Narwar whereas EV1 for Banswara.
ii) The KS test results confirmed that the EV1, EV2 and GEV distributions are acceptable for RFA for Narwar
and Banswara.
C) Analysis Based on Diagnostic Test
For the selection of most suitable distribution for estimation of rainfall, D-index values were computed by EV1,
EV2 and GEV distributions through Eq. (6) and the results are presented in Table 6.
Table 6. D-index values of EV1, EV2 and GEV distributions
Rain gauge station
EV1
EV2
GEV
Narwar
1.492
1.523
1.463
Banswara
0.523
1.501
0.536
From the diagnostic test results, as given in Table 6, it may be observed that:
i) The D-index values of GEV for Narwar and EV1 for Banswara were found as minimum when compared to
the corresponding values of other probability distributions.
ii) 2 test results didnt support use of GEV distribution for RFA for Narwar though the D-index value of the
distribution is observed as minimum.
iii) There is no much variation between the D-index values computed from EV1 and GEV distributions for
Banswara.
From the above, it may be noted that the quantitative assessment made through GoF and diagnostic tests are
inconclusive. Therefore, the selection of most suitable distribution for estimation of rainfall was made through
qualitative assessment using probability plots, as presented in Figures 1 and 2.
Figures 1 and 2, it can be seen that the fitted curves using EV1 (for Narwar) and GEV (for Banswara)
distribution are in the form of linear and very close to the line of agreement with the recorded data. Based on
qualitative and quantitative assessment, it was identified that the EV1 is better suited probability distribution for
estimation of rainfall for Narwar whereas GEV for Banswara. By considering the design-life of the structure over the
entire intended economic life time, the study recommended that the 1000-year return period estimated rainfall of
about 507 mm for Narwar (using EV1) and 700 mm for Banswara (using GEV) could be used for design purposes.

69

N. Vivekanandan / International Journal of Engineering Issues

Figure 1. Plots of recorded and estimated rainfall using EV1, EV2 and GEV distributions for Narwar

Figure 2. Plots of recorded and estimated rainfall using EV1, EV2 and GEV distributions for Banswara

70

N. Vivekanandan / International Journal of Engineering Issues

V. CONCLUSIONS
The paper describes the study on evaluation of EV1, EV2 and GEV distributions for RFA for Narwar and
Banswara through GoF and diagnostic tests. The following conclusions are drawn from the study:
i) For return period 20-year and above, it is observed that the estimated rainfall by EV2 (MLM) is
consistently higher than the corresponding values of EV1 and GEV distributions for Narwar and Banswara.
ii) The 2 test results support the use of EV2 distribution for EVA of rainfall for Narwar whereas EV1 for
Banswara.
iii) The KS test results confirm the EV1, EV2 and GEV distributions are found to be acceptable for RFA for
Narwar and Banswara.
iv) Based on qualitative and quantitative assessment, the study identifies the EV1 distribution is better suited
for estimation of rainfall at Narwar whereas GEV for Banswara.
v) The study suggests the 1000-year return period estimated rainfall of about 507 mm for Narwar (using EV1)
and 700 mm for Banswara (using GEV) could be used as the design parameter for planning and design of
hydraulic structures.
ACKNOWLEDGEMENTS
The author is grateful to the Director, Central Water and Power Research Station, Pune, for providing the
research facilities to carry out the study. The author is thankful to M/s Nuclear Power Corporation of India Limited,
Mumbai and India Meteorological Department, Pune, for supply of rainfall data.
REFERENCES
[1] BHAKAR, S. R., BANSAL, Anil Kumar, CHHAJED, Neeraj, et al. Frequency analysis of consecutive days
maximum rainfall at Banswara, Rajasthan, India.ARPN Journal of Engineering and Applied Sciences, 2006,
vol. 1, no 3, p. 64-67.
[2] DABRAL, P. P., PAL, Mautushi, et SINGH, R. P. Probability analysis for one day to seven consecutive days
annual maximum rainfall for Doimukh (Itanagar), Arunachal Pradesh. Journal of Indian Water Resources,
2009, vol. 2, p. 9-15.
[3] RAO, A.R., et HAMEED, K.H. Flood frequency analysis, CRC Press, Boca Raton, Florida (USA), 2000.
[4] HOSKING, J.R.M. L-moments: Analysis and estimation of distributions using linear combinations of order
statistics, Royal Statistical Society, Series-B, 1990, Vol. 52, No. 1, p. 105-124.
[5] TOPALOGLU, F. Determining suitable probability distribution models for flow and precipitation series of the
Seyhan River basin, Turkish Journal of Agriculture and Forestry, 2002, vol. 26, no. 1, p. 189 194.
[6] E. Guevara, Engineering design parameters of storms in Venezuela, Hydrology Days, pp. 80-91, 2003.
[7] KUMAR, Rakesh et CHATTERJEE, Chandranath. Regional flood frequency analysis using L-moments for
North Brahmaputra region of India. Journal of Hydrologic Engineering, 2005, vol. 10, no 1, p. 1-7.
[8] BALDASSARRE, G. Di, CASTELLARIN, A., et BRATH, A. Relationships between statistics of rainfall
extremes and mean annual precipitation: an application for design-storm estimation in northern central
Italy. Hydrology and Earth System Sciences, 2006, vol. 10, no 4, p. 589-601.
[9] ESLAMIAN, S. Saeid et FEIZI, Hussein. Maximum monthly rainfall analysis using L-moments for an Arid
Region in Isfahan Province, Iran. Journal of Applied Meteorology and Climatology, 2007, vol. 46, no 4, p. 494503.
[10] GONZLEZ, Javier et VALDS, Juan B. A regional monthly precipitation simulation model based on an Lmoment smoothed statistical regionalization approach. Journal of Hydrology, 2008, vol. 348, no 1, p. 27-39.
[11] YUREKLI, Kadri, MODARRES, Reza, OZTURK, Fazli, et al. Regional daily maximum rainfall estimation for
Cekerek Watershed by L-moments.Meteorological Applications, 2009, vol. 16, no 4, p. 435-444.
[12] GUBAREVA, T. S. et GARTSMAN, B. I. Estimating distribution parameters of extreme hydrometeorological
characteristics by L-moments method. Water resources, 2010, vol. 37, no 4, p. 437-445.
71

N. Vivekanandan / International Journal of Engineering Issues

[13] BADRELDIN, G. H. H. et FENG, P. Regional rainfall frequency analysis for the Luanhe Basin using Lmoments and cluster techniques. In : International Conference on Environmental Science and Development
(ICESD 2012). 2012. p. 126-135.
[14] HOSKING, J. R. M. et WALLIS, J. R. Some statistics useful in regional frequency analysis. Water Resources
Research, 1993, vol. 29, no 2, p. 271-281.
[15] HOSKING, Jonathan Richard Morley et WALLIS, James R. Regional frequency analysis: an approach based
on L-moments. Cambridge University Press, 2005.
[16] ANNIS, C. Goodness-of-fit tests for statistical distributions. Statistical Engineering (www.
statisticalengineering. com), 2007.
[17] ZHANG, Jin. Powerful goodness-of-fit tests based on the likelihood ratio.Journal of the Royal Statistical
Society: Series B (Statistical Methodology), 2002, vol. 64, no 2, p. 281-294.
[18] United States Water Resources Council (USWRC), Guidelines for determining flood flow frequency, Bulletin
No. 17B, 1981.

72