CHAPTER

7
Flood Frequency
Analysis

7.1 INTRODUCTION
There are numerous problems in hydrology wherein the data mostly contains
measurements on a single random variable. Hence, univariate analysis and
univariate estimation are important keys to solve hydrological problems. The
objective of univariate analysis is to analyse measurements on the random
variable, which is called sample information; and identify the statistical
population from which we can reasonably expect the sample measurements to
have come from. After the underlying population has been identified, one can
make probabilistic statements about the future occurrences of the random
variable; this represents univariate estimation. It is important to remember that
univariate estimation is based on the assumed population and not the sample;
the sample is used only to identify the population.
Hydrologic processes, such as rainfall, snowfall, floods, droughts, etc. are
usually investigated by analysing their records of observations. Many
characteristics of these processes may not represent definite relationship. For
example, if annual instantaneous peak discharges of a river are plotted, then a
rather erratic graph would be obtained. The variation of peak discharge from
one year to another cannot be explainedby fitting a definite relationship,which
we call as deteministic relationship. In hydrologic analysis, the annual peak
discharge is considered to be a random variable. Probability and statistical
methods are employed for analysis of random variables. In this chapter, some
elementary probability distributions are presented, which are used for frequency
analysis in hydrology.
In order to have meaningful estimates from flood frequency analysis, the
following assumptions are implicit.
 Flood Frequency Analysis 249

0 The data to be analysed describe random events.

0 The data is homogeneous.
0 The population parameters can be estimated from the sample data.
0 It is of good quality.

If the data available for analysis do not satisfy any of the listed assumptions,
then the estimates are not considered reliable. Moreover, the data should be
(i) relevant, (ii) adequate,and (iii) accurate. For flood frequency analysis, either
annual flood series or partial duration series may be used.
In general, an array of annual peak flood series may be considered as a
sample of random and independent events. The non-randomness of the peak
series will increase the degree of uncertainty in the derived frequency
relationship. Various tests are available to check the randomness of the peak
flow data. The annual maximum flood series can generally be regarded as
consisting of random events as the mean interval of each observed flood peak
is 1 year. However, in case data is used for partial duration series analysis,
then independence among the data is doubtful. The peaks are selected in such
a way that they constitute a random sample.
The term relevant means that the data must deal with the problem. For
example, if the problem is of the duration of flooding, then the data series
should represent the duration of flows in excess of some critical value. If the
problem is of interior drainage of an area, then the data series must consist of
the volume of water above a particular threshold.
The term adequate primarily refers to the length of data. The length of
data primarily depends upon variability of data; and hence, there is no guideline
for the length of data to be used for frequency analysis.
The term accurateprimarily refers to the homogeneity of data and accuracy
of the discharge figures. The data used for analysis should not have any effect
of man-made changes. Changes in the stage discharge relationship may render
stage records non-homogeneous and unsuitable for frequency analysis. It is
therefore preferable to work with discharges; and if stage frequencies are
required, then most recent rating curve is used.
Watershed history and flood records should be carefully examined to ensure
that no major watershed changes have occurred during the period of record.
Only those records, which represent relatively constant watershed conditions,
should be used for frequency analysis. The fundamental terms related to
statistics are explained in Chapter 6. In the following section, some of the
popular methods used in flood frequency analysis are explained with suitable
examples.
250 Engineering Hydrology

7.2 METHODS OF PARAMETER ESTIMATION

Parameter estimation is important in flood frequency analysis. Any set of flood
occurrences can be described using some statistical distribution. Statistical
distributionmay be mathematicallyrepresented through functional relationships
which may contain a set of parameters. Some of these distributions have been
described in preceding chapters. There are four well-known parameter
estimation techniques, viz.:
0 Graphical method,

0 Frequency factor method,

0 Method of moments, and

0 Method of probability weighted moments and L-moments.

7.2.1 Graphical M e t h o d
In graphical method of parameter estimation, the variate under consideration
is regarded as a function of a reduced variate of a known distribution. The
steps involved in the graphical method are as follows:

(i) Arrange the variates of annual maximum flood series in ascending order
and assign different ranks to individual variates.
(ii) Assign plotting positions to each of the variates. The plotting position
formula may be used depending upon the type of distribution being fitted.
The recommended plotting position formulae for normal, lognormal,
Gumbel EV-I, Pearson type-111, and log-Pearson type-I11 distributions
are given in Table. 7.1
(iii) Estimate the reduced variates for the selected distribution corresponding
to different plotting positions, which represent the probability of non-
exceedence.
Table 7.1 Commonly used unbiased plotting position formulae
Distributions Recommended plotting Form of the plotting position
position formula formulae F (X Ix) = 1 - - 1
T

Normal and Lognormal Blom (i - 0.375)/(N + 0.25)

Gumbel EV-I Gringorton (i - 0.4)/(N + 0.12)
Any distribution Weibull i/(N + 1)
PT3 and LP3 Cunnane (i - 0.4)/(N + 0.2)

[i indicates the rank number. i = 1 for the smallest observation, and i = N for the largest
observation, when flood series is arranged in ascending order.]

The reduced variates for normal and lognormal distributions are computed
with the help of the table given in Appendix 11.For Gumbel (EV-I) distribution,
the reduced variates are computed using the relationships given in Chapter 6.
In case of Pearson type-I11and log-Pearsontype-I11which are three-parameter
 Flood Frequency Analysis 251

distributions, different sets of reduced variates are obtained for different

coefficients of skewness. Graphical method is not widely used nowadays.
However, to illustrateits applicationin parameter estimation,an example using
Gringrorton plotting position formula is given here. Let the relation among xT
(discharge correspondingto a return period T), location parameter u, and scale
parameter P be of the type:
=u+PYT
XT (7.1)
For Gumbel distribution,

For two dischargesxT1 and x, corresponding to return period Tl and T2,let yT

be yTl and y,. Thus,
xT1 = p +YTlP (7-3)
xn = iu +Y,P (7-4)
From Eqs. (7.3) and (7.4),
P= xT1- x T 2
= slope of straight line (7.5a)
(YTl-YT2)

u = xT corresponding to yT = 0 (7.5b)

Example 7.1 The annual maximum series of a certain gauge and discharge
site are given below. Table 7.2(a) shows the ranks of the AMS in descending
order. Estimate the flood and the parameters for a return period of 1000 years
using Gringrorton plotting position formula.

Table 7.2(a)
m 42 m 42 m 42
1 2294 11 1095 21 675
2 1880 12 915 22 467
3 1850 13 889 23 467
4 1700 14 869 24 392
5 1500 15 867 25 371
6 1466 16 860 26 325
7 1395 17 792 27 288
8 1366 18 720 28 212
9 1175 19 693 29 170
10 1100 20 675 30 120
252 Engineering Hydrology

Solution
Table 7.2(b) Calculation of recurrence interval T using Gringrorton plotting
position formula

1 I 120 I 0.01859 I 1.018945 I 16 I 867 I 0.5166 12.068681

2 I 170 I 0.05179 I 1.054622 I 17 I 869 I 0.5498 12.221239
3 212 0.08499 1.092888 18 889 0.583 2.398089
4 288 0.11819 1.134036 19 915 0.6162 2.605536
5 I 325 I 0.15139 I 1.178404 I 20 I 1095 I 0.6494 12.852273
6 I 371 I 0.18459 I 1.226384 I 21 I 1100 I 0.6826 13.150628
7 I 392 I 0.21779 I 1.278438 I 22 I 1175 I 0.7158 13.518692
8 I 467 I 0.25099 I 1.335106 I 23 I 1366 I 0.749 13.984127
9 467 0.28419 ~ 24 1395 0.7822 4.591463
10 675 0.31739 1.464981 25 1466 0.8154 5.417266
11 I 675 I 0.35059 I ~ I 26 I 1500 I 0.8486 16.605263
12 I 693 I 0.38379 I 1.622845 I 27 I 1700 I 0.8818 18.460674
13 I 720 I 0.41699 I 1.715262 I 28 I 1850 I 0.915 111.76563
14 I 792 I 0.45019 I 1.818841 I 29 I 1880 I 0.9482 119.30769
15 860 0.48339 1.935733 30 2294 0.9814 53.78571

For i = 1, (1 1/T) = (1 -0.44) /(30 + 0.12) = 0.01859. Therefore, T = 1.018945.

-

Similarly, return periods for all the discharge can be determined. Since the
discharge values are in ascending order, the lower rank of a discharge is taken
into account where repetition of discharge has taken place.
The graph is plotted on a Gumbel probability paper between Q and T. In
Gumbel probability paper, the graph will always be a straight line because
there exists a linear relationship between return period T and discharge Q.
Reduced variate values are shown below the scale of return period.
Using Eq. (7.2),
For T = 53.78, yT= 3.9756
For T = 1.018945, yT=-1.3825
Using Eq. (7.5a),
(2294 - 120)
slope = p = = 405.74
[3.9756 - (-1.3825)]
Corresponding to yT = 0, intercept u = 688.0 m3/s
From Fig. 7.1, it can be seen that after extending the straight line, the 1000
year return flood comes to be 3800 m3/s.
 Flood Frequency Analysis 253

5.0
4.5 - -

4.0 - -

$ 3.5
0
- -

8 3.0 - -
s

-2 -1 0 1 2 3 4 5 6 7
Reduced variate

Fig. 7.1 Gumbel probability paper using Gnngrorton plotting

position formula

Example 7.2 Estimate u and p in the previous example by considering the

following plotting position formula T = (N + l)h, where N is the sample size
and i is the order of flood arranged in descending manner.
Solution
Table 7.3 contains the necessary computations and Fig. 7.2 shows the plot of
xT versus reduced variate.
Table 7.3

Rank Q (rn3/s) ( N + l )
~ Rank Q (rn3/s) ( N
~ Rank Q (rn3/s) ( N + 1)
~

~~~~~~~
i i i

~~~~~~~
1 2294 31 11 1095 2.818 21 675 1.4762
2 1880 15.5 12 915 2.583 22 467 -
~~~~~~~

~~~~~~~
3 1850 10.33 13 889 2.384 23 467 1.3478
~~~~~~~
4 1700 7.75 14 869 2.214 24 392 1.2916
~~~~~~~
5 1500 6.2 15 867 2.067 25 371 1.24
~~~~~~~
6 1466 5.166 16 860 1.9375 26 325 1.1923
~~~~~~~
7 1395 4.428 17 792 1.823 27 288 1.1481
~~~~~~~
8 1366 3.875 18 720 1.722 28 212 1.1071
~~~~~~~
9 1175 3.444 19 693 1.6315 29 170 1.0689
10 1100 3.1 20 675 - 30 120 1.0333
254 Engineering Hydrology

4.5
5 . 0

1.01 2 10 50 100 250 500 1000

Recurrence internal in T (years)

I I I I I I I I I I
-2.0 -1.0 0 1 2 3 4 5 6 7
Reduced variate

Fig. 7.2 Gumbelprobability paper using Weibullplottingposition

formula

Using Eq. (7.2),

For T = 31, y,= 3.4176 and
For T = 1.033, y,=-l.234
Using Eq. (7.5a),

slope = fi = (2294 - 120) = 467.36

[3.4176 - (- 1.234)]

Corresponding to y , = 0, intercept u = 682.0 m3/s

From Fig. 7.2, it can be seen that after extending the straight line, the 1000
year return flood comes to be 4200 m3/s.
It is to note that plotting position formula influences estimates of model
parameters. Thus, one has to be careful in selecting a plotting formula for a
particular distribution.
 Flood Frequency Analysis 255

7.2.2 Frequency Factor

Similar to Eq. (7. l), a general frequency equation proposed by Chow (1964),
and applicable for different distributions, is in the following form:
XT= m KTs (7.6)
where, xT = magnitude of the different peak flood series, or magnitude of the
flood at the required return period T; KT = frequency factor corresponding to
xT; rn and s = mean and standard deviation of the population, respectively,
which would be replaced by the sample statistics. The similarity between Eqs.
(7.1) and (7.6) is evident. Table 7.4 provides expressions to determine KT for
certain probability distributions.

Table 7.4 Frequency factor for different distributions

112
Normal distribution

2.515517 +0.802853w+0.010328w2
z=w-
1 +1.432788w+0.189269w2+0.001308w3
KT=z

EV-I distribution 1 KT=-*(,,772+hk(&)]}

Log-Pearson type-I11 KT = z + (z2

- l)k + 1 (z3
- 6z)P - (2
- l)k3+ zk4 + 1 K5
3 3
distribution where k = 816 and g is skewness

Example 7.3 Calculate the frequency factor for an event with a return period
of 100 years using normal distribution and EV-I distribution.
Solution
Normal distribution
T = 100 years, p = 1/100 = 0.01
112

w = [ln($]]
112

= [ln[&]]
= 3.034
256 Engineering Hydrology

2.5 15517 + 0.802853~

+0.010328 w 2
Z=W-
1+ 1.432788w+O.189269w2+O.O01308d

2.515517 + (0.802853 x3.034)

= 3.034 -
+ [0.010328 x (3.034) '1
1+ (1.432788 x 3.034) +[0.189269 x(3.034) '1
+ [0.001308 x (3.034)3]
= 2.325

EV-I distribution

K T = -~
"1
7c
0.5772+In h 100-1
( qJ

= 3.136

7.2.3 M e t h o d of M o m e n t s (MOM)
The method of moments makes use of the fact that if all the moments of a
distribution are known then everything about the distribution is known. For all
the distributions in common usage, four moments or fewer are sufficient to
specify all the moments. For instance, two moments, the first together with
any moment of even order are sufficient to specify all the moments of the
normal distribution, and therefore the entire distribution. Similarly, in
Gumbel EV-I distribution, the first two moments are sufficient to specify all
the moments and hence the distribution.For Pearson type-I11distribution three
moments are sufficient, always taken as the first three, to specify all the
moments. In these cases, the number of moments needed to specify all the
moments and hence the distribution equals the number of parameters.
This method is dependent on the assumption that the distribution of variate
values in the sample is representative of the population distribution.Therefore,
a representation of the former provides an estimate of the later. Given that the
form of the distribution is known or assumed, the distribution that the sample
follows is specified by its first two or three moments calculated from the data.
Table 7.5 provides relationships between model parameters and moments for
a variety of distributions.
 Flood Frequency Analysis 257

Table 7.5 Probability distribution parameters in relation to sample moments

Distribution Pvobability Density Function Range Equation of
pavametevs in tevms
of the sample moments

Exponential x2O a= x
Ixtreme value (-w < u= X - 0.57728
type-I
x < w)

Gamma x20

Normal u=x, o=sx

Lognormal

Pearson x2u u= x- S X h ,
type-I11

Log-Pearson lnx 2 u u=y-syfi

type-I11

Table 7.5 does not include other GEV distributions. Stedinger et al. (1993)
have given MOM estimators for three types of GEV distributions.

Parameter E s t i m a t e s using MOM for GEV Distribution

[ ( ;)I
F(q) = exp -exp -- fory=O (7.8)

where, q is the flood exceedence value, u is a location parameter, p is a scale

parameter, and yis the shape parameter. The range of the variable, q, depends
on the sign of y
258 Engineering Hydrology

When yis negative (g > 1.1369), then the variable q can take values in the
range of [(u + P/.i, < q < -1 which make it suitable for flood frequency analysis.
When yis positive, then the variable q becomes upper bounded; and for
y = 0, the distribution reduces to two-parameter extreme value type-I or
Gumbel's distribution. The MOM estimators of the GEV parameters are given
by Stedinger et al., (1993) as:

(7.10)

where, sign is '+' or '-' depending upon the value of y-estimate, and g is the
skewness.
P arame t e r E s t i m a t e s u s i n g MOM for Generalized P a r eto (GP)
Distribution
The GP distribution introduced by Pickands (1979, has the following CDF:

(7.12)

where, P is the scale parameter, u is the location parameter, and yis the shape
parameter. For y= 0, exponential distribution is obtained as a special case. For
y I 0, the range of u is (u I q I -) and for y 2 0 the upper bound exists:
u I q I (u + P / y ) . The parameter estimates of the GP distribution using MOM
estimation method are given in the following relations:

2(1- y)(l + 2yy2

(7.13)
g= (1+3y)
Given the skewness of the AMS data, ycan be computed numerically. The
other parameters are then obtained as follows:
P = o [ ( l +y)2(1 +2y)]ll2 (7.14)

(7.15)

where, p and o are the mean and standard deviation, respectively.

 Flood Frequency Analysis 259

Example 7.4 Estimate the parameters using method of moments for the given
data in Example 7.1 considering GEV distribution and GP distribution.
Solution
Parameter Estimates for GEV Distribution
Using the given data, the following statistics are calculated:
Mean Q = 919.6 m3/s
Standard deviation (0)= 561.88 m3/s
Skewness coefficient (g) = 0.6348
Using skewness coefficient g and Eq. (7.1 l), one can estimate yto decide the
type of GEV distribution.
Skewness coefficient (g)for AMS is 0.6348.
Using the equation given by Stedinger et al., (1993), the value of yis computed
as 0.10083 using the method of iteration.
Equation (7.10) yields P/o= 0.8743.
The value of (r which is the standard deviation of AMS data is 561.88 m3/s,
therefore P = 491.3 1 m3/s.
Hence, using Eq. (7.9), we have:

491.31 x [r(i.i0083) - 11
u = 919.6 +
0.10083
= 680.91 m3/s

Thus, the parameters estimated using method of moments u, p, and yare 680.91,
491.31, and 0.10083, respectively.

Parameter Estimates for GP Distribution

Skewness coefficient (g) for AMS is 0.6348.
Using Eq. (7.13), the value of yis computed as 0.456 using the method of
iteration.
Using P = o[(l + ~ ) ~+(2y)I1l2
1 = 561.SS[(1.456)2(1 + 2 x 0.456)I1l2

= 1131.22 m3/s
u =p-PPl(l + y ) = 919.6 - [1131.22/1.456]
= 142.66 m3/s
260 Engineering Hydrology

7.2.4 Probability Weighted Moment (PWM) and Linear (L)

Moments
Greenwood et al. (1979) introduced the method of probability weighted moment
(PWM) and showed its usefulness in deriving explicit expressions for
parameters of distributions whose inverse forms x = x(F) can be explicitly
defined. If PDF be denoted by F = F(x) =P[XI x], then PWMs are the moment
of the function x(F); and they are expressed as:
1
Mij, k E[xiFJ(1 - F ) k ] = [x(F)FJ(1 - F ) k dF (7.12)
0
It can be shown that,

(7.13)

where, Mi,.i,k is the PWM of the order of i, j , and k; and E is the expectation
operator. Mi, o, represents the conventional moment about the origin of
order i. These can than be related to the distribution parameters and the resultant
relation may be a simpler structure than the conventional moments and the
parameters. For example, the parameters of Gumbel EV-I distribution using
PWMs are as follows:
x ( F ) = u P In[-ln ( F ) ]
- (7.14)
1
Mi,.i, =
[u - P In(-ln F)Ii Fi (1 F) dF
- (7.15)
0
1 1
= j [u F ~ I ~ j Fp h - l n F ) ]~i
- ~ d~ (7.16)
0 0
-u ln(l+j)+&
- -+p (7.17)
1 +j 1 +j
where, E = Eulers constant = 0.5772
Ml,O,O= + PE (7.18)

Ml,I,O= ;P [ ln(2)2 + ]
+
E
(7.19)

Solving Eqs. (7.18) and (7.19):

(7.20)
and - 0.5772 P
u = Ml,o,o (7.21)
L-moments are linear combination of order statistics which are robust to outliers
and are unbiased for small samples, making them suitable for flood fi-equency
analysis, including identification of distribution and parameter estimation
(Hosking, 1990, Hosking and Wallis, 1997). L-moments are identified as linear
combination of probability weighted moment (PWM)
 Flood Frequency Analysis 261

q y , 0 = P, =EV[F(x)IY} (7.22)
where, F(x) is the cumulative distribution function of x. When r = 0, Po is the
mean. The first four L-moments expressed as linear combination of PWMs
are:
4 =Po (7.23a)
n, = 2 P1- P o (7.23b)
= P2 - P1+ P O (7.23~)
a, = 20 p3 - 30 p2 + 12 p1- p o (7.23d)
Landwehr et al. (1979) recommends the use of biased estimates of PWMs and
L-moments, since such estimates often produce quintile estimates with lower
root mean square error than unbiased alternatives. Nevertheless, unbiased
estimators are preferred in moment diagrams for goodness-of-fit for less bias
and are invariant when the data is multiplied by a constant. Unbiased estimates
of PWMs for any distribution can be computed as follows:

b, =
l n
- z
,=I
X, (7.24a)

b, = kl[( ( n - j ) ]x, (7.24b)

j=1 n>(n-1)

z[(nn(n- j )-l)(n -2) 1xJ

n-2 ( n - j 1) -
b,= (7.24~)
j=l

z[( n-n(nj ) ( n l)(n 2)(n 3) 1

n-3 - j l)(n - j 2)
- -
b,= (7.24d)
j=l - - - xJ
where, xirepresents the ordered streamflow with x1 as the largest streamflow
and x, as the smallest. The L-moment ratios which are used for expressing the
parameter estimates are expressed as:
L-coefficient of variation (z) = A,
L-skewness (z,) = A3/A2

and L-kurtosis (z,) = A4/A,

The sample estimates of L-moments are calculated by replacing b,, b,, b,, and
b, for Po, P,, P,, and P2 in Eq. (7.23). So, the parameters of Gumbel EV-I
distribution in the above example can be written in terms of L-moment as
follows:
p = &/log 2
u = a, - 0.5772p
The relationship equationsbetween sample moment parameters and population
parameter estimation for different distributions using L-moment are given in
Hosking and Wallis (1997). Although the theory and application of L-moment
parallels those of conventional moment, there are certain advantages of
262 Engineering Hydrology

L-moment. Since sample estimators of L-moments are always a combination

of ranked observations, they are subjected to less bias than ordinary product
moments. This is because ordinary product moments require squaring and
cubing, giving greater weight to observations far from the mean resulting in
bias.
Parameter Estimates using PWM Method for GEV Distribution
The PWM estimators for GEV distribution are given by Hosking et al., (1985)
as:

(7.25)

(7.26)

2
y= 7.859~+ 2 . 9 5 5 4 ~; ~ c = (7.27)
+
(r3 3) - ln(2)/ln(3)
Parameter Estimates using PWM Method for GP Distribution
Similarly, the parameter estimates by the PWM method is given by Hosking
and Wallis (1997) as follows:
y = (1 3z3)/(1+ z3); p = h ( i
- + y)(2 + y); u= a,
h ( 2 + y) -

(7.28)
where, z3 is the L-skewness, A, and are the first and second L-moments for
the sample data. Hosking and Wallis (1997) have given the PWM estimator
equations as follows:
Y = (4 4 0 Y h 2 P = (1 + Y)(4 4 0 )
- - -
(7.29)
where, Ai are the L-moments, which are defined for y> -1.

Example 7.5 Estimate the parameters using probability weighted moments

for the given data in Example 7.1 considering GEV distribution and GP
distribution.
Solution
Parameter estimates using GE V distribution
Table 7.6 Calculation of probability weighted moments
Rank Q (m3/s)
~~~~~~~~~~
b, b2 b3 Rank Q (m3/s) b, b2 b3
1
~~~~~~~~~~
2294 76.467 76.467 76.47 16 860 13.839 6.425 2.86
2
~~~~~~~~~~
1880 60.506 58.345 56.18 17 792 11.834 5.072 2.07
3
~~~~~~~~~~
1850 57.414 53.313 49.36 18 720 9.931 3.901 1.44
4 1700 50.805 45.361 40.32 19 693 8.762 3.129 1.04
(Contd.)
 Flood Frequency Analysis 263

(Contd.)

~~~~~~~~~~
5 1500 43.103 36.946 31.47 20 675 7.759 2.494 0.74
~~~~~~~~~~
6 1466 40.441 33.220 27.07 21 675 6.983 1.995 0.52
~~~~~~~~~~
7 1395 36.879 28.977 22.54 22 467 4.294 1.074 0.24
~~~~~~~~~~
8 1366 34.543 25.907 19.19 23 467 3.757 0.805 0.15
~~~~~~~~~~
9 1175 28.362 20.259 14.26 24 392 2.703 0.483 0.07
~~~~~~~~~~
10 1100 25.287 17.159 11.44 25 371 2.132 0.305 0.03
~~~~~~~~~~
11 1095 23.914 15.373 9.68 26 325 1.494 0.160 0.01
~~~~~~~~~~
12 915 18.931 11.494 6.81 27 288 0.993 0.071 0.00
~~~~~~~~~~
13 889 17.371 9.926 5.51 28 212 0.487 0.017 0.00
~~~~~~~~~~
14 869 15.982 8.562 4.44 29 170 0.195 0.000 0.00
~~~~~~~~~~
15 867 14.948 7.474 3.60 30 120 0.000 0.000 0.00
620.118 474.713 387.516

First of all, the probability weighted moments are calculated using Eq. (7.24)
and the calculations are shown in Table 7.6. From the table, the summation of
the values in columns (3), (4), and (5) gives the values of b,, b,, and b,,
respectively; and b, is the mean of AMS. The L-moments are calculated as
under.
A, = b, = 919.6
& = 2 b, ~ b, = 320.63

A, = 6 b, 6 b , + b, = 47.17
~

A4 = 20 b, 30 b, + 12 b , b, = 30.75
~ ~

2, = a,/& = 0.147
Therefore, c = [2/(2, + 3)] [ln(2)/ln(3)] = [2/(3.147)]
~ ~ 0.6309 = 0.00459
y= 7.859~+ 2.9554~~= 0.03617

320.63 x0.03617
P= = 476.20 m3/s
r(1+ 0.03617)(1 -2 -0.03613

Thus, the parameters estimated using probability weighted moments u, P, and

yare 705.0,476.20, and 0.03617, respectively.
Parameter estimates using GP distribution

Y= (1-323) = [1 -(3 xO.147)]/(1.147)=0.487

(1 + 7 3 )
P = h(1+ y)(2 + y) = 320.63 (1.487) (2.487) = 1185.74 m3/s
u = A, - h ( 2 + y) = 919.6 - 320.63 (2.487) = 122.19 m3/s
264 Engineering Hydrology

7 . 3 RETURN PERIOD FLOOD ESTIMATION

A flood of certain magnitude is expected to occur after a certain recurrence
interval. Expected or average value of this recurrence interval is also known
as return period. Given the flood records at any site, it is of interest to compute
the magnitude of floods which correspond to different return periods. One
possible and most widely used approach is to identify the particular statistical
distributions, and use these to predict a flood of desired return period. In this
section, we discuss the use of different statistical or probabilistic distributions
in return period flood estimation.
7.3.1 Normal Distribution
The parameters of the normal distribution, p and o,which describe the
characteristics of the given set, are computed as:

(7.30)

l(7.31)
After computing the parameters, p and o,the T-year flood estimates can be
obtained using the normal distribution. The steps are:
Compute X and s from the sample data. X and s provide estimates for p
and o.
Compute the probability of non-exceedence using the relation:
F(x <x>= 1 1/T ~ (7.32)
Compute the normal reduced variate (z,) from the statistical table,
corresponding to the probability of non-exceedence computed in
step (2).
Estimate the flood for T-year recurrence interval using the following
equation:
x,=p+oz, (7.33)

Example 7.6 Using the data given in Example 7.1, estimate the flood for a
return period of 1000 years using the formula given by Swamee and Rathie
(2007) for normal distribution.
If the data follows normal distribution, then according to Swamee and Rathie
(2007):

k= { -0.6144[-ln(T - 1)]11.12 for (1 I T 5 2 )

-0.6144[ln(T - 1)]11.12 for (T 2 2)

k= X-P
~

o
 Flood Frequency Analysis 265

Solution
Since T = 1000 years, using the formula given above k = 3.449

Using the formula k = ~ - ,where p = 919.6 and s = 561.8827, the value of

0
flood discharge is equal to x = 3.449 x 561.8827 + 919.6 = 2857.53 m3/s.

7.3.2 Lognormal Distribution (Two parameter)

In lognormal two-parameter distribution, the variates are transformed to the
log domain by taking logarithm of each variate, and then the mean and standard
deviation of the transformed series are computed. The flood at required
recurrence interval can be computed in the following steps after fitting the
lognormal distribution with the sample data.

(i) Transform the original series ofpeak flood data into log domain by takmg
logarithm of each variate to base e.
(ii) Compute the mean ( j )and the standard deviation (s), from the log-
transformed series. 7 ands, provide estimates for py and o, respectively.
(iii) Compute the probability of non-exceedence for the given recurrence
interval (T).
(iv) Compute the normal reduced variate (zT)corresponding to the probability
of non-exceedence computed in step (3).
(v) Estimate the flood for T-year recurrence interval in log domain using the
following equation:
YT=Py+PyzT (7.34)
(vi) Transform the estimated T-year flood in original domain by computing
its exponent, i.e.,
X T = exP(YT) (7.35)

Example 7.7 Estimate 1000-years floods assuming that the peak discharge
data follow lognormal distribution for the data given in Example 7.1.
Solution
Since p, = 6.60 and oy = 0.743
Probability of non-exceedence, F(z) = 1 - 1/1000 = 0.999
z T = 3.10 for T = 1000 years
YT=Py+oyzT
ylooo= 6.60 + (0.743 x 3.10) = 8.9033
xlooo= (e)ylooo = (e = 7356.2 m3/s
266 Engineering Hydrology

7.3.3 Gumbels Extreme Value Type-I Distribution

The relationship between the parameters and the statistical moments of the
data are given by the following equations:

P= 7
JSO (7.36)
u =/A- 0.450 (7.37)
The population statistics, /A and o would be replaced by the sample statistics,
X and s, respectively. The step-by-step procedure for computing the T-year
flood using EV-I distribution is given below:

Compute mean ( X ) and standard deviation (s) from the sample. X and s
provide estimates for p and O,respectively.
Compute the parameters, u and P, using Eqs. (7.36) and (7.37).
Compute the probability of non-exceedence corresponding to T-year
recurrence interval.
Compute the EV-I reduced variate bT) using the relationship:

(7.38)

Estimate T-year recurrence interval flood using the EV-I distribution as

follows:
xT= + PYT (7.39)
Sample Size Correction
To estimate frequency factor, Chow (1964) has given the frequency factor
which is given in Table 7.4. Literature does indicate that frequency factor
depends upon the sample size (Subramanya, 1984, Raghunath, 2007).

KT= YT - Yn
~

Sn
xT= /A + OK,
where, /A and o are mean and standard deviation respectively, y , = reduced
mean and s, = reduced standard deviation which is a function of sample size.
As N + 00, y , = 0.577 and s, = 1.2825. Tables 7.7 and 7.8 give the values of
reduced mean and reduced standard deviation as per the sample size,
respectively.

Table 7.7 Reduced mean (y,) in Gumbel EV-I distribution

N O 1 2 3 4 5 6 7 8 9
10 0.4952 0.4996 0.5035 0.507 0.51 0.5128 0.5157 0.5181 0.5202 0.522
20 0.5236 0.5252 0.5268 0.5283 0.5296 0.5309 0.532 0.5332 0.5343 0.5353
(Contd.)
 Flood Frequency Analysis 267

(Contd.)

~~~~~~~~~~~
30 0.5362 0.5371 0.5380 0.5388 0.5396 0.5402 0.5410 0.5418 0.5424 0.5430
~~~~~~~~~~~
40 0.5436 0.5442 0.5448 0.5453 0.5458 0.5463 0.5468 0.5473 0.5477 0.5481
~~~~~~~~~~~
50 0.5485 0.5489 0.5493 0.5497 0.5501 0.5504 0.5508 0.5511 0.5515 0.5518
~~~~~~~~~~~
60 0.5521 0.5524 0.5527 0.5530 0.5533 0.5535 0.5538 0.5540 0.5543 0.5545
~~~~~~~~~~~
70 0.5548 0.5550 0.5552 0.5555 0.5557 0.5559 0.5561 0.5563 0.5565 0.5567
~~~~~~~~~~~
80 0.5569 0.5570 0.5572 0.5574 0.5576 0.5578 0.5580 0.5581 0.5583 0.5585
~~~~~~~~~~~
90 0.5586 0.5587 0.5589 0.5591 0.5592 0.5593 0.5595 0.5596 0.5598 0.5599
100 0.56

Table 7.8 Reduced standard deviation (sn)in Gumbel EV1 distribution

N O I 2 3 4 5 6 7 8 9
~~~~~~~~~~~

~~~~~~~~~~~
10 0.9496 0.9676 0.9833 0.9971 1.0095 1.0206 1.0316 1.0411 1.0493 1.0565
~~~~~~~~~~~
20 1.0628 1.0696 1.0754 1.0811 1.0864 1.0915 1.0961 1.1004 1.1047 1.1086
~~~~~~~~~~~
30 1.1124 1.1159 1.1193 1.1226 1.1255 1.1285 1.1313 1.1339 1.1363 1.1388
~~~~~~~~~~~
40 1.1413 1.1436 1.1458 1.1480 1.1499 1.1519 1.1538 1.1557 1.1574 1.1590
~~~~~~~~~~~
50 1.1607 1.1623 1.1638 1.1658 1.1667 1.1681 1.1696 1.1708 1.1721 1.1734
~~~~~~~~~~~
60 1.1747 1.1759 1.1770 1.1782 1.1793 1.1803 1.1814 1.1824 1.1834 1.1844
70 1.1854 1.1863 1.1873 1.1881 1.1890 1.1898 1.1906 1.1915 1.1923 1.1930
80 1.1938 1.1945 1.1953 1.1959 1.1967 1.1973 1.1980 1.1987 1.1994 1.2001
90 1.2007 1.2013 1.2020 1.2026 1.2032 1.2038 1.2044 1.2049 1.2055 1.2060
100 1.2065

Example 7.8 Using the data given in Example 7.1, estimate 1000-yearsfloods
assuming that the peak discharge data follow Gumbels EV-I distribution.
Solution
Since p = 919.6

O= 561.8827, and
a = 0.7797 x o = 0.7797 x 561.8827 = 438.09
u = p 0.450
~

u = 919.6 - (0.45 x 561.8827) = 666.75

y,=-In [ A]
In-

ylooo= -In [In (1000/999)] = 6.907255

XlOOO = + PYIOOO

xloo0= 666.75 + (438.09 x 6.907255) = 3692.75 m3/s

268 Engineering Hydrology

Consideration of sample size

p = 919.6, O = 561.8827,
ylooo= -In [ln(l000/999)] = 6.907255
F o r N = 30,y, = 0.5362 ands, = 1.1124
K,= (6.907255 - 0.5362)/1.1124 = 5.7273
xlooo= 919.6 + (561.8827 x 5.7273) = 4137.67 m3/s
It is to note that estimation of return flood is influenced by the sample size.

7.3.4 Pearson Type-I11 (PT-111) Distribution

PT-I11 distribution is a three-parameter distribution. Therefore, three moments
are needed for computing the parameters. Mean, standard deviation, and
skewness computed from the sample data describe the measures for the first
three moments of the sample data. The following steps are usually involved in
computing the T-year flood using the PT-I11 distribution.

(i) Compute the mean, p and standard deviation O.Compute the coefficient
of skewness, g , from the sample using the following equation:

(7.40)

(ii) Compute the probability of exceedence for the given recurrence interval,
T, which equals to 1/T.
(iii) Estimate the frequency factor (K,) from the table corresponding to the
computed coefficient of skewness Cg) and the probability of exceedence.
(iv) Estimate T-year flood using the equation:
x , = p + OK, (7.41)

Example 7.9 Using the data given in Example 7.1, estimate 1000-yearsfloods
assuming that the peak discharge data follow PT-I11 distribution.
Solution
Since p = 919.6

p = 561.8827, and
g = 0.634
Probability of exceedence = 1/1000 = 0.001
Frequency factor, K, = 4.00 (using table given in Appendix VI)
X, = p + OK,
= 919.6 + (561.8827 x 4.00)

= 3 172.35 m3/s
 Flood Frequency Analysis 269

7.3.5 Generalized Extreme Value (GEV) Distribution

The L-moment estimators A,, &, A,, and 7, are obtained by unbiased estimators
of the first three PWMs. The maximum likelihood estimators are determined
numerically using a modified Newton-Raphson algorithm (Hosking et al.,
1985). The return period flood is given as:

1/T (7.42)

Example 7.10 Using the data of Example 7.1 estimate the flood for a return
period of 1000 years using (i) method of moments, and (ii) PWM considering
both GEV and GP distributions.
Solution
Using MOM
From Example 7.4, the parameters estimated using method of moments u, p,
and yare 680.91, 491.31, and 0.10083, respectively. Using these parameter
values in Eq (7.42):
p[1- (-logF)Y]
q ( T )= u +
Y
491.31[1- (-10gO.999)~~'~~*~]
q ( T ) = 680.91 +
0.10083
= 3321.14 m3/s

Using P WM
From Example 7.5, the parameters estimated using probability weighted
moments u, p, and yare 705.0,476.20, and 0.03617, respectively. Using these
values of the parameters in Eq (7.42), we get:
476.2[1- ( - 1 0 g O . 9 9 9 ) ~ . ~ ~ ~ ' ~ ]
q ( T ) = 705.0 +
0.036 17
= 3920.27 m3/s

7.3.6 Generalized Pareto (GP) Distribution

The return period flood is given as follows:

q ( T ) = u + P[l-(l-F)Y1 ; F=l-l/T (7.43)

Y

Example 7.11 Estimate the flood for a return period of 1000 years using
(i) method of moments and (ii) PWM using GP distributions using the data
given in Example 7.1.
270 Engineering Hydrology

Solution
Using MOM
From Example 7.4, the parameters estimated using method of moments u, p,
and yare 142.66, 1131.22, and 0.456, respectively. Using these values of the
parameters in Eq. (7.43), we get:

F = 1 - (1/1000) = 0.999
1131.22[1- (1- 0.999)0.456]
q ( T ) = 142.66 +
0.456
= 251.7 m3/s
Using P WM
From Example 7.5, the parameters estimated using probability weighted
moments u, p, and yare 122.19, 1185.74,and 0.487, respectively. Using these
values of the parameters in Eq (7.43), we get:
1185.74[1- (1 - 0.999)0.487]
q(T) = 122.19 +
0.487
= 2472.74 m3/s

7.4 ESTIMATION OF FLOOD DISCHARGE FOR A

CONFIDENCE INTERVAL
It is oRen of interest to estimate confidence interval for a predicted return
flood. The procedure to estimate confidence interval involves the following
steps.
1. Compute standard error (S,) for a given distribution as given below.
For normal distribution

(7.44)

For extreme value type-1

(7.45)
where, KT is the fi-equencyfactor.
2. Compute flood discharge for a particular confidence limits as the formula
given below. Table 7.9 provides the value offc) for a particular confidence
interval.
 Flood Frequency Analysis 271

Confidence Limits
xconf = xT *Ac)sE (7.46)
where,Ac) is the function of confidence probability.
Table 7.9 Determination of f ( c ) for a particular confidence interval
c(%) 50 68 80 90 95 99
Ac) 0.674 1.0 1.282 1.645 1.96 2.58

Example7.12 Data covering a 50 years for a river at a station yielded: mean

= 5000 m3/s and standard deviation = 1200 m3/s. Using Gumbel's method,
estimate the flood discharge for T = 250 years. What are the 95% and 80%
confidence limits for this estimate?
Solution
F(z) = 1 ~ 1/250 = 0.996

Using the formula K, = -~ "(

7-c
[(
0.5772 + In In T
' ~ 2 we have:

K, = -&{0.5772
7-c

= 3.853
)I)&(
+ ln[ ln
xT= p + K,o= 5000 + 3.853 x 1200 = 9624.15 m3/s
Calculation of standard error

[
SE = 1200 -(1+1.1396 ~ 3 . 8 5 3+1.1 ~ 3 . 8 5 3
0:
For c = 95 % , f ( c ) = 1.96
= 790.22>
r2
xg5%= 9624.15 f (1.96 x 790.92) = 11174.35 m3/s and 8073.94 m3/s
xs0%= 9624.15 f (1.282 x 790.92) = 10638.1 m3/s and 8610.19 m3/s

7.5 REGIONAL FLOOD FREQUENCY ANALYSIS

The purpose of frequency analysis is to estimate the design flood for desired
recurrence interval, assuming the sample data follow a theoretical frequency
distribution. It is assumed that the sample data is a true representative of the
272 Engineering Hydrology

population. It is generally seen that a minimum of 30 to 40 years of records are

needed in order to carry out flood frequency analysis to the at-site data for
estimating the floods in extrapolation range, somewhat within the desired
accuracy.
In case the length of records is too short, it represents inadequate data
situation; then at-site flood frequency analysis fails to provide the reliable and
consistent flood estimates. Under inadequate data situation, regional flood
frequency curves together with at-site mean provides more reliable and
consistent estimates of floods. For ungauged catchments, the regional flood
frequency analysis approach is the only way to estimate the flood for a desired
recurrence interval, for which a regional relationship between mean annual
peak flood and catchment characteristics is developed along with the regional
frequency curves.
7.5.1 Index-Flood Method
Basically, the index-flood method (Dalrymple, 1960) extrapolates statistical
information of runoff events for flood frequency analysis from gauged
catchments to ungauged catchment in the vicinity having similar catchments
and hydrologic characteristics. Given below are the sequential steps which are
to be followed to estimate return period flood using index-flood method.

(i) Select the gauged catchments within the region having similar
characteristics to the ungauged catchment.
(ii) Determine the base period to be used for the study.
(iii) Establish the flood frequency curve prepared with the ranked annual
series and estimate the correspondingplotting position using Gringortan
plotting formula.
(iv) Q,= 2.33 yr (annual mean flood) is estimated using EV-I parameters.
(v) Establish the relationship between the mean annual flood and the basin
area.
(vi) Rank the ratios of floods of selected return periods to the mean annual
flood at each site and for a given return period, obtain the ratio of flood
(having return period T ) to mean annual flood for each site.
(vii) For N sites, there will be Nvalues of such ratios for a given return period.
(viii) Use these ratios to compute median flood ratio for each of the selected
return period and multiply each median flood ratio by the estimated mean
annual flood of the ungauged catchment, and plot these values versus
recurrence interval on Gumbel probability paper.

The curves obtained in the last step are the flood frequency curves for ungauged
catchments.
 Flood Frequency Analysis 273

Example 7.13 Flood fi-equency analysis for five homogeneous catchments

provides the following estimates for mean annual flood and 50-year return
flood.

Catchment Mean annualflood I 50-year return flood

I A I 20 I 100 I
I B I 35 I 150 I
I C I 25 I 120 I
D 50 170
E 10 90

An ungauged catchment which is having similar characteristics to catchments

A to E is having a mean annual flood of 40 m3/s. Estimate the 50-year return
flood for the ungauged catchment.
Solution

Table 7.11 Solution table for Example 7.13

Catchment Ratio of 50-year returnflood to mean annualflood
A 5
B 4.285
C 4.8 I
I D 1 3.4 I
9.0

Thus, median value of ratio of 50-year return flood to mean annual flood = 4.8
As the mean annual flood for ungauged catchment = 40 m3/s
The 50-year return flood for ungauged catchment = 40 x 4.8 = 192 m3/s

7.5.2 M e t h o d of L - M o m e n t s
The following sequential steps are followed:
(i) Test regional homogeneity for the selected gauged catchments using the
procedure described by Dalrymple (1960), and discard the catchments
which are not homogeneous.
(ii) Arrange the flood series, and compute L-moments.
(iii) Let the L-moments are Al,i, &,i, h,i,A4,? where i is the index of sites. At
each site i, the no. of records of flood data are ni.
(iv) The total data of all the site is N where N is equal to E n i where m is the
i=l
number of sites.
(v) Scale each of the linear moments by multiplying with the ratio njN.
(vi) Sum all such scale and this gives the mean annual flood.
274 Engineering Hydrology

(vii) Normalize the mean annual flood so that it becomes unity. In other words
assurne the mean flood is unity. This implies aggregated value of A, is 1.
(viii) Corresponding to A, = 1, obtain & and A,. Use these A values to estimate
PDF parameters.
(ix) Use PDF or CDF with the known parameters to compute a flood of return
period T such that the mean of such computed return period floods is
unity.
(x) Multiply the flood in step (ix) with the mean annual flood of ungauged
catchment. The product is the magnitude of the flood of ungauged
catchment having return period T.
To explain steps (i) to (x), an example is provided here.

Example 7.14 For North Cascades drainage basin, the lengths and L-moment
ratios for 19 sites are given in Table 7.12. Estimate regional parameters
assuming that PT-I11 distribution fits well to the given data.

Table 7.12 Summary statistics of North Cascades precipitation data

(Source: Hoskings and Wallis, 1997)
n 11 z z3 Col (2)x Col(3) x Col(4) x
Col (1)/1378 Col(1)/1378 Col(1)/1378
~~~

~~~
98 19.69 0.1209 0.0488 1.400305 0.008598 0.003471
~~~
59 62.58 0.0915 0.0105 2.679405 0.003918 0.00045
~~~
90 40.85 0.1124 0.0614 2.667997 0.007341 0.00401
~~~
61 46.05 0.1032 0.0417 2.038498 0.004568 0.001846
~~~
65 45.02 0.0967 -0.0134 2.123585 0.004561 -0.00063
~~~
86 31.04 0.1328 4.0176 1.937184 0.008288 -0.001 1
~~~
78 80.14 0.1008 0.0943 4.536226 0.005706 0.005338
72 41.31 0.1143 0.0555 2.158433 0.005972 0.0029
~~~
67 30.59 0.1107 0.0478 1.487322 0.005382 0.002324
~~~
99 32.93 0.1179 0.0492 2.365798 0.00847 0.003535
~~~
49 17.56 0.1308 0.094 0.624412 0.004651 0.003343
~~~
61 69.52 0.11 19 -0.0429 3.077446 0.004953 -0.0019
~~~
69 47.65 0.1018 0.0435 2.385958 0.005097 0.002178
~~~
73 102.5 0.1025 0.0182 5.429971 0.00543 0.000964
~~~
70 52.41 0.1054 -0.0224 2.662337 0.005354 -0.00114
~~~
66 79.7 0.1174 0.0124 3.817271 0.005623 0.000594
~~~
59 44.64 0.1115 -0.0346 1.911292 0.004774 -0.00148
~~~
74 58.66 0.1003 0.0446 3.150102 0.005386 0.002395
~~~
82 39.02 0.1046 0.0128 2.321945 0.006224 0.000762
Sum 941.86 Avg. Avg. Avg.
=1378 a, = 48.775 7 = 0.11029 7, = 0.027859

The regional normalized average mean, L-CV, L-skewness are: A, = 1, z = 0.11029,

z, = 0.027859
 Flood Frequency Analysis 275

Solution
For the data of Table 7.12, it is reported that PT-I11 distribution gives close
fit to the regional average L-moments. Thus, it is of interest to estimate
parameters of PT-I11 distribution.
To estimate these parameters, we use the following set of equations given
by Hoskings and Wallis (1997):
1 + 0.29062
I z + 0. 1882z2 + 0 . 0 4 4 2 ~ ~
(7.47)
0.360672 - 0 . 5 9 5 6 7 ~+~0 . 2 5 3 6 1 ~ ~
1- 2.788612 + 2 . 5 6 0 9 6 ~- 0~. 7 7 0 4 5 ~ ~
(i I
5 123 < 1)

where, z =3 4 ~ ~ ) ~

We find,
z = 3 z (0.027859)2 = 0.00731479
Using the value of z in Eq (7.47), the value of p = 136.811
Using the value of p in Eq (7.48), the parameters are obtained as p = 1.0,
(r= 0.1196, andg = 0.1709

[Hint: Using skewness value and recurrence interval, determine K, and thus
xT= 1 + K, 0. For 50-year return flood, K, = 2.143.
Thus, x, = 1.2563. If the mean annual flood of ungauged catchment is Q
m3/s,then the 50-yearreturn flood for the ungauged catchment will be 1.2563Q
m3/s.]

7.6 RISK AND RELIABILITY CONCEPTS

Basic questions to be answered in design are as follows:
1. Why should we do risk analysis?
2. What should be the return period for which the structure should be
designed?
3. What is the risk involved when we design a structure having a design life
of n years for a T-year return period flood?
4. How much risk is permissible?
7.6.1 Risk Analysis Parameters
If this design event has a return period of T years and a corresponding annual
probability of exceedence ofp, then:
276 Engineering Hydrology

1
P= (7.49)
The probability of non-occurrence in any one year is:
q=1-- 1
(7.50)
T
The probability of non-occurrence in n years is:

q = (1- +)n (7.51)

Hence, the probability that Xwill occur at least once inn years, i.e., the risk of
failure R is:

R=l-(l- +)n (7.52)

where, n is the design life of the structure.

Example 7.15 What will be the risk involved for a hydraulic structurehaving
a design life of 100 years if it is designed for: (i) 50-year return period flood,
and (ii) 1000-yearreturn period flood?
Solution
(i) n = 100 years, T = 50 years, and the risk (R) involved may be computed by
Eq. (7.52) as:

R = 1 - (1 - &) 100

= 0.867 = 86.7%
(ii) n = 100 years, T = 1000 years, and the risk involved is:

= 0.095 = 9.5%

Table 7.13 Return periods associated with various degrees of risk and
expected design life
Expected Design Life
Riskyo ~~~~~~~~

~~~~~~~~
2 S 10 15 20 25 so 100
75 ~~~~~~~~
2.00 4.02 6.69 11.0 14.9 18.0 35.6 72.7
50 ~~~~~~~~
3.43 7.74 11.9 22.1 29.4 36.6 72.6 144
40 ~~~~~~~~
4.44 10.3 20.1 29.9 39.7 49.5 98.4 196
30 6.12 14.5 28.5 42.6 56.5 70.6 140 281
(Contd.
 Flood Frequency Analysis 277

(Contd.)
25 ~~~~~~~~
7.46 17.9 35.3 52.6 70.0 87.4 174 348
20 ~~~~~~~~
9.47 22.9 45.3 67.7 90.1 112 224 449
15 ~~~~~~~~
12.8 31.3 62.0 90.8 123 154 308 616
10 ~~~~~~~~
19.5 48.1 95.4 142 190 238 475 950
5 ~~~~~~~~
39.5 98.0 195 292 390 488 976 194
2 99.5
~~~~~~~~
248 496 743 990 1238 2475 4950
1 198 498 996 1492 1992 2488 4975 9953

Based on the risk acceptable, the return period for which the structure should
be designed can be ascertained.

Example 7.16 What is the return period that a highway engineer must use in
his design of a critical underpass drain if he is willing to accept only:
(i) 10% risk that the flooding will occur in the next 5 years?
(ii) 20% risk that the flooding will occur in the next 2 years?
Solution
(i) The risk involved (R)is 0.10 and n = 2 years. Substitutingthese values in
the following equation:

R = 1 ~ (1 - + ) n

Or, T = 48.1 years

This means that there is 10% chance that a 48.1-year flood will occur
once or more in the next 5 years.
(ii) The risk involved (R) is 0.20 and n = 2 years.

0.20=1-
(1-- 3
Or, T = 9.47 years
It shows that there is a chance of 20% that a flood corresponding to a
return period of 9.47 years will occur in the next 2 years. These values
may also be read from the Table 7.13.
278 Engineering Hydrology

7.7 BINOMIAL DISTRIBUTION

The Binomial distributionis a discrete distribution and is based on the Binomial
theorem which states that probability of exactly x successes in n trials is:

(7.53)

where, (:)= n!
x!(n - x)! '
q=l-p (7.54)

p = probability of exceedence/success
q = probability of non-exceedence/failure
x = number of exceedence/successes
n = total number of events
The assumptions for binomial distribution are same as those for Bernoulli
trials. Tossing a coin and drawing a card from a pack are the common examples
of Bernoulli trials which operate under the following three conditions:

(i) Any trial can have success or failure (not both), true or false, rain or no
rain.
(ii) Successive trials are independent.
(iii) Probabilities are stable.

Binomial distribution is valid under the above three conditions.

Example 7.17 If a dam is having a project life of 50 years, then what is the
probability that a flood with a return period of 100 years will occur (i) once,
and (ii) twice during the life of the dam?
Solution
(i) Given, n = 50 years, T = 100 years, andp = probability of exceedence
1 1
Here, p = -= ~ = 0.01
T 100
= (1 -p) = (1 - 0.01) = 0.99

x= 1
~1 -X = 50 - 1=49

P(1) = :)(' (0.01)' (0.99)49

- 50! (0.01)' (0.99)49

1!(49)!
= 0.306 = 30.6%
 Flood Frequency Analysis 279

(ii) Given, n = 50 years, T = 100 years, and x = 2

n -x = 5 0 - 2 =48

P(2) = (2) (0.01)2(0.99)48

= 0.0756 = 7.56%

This means that there is a 30.6% chance that a 100-year return period
flood will occur once during the project life, and the chances of its twice
occurrence is 7.55%.

7.8 POISSON DISTRIBUTION

The terms of binomial expansion are a little inconvenient to compute for large
numbers. If n is large (> 30) andp is small (< 0. l), then binomial distribution
tends to Poisson distribution.
axe- (7.55)
P(x) = ~

X!
where, A = n p
The conditions for this approximation are:

(i) The number of events is discrete.

(ii) Two events cannot coincide.
(iii) The mean number of events in unit time is constant.
(iv) Events are independent.

Example 7.18 Solve Example 7.16 using Poisson distribution.

Solution
n=50 and p = 1/100

(50 x A) 1
e-(50/100)
= 0.303 = 30.3%
(i) P(l) = l!
2
(50x6) e- (50/100)
(ii) P(2) = = 0.0758 = 7.58%
2!
The results obtained by Binomial distribution and Poisson distribution are
almost same.
280 Engineering Hydrology

7.9 DESIGN FREQUENCIES

The designer is generally concerned with the return period for which the
structure should be designed. The design frequencies shown in Table 7.4 are
typical of levels generally encountered in minor structure design (Viessman
et al., 1977).
Probable Maximum Flood (PMP) Probable maximum flood is the flood
caused by probable maximum precipitation. Probable maximum flood is
generally obtained using unit hydrograph and rainfall estimates of the PMP.
Standard Project Floods (SPF) Standard project flood is the flood caused
by standard project storm which is generally obtained from a survey of severe
storms in the general vicinity of the drainage basin or severe storms experienced
in meteorologically similar areas.
Table 7.14 Minor structure design frequencies

I Type of Minor Structures I Return Period (in T years) I Frequency (f'= +] I

Highway crossroad drainage 10 0.1
0 4 0 0 ADT* 10-25 0.1-0.04
400-1700 ADT 25 0.04
170&5000ADT 50 0.02
5000 ADT 5 0.2
Airfields 2-1 0 0.5-0.1
Storm drainage 2-5 05-0.02
Levees 5-50 0.20-0.02
*AverageDaily Traffic

Category Impoundment danger Failure damage

potential
Storage
- I Height I Potential Spillway design
(Acve-feet) cft.) Loss of life Damage flood
Major 50,000 60 Considerable Excessive or Probable
as a matter of maximum
policy flood (a)
Intermediate 1000-50,000 40-100 Possible but - Standardproject

none
small flood Ib)
Minor 1000 Of same 5&100 years
magnitude as occurrence
~ 50 cost of dam interval
 Flood Frequency Analysis 281

Table 7.16 Frequency of return period advocated by IBS code

Structure Return Period of
design flood (years)

Minor dams with storage less than 6000 ha.m. 1oo*

I Barrages and pick up weirs I I
(a) Free board 500+
(b) Items other than free board 50-100
River Training Works (Calculation of Scour) 50-loo@
Water way of bridges 50
* IS 5477 Method of fixing capacities of reservoirs-Part IV
-

+IS 6966 Criteria for hydraulic design of barrages and weirs

-

@ IS 3408 Criteria for river training works for barrages and weirs in alluvium
-

In case of spillwaysfor major and medium reservoirs with storage capacities

of more than 6000 ha.m., designing is generally done for the probable maximum
flood which is the physical upper limit of the flood in the catchment.

Table 7.17 CWC, India recommends the following design frequencies

Structure I Return period
Major dam with storage more than Probable maximum flood or frequencynot
50,000 acre feet less than once in 100 years
Permanent barrages and minor dams with Standard project flood or 100 year flood
storage less than 50,000acre feet whichever is higher
Pick up weirs 50-100 years
Canal aqueducts:
(a) Waterways 5C100 years
(b) Foundations and freeboard 5C500 years

7.10 PEAK OVER THRESHOLD (POT) MODELS

Two approaches commonly used for probabilistic analysis of extreme flood
magnitudes are based on annual maximum series (AMS) and partial duration
series (PDS). The AMS model considers the annual maximum flood of each
year in a series that has as many elements as there are years in the data record.
In contrast, the PDS model considers all flood peaks above a certain threshold
level. Clearly, in PDS model, more than one flood per year may be included;
hence, it is also referred as peak over threshold (POT) method. For this reason,
PDS models are used for flood frequency analysis (FFA) where the flood record
is short.
In the limit, when the threshold is increased, recurrence interval of large
events computed using AMS and PDS models tend to converge. The methods
followed for formulating these models are quite different; e.g., an AMS model
282 Engineering Hydrology

uses a cumulative distribution function (CDF) to model the exceedences,

whereas the PDS model uses two probabilistic models: (a) one for the
probability of occurrence of peaks above a threshold, and (b) the CDF, for
modelling the flood exceedences. Thus, for a fixed size of observation years,
the sample size in PDS method is a random variable. We shall consider the
approach that is employed in formulationof such a model (Madsen et al. 1997).
7.10.1 Models for PDS Analysis
In PDS model, all peak events above a threshold level are considered, and the
number of these occurrences (x) in a given year is assumed to follow Poisson
distribution having the probability mass function:
e-pp'"
P(x = ). = -; r = 0 , 1 , 2 , ... (7.56)
r!
where, P(x = r) is the probability of r events in a year, p is the average number
of threshold exceedences. For n number of observed exceedences in N years,
p is equal to nlN, however, Eq. (7.56) is true as long as the flood peaks are
independent. For Poisson distribution, PWM estimations and the population
estimate of p is given by:
n
iu= (7.57)
The mean and variance of r are defined as follows:
E(r) = p = n/N = Var(r) = o2; (n 2 5) (7.58)
For p > 5, the distributionof r is symmetrical,and it asymptoticallyapproaches
a normal distribution (Stark and Woods, 1986). For no flood exceedences to
occur in a given year, substituting r = 0 in Eq. (7.56), the following is obtained.

(7.59)
which is the PDF of an exponential distribution. The cumulative mass function
of Eq. (7.56) for N years is given by:
F=l-pC@ ; (N>O) (7.60)
Assume that a threshold level qo is chosen, corresponding to a mean annual
number of exceedences of Poisson distribution p ; then at any higher threshold
(q > qo),the number of exceedencesr' is Poisson distributedwith the following
parameter (Flood Studies Report, 1975):
P' = P[l - F(q)l (7.61)
Since F(q) I 1, it follows that (p' I p ) , i.e., the truncated mean of exceedences
above the new threshold q decreases, which is quite obvious.
The AMS model used for determinationof T-year return period flood [q(T)]
requires identification of a statistical distribution so as to obtain the [q(T) - r]
relationship in explicit form.
 Flood Frequency Analysis 283

If n floods exceed qo in N years of record in PDS, the average rate of

occurrence is p = n/N. Consequently, the flood which has a return period T
sampling units in this PDS has a return period of approximately T = T/p
years. Thus, the PDS method uses two probabilistic models:

(a) one for the probability of occurrence (PA)for p exceedences in each year,
and
(b) the statistical distribution Ffi,yof the maximum of n exceedences, where
Y = 4 -40.
The probability of annual maxima is obtained using the CDF, computed using
the total probability theorem (Shane and Lynn, 1964) as:

(7.62)
r=O

If Pfifollows Poisson distribution, and Ffi,yfollows an exponential distribution

= 1 exp [+q
- q0)/P]with parameter P, then
-

(7.63)
F(q) given by Eq. (7.63) is the Gumbel or EV-I distribution that is widely used
in FFA. From Eq. (7.63), the following PDS model can be obtained.
q(T) = 40 + Pun ( 4 1 + PYT (7.64)
where, yT = -In [-ln( 1 l/T)] is the Gumbel reduced variate.
-

Since the T-year return period is defined as the average interval of time within
which the magnitude of q of the flood event will be equaled or exceeded once,
the POT model should be preferred over the AMS model for estimating the
q-Trelationship. Thus, a comparison of the AMS and PDS model is of practical
interest.

Example 7.19 The daily flow records of a catchment were available for
18 years. From the data record, all the flow values crossing a threshold of
40 m3/s are given in Table 7.18 These thresholds are to be used to develop a
PDS model using Poisson distribution (PD) for counting the peaks exceeding
the threshold and exponential distribution (ED) is to be used for modelling the
exceedences. Use the derived PDSPD-ED model to find the 50-year return
period flood. Assume mean exceedences @) = 3.
284 Engineering Hydrology

Rank (i) Q (m3/s) F yi Rank (i) Q (m3/s) F Yi

1 276.92 0.991 4.659 31 63.36 0.483 0.660
2 237.5 0.974 3.635 32 63.36 0.466 0.628
3 I 136.35 I 0.957 I 3.140 I 33 I 62.95 I 0.449 I 0.596
4 I 131.35 I 0.940 I 2.810 I 34 I 62.95 I 0.432 I 0.566
5 124.04 0.923 2.562 35 62.54 0.415 0.537
6 118.12 0.906 2.364 36 61.31 0.399 0.508
7 116.38 0.889 2.199 37 61 0.382 0.481
8 I 114.07 I 0.872 I 2.057 I 38 I 61 I 0.365 I 0.454
9 I 109.54 I 0.855 I 1.932 I 39 I 61 I 0.348 I 0.427
10 I 97.61 I 0.838 I 1.822 I 40 I 59.3 I 0.331 I 0.402
11 96.04 0.821 1.722 41 59.3 0.314 0.377
12 93.58 0.804 1.632 42 59 0.297 0.352
13 91.91 0.788 1.549 43 57.32 0.280 0.329
14 I 90.91 I 0.771 I 1.472 I 44 I 56.15 I 0.263 I 0.305
15 90.91 0.754 1.401 45 55 0.246 0.283
16 86.92 0.737 1.335 46 55 0.229 0.261
17 85.94 0.720 1.273 47 54.23 0.212 0.239
18 I 84.96 I 0.703 I 1.214 I 48 I 54.23 I 0.196 I 0.218
19 I 82.07 I 0.686 I 1.159 I 49 I 54.23 I 0.179 I 0.197
20 81.12 0.669 1.106 50 52.35 0.162 0.176
21 80.17 0.652 1.056 51 51.61 0.145 0.156
22 80.17 0.635 1.009 52 51.61 0.128 0.137
23 I 78.3 I 0.618 I 0.963 I 53 I 51.24 I 0.111 I 0.118
24 I 71.94 I 0.601 I 0.920 I 54 I 50.87 I 0.094 I 0.099
25 70.61 0.585 0.878 55 50.14 0.077 0.080
26 67.57 0.568 0.839 56 49.78 0.060 0.062
27 67.14 0.551 0.800 57 49.06 0.043 0.044
28 I 67.14 I 0.534 I 0.763 I 58 I 49.06 I 0.026 I 0.027
29 I 65.45 I 0.517 I 0.728 I 59 I 48 I 0.009 I 0.010
30 I 64.61 I 0.500 I 0.693 I I I 1
Solution
Mean of the partial duration series ( 4 ) = Mean of the data in Column (2)
= 82.23 m3/s.

N = (No. of years of AMS data) x ,u = 3 x 18 = 54

That means the threshold (qo)from the table is equal to 50.87 m3/s (marked as
bold in the Table).
By MOM estimation: p = N ( 4 - qo)/(N- 1) = 3 1.95 m3/s
Hence, q (50y) = 50.87 + 31.95 [In (3)] + 31.95y5, = 210.3 m3/s.
b 5 0 =-In {-ln (1 - 1/50)} = 3.91
 Flood Frequency Analysis 285

SUMMARY

In this chapter, different methods of flood parameter estimation have been

discussed. Flood estimation using normal, lognormal, Pearson type,
log-Pearson, and Gumbel distribution have been explained in detail. For
designing any hydraulic structure, flood fi-equency analysis and risk factor
and reliability of the structure in frequency analysis are generally considered.
All these areas have been covered in this chapter.
The designer is generally concerned with the return period for which the
structure should be designed. Depending on the available data at any site, it is
possible to fit a suitable probability distribution function. Subsequently, this
function can be used to estimate flood of an assumed return period. Use of
PDS models is also highlighted. Using such models, all flood peaks above a
certain threshold level can be analysed. This approach can be more useful
when the data on annual maximum flood is limited.

EXERCISES

350 50
275 25

Find out the amount of peak flood corresponding to 100-years return

period. Ans. 384.76 m3/s
7.2 A bridge is to be constructed over a river for a design flood of
500 m3/s. The mean and the standard deviation for annual flood series
are 250 m3/s and 100 m3/s, respectively. Calculate the return period
corres-ponding to this design flood using Gumbel method.
Ans. 44.47 years
7.3 Let the probability of rainfall occurrence on any day in the month of
May is 0.3. If you are wishing a no-rainfall 7-day period as you are
planning for a 7-day study tour, (i) find the probability that the rainfall
will not occur in these 7 days; (ii) find the probability of 1 rainy day
during these 7 days. Ans. (i) 0.0824 (ii) 0.2471
7.4 What would be the return period of a design storm to be used for the
design of a hydraulic structure? Let there be a probability of 25% that
the storm will occur in the next 25 years. Ans. 87.40 years
286 Engineering Hydrology

7.5 The return period ( T ) and risk of failure (R) are related as:

R = 1 - (1 - $r
where, n is the number of years. Plot a graph between R and T for
n = 10,25,50, 100, and 150.
(i) Observe the nature of the graphs.
(ii) Find the risk of failure that a 50-year flood will occur at least once
in the next 10 years. Ans. 0.183 or 18.3%
7.6 A stream has the flood peaks of 1200 m3/s and 1060 m3/s for the return
periods of 100 and 50 years, respectively, using Gumbel method for a
dataset spanning 30 years. Find (i) mean and standard deviation; and
(ii) flood peak corresponding to a return period of 500 years.
Ans. 425.85 m3/s, 257.24 m3/s, 1523.7 m3/s
7.7 The mean, standard deviation, and coefficient of skewness of
log-transformedannual maximum peak flood series of a typical gauging
site are given below:

MI
l
j Mean (m3/s)
Standard deviation (m3/s)
Coefficient of skewness
700.25
255.6

Estimate 100-year floods assuming that the peak discharge data follow
EV-I distribution.Assume the peak discharge follows other distributions
and compare their results. Ans. 1501.97 m3/s
7.8 Using Poisson distribution, find: (i) the probability that a 6-year flood
will occur 7 times in a period of 20 years, (ii) the probability that it will
not occur at all, and (iii) how many floods of this magnitude will occur
on an average during 20 years? Compare the results with those obtained
from binomial distribution. Ans. (i) 0.0322 (ii) 0.0358 (iii) 0 3.33
7.9 Suppose you are going to design a check dam for a flood having a return
period of 100 years. What are the risks that you are going to accept for
flooding to occur in the next 5 years? Ans. 0.049
7.10 A temporary dam is to be built to protect a 5-year construction activity
for a hydroelectric power plant. If the dam is designed to withstand
25-year flood, determine the risk that the structure will be overtopped:
(i) at least once in the 5-year construction period, and (ii) not at all
during the 5-year period. Ans. (i) 0.185 (ii) 0.815
 Flood Frequency Analysis 287

OBJECTIVE QUESTIONS

1. Which is not a parameter estimation technique?

(a) Graphical method (b) Least squares method
(c) Method of moments (d) None of these
2. For normal and lognormal distribution,the recommended plotting position
formula is
(a) Blom (b) Gringorton
(c) Cunnane (d) None of these
3. Coefficient of correlation is
(a) Equal to the coefficient of determination
(b) Square root of the coefficient of determination
(c) Cubic root of the coefficient of determination
(d) None of the above
4. The relationship between the return period ( T ) and the probability of
exceedence (p) is
1
(a) p = 1 - T (b) P =
(c) T = 1 - p (d) None of these
5. A dam is having a project life of 50 years. What is the probability of flood
that does not exceed having a return period of 100 years?
(a) 0.01 (b) 0.99
(c) 0.9 (d) None of these
6. What is the return period that a designer must use in a dam design if
he/she is willing to accept: (i) only 5% risk that the flooding will occur in
the next 25 years.
(a) 100 (b) 250 (c) 487 (d) 500
7. For an annual flood series arranged in decreasing order of magnitude, the
return period for a magnitude listed at position m in a total of N entries is
(a) m/N (b) m/(N + 1 ) (c) ( N + l)/m (d) N/(m + 1 )
8. Probable maximum flood is
(a) An extremely large flood
(b) The flood caused by probable maximum precipitation
(c) Standard project flood
(d) All of the above
288 Engineering Hydrology

9. As per the CWC recommendation, in the design of permanent barrages

and minor dams with storage capacity of less than 50,000 acre feet, the
return period is considered as
(a) Probable maximum flood
(b) Standard project flood
(c) Standard project flood or 100-year flood, whichever is higher
(d) None of the above
10. The regional flood frequency analysis approach is used to estimate the
flood
(a) For gauged catchment (b) For ungauged catchment
(c) Both (a) and (b) (d) None of these

REFERENCES

Bhunya, P.K., S.K. Jain, C.S.P. Ojha, and A. Agarwal, A simple estimation
technique for three parameter GEV distribution, J. Hydrologic
Engineering, ASCE, 12(6), pp. 682-689, 2007.
Chow, V.T., 1964,Handbook ofApplied Hydrology. McGraw-Hill, New York.
Central Water Commission, India, Estimation of design Flood Peaks, Flood
Estimation Directorate, Report No. 1/73, New Delhi, 1973.
Dalrymple, T., 1960,FloodFrequency Methods, U. Survey Water Supply Paper
1543-A, pp. 11-5 1, Washington.
Greenwood, J.A., J.M. Landwehr, N.C. Matalas, and J.R. Wallis, 1979,
Probability Weighted Moments: Definition and Relation to Parameters
of Several Distibutions Expressible in Inverse Form, Water Resources
Research, Vol. 15, No. 5, pp. 1049-1054.
Hosking, J.R.M., and J.R. Wallis, 1997, Regional frequency analysis: An
approach based on L-moments, Cambridge University Press, UK.
Hosking, J.R.M., J.R. Wallis, and E.F. Wood, 1985, Estimation of the
Generalized Extreme Value Distribution by the Method of Probability
Weighted Moments, Technometrics, 27, pp. 25 1-26 1.
Hosking, J.R.M., 1990, L-Moments:Analysis and Estimation of distributions
using Linear Combinations of Order Statistics, J. of Royal Statistical
Society B, Vol. 52, pp. 105-124.
Landwehr, J. M., N. C. Matalas., and J. R. Wallis, 1979, Probability weighted
Moments Compared with Some Traditional Techniques in Estimating
Gumbel Parameters and quantiles, Water Resources Research, 15(5),
pp. 1055-1064.
Madsen, H., P.F. Rasmussen, and D. Rosbjerg, 1997, Comparison of annual
maximum series and partial duration series methods for modelling extreme
hydrological eventsIAt-site modelling, WaterResources Resarch., 33(4),
pp. 747-757.
 Flood Frequency Analysis 289

National Environmental Research Council (NERC), 1975,Flood studies report,

Vol-1, NERC, UK.
Pickands, J., 1975, Statistical inference using extreme order statistics,
Ann. Stat. 3(1), pp. 119-131.
Raghunath, H.M., 2007, Hydrology Principles, Analysis and Design, New
Age International (P) Limited publishers.
Shane, R.M., and Lynn, W.R., 1964. Mathematical model for flood risk
evaluation, J. Hydraulics Div., ASCE, 90 (HY6), pp. 1-20.
Stark, H. and J. Woods, 1986,Probability, Random Processes, andEstimation
Theoryfor Engineers, Prentice-Hall, Englewood Cliffs.
Stedinger, J.R., R.M. Vogel, and E. Foufoula-Georgiou, 1993, Frequency
analysis of Extreme Events,Handbook ofHydrology, ed. D.R.Maidment,
McGraw-Hill, New York.
Subramanya,K., 1994,Engineering Hydrology, Tata McGraw-Hill,New Delhi.
Swamee, P.K. and P.N. Rathie, 2007, Invertible alternatives to normal and
lognormal distributions, J. of Hydrologic Engineering, ASCE, 12(2),
pp. 218-221.
Viessman Jr., W., G. L. Lewis, and J.W. Knapp, 1989, Introduction to
Hydrology, 3rd Ed., Harper and Row, New York.