TOPIC: ANALYSIS OF HYDROLOGIC DATA

 Frequency Analysis
 Normal Distribution
 Gumbel Distribution
 Lognormal Distribution
 Gamma/Pearson Type III Distribution
 Log Pearson Type III Distribution

PROBABILITY AND STATISTICS IN WATER-RESOURCES ENGINEERING

Introduction
 All hydrologic phenomena obey the laws of nature and can, in theory, be described by solutions of the relevant
fundamental equations. In most cases, however, the uncertainty and amount of spatial and temporal detail that
needs to be known about the physical system, combined with the limitations of analytical and numerical
techniques for solving systems of equations, preclude an exact description of most natural phenomena.
Processes that can be specified with certainty are called deterministic processes, while those that cannot are
called stochastic processes.
 The two types of stochastic processes commonly encountered in engineering practice are:
(1) processes where the outcome is uncertain because of the lack of information on the process or
parameters affecting the process, called epistemic uncertainty, and
(2) processes where the outcome is uncertain because the process generating the outcome is
fundamentally random, called natural uncertainty.

 In the first case, epistemic uncertainty can be reduced, or even eliminated, by additional measurements of the
variables affecting the process or by the use of more accurate models.
CE 521In–the second
WATER case, naturalENGINEERING
RESOURCES randomness is intrinsic to the process and uncertainty in the outcome might not be 1
reduced by additional measurements.
 In many cases, recorded data include both epistemic and natural uncertainty, and separation of these
components is a challenging task. An example of this is in the analysis of climatic data, where separation of
random climate variability from more predictable long-term climate changes is a common challenge. Whether
uncertainty is epistemic, natural, or a combination of both, an uncertain outcome is generally defined by a
probability distribution that describes the relative likelihood of all possible outcomes. In practice, uncertain
variables are collectively referred to as random variables.
 A common application of probability theory in water-resources engineering involves the assignment of an
exceedance probability, Pe, to hydrologic events.
 A system designed to handle a particular hydrologic event can be expected to fail with a probability equal to the
probability of exceedance, Pe, of the design event. This probability of system failure is called the risk of failure,
and the probability that the system will not fail, 1−Pe, is called the reliability of the system. Exceedance
probabilities of hydrologic events are usually stated in terms of the probability that a given event is exceeded in
any given year. The average number of years between exceedances is called the return period, T, which is
related to the annual exceedance probability, Pe, by
1
T=
Pe
The return period, T, is sometimes called the average recurrence interval (ARI) and is usually a criterion for
selecting design events in water-resource systems. In cases where water resource systems are to be in place for
only a short period of time (less than a year), it might be appropriate to calculate return periods of design events
based on their occurrence within a specific season of the year (e.g., McCuen and Beighley, 2003).
 The determination of flows corresponding to a given return period is also of central importance in the
restoration of natural channels, since it is commonly assumed that natural channels evolve to contain surface
runoff having approximately a 2-year return period (e.g., He and Wilkerson, 2011).

1|Page
  If a water-resource system is designed for a hydrologic event with exceedance probability, Pe, and the system
has a design life of N years, then the risk that the system will fail within its design life, Pf, is given by
Pf =1−(1−Pe) N
This formulation is particularly useful in assessing the cost of failure, since the expected damage cost is equal to
the probability of failure multiplied by the cost associated with the failure.

8.3 FREQUENCY ANALYSIS

o Frequency analysis is concerned with estimating the relationship between an event, x, and the return period, T,
of that event.
o Recalling that T is related to the exceedance probability of x, P X(X≥ x), by the relation
1
T=
Px
then T is related to the cumulative distribution function of x, P X(X < x), by the relation
1
T=
1−P X ( X−x )
o A fundamental component of frequency analysis is the assumption that the underlying probability distribution of
an event is known. However, in reality, the underlying probability distribution of an observed event can only be
hypothesized, and the adequacy of this hypothesis must be assessed using tests such as the Chi-Square and
Kolmogorov–Smirnov tests before proceeding with a frequency analysis based on an assumed probability
distribution.
o The uncertainty of sample estimates of parameters in probability distributions requires caution when using data
records shorter than 10 years, and caution should also be exercised when estimating the values of variables with
recurrence intervals longer than twice the record length (Viessman and Lewis, 1966; Davie, 2002). The
conventional frequency analyses presented here are based on an assumed population probability distribution,
CE 521however,
– WATER alternative,
RESOURCES and ENGINEERING
sometimes more accurate, nonparametric methods using artificial neural networks 2
have also been proposed for use in frequency analysis (e.g., He and Valeo, 2009).
o Many cumulative distribution functions cannot be expressed analytically and are tabulated as functions of
normalized variables, X’, where
X −μ x
X'=
σx
and μ x and σ x are the mean and standard deviation of the population, respectively. The cumulative probability
distribution of X’, P X ' (X’ < x’), for many distributions depends only on x’ and the skewness coefficient, g x, of the
population and can be readily obtained from statistical tables. Denoting the value of X’ with return period T by
x’T, for many distributions
x 'T =K T ( T , g x )
where K T ( T , g x )is derived from the cumulative distribution function of X’ and is commonly called the frequency
factor. Combining Equations 8.140 and 8.141 leads to
x T =μ x + K T σ x
where x T is the realization of X with return period T. The frequency factor is applicable to many, but not all,
probability distributions. The frequency factors for a few probability distributions that are commonly used in
hydrologic practice are described here.

8.3.3.1 NORMAL DISTRIBUTION

o In the case of a normal distribution, the variable X’ is equal to the standard normal deviate and has a N(0, 1)
distribution. The cumulative distribution function of the standard normal deviate is given in Appendix C.1, and
the frequency factor is equal to the standard normal deviate corresponding to a given exceedance probability or
return period.
o The frequency factor, K T , can also be approximated by the empirical relation (Abramowitz and Stegun, 1965)

2|Page
 2.515517+ 0.802853 w+0.010328 w 2
K T =w− ( Eq . 8.143)
1+ 1.432788 w+0.189269 w 2+0.001308 w3

Where
w=¿

and p is the exceedance probability (= 1/T). When p > 0.5, 1 − p is substituted for p in Equation above and the
value of KT computed using Equation 8.143 is given a negative sign. According to Abramowitz and Stegun (1965),
the error in using Equation 8.143 to estimate the frequency factor is less than 0.00045. Alternative expressions
for the normal-distribution frequency factor (Equation 8.143) having comparable or lesser accuracy have also
been proposed (e.g., Swamee and Rathie, 2007b).

EXAMPLE 8.24
Annual rainfall at a given location can be described by a normal distribution with a mean of 127 cm and
a standard deviation of 19 cm. Estimate the magnitude of the 50-year annual rainfall.
Solution:
 The 50-year rainfall, x 50, can be written in terms of the frequency factor, K 50, as
x 50=μ x + K 50 σ x
 From the given data, μ x =127 cm, σ x =19 cm, and the exceedance probability, p, is given by
1 1
P= = =0.02
T 50
 The intermediate variable, w, is given by Equation 8.144 as
w=¿
 Equation
CE 521 – WATER 8.143 gives
RESOURCES the frequency factor, K 50, as
ENGINEERING 3
2.515517+ 0.802853 w+0.010328 w 2
K T =w−
1+ 1.432788 w+0.189269 w 2+0.001308 w3
2.515517+0.802853 w+ 0.010328 ( 2.797 )2
K 50=2.797− 2 3
1+ 1.432788 ( 2.797 )+ 0.189269 ( 2.797 ) +0.001308 ( 2.797 )
¿ 2.054
 The 50-year rainfall is therefore given by
x 50=μ x + K 50 σ x =127+2.054 ( 19 ) =166 cm

8.3.3.2 LOG-NORMAL DISTRIBUTION

 In the case of a log-normal distribution, the random variable is first transformed using the relation
Y =ln x( Eq . 8.145)
y
and the value of Y with return period T, T , is given by
y T =μ y + K T σ y ( Eq. 8.146)
where μ y and σ y are the mean and standard deviation, respectively, of Y, and K T is the frequency factor of the
standard normal deviate with return period T. The value of the original variable, X, with return period T, x T , is
then given by
y T =e yt
−1
x T =ln
EXAMPLE 8.25
The annual rainfall at a given location has a log-normal distribution with a mean of 127 cm and a standard
deviation of 19 cm. Estimate the magnitude of the 50-year annual rainfall.

Solution:
From the given data, μ x = 127 cm, σ x = 19 cm, and the mean, μ y , and standard deviation, σ y, of the log-
transformed variable, Y = ln X, are related to μ x and σ x by Equation 8.75, where
3|Page
 σ 2y
(
μ x =exp μ y +
2 ),σ 2x =μ2x ¿ ¿
These equations can be put in the form
μ 4x σ 2x + μ2x
1
2 ( )
μ y = ln 2 2 , σ y = ln
μ x +σ x
which lead to
μx
2
√
1 127 4 19 2+127 2
μ y = ln
2 (
1272 +192
=4.83 ,σ y
)= ln
127 2 √=0.149

Since the rainfall, X, is log-normally distributed, Y (=lnX) is normally distributed and the value of Y corresponding
to a 50-year return period, y50, is given by
y 50=μ y + K 50 σ Y
where K50 is the frequency factor for a normal distribution and a return period of 50 years. In the previous
example it was shown that K50 =2.054, therefore
y 50=4.83+2.054 ( 0.149 )=5.14
and the corresponding 50-year rainfall, x 50, is given by
x 50=e =e 5.14=171 cm
y 50

8.3.3.3 Gamma/Pearson Type III distribution

 In the case of the Pearson Type III distribution, also called the three-parameter gamma distribution, the
frequency factor depends on both the return period, T, and the skewness coefficient, g x. If the skewness
coefficient falls between −1 and +1, approximate values of the frequency factor for the gamma/Pearson Type III
distribution, KT, can be estimated using the relation (Kite, 1977; Viessman and Lewis,2003)
1 '
CE 521K– =
WATER
T −k ) k +1 ¿3−1
{[
xRESOURCES T
ENGINEERING
( Eq .8.148) } 4
3k
where x’T is the standard normal deviate corresponding to the return period T and k is related to the skewness
coefficient by
gx
k= ( Eq. 8.149)
6
when the skewness, gx , is equal to zero, the expanded form of equation 8.148 indicates that K T =x ' T , and the
gamma/Pearson Type III distribution is identical; to the normal distribution. In cases where the skewness
coefficient falls outside the -1 to +1 range, the frequency factors given in Appendix C.2 should be used. Methods
for calculating the confidence limits of the Pearson Type III distribution can be found in Ames (2006).

EXAMPLE 8.26
The annual-maximum discharges in a river show a mean of 811 m 3/s, a standard deviation of 851 m 3/s,
and a skewness of 0.94. Assuming a Pearson Type III distribution, estimate the 100-year discharge. If subsequent
measurements indicate that the skewness coefficient is actually equal to 1.52, how does this affect the estimated
100-year discharge?
Solution:
From the given data: μ x = 811 m3/s, σ x = 851 m3/s, and g x = 0.94. The standard normal deviate corresponding
g x 0.94
to a 100-year return period, x ' 100, is 2.33 (Appendix C.1), k = = =0.157 , and the frequency factor K100 is
6 6
given by Equation 8.148 as
1 1
K 100 = { [ x 'T −k ) k +1 ¿3−1 }= ¿
3k 3(0.157)
The 100-year discharge, x100, is therefore given by
x 100 =μ x + K 100 σ x =811+ ( 3.00 ) ( 851 )=3 , 364 m3 /s

4|Page
 If the skewness is equal to 1.52, this is outside the [−1, +1] range of applicability of Equation 8.148, and the
frequency factor must be obtained from the tabulated frequency factors in Appendix C.2. For a return period of
100 years and a skewness coefficient of 1.52, Appendix C.2 gives K 100 =3.342, and the 100-year discharge is given
by
x 100 =μ x + K 100 σ x =811+ ( 3.342 )( 851 ) =3 ,655 m 3 / s
Therefore, the 62% change is estimated skewness from 0.94 to 1.52 causes a 9% change in the 100-year
discharge from 3364 m3/s to 3655 m3/s.
8.3.3.4 LOG-PEARSON TYPE III DISTRIBUTION
 In the case of a log-Pearson Type III distribution, the random variable is first transformed using the relation
Y = ln X ( Eq. 8.150)
 The value of Y with return period T, y T, is given by
y T =μ y + K T σ y (Eq 8.151)
where μ y and σ yare the mean and standard deviation, respectively, of Y, and KT is the frequency factor of the
Pearson Type III distribution with return period T and skewness coefficient g y . The value of the original
variable, X, with return period T, x T ,is then given by
x T =ln−1 y T =e y (Eq . 8.152)
T

whenever the skewness of Y (=lnX) is equal to zero, the log-Pearson Type III distribution is identical to the log-
normal distribution.

EXAMPLE 8.27
The natural logarithms of annual-maximum discharges (in m 3/s) in a river show a mean of 6.33, a standard
deviation of 0.862, and a skewness of −0.833. Assuming a log-Pearson Type III distribution, estimate the 100-year
discharge.
Solution:
CE 521From
– WATER RESOURCES g x −0.833
the given data, μ y =ENGINEERING
6.33, σ y = 0.862, and g y =− 0.833, which yields ¿ = =−0.139. The 5
6 6
standard normal deviate corresponding to a 100-year return period, x ' 100, is 2.33 (Appendix C.1), and the frequency
factor for a 100-year return period, K100, is given by Equation 8.148 as
1 1
K 100 = { [ x 'T −k ) k +1 ¿3−1 }= ¿
3k 3(−0.139)
The 100-year log-discharge, y100, is therefore given by
y 100 =μ y + K 100 σ y =6.33+ ( 1.72 ) ( 0.862 )=7.81
and the 100-year discharge, x100, is
x 100 =e =e7.81=2,465m 3 / s
y100

The U.S. Interagency Advisory Committee on Water Data published an important document called Bulletin
17B(USIAC,1982), which provides detailed guidance on the application of the log-Pearson Type III distribution to
annual-maximum flood flows. Of particular concern in applying the log-Pearson Type III distribution to observed data
is that skew coefficients estimated from small samples might be highly inaccurate. In cases where this is a concern,
Bulletin 17B outlines a procedure for developing regional skew coefficients based on at least 40 gaging stations
located within a 160-km (100-mi) radius of the site of concern, with each station having at least 25 years of data. If
this procedure is not feasible, the generalized skew map shown in Figure 8.12 is provided as an easier but less
accurate alternative. This map of generalized logarithmic skew coefficients for peak annual flows was developed
from skew coefficients computed at 2972 gaging stations, all having at least 25 years of record, following the
procedures outlined in Bulletin 17B. Depending on the number of years of record, regionalized skew coefficients are
used either in lieu of or in combination with values computed from observed flows at the particular stream gage of
concern. Equation 8.153 allows a weighted skew coefficient G W to be computed by combining a regionalized skew
coefficient GR and station skew coefficient G,
( MSE¿¿ S)(G R )
G w =( MSE R ) ( G )+ ( Eq .8.153)¿
MSE R + MSE S

5|Page
 where the station skew G is computed from observed flows at the station of interest; the regional skew, G R, is either
developed from multiple stations following procedures outlined in Bulletin 17 Borread from Figure 8.12; MSE S
denotes the mean-square error of the station skew, which can be estimated for N years of data using the relation
(Wallis et al., 1974)
MSE s=10¿ ¿
where
A={−0.33+ 0.08|G|if |G|≤0.90
−0.52+0.30|G|if |G|> 0.90

B={ 0.94−0.26|G|if |G|≤ 1.50

0.55if |G|>1.50

and MSER is the mean-square error of the regional skew. If G R is taken from Figure 8.12, then MSE R is equal to 0.302.
Recent studies have further validated using this method to estimate the skewness of log-transformed annual peak
flows in the United States (Griffisand Stedinger, 2007c; 2009). However, issues still remain as to the fundamental
validity of using regional skews that neglect local storage effects (McCuenandSmith,2008), and the need for more
formal criteria in selecting the most appropriate regional data to supplement local data (Chebana and Ouarda, 2008)

CE 521 – WATER RESOURCES ENGINEERING 6

Figure 8.12: Generalized skew coefficients of the logarithms of annual-maximum streamflow Source: Interagency Advisory Committee on Water
Data (Bulletin 17B, 1982).

EXAMPLE 8.28
Observations of annual-maximum flood flows in the Ocmulgee River near Macon, Georgia, from1980– 2000
yield a skew coefficient of the log-transformed data equal to −0.309. Determine the (weighted) skew coefficient that
should be used for flood-frequency analysis.
Solution:
In central Georgia, where Macon is located, Figure 8.12 gives a generalized skew coefficient G R of−0.1, and MSER
is equal to 0.302. From the observed annual-maximum flood flows, G=− 0.309, and for the 21-year period of record
(1980–2000), N =21. By definition
A =− 0.33 + 0.08|G|
=− 0.33 + 0.08|−0.309|

6|Page
 =− 0.305
B = 0.94 − 0.26|G|
= 0.94 − 0.26|−0.309|
= 0.860
¿¿
MSES = 10
and the weighted skew coefficient is given by
( MSE¿¿ S)(GR ) ( 0.302 )(−0.309 )+(0.262)(−0.1)
Gw =( MSE R ) ( G )+ = =−0.212 ¿
MSE R + MSE S 0.302+ 0.262
Therefore, a skew coefficient of −0.212 is more appropriate than the measured skew coefficient of −0.309 for
frequency analyses of flood flows in the Ocmulgee River at Macon, Georgia. As the period of record increases, it is
expected that the measured skew coefficient will approach the generalized skew coefficient derived from Figure
8.12.

STATISTICAL TABLES
C.1 AREAS UNDER STANDARD NORMAL CURVE

CE 521 – WATER RESOURCES ENGINEERING 7

7|Page
CE 521 – WATER RESOURCES ENGINEERING 8

8|Page
CE 521 – WATER RESOURCES ENGINEERING 9

9|Page
 CE 521 – WATER RESOURCES ENGINEERING 10

10 | P a g e
PROBLEMS:
1. The annual rainfall at a given location is normally distributedwithameanof152 cm and a standard deviation of 30
cm. Estimate the magnitude of the 10-year and 100-year annual rainfall amounts.
2. The annual rainfall at a given location has a log-normal distribution with a mean of 152 cm and a standard
deviation of 30 cm. Estimate the magnitude of the 10-year and 100-year annual rainfall amounts. Compare your
results with those obtained in Problem 1.
3. The annual-maximum discharges in a river show a mean of 480 m 3/s, a standard deviation of 320 m3/s, and a
skewness of 0.87. Assuming a Pearson Type III distribution, use the frequency factor to estimate the 50-year and
100-year annual-maximum discharges.
4. Compare the Pearson Type III frequency factor for a 100-year event estimated using Equation 8.148 with the
same frequency factor estimated using Appendix C.2, for skewness coefficients of 3.0, 2.0, 1.0, 0.5, and 0.0. How
do these results support the assertion that Equation 8.148 should be used only when the skewness coefficient is
in the range [−1, +1]?
5. The annual-maximum discharges in a river show a mean of 480 m 3/s, a standard deviation of 320 m3/s, and a
log-transformed skewness of 0.1. Assuming a log-Pearson Type III distribution, estimate the 100-year annual
maximum discharge.
6. Observations of annual-maximum flood flows in the Mississippi River near Baton Rouge, Louisiana, from 1985 to
2000 yield a skew coefficient of the log transformed data equal to −0.210. Determine the (weighted) skew
coefficient that should be used for flood-frequency analysis.

Source: David Chin, Water Resources Engineering, PEARSON, Singapore, 3rd Ed, 2013

CE 521 – WATER RESOURCES ENGINEERING 11

11 | P a g e