FREQUENCY ANALYSIS IN HYDROLOGY

I. INTRODUCTION
Information on different hydrological events does not depend on known physical or
chemical laws, but is governed by probabilities, which is important in the planning and
design of projects related to water.
In the case of the flow of a river, there are daily and annual variations and it is not
possible to predict exactly what its value will be at any given time.
For example, if a bridge is to be designed, the hydrological study will determine the
river's flooding with a critical probability (which seeks to determine the most critical
case), which represents a risk for said construction. These data can only be determined
through probability and statistical analysis based on past hydrological records of the river.
It should be noted that hydrological systems (rivers, lakes, snow-capped mountains, etc.)
are sometimes affected by extreme events, such as storms, rainfall, floods and droughts.
The magnitude of the event is inversely related to its frequency of occurrence, meaning
that extreme events occur less frequently compared to moderate events.
Frequency analysis of hydrological information seeks to relate the magnitude of extreme
events with their frequency of occurrence, through the use of probability distribution
functions.

II. THEORETICAL FRAMEWORK

FREQUENCY ANALYSIS
Frequency analysis is a tool used to predict future flow behavior at a site of interest based
on historical flow information. It is a method based on statistical procedures that allows
calculating the magnitude of the flow associated with a return period. Its reliability
depends on the quality of the historical series, in addition to the uncertainty inherent in
the selected probability distribution. When extrapolations are intended to be used, the
return period is greater than the length of the available series, the relative error associated
with the probability distribution used is more important, while in interpolations the
uncertainty is mainly associated with the quality of the data to be modeled; in both cases
the uncertainty is high depending on the amount of data available (Ashkar, et al. 1994).
Extrapolation of extreme frequencies in an empirical distribution of floods is extremely
risky (Garcon, 1994).
To determine the magnitude of extreme events when the probability distribution is not an
easily invertible function, it is necessary to know the variation of the variable with respect
to the mean. Chow in 1951 proposed determining this variation from a frequency factor
KT that can be expressed as:

and can be estimated from the data

For a given distribution, a relationship between K and the return period Tr can be
determined. This relationship can be expressed in mathematical terms or through the use
of a table.

Frequency analysis consists of determining the parameters of the probability

distributions and determining with the frequency factor the magnitude of the event for a
given return period.

The main probability distributions used in hydrology, the way of estimating their
parameters, the frequency factor and the confidence limits are described below. The
latter are indicators of how much uncertainty there is with the extrapolations, since
determining the range of values where the variables would really be, if the range is very
large the uncertainty is very high and if it is small, on the contrary, there will be a lot of
confidence in the estimated value.

PROBABILITY DISTRIBUTIONS FOR CONTINUOUS VARIABLES

1. NORMAL DISTRIBUTION

The normal distribution is a symmetric bell-shaped distribution, also known as the

Gaussian Bell. Although it often does not fit hydrological data, it has wide application,
for example, to transformed data that follow the normal distribution.

 Density function:

The density function is given by:

The two parameters of the distribution are the mean m and standard deviation s
for which (mean) and s (standard deviation) are derived from the data.

 Parameter estimation:

 Frequency factor:

If you work with the X without transforming the K it is calculated as:

 This factor is the same as the standard normal
variable

 Confidence limits:

where a is the probability level and is the quantile of the standardized normal
distribution for a cumulative probability of 1-a and Se is the standard error

2. TWO-PARAMETER LOGNORMAL DISTRIBUTION

If the logarithms Y of a random variable X are normally distributed, then X is said to

be normally distributed.

This distribution is widely used for calculating extreme values, for example Q max, Q
minimum, P max, P minimum (excellent results in Antioquia). It has the advantage
that X>0 and that the Log transformation tends to reduce positive asymmetry since
when taking logarithms, the larger data are reduced in a greater proportion than the
smaller ones.

Limitations: It has only two parameters, and requires that the logarithms of the
variables be centered on the mean.

 Density function:

y = ln x
where:
my : mean of the population logarithms (scalar parameter), estimated y.
sy : Standard deviation of the population logarithms, estimated sy.

 Parameter estimation:

 Frequency factor:
It can be worked in the original field and in the transformed field.
 Transformed field: If you work in the transformed field, you work with the mean and
standard deviation of the logarithms, like this:

Ln(XTr) = xTr+KSy

where: XTr = eln (xTr)

with K being the standardized normal variable for the given Tr, Xy being the
mean of the logarithms and Sy being the standard deviation of the logarithms.

Original field: If you work with the X without transforming the K it is calculated as

K is the standardized normal variable for the given Tr.

 Confidence limits:

In the transformed field.

where, n number of data, Se standard error, KT standardized normal variable.

3. GUMBEL OR EXTREME DISTRIBUTION TYPE I

An important family of distributions used in hydrological frequency analysis is the

general distribution of extreme values, which has been widely used to represent the
behavior of floods and droughts (maximums and minimums).

 Density function:

Where a and b are the parameters of the distribution.

 Parameter estimation
 Where: x and s are the mean and standard deviation estimated from the sample.

 Frequency factor:

Where Tr is the return period. For the Gumbel distribution we have that flow
for a return period of 2.33 years is equal to the average of the flows maximums.

 Confidence limits:

KT is the frequency factor and t(1-a) is the standardized normal variable for a
probability of non-exceedance of 1-a.

4. THREE-PARAMETER GAMMA DISTRIBUTION OR PEARSON TYPE 3

This distribution has been one of the most used in hydrology. Since most hydrologic
variables are skewed, the Gamma function is used to fit the frequency distribution of
variables such as annual maximum floods, minimum streamflows, annual and seasonal
flow volumes, extreme rainfall values, and short-duration rainfall volumes. The
Gamma distribution function has two or three parameters.

 Density function:

where,

x0 £ x < a for a > 0

a < x £ x0 for a < 0

a and b are the scale and shape parameters, respectively, and x0 is the parameter
of location
 Parameter estimation:
 Cs is the coefficient of skewness, x and s are the mean and standard deviation of
the sample respectively.

 Frequency factor:

where z is the standardized normal variable

This K value is tabulated according to the Cs value calculated with the sample.

 Confidence intervals:

Xt ± t(1-a) It

Where S is the sample standard deviation, n is the number of data and d is

find tabulated in terms of Cs and Tr.

5. 3-PARAMETER LOG GAMMA OR LOGPEARSON DISTRIBUTION

If the logarithms Y of a random variable X fit a Pearson type III distribution, then the
random variable X is said to fit a Log Pearson Type III distribution. This distribution
is widely used around the world for frequency analysis of peak flows. This is worked
in the same way as for the Pearson Type III but with Xy and Sy as the mean and
standard deviation of the logarithms of the original variable X.

 Density function:

where,
y0 £ y < a for a > 0
a £ and £ and0 for a < 0

a and b are the scale and shape parameters, respectively, and y0 is the parameter
of location.

 Parameter estimation:

Cs is the coefficient of skewness, Xy and Sy are the mean and standard deviation of
the sample logarithms respectively.
  Frequency factor:

where z is the standardized normal variable

This K value is tabulated according to the Cs value calculated with the sample.

 Confidence intervals:

Xt ± t(1-a) It

Where Sy is the standard deviation of the sample logarithms, n is the number of

data, and d is tabulated as a function of Cs and Tr.

FIT TESTS

To determine how adequate the fit of the data is to a probability distribution, a series of
statistical tests have been proposed to determine whether the fit is adequate. These are
statistical analyses and must be understood as such, that is, the physical significance of
the adjustments cannot be ignored.

a) Smirnov Kolmogorov test

The statistician Smirnov Kolmogorov D considers the deviation of the sample

probability distribution function P(x) from the theoretical probability function, chosen
Po(x) such that:

The test requires that the Dn value calculated using the above expression be less than
the tabulated Dn value for a required probability level.

This test is easy to perform and involves the following steps:

 The Dn statistic is the maximum difference between the sample cumulative

distribution function and the chosen theoretical cumulative distribution
function.
 The probability level is set to a, values of 0.05 and 0.01 are the most common.
 The critical value Da of the test must be obtained from tables based on a and n.
 If the calculated value Dn is greater than Da, the chosen distribution should be
rejected.

b) Chi Square Test

 A measure of the discrepancies between the observed frequencies (fo) and the
calculated frequencies (fc) by means of a theoretical distribution is given by the χ²
statistic
where:

If the statistic χ²=0 it means that the theoretical and empirical distributions fit exactly,
while if the statistic χ²>0, they differ. The distribution of the χ² statistic can be likened
to a Chi-square distribution with (kn-1) degrees of freedom, where k is the number of
intervals and n is the number of parameters of the theoretical distribution. The χ²
function is tabulated. Suppose that a hypothesis Ho is to accept that an empirical
distribution fits a Normal distribution. If the calculated value of χ² by the previous
equation is greater than some critical value of χ², with significance levels of 0.05 and
0.01 (the confidence level is 1-a) it can be said that the observed frequencies differ
significantly from the expected (or calculated) frequencies and then the hypothesis Ho
is rejected, if the opposite occurs then it is accepted.

III. LITERATURE

 http://www.ramos.utfsm.cl/doc/1393/sc/Hidrologia-UTFSM-V3.pdf
 http://hidrologia.usal.es/temas/calculos_esta.pdf
 http://www.bdigital.unal.edu.co/4578/1/AA_4044.pdf