Engineering Hydrology; Lecture Note

CHAPTER SIX

STOCHASTIC HYDROLOGY

5 Stochastic Hydrology

5.4 Introduction
Stochastic hydrology describes the physical processes involved in the movement of water onto,
over, and through the soil surface. Quite often the hydrologic problems we face do not require a
detailed discussion of the physical process, but only a time series representation of these
processes. Stochastic models may be used to represent, in simplified form, these hydrologic time
series. Some background in probability and statistics is necessary to fully understand this
concept.

5.5 Time Series

The measurements or numerical values of any variable that changes with time constitute a time
series. In many instances, the pattern of changes can be ascribed to an obvious cause and is
readily understood and explained, but if there are several causes for variation in the time series
values, it becomes difficult to identify the several individual effects. In Fig. 5.1, the top graph
shows a series of observations changing with time along the abscissa; the ordinate axis represents
the changing values of y with time, t. From visual inspection of the series, there are three
discernible features in the pattern of the observations. Firstly, there is a regular gradual overall
increase in the size of values; this trend, plotted as a separate component y1(t), indicates a linear
increase in the average size of y with time. The second obvious regular pattern in the composite
series is a cyclical variation, represented separately by y 2(t), the periodic component. The third
notable feature of the series may be considered the most outstanding, the single high peak half
way along the series. This typically results from a rare catastrophic event which does not from
part of a recognizable pattern. The definition of the function y 3(t) needs very careful
consideration and may not be possible. The remaining hidden feature of the
series is the random stochastic component, y4(t), which represents an irregular but continuing
variation within the measured values and may have some persistence. It may be due to
instrumental of observational sampling errors or it may come from random unexplainable
fluctuations in a natural physical process. A time series is said to be a random or stochastic
process if it contains a stochastic component. Therefore, most hydrologic time series may be
thought of as stochastic processes since they contain both deterministic and stochastic
components. If a time series contains only random/stochastic component is said to be a purely
random or stochastic process.

The complete observed series, y(t), can therefore be expressed by:

y(t) = y1(t) + y2(t) + y3(t) + y4(t) (5.1)

The first two terms are deterministic in form and can be identified and quantified fairly easily;
the last two are stochastic with major random elements, and some minor persistence effects, less
easily identified and quantified.

~1~
 Engineering Hydrology; Lecture Note

Figure 5.1: The time series components

5.6 Properties of Time Series

The purpose of a stochastic model is to represent important statistical properties of one or more
time series. Indeed, different types of stochastic models are often studied in terms of the
statistical properties of time series they generate. Examples of these properties include: trend,
serial correlation, covariance, cross-correlation, etc. Therefore, before reviewing the different
types of stochastic models used in hydrology, some distribution properties of stochastic
processes need to be discussed. The following basic statistics are usually used for expressing the
properties/characteristics of a time series.

_ 1
Mean, µ=E(X1) = X = ∑ n Xt (5.2) nt=1

1 −
2 2 2
Variance, σ = S = ∑ n (Xt − X ) (5.3) n
−1 t=1 n−L − _ Co var
1
iance, λ=Cov( Xt , Xt+L ) =∑(
Xt −X )(Xt+L −X ) (5.4)

~2~
 Engineering Hydrology; Lecture Note

n−L
t=1

Where L is the time lag.

Stationary time series:

If the statistics of the sample (mean, variance, covariance, etc.) as calculated by equations (5.2)-
(5.4) are not functions of the timing or the length of the sample, then the time series is said to be
stationary to the second order moment, weekly stationary, or stationary in the broad sense.
Mathematically one can write as:
E(X1) =µ
2
Var(Xt ) =σ
Cov(Xt , Xt+L ) =λL

In hydrology, moments of the third and higher orders are rarely considered because of the
unreliability of their estimates. Second order stationarity, also called covariance stationarity, is
usually sufficient in hydrology. A process is strictly stationary when the distribution of Xt does
not depend on time and when
all simultaneous distributions of the random variables of the process are only dependent on their
mutual time-lag. In another words, a process is said to be strictly stationary if its n-th (n for any
integers) order moments do not depend on time and are dependent only on their time lag.

Non-stationary time series:

If the values of the statistics of the sample (mean, variance, covariance, etc.) as calculated by
equations (5.2)-(5.4) are dependent on the timing or the length of the sample, i.e. if a definite
trend is discernible in the series, then it is a non-stationary series. Similarly, periodicity in a
series means that it is non-stationary. Mathematically one can write as:
E(X1) =µt
2
Var(Xt ) =σt
Cov(Xt , Xt+L ) =λL,t
White noise time series:
For a stationary ties series, if the process is purely random and stochastically independent, the
time series is called a white noise series. Mathematically one can write as:
E(X1) =µ
2
Var(Xt ) =σ
Cov(Xt , Xt+L ) =0 for all L ≠ 0
Gaussian time series:
A Gaussian random process is a process (not necessarily stationary) of which all random
variables are normally distributed, and of which all simultaneous distributions of random
variables of the process are normal. When a Gaussian random process is weekly stationary, it is

~3~
 Engineering Hydrology; Lecture Note

also strictly stationary, since the normal distribution is completely characterized by its first and
second order moments.

5.7 Analysis of Hydrologic Time Series

Records of rainfall and river flow form suitable data sequences that can be studied by the
methods of time series analysis. The tools of this specialized topic in mathematical statistics
provide valuable assistance to engineers in solving problems involving the frequency of
occurrences of major hydrological events. In particular, when only a relatively short data record
is available, the formulation of a time series model of those data can enable long sequences of
comparable data to be generated to provide the basis for better estimates of hydrological
behaviour. In addition, the time series analysis of rainfall, evaporation, runoff and other
sequential records of hydrological variables can assist in the evaluation of any irregularities in
those records. Cross-correlation of different hydrological time series may help in the
understanding of hydrological processes. Tasks of time series analysis include:
(1) Identification of the several components of a time series.
(2) Mathematical description (modelling) different components identified. If a
hydrological time series is represented by X1, X2, X3, ..., Xt, ..., then symbolically, one can
represent the structure of the Xt by: X ⇔ [Tt, Pt, Et] Where Tt is the trend component, Pt is the
periodic component and Et is the stochastic component. The first two components are specific
deterministic features and contain no element of randomness. The third, stochastic, component
contains both random fluctuations and the self-correlated persistence within the data series.
These three components form a basic model for time series analysis. The aims of time series
analysis include but not limited to:
(1) description and understanding of the mechanism,
(2) Monte-Carlo simulation,
(3) forecasting future evolution, Basic to stochastic analysis is the assumption that the
process is stationary. The modelling of a time series is much easier if it is stationary, so
identification, quantification and removal of any non-stationary components in a data series is
under-taken, leaving a stationary series to be modelled.

5.7.1 Trend component

This may be caused by long-term climatic change or, in river flow, by gradual changes in a
catchment's response to rainfall owing to land use changes. Sometimes, the presence of a trend
cannot be readily identified.
Methods of trend identification:

Different statistical methods, both nonparametric tests and parametric tests, for identifying trend
in time-series are available in the literature. Two commonly used methods for identifying the
trend are discussed briefly in this section.
(1) Mann-Kendall test The test uses the raw (un-smoothed) hydrologic data to detect possible
trends. The Kendall statistic was originally devised by Mann (1945) as a non-parametric test for
trend. Later the exact distribution of the test statistic was derived by Kendall (1975).

The Mann-Kendall test is based on the test statistic S defined as follows:

n−1 n

~4~
 Engineering Hydrology; Lecture Note

S =∑ ∑sgn(xj − Xi ) (5.5)
i=1 j=i+1
Where the Xj are the sequential data values, n is the length of the data set, and
⎧1if θ>0
⎪
sgn(θ) = ⎨0if θ=0 (5.6) ⎪

⎩−1if θ<0
Mann (1945) and Kendall (1975) have documented that when, the statistic S is approximately
normally distributed with the mean and the variance as follows:

E(S) =0 (5.7)
q
n(n −1)(2n +5) −∑tp (tp −1)(2tp + 5)
p=1
V (S) = (5.8)
18
th th
Where n = number of datatp = the number of ties for the p value (number of data in the p
group)q = the number of tied values (number of groups with equal values/ties) The standardized
Mann-Kendall test statistic ZMK is computed by

⎧S −1

⎪ S > 0
Var(s)
⎪
⎪
sgn(θ) = ⎨0 S =0 (5.9) ⎪ S +1
⎪S <0⎪Var(s)
⎩
The standardized MK statistic Z follows the standard normal distribution with mean of zero and
variance of one. The hypothesis that there has not trend will be rejected if
Zmk

> Z1−α/2
(5.10)
Where Z1−α/2 is the value read from a standard normal distribution table with αbeing the
significance level of the test.

(2) Linear regression method

~5~
 Engineering Hydrology; Lecture Note

Linear regression method can be used to identify if there exists a linear trend in a hydrologic
time series. The procedure consists of two steps, fitting a linear regression equation with the time
T as independent variable and the hydrologic data, Y as dependent variable, i.e. Y = α + β.T
(5.11) and testing the statistical significance of the regression coefficient β. Test of hypothesis
concerning β can be made by noting that (β – β0)/Sβ has t distribution with n-2 degrees of
freedom. Thus the hypothesis H0: β = β0 versus H0: β≠β0 is tested by computing
β−β0
t= (5.12)
Sβ
Where Sβ is the standard deviation of the coefficient β with
SSβ=

(5.13)
n_
2
∑(Ti −T )
i=1
1 ^
2
and S = ∑ n (Y i − Y i ) (5.14) n − 2 i=1
^
Where S is the standard error of the regression, Yi and Y are observed and estimated hydrologic
variable from the regression equation, respectively. The hypothesis, i.e. no trend, is rejected if

t
> t1−α/2, n−2
Models for trend:
The shape of the trend depends on the background of the phenomenon studied. Any smooth trend
that is discernible may be quantified and then subtracted from the sample series. Common
models for trend may take the following forms: Tt = a + bt (a linear trend, as in Fig. 5.1) (5.15)
2 3
or Tt = a + bt + ct + dt + ... (a non-linear trend) (5.16) The coefficients a, b, c, d ... are usually
evaluated by least-squares fitting. The number of terms required in a polynomial trend being
primarily imposed by the interpretation of the studied phenomenon. The number of terms is
usually based on statistical analysis, which determines the terms contributing significantly to the
description and the interpretation of the time series. Restriction is made to the significant terms
because of the principle of parsimony concerning the number of unknown parameters (constants)
used in the model. One wishes to use as small a number of parameters as possible, because in
most cases the addition of a complementary parameter decreases the accuracy of the other
parameters. Also prediction- and control procedures are negatively influenced by an exaggerated
number of parameters. This principle of parsimony is not only important with respect to the
selection of the trend function but also with respect to other parts of the model.

~6~
 Engineering Hydrology; Lecture Note

5.7.2 Periodic component

In most annual series of data, there is no cyclical variation in the annual observations, but in the
sequences of monthly data distinct periodic seasonal effects are at once apparent. The existence
of periodic components may be investigated quantitatively by (1) Fourier analysis, (2) spectral
analysis, and (3) autocorrelation analysis. Of which, the autocorrelation analysis method is
widely used by hydrologists and will be discussed briefly in this section.
Identification of periodic component by autocorrelation analysis:
The procedure consists of two steps, calculating the autocorrelation coefficients and testing their
statistical significance. For a series of data, Xt, the autocorrelation coefficient rL between Xt and
Xt+L are calculated and plotted against values of L (known as the lag), for all pairs of data L
time units apart in the series:
1 n−L _ _1 n _2
rL =∑(Xt −X )(Xt+L −X )/ ∑(Xt −X ) (5.17) n − Lt=1 nt=1 _
Where X is the mean of the sample of n values of Xt and L is usually taken for values from zero
up to n/4. A plot of rL versus L forms the correlogram. The characteristics of a time series can be
seen from the correlogram. Examples of correlograms are given in Fig 54.2. Calculation of
equation (5.17) for different L gives the following cases:
If L = 0, rL = 1. That is, the correlation of an observation with itself is one.
If rL ≈ 0 for all L ≠0, the process is said to be a purely random process. This indicates
that the observations are linearly independent of each other. The correlogram for such a complete
random time series is shown in Fig 5.2(a).
If rL ≠ 0 for some L ≠0, but after L > τ, then , the time series is still referred to as simply
a random one (not purely random) since it has a ‘memory’ up to L = τ. When rτ , the process is
said to have no memory for what

occurred prior to time t-τ. The correlogram for such a non-independent stochastic process is
shown in Fig 5.2(b). This is representative of an auto regressive process. Typically, such a
correlogram could be produced from a series described by the Autoregressive model:
Xt = a1Xt-1 + a2Xt-2 +a3Xt-3 + … + εt
(4.18) where ai are related to the autocorrelation coefficients ri and εt is a random
independent element.
• In the case of data containing a cyclic (deterministic) component, then rL ≠0 for all L ≠0, the
correlogram would appear as in Fig. 5.2(c). Where T is

~7~
 Engineering Hydrology; Lecture Note

Figure 5.2 Examples of Correlograms

Modelling of periodic component:
A periodic function Pt is a function such that Pt+T = Pt
for all t

The smallest value of T is called the period. The dimension of T is time, T thus being a number
of time-units (years, months, days or hours, etc.) and we also have Pt+nT = Pt for all t and for all
integer n.

The frequency is defined as the number of periods per time-unit:

1
frequency = period
Trigonometric functions are simple periodic functions. For example, α sin (ωt + β)

has a period of 2π/ω, becauseα sin [ω(t+2π/ω)+β] = α sin(ωt+2π+β) = α sin(ωt + β)

The pulsation or angular frequency is defined as

~8~
 Engineering Hydrology; Lecture Note

2π1
ω==2π. frequency
period

the constant α is termed the amplitude and βthe phase (with respect to the origin) of the sine-
function.

A simple model for the periodic component may be defined as (for morediscussions refer to the
literature of Time Series Analysis):Pt = m + Csin(2πt/T) (5.19)

Where C is the amplitude of the sine wave about a level m and of wavelength T.The serial (auto)
correlation coefficients for such a Pt are given by:rL = cos (2πL/T) (5.20)

The cosine curve repeats every T time units throughout the correlogram with r L
= 1 for L = 0, T, 2T, 3T,… Thus periodicities in a time series are exposed by regular cycles in
the corresponding correlograms.

Once the significant periodicities, Pt, have been identified and quantified by µt
(the means) and σt (the standard deviations) they can be removed from the original times series
along with any trend, Tt, so that a new series of data, Et, is
formed:

Xt − Tt − mt
Et = (5.21)
St Simple models for periodic component in hydrology can be seen in the literature.
For example, in many regions, typical monthly potential evapotranspiration variation during the
year can be modelled more or less by a sinusoidal function, with a couple of parameters to tune
the annual mean and the amplitude (Xu and Vandewiele, 1995).
This behavior leads to the idea to model by a truncated Fourier series:
ept = {a + bsin[(2π /12)(t-c)]}+
where again t is time in month. The plus sign at the end is necessary for avoiding negative values
of ep which otherwise may occur in rare cases. Again parameters a, b and c are characteristics of
the basin.

5.7.3 Stochastic component

Et represents the remaining stochastic component of the time series free from non-stationary
trend and periodicity and usually taken to be sufficiently stationary for the next stage in simple
time series analysis. This Et component is analysed to explain and quantify any persistence
(serial (auto) correlation) in the data and any residual independent randomness. It is first
standardized by:
_

~9~
 Engineering Hydrology; Lecture Note

E −E
Zt = t (5.22) SE
_
Where E and sE are the mean and standard deviation of the Et series. The series, Zt, then has
zero mean and unit standard deviation. The autocorrelation coefficients of Zt are calculated and
the resultant correlogram is examined for evidence and recognition of a correlation and/or
random structure.

For example, in Fig. 5.3a for a monthly flow, the correlogram of the Zt stationary series (with the
periodicities removed) has distinctive features that can be recognized. Comparing it with Fig.
5.2, the Zt correlogram resembles that of an auto regressive (Markov) process. For a first order
Markov model:
Zt = r1Zt−1 + et (5.23)

Where r1 is the autocorrelation coefficient of lag 1 of the Zt series and et is a random

independent residual. A series of the residuals et may then be formedfrom the Zt series and its
known lag 1 autocorrelation coefficient, r1: et = Zt − r1Zt−1 (5.24)

The correlogram of residuals is finally computed and drawn (Fig.5.3b). For this data this
resembles the correlogram of 'white noise', i.e. independently distributed random values. If there
are still signs of autoregression in the e t correlogram, a second-order Markov model is tried, and
the order is increased until a random et correlogram is obtained. The frequency distribution
diagram of the first order evalues (Fig. 5.3c) demonstrates an approximate approach to the
t
normal (Gussian) distribution.

At this stage, the final definition of the recognizable components of the time series has been
accomplished including the distribution of the random residuals. As part of the analysis, the
fitted models should be tested by the accepted statistical methods applied to times series. Once
the models have been formulated and quantified to satisfactory confidence limits, the total
mathematical representation of the time series can be used for solving hydrological problems by
synthesizing non-historic data series having the same statistical properties as the original data
series.

~ 10 ~
 Engineering Hydrology; Lecture Note

Figure 5.3 River Thames at Teddington Weirs ( 82 years of monthly flows, from Shaw, 1988)

5.8 Time Series Synthesis

The production of a synthetic data series simply reverses the procedure of the time series
analysis. First, for as many data items as are required, a comparable sequence of random
numbers, drawn from the et distribution, is
generated using a standard computer package. Second, the corresponding synthetic Zt values are
recursively calculated using equation 5.23 (starting the series with the last value of the historic Z t
series as the Zt-1 value). Third, the Et series then derives from equation 5.22 in reverse:
_
Et =Zt SE +E (5.25)

The periodic component Pt represented by mt and st for time period t is then added to the Et
values to give:
Xt = Tt + EtSt + mt (from equation 5.21) (5.26)

The incorporation of the trend component Tt then produces a synthetic series of Xt having
similar statistical properties to the historic data series.

5.9 Some Stochastic Models

Ultimately design decisions must be based on a stochastic model or a combination of stochastic
and deterministic models. This is because any system must be designed to operate in the future.
Deterministic models are not available for generating future watershed inputs in the form of

~ 11 ~
 Engineering Hydrology; Lecture Note

precipitation, solar radiation, etc., nor is it likely that deterministic models for these inputs will
be available in the near future. Stochastic models must be used for these inputs.

5.9.1 Purely random stochastic models

Possibly the simplest stochastic process to model is where the events can be assumed to occur at
discrete times with the time between events constants, the events at any time are independent of
the events at any other time, and the probability distribution of the event is known. Stochastic
generation from a model of this type merely amounts to generating a sample of random
observations from a univariate probability distribution. For example, random observations for
any normal distribution can be generated from the relationship, y = σRN + µ(5.27)

Where RN is a standard random normal deviate (i.e. a random observation from a standard
normal distribution) and µ and σ are the parameters of the desired normal distribution of y.
Computer routines are available for generating standard random normal distribution.

5.9.2 Autoregressive models

Where persistence is present, synthetic sequences cannot be constructed by taking a succession
of sample values from a probability distribution, since this will not take account of the relation
between each number of sequences and those preceding it. Consider a second order stationary
time series, such as an annual time series, made up of a deterministic part and a random part. The
deterministic part is selected so as to reflect the persistence effect, while it is assumed that the
random part has a zero mean and a constant variance. One of the models to simulate such a series
is the Autoregressive model. The general form of an autoregressive model is:
( yt −µ) =β1( yt−1 −µ) +β2( yt−2 −µ) + ... +βk ( yt−k −µ) +εt (5.28)

Where µ is mean value of the series, β is the regression coefficient, the {y1, y2, …, yt,…} is the
observed sequence and the random variables εt are usually assumed to be normally and
independently distributed with zero mean and variance . In order to determining the order k of
autoregression required to describe the persistence adequately, it is necessary to estimate k+2
parameters: β1, β2, …βk, µ and the variance of residuals . Efficient methods for
estimating these parameters have been described by Kendall and Stuart (1968), Jenkins and
Watts (1968).

The first order autoregression: yt −µ =β1( yt−1 −µ) +εt (5.29)

has found particular application in hydrology. When equation (5.29) is used to model annual
discharge series, the model states that the value of y in one time period is dependent only on the
value of y in the preceding time period plus a random component. It is also assumed that ε t is
independent of yt.

Equation (5.29) is the well-known first order Markov Model in the literature. It has three
2
parameters to be estimated: µ, β1, and σ E .
For the moment method of parameter estimation, parameter µ can be computedfrom the time
series as the arithmetic mean of the observed data. As for β 1, the Yule-Walker equation (Delleur,

~ 12 ~
 Engineering Hydrology; Lecture Note

1991) shows that:

P
ρk =∑β j ρk −jk >0 (5.30)
j=1
the above equation, written for k = 1, 2, …, yields a set of equations. Where ρ k is the
autocorrelation coefficient for time lag k. As the autocorrelation coefficients ρ 1, ρ2, …, can be
estimated from the data using equation (4.17), these equations can be solved for the
autoregressive parameters β1, β2, …, β p. This is the estimation of parameters by the method of
moments. For example, for the first order autoregressive model, AR(1), the Yule-Walker
equations yield
ρ1 =β1.ρ0 =β1 sin ceρ0 =1 (5.31)
Similarly we can derive the equations for computing β1 and β2 for the AR(2) model as

ρ (1 − ρ )
β1 = 1 12 2 (5.32)
1 − ρ

2
It can be shown that σ E . is related to (the variance of the yt series) by:
2 2 2
σε =σ y (1−β1 ) (5.33)

2 2
If the distribution of y is N (µy ,σ y ) then distribution of ε is N(0, σE .). Random
2
values yt can now be generated by selecting εt randomly from a N(0, σ E .)

2 2
distribution. If z is N(0,1) then Zσφ or Zσ y 1−β1 is N (0,σε . Thus, a model for
2
generating Y’s that are N(µy ,σ y ) and follow the first order Markov model is

2
yt =µ y +β1( yt−1 −µ y ) + Ztσ y 1−β1 (5.34)
The procedure for generating a value for yt is:
_
(1) estimate µy, σ y, and β1 by y , sx and r1(eq.5.17) respectively,
(2) select a zt at random from a N(0, 1) distribution, and

_
(3) calculate yt by eq. (5.34) based on y , s and β1, and yt-1.
x

~ 13 ~
 Engineering Hydrology; Lecture Note

2
The first value of yt, i.e. y1, might be selected at random from a N(µy, σy ).
To eliminate the effect of y1 on the generated sequence, the first 50 or 100 generated values
might be discarded. Equation (5.34) has been widely used for generating annual runoff from
watersheds

5.9.3 First order Markov process with periodicity: Thomas - Fiering model
The first order Markov model of the previous section assumes that the process is stationary in its
first three moments. It is possible to generalise the model so that the periodicity in hydrologic
data is accounted for to some extent. The main application of this generalisation has been in
generating monthly streamflow where pronounced seasonality in the monthly flows exists. In its
simplest form, the method consists of the use of twelve linear regression equations. If, say,
twelve years of record are available, the twelve January flows and the twelve December flows
are abstracted and January flow is regressed upon December flow; similarly, February flow is
regressed upon January flow, and so on for each month of the year.

q = q jan +b (q − qdec ) +ε
jan jan dect jant

q feb = q feb + b feb (q jan − q jan ) +εfeb

………. Fig.4.4 shows a regression analysis of qj+1 on qj, pairs of successive monthly flows for
the months (j+1) and j over the years of record where j = 1, 2, 3, ..., 12 (Jan, Feb, ... Dec) and
when j = 12, j+1 = 1 = Jan (there would be 12 such regressions). If the regression coefficient of
month j+1 on j is bj, then the
^
regression line values of a monthly flow, yj+ , can be determined from the
1
previous months flow qj, by the equation:
^_ _
qj +1 = qj +1+ bj (qj − qj ) (5.35)
To account for the variability in the plotted points about the regression line reflecting the
variance of the measured data about the regression line, a further component is added:

Z.Sj +1 (1− r 2 j )

where is the standard deviation of the flows in month j+1, rj is the correlation coefficient
between flows in months j+1 and j throughout the record, and Z = N(0, 1), a normally distributed
random deviate with zero mean and unit standard deviation. The general form may written as
^
qj = qj + bj (qj − qj ) + Zj+1.i .Sj+1 (1− rj 2) (5.36)

+1 +1

~ 14 ~
 Engineering Hydrology; Lecture Note

Where bj = rj * Sj 1/ Sj , there are 36 parameters for the monthly model (q, for

+
each month). The subscript j refers to month. For monthly synthesis j varies from 1 to 12
throughout the year. The subscript i is a serial designation from year 1 to year n. Other symbols
are the same as mentioned earlier.

Fig.5.4: Thomas-Fiering model The procedure for using the model is as

follows:
(1) For each month, j = 1, 2, … 12, calculate
_ 1
(a) The mean flow, q β= ∑qj,i;(i =j,12 j,24 +j,...)
n
_
2
∑(qj,i − qj )
(b) The standard deviation, Sj =

n −1

(c) The correlation coefficient with flow in the preceding month,

∑(qj,i −q )(qj+1,i − qj+1)

j
rj =

_
_
2 2
∑(qj,i − qj ) ∑(qj+1,i − qj+1)
ii

~ 15 ~
 Engineering Hydrology; Lecture Note

(d) The slope of the regression equation relating the month’s flow to flow in the preceding
month:
Sj+1

bj =rj Sj

where Z is a random Normal deviate N(0, 1).

(3) To generate a synthetic flow sequence, calculate (generate) a random number sequence {Z 1,
Z2, … }, and substitute in the model.

5.9.4 Moving average models

The model form:
The moving average has frequently been used to smooth various types of hydrologic time series
such as daily or weekly air temperature, evaporation rates, wind speed, etc. The moving average
process used in the stochastic generation hydrologic data is somewhat different. In this use, the
moving average process describes the deviations of a sequence of events from their mean value.
A process {x1} defined as xt = et + Φ 1et-1 + Φ 2et-2 + ...+ Φ qet-q (5.37) Where {xt} is an
uncorrelated stationary process, is called a moving average process of order q, denoted MA(q)-
process. It can also be written as xt = et -θ1et-1 -θ2et-2 - ...- θqet-q (5.38) with Φ1 = -θ1, Φ 2 =-
θ2, ..., Φ q = -θq. The properties of the moving average process:
The autocovariance of the process is obtained by forming the product and

For k = 0 we obtain the variance of the process

With the convention θo = -1

~ 16 ~
 Engineering Hydrology; Lecture Note

The autocorrelation function is then

(5.43)

Equations (5.40) and (5.41) can be used for the estimation of the parameters by method of
moments. For this purpose they are rewritten as follows:

(5.44)

(5.45) Equ. (5.44) and

(5.45) are used recursively. For example for the MA(1) model

xt = et -θ1et-1 (5.46) we have

data.

~ 17 ~
 Engineering Hydrology; Lecture Note

5.9.5 ARMA models

Model form:
In stochastic hydrology ARMA models are known as Auto-Regressive Moving Average
(ARMA) models. They combine any direct autocorrelation properties of a data series with the
smoothing effects of an updated running mean through the series. The two components of the
model for a data series xt, e.g. annual river flows, are described by:

(5.48) Moving average

(MA(q)) xt = et -θ1et-1 -θ2et-2 -...-θqet-q (5.49) Where et are random numbers with zero mean
and variance σe 2 .

One of the merits of the ARMA process is that, in general, it is possible to fit a model with a
small number of parameters, i.e. p+q. This number is generally smaller than the number of
parameters that would be necessary using either an AR model or a MA model. This principle is
called the parsimony of parameters.

Properties of ARMA model: Consider in the ARMA(1, 1) model which has been used
extensively in

(5.52) Multiplying
both sides of (5.52) by Xt-k

~ 18 ~
 Engineering Hydrology; Lecture Note

and taking the expectation of both sides we obtain the autocovariance

(5.53)
For k = 0, equ (5.53) becomes

but

and

(5.54)
Thus

(5.55)

For k = 1 equ (5.53) becomes

Combining with the previous equation or

(5.56)
and

(5.57)
For k ≥ 2

~ 19 ~
 Engineering Hydrology; Lecture Note

k ≥ 2 (5.58)
the autocorrelation function (ACF) is obtained by dividing (5.56), (5.57) and

Observe that the MA parameter θ 1 enters only in the expression for ρ 1. For ρ2
and beyond the behaviour of the autocorrelation is identical to that of the AR(1) model.
Estimates of the parameters θ1 and β2 and can be obtained from equations (5.59b) and (5.59c),
since the serial (auto) correlation coefficients ρ1 and ρ2 can be computed from data.

ACF depends on AR and MA parameters.

Hydrologic justification of ARMA models
A physical justification of ARMA models for annual streamflow simulation is asfollows.
Consider a watershed with annual precipitation Xt, infiltration aXt and

evapotranspiration bXt. The surface runoff is (1-a-b)Xt = dXt. (See Fig 5.5).

~ 20 ~
 Engineering Hydrology; Lecture Note

Fig.5.5 Conceptual representation of the precipitation-streamflow process after Salas and

Smith (1980)
Let the groundwater contribution to the stream be cSt-1.

(5.62)

(5.63)
or

(5.64)
Rewriting (5.62) or

~ 21 ~
 Engineering Hydrology; Lecture Note

(5.65) and

rewriting (5.64) as

(5.66) Combining
(5.62), (5.66) and (5.65) we obtain

which has the form of an ARMA (1, 1), i.e. equation (5.52) model when the precipitation, X t is
an independent series and when (1-c) = β1, d = 1, and [d(1-c)ac)] = θ1..

5.10 Uses of Stochastic Models

(1) To make predictions of frequencies of extreme events
Stochastic models have been used to make predictions about the frequency of occurrence of
certain extreme events of interest to the hydrologist. Models such as that given by equation
(5.29) are selected, and the residual is taken to be random variable with probability distribution
whose parameters are specified. The parameters are estimated from data; so-called "synthetic"

~ 22 ~
 Engineering Hydrology; Lecture Note

sequence {yt} can then be constructed, and the frequency with which the extreme event occurs in
them can be taken as an estimate of the "true" frequency with which it would occur in the long
run.
(2) For the investigation of system operating rules
A further use for synthetic sequences generated by stochastic models is in reservoir operation,
such as the investigation of the suitability of proposed operating rules for the release of water
from complex systems of interconnected reservoirs. By using the generated sequence as inputs to
the reservoir system operated according to the proposed rules, the frequency with which
demands fail to be met can be estimated. This may lead to revision of the proposed release rules;
the modified rules may be tested by a similar procedure.
(3) To provide short-term forecasts
Stochastic models have been used to make forecasts. Given the values xt, xt-1, xt-2, ...; yt, yt-1,
yt-2, ... assumed by the input and output variables up to time t, stochastic models have been
constructed from this data for forecasting the output from the system at future times, t+1, t+2, ...,
t+k, .... In statistical terminology, k is the lead-time of the forecast. Many stochastic models have
a particular advantage for forecasting purposes in that they provide, as a by-product of the
procedure for estimating model parameters, confidence limits for forecasts (i.e. a pair of values,
one less than the forecast and one greater, such that there is a given probability P that these
values will bracket the observed value of the variable at time t+k). Confidence limits therefore
express the uncertainty in forecasts; the wider apart the confidence limits, the less reliable the
forecast. Furthermore, the greater the lead-time k for which forecasts is required, the greater will
be the width of the confidence interval, since the distant future is more uncertain than the
immediate.
(4) To "extend" records of short duration, by correlation
Stochastic models have been used to "extend" records of basin discharge where this record is
short. For example, suppose that it is required to estimate the instantaneous peak discharge with
a return period of T years (i.e. such that it would recur with frequency once in T years, in the
long run). One approach to this problem is to examine the discharge record at the site for which
the estimate is required, to abstract the maximum instantaneous discharge for each year of
record, and to represent the distribution of annual maximum instantaneous discharge by a
suitable probability density function. The abscissa, Yo, say, that is exceeded by a proportion 1/T
of the distribution then estimates the T-year flood. It, however, frequently happens that the
length of discharge record available is short, say ten years or fewer. On the other hand, a much
longer record of discharge may be available for another gauging site, such that the peak
discharges at the two sites are correlated. In certain circumstances, it is then permissible to
represent the relation between the annual maximum discharges at the two sites by a regression
equation and to use this fitted equation to estimate the annual maximum instantaneous discharges
for the site with short record.
(5) To provide synthetic sequences of basin input
Suppose that the model has been developed for a system consisting of a basin with rainfall as
input variable, streamflow as output variable. If a stochastic model were developed from which a
synthetic sequence of rainfall could be generated having statistical properties resembling those of
the historic rainfall sequence, the synthetic rainfall sequence could be used as input to the main
model for transformation to the synthetic discharge sequence. The discharge so derived could
then be examined for the frequency of extreme events.

This approach to the study of the frequency of extreme discharge events is essentially an

~ 23 ~
 Engineering Hydrology; Lecture Note

alternative to that described in paragraph (1) above. In the latter, a synthetic sequence is derived
from a stochastic model of the discharge alone; in the former, a synthetic discharge sequence is
derived by using a model to convert a synthetic sequence of rainfall into discharge.

~ 24 ~