Rainfall Process

1

Applied Hydrology

RSLAB-NTU

Lab for Remote Sensing Hydrology and Spatial Modeling

Rainfall Process

Professor Ke-Sheng ChengDept. of Bioenvironmental Systems Engineering

National Taiwan University

2Lab for Remote Sensing Hydrology and Spatial ModelingRSLAB-NTU

Types of Rainfall ProcessTemporal rainfall process

Storm-rainfall HourlySubhourly

Daily rainfallTDP-rainfall, monthly rainfall

Spatial rainfall processHourly rainfallSeasonal rainfallAnnual rainfall

Spatio-temporal rainfall process

Hyetograph


Time series of point-rainfall data

Intermittency of storms


Time series of point-rainfall data

Occurrences (arrivals) of storm events

Duration & total depth of storm events


Probability distribution of storm-rainfall parameters

Arrival (or occurrence) of storm eventsPoisson or geometrical

Storm durationExponential of gamma

Total depth or average intensityGamma


Storm-rainfall sequence models

Bartlett-Lewis rectangular pulses (BLRP) model

Neyman-Scott rectangular pluses (NSRP) model

Simple-scaling Gauss-Markov (SSGM) model


Previous findings

Eagleson (1970) pointed out that for given climatic conditions, storm events of a given scale (microscale, mesoscale, or synoptic scale) exhibit similar time distributions when normalized with respect to total rainfall depths and storm durations.

Convective cells and thunderstorms are dominant types of storms at the microscale and the mesoscale, respectively. Events of synoptic scale include frontal systems and cyclones that are typically several hundred miles in extent and often have series of mesoscale subsystems.


In general, convective and frontal-type storms tend to have their peak rainfall rates near the beginning of the rainfall processes, while cyclonic events have the peak rainfall somewhere in the central third of the storm duration.

Koutsoyiannis and Foufoula-Georgiou (1993) proposed a simple scaling model to characterize the time distribution of instantaneous rainfall intensity and incremental rainfall depth within a storm event.


Storm-rainfall sequence models

Simple-Scaling Gauss-Markov Model


Simple Scaling Model for Storm Events

A natural process fulfills the simple scaling property if the underlying probability distribution of some physical measurements at one scale is identical to the distribution at another scale, multiplied by a factor that is a power function of the ratio of the two scales (Gupta and Waymire, 1991).


Let X(t) and X(t) denote measurements at two distinct time or spatial scales t and t, respectively. We say that the process {X(t), t0} has the simple scaling property if there is some real number H such that

for every real > 0. The denotes equality in distribution, and

H is called the scaling exponent.

)}({)}({ tXtX Hd

d


Scaling (Scale Invariant) Properties

Strict Sense Simple Scaling

Wide Sense Simple Scaling (Second-Order Simple Scaling)

Multiple Scaling

)()( sZsZ vn

d

v

2,1)],([)]([ ksZEsZE kv

knkv

1)],([)]([ kkv

nkkv sZEsZE k


Simple Scaling Model for Storm Events – Instantaneous Rainfall

Let represent the instantaneous rainfall intensity at time t of a storm with duration D.

The simple scaling relation for is

for every .

),( Dt

)},({)},({ DtDt Hd

}0,0),,({ DtDt

0


In practice, cumulative rainfall depths over a fixed time interval, for example one hour, are recorded and we shall refer to them as the incremental rainfall depths.

The i-th incremental rainfall depth can be expressed by

),( DiX

),( DiX

i

idtDt

)1(),(


The above equations are valid even for nonstationary random processes.


Incremental rainfall percentage or normalized (dimensionless) rainfall rate.


The i-th incremental rainfall percentages of storms of durations D and D are identically distributed if the time intervals areand , respectively.


Assuming that storm events are independent, the identical distribution property allows us to combine the i-th (i = 1, 2, …, ) incremental rainfall percentages of any storm durations to form a random sample, and the parameters (e.g. mean and variance) of the underlying distribution can be easily estimated.

]/[]/[ DD


Gauss-Markov model of dimensionless hyetographs


By the Markov property, the joint density

of Y(1), Y(2),…, Y(n) factorizes

),,,( 21 nyyyf

)|(

)|()(),,,(

1)1(|)(

12)1(|)2(1)1(21

nnnYnY

YYYn

yyf

yyfyfyyyf


Application of the SSGM model

Design storm hyetographBecause the design storm hyetograph

represents the time distribution of the total storm depth determined by annual maximum rainfall data, the design storm hyetograph is optimally modeled when based on observed storm events that actually produced the annual maximum rainfall, i.e. the so-called annual maximum events .


Event duration the duration of an observed storm event.

Design duration the designated time interval for a design storm.

In general, the design durations do not coincide with the actual durations of historic storm events. The design durations are artificially designated durations used to determine the corresponding annual maximum depths; whereas event durations are actual raining periods of historic storm events.


Annual maximum events tend to occur in certain periods of the year (such as a few months or a season) and tend to emerge from the same storm type. Moreover, annual maximum rainfall data in Taiwan strongly indicate that a single annual maximum event often is responsible for the annual maximum rainfall depths of different design durations. In some situations, single annual maximum event even produced annual maximum rainfalls for many nearby raingauge stations.

Dept of Bioenvironmental Systems EngineeringNational Taiwan University

Examples of annual max events in Taiwan

Design durations


Using only annual maximum events for design hyetograph development enables us not only to focus on events of the same dominant storm type, but it also has the advantage of relying on almost the same annual maximum events that are employed to construct IDF curves.


Because the peak rainfall depth is a key element in hydrologic design, an ideal hyetograph should not only describe the random nature of the rainfall process but also the extreme characteristics of the peak rainfall.

Therefore, our objective is to find the incremental dimensionless hyetograph that not only represents the peak rainfall characteristics but also has the maximum likelihood of occurrence.


Our approach to achieve this objective includes two steps: Determine the peak rainfall rate of the

dimensionless hyetograph and its time of occurrence, and

Find the most likely realization of the normalized rainfall process with the given peak characteristics.


Suppose that N realizations {y(i,j): i = 1, 2, …, n; j =1, 2, …, N } of the random process Y are available.

N

jp jy

Ny

1

* )(1

N

jp jt

Nt

1

* )(1

The value of t* is likely to be non-integer and should be rounded to the nearest integer.


m

y

y

y

y

y

y

y

y

DD

CD

C

D

C

DD

C

D

C

D

C

DD

C

D

C

D

C

DD

C

D

C

D

C

DD

C

D

C

D

C

DD

CD

C

D

C

DD

CD

C

D

C

D

n

n

t

t

t

nn

n

n

n

n

n

nn

n

t

t

t

t

tt

t

t

t

t

t

tt

t

t

t

t

t

tt

t

1

1

1

3

2

1

2

11

1

2

2

2

2

2

11

1

1

1

1

2

1

2

11

1

4

4

4

24

33

3

3

3

3

23

22

2

2

2

2

22

1

*

*

*

*

*

*

*

**

*

*

*

*

*

**

*

*

*

*

*

**

*

0000010000

0011111111

011

0

01)1

(0

010)1

(

110)1

(

010)1

(

0100)1

(0

0100)1

(

0100)1

(


*

1

1

2

12

1

1

22

2

12

2

2

11

1

11

1

1

2

1

1

1

2

12

1

1

44

43

4

24

32

3

3

33

32

3

23

21

2

2

22

21

2

22

1

1

1

)1

(

)1

(

)1

(

)1

(

)1

(

)1

(

)1

(

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

*

y

DD

CD

C

D

C

DD

C

D

C

D

C

DD

C

D

C

D

C

DD

C

D

C

D

C

DD

C

D

C

D

C

DD

CD

C

D

C

DD

CD

C

D

C

D

nn

nn

n

nn

nn

n

n

nn

n

n

tt

t

tt

t

tt

t

t

tt

t

tt

t

tt

t

t

tt

t

tt

t

tt

t

t

The dimensionless hyetograph {yi, i =

1,2,…, n} is determined by solving the matrix equation.


Model Application

Scale-invariant Gauss-Markov model Two raingauge stations Hosoliau and Wutu

h, located in Northern Taiwan.


Annual maximum events that produced annual maximum rainfall depths of 6, 12, 18, 24, 48, and 72-hour design durations were collected.

All event durations were first divided into twenty-four equal periods i (i=1,2,…,24, D = event duration, =D/24).

Rainfalls of each annual maximum event were normalized, with respect to the total rainfall depth and event duration.


Parameters for the distributions of

normalized rainfalls


Normality check for normalized rainfalls

The Gauss-Markov model of dimensionless hyetographs considers the normalized rainfalls {Y(i), i=1,2,…, n} as a multivariate normal distribution.

Results of the Kolmogorov-Smirnov test indicate that at = 0.05 significance level, the null hypothesis was not rejected for most of Y(i)’s.

The few rejected normalized rainfalls occur in the beginning or near the end of an event, and have less rainfall rates.


Evidence of Nonstationarity

In general,

Autocovariance function of a stationary process:

For a non-stationary process, the autocovariance function is NOT independent of t.


Calculation of Autocorrelation Coefficients of a Nonstationary Process


Design storm hyetographs


Translating hyetographs between storms of different durations

Some hyetograph models in the literature are duration-specific (the SCS 6-hr and 24-hr duration hyetographs) and return-period-specific (the alternating block method (Chow, et al., 1988).

The simple scaling property enables us to translate the dimensionless hyetographs between design storms of different durations.


Translating the dimensionless hyetographs between design storms of durations D and D is accomplished by changing the incremental time intervals by the duration ratio. Values of the normalized rainfalls Y(i) (i=1,2,…,n) remain unchanged.


For example, the incremental time intervals () of the design storms of 2-hr and 24-hr durations are five (120/24) and sixty (1440/24) minutes, respectively.

The changes in the incremental time intervals are important since they require the subsequent rainfall-runoff modeling to be performed based on the “designated” incremental time intervals.


A significant advantage of the simple scaling model is that developing two separate dimensionless hyetographs for design storms of 2-hr and 24-hr durations can be avoided.

The incremental rainfall depths of a design storm are calculated by multiplying the y-coordinates of the dimensionless hyetograph by the total depth from IDF or DDF curves.


IDF Curves and the Scaling Property

The event-average rainfall intensity of a design storm with duration D and return period T can be represented by

From the scaling property of total rainfall


IDF Curves and Random Variables

is a random variable and represents the total depth of a storm with duration D.

is the 100(1-p)% percentile (p =1/T) of the random variable, i.e.,

),( DDh

),( DDhT


Random Variable Interpretation of IDF Curves


IDF Curves and the Scaling Property

Horner’s Equation:

D >> b , particularly for long-duration events.

Neglecting b

C = - H

c

m

T bD

aTDi

)()(

)()( DiDi Tc

T


Regionalization of Design Storm Hyetographs

Cluster analysis


DDF (IDF) curves under scaling property

Development of scaling , or scale-invariant, models enable us to transform data from one temporal or spatial model to another and help to overcome the problem of inadequacy of the available data.

Let X(t) and X(t) be the annual maximum rainfall depths corresponding to design durations t and t, respectively. The scale-invariant property of the annual maximum depths implies

)()( tt qn

q

,2,1,)]([)]([ rtXEtXE rnrr

))]()([)]()([( qttXPttXP qq


The above equation is referred to as the “wide sense simple scaling” (WSSS). The property of WSSS can be easily checked from the data since

nrnr

r

r

t

t

tXE

tXE

)]([

)]([

nrtt

tXEtXE rr

log)log(

)]([log)]([log


Plot of log(Xr(t))~log(t) has a slope of rn, if the WSSS holds for X(t).


Empirical evidence indicates that temporal rainfall and its extreme events sometimes may be far from simple scaling, and can exhibit a multiple scaling behavior.

The concept of wide sense multiple scaling (WSMS) is characterized by ,2,1,)]([)]([ rtXEtXE rnrr r

)1( 1


Simple scaling DDF


The above equation implies the coefficient of skewness and coefficient of kurtosis are invariant with duration.


Parameters estimation


Multiple scaling DDF


Parameters estimation

Rainfall Process

Documents

Transcript of Rainfall Process