VNU JOURNAL OF SCIENCE, Nat., Sci.. & Tech., T .x x , N03, 2004

A P P L Y IN G A R T IF IC IA L N E U R A L N E T W O R K S (A N N ) M O D E L IN
F L A S H F L O O D S IM U L A T IO N A N D F O R E C A ST
N g u y e n H u u K h ai
D epartm ent o f Hydro-M eteorology & Oceanography, College o f Science ,V N U
Le X u a n Cau
M in istry o f Resources & Environm ent

Flash floods u sually h ap pen in small basins where the m easured data of flows is not
enough. Determining p a ra m e te rs of conceptual models such as: TANK, NAM, SSARR or
HEC meets a lot of difficulties. Artificial N eural Networks (ANN) model is one suitable
solution to sim ulate and forecast flash flood.
1. P relim in ary in tr o d u c t io n to ANN m o d el
ANN model establish es a combination of some am ount of input and output variables
through a form of sem i-character. ANN activities are similar to activities of the brain. Input
is considered as stim ulation, b ut o u tpu t m eans a responses. ANN can study through
examples, then generalize its characteristics to m eet optimal responses.
ANN uses N euron as a basic fundam ental correctional unit (fig.l). Each neuron is
specifies by some com ponents as follows:
- Active level.
- Connecting com bination of neuron inputs .
- O utput
- Threshold value.

O utput

---------- ►

Figure 1: Fundamental correctional unit of ANN
Where:

(gradient method).
(Input Layer)

(Hidden Layer)

(Output Layer)

XI

Input signal Xị

O utput signal Xj

O utput signal y-j O utput signal z k

Figure 2: Multi-layer ANN (Network of Dinh river-Binh Thuan Prov.)
In ANN, input signs are distributed among hidden nods. After that, these hidden
nods change them into output signs. These signs are transm itted to output of ANN.
If the input of ANN are Xị, (i = 1-Ninp), then output of hidden layers will be y-j, G=l-Nhitl):
Nmp
N
Yj = G

a o j + ^ a ij y i
i=i

( 1)
J

and output will be z k , k = 1-Nout'

process to find out the minimum error function, th a t is a mean square error MSE:
-I

MSE = —...... i —---^ exam

^fexam^N'out
£
(obsj - mod; )2

out

(4)

i=i

Where:
N,.xani: examples patterns for studying,
obs, :output observed patterns,
mod, : output patterns calculated from ANN .
a, b are the param eters optimized by the gradient method.
Versions of ANN are built up to create advantages for running models and outputting
results. WinNN version 0.97, built by Y.Danon, April 1995, has been used in this document.
ANN model doesn’t require continuous data. It allows analyzing and choosing param eters
of all floods a t the same time. That is a real advantage as compared to the black-box and
conceptual models in hydrology. It also allows establishing directly the relation between
rainfalls and w ater levels without using flows and overcoming principal difficulties, such as
maintaining a station observing discharge in small basins.
2. A p p lyin g ANN to sim u la te and foreca st flash flood
We apply ANN to sim ulate and to forecast flash floods and great floods for some
basins, in which there are Dinh river in Binh T huan province (F=435krrr for Z30D station),


1997

1998

1999

1055

1004

1041

1100

1355

1040.6

1099.1

1351.8

1054.2

932.2

I4(X)

Tliucdo

Pat. Index

Figure 3: Comparison to simulated and observed process
Verification with independent data series, chain 5 patterns from 1 to 5 are chosen,
responding to a flood. ANN will study the rest patterns to determine param eters, and then
compute for 5 chosen patterns. Result will be shown in fig.4:

T arget

N c t- I

60

cL

A
P a t. I n d e x

Figure 4: Verification by ANN for flash flood of Dinh river

V N U . Journal o f Science. N a t., Sri., á Tech.. T. XX. N lt3 . 2004


A p p lyin g a rtific ia l neural networks.

61

If foreseeing period was chosen as 12h (equivalent to observed rainfall data) the
forecasting result is rath er well as shown bellow:
+ With an error of 5%, it will have good patterns of p=88%,

8j_

10

1^

16ị

2ŨỊ

24,

28,___ 32

36

PdL Index

Figure 5. Comparison to flash flood process in Nam La basin

V N U . Journal o f Science. N at.. Sri., á Tech..

T.xx,

N t>3, 2004


N g u yen H ull K h ai, Lc X u an Cau

62

+ RMSE = 0.0401,
+ Maximum error = 8,648%.

I(X »

/

Target

Net-out

Net-1

« 1 ___6 , _______ 9ị

12,

I5|

18,

211

24|

27j

30|

33|

+ If initial param eters are incompatibly chosen (The num ber of hidden nodes, inputoutput variables), it will not give excellent results or spend much time.
+ ANN’s param eters are directly determined and adjusted through observed data,
therefore when computing for basin, which has no data can meet difficulties.
However, comparison to the models being used in hydrology nowadays, the results
of ANN is more optimistic, including independent forecasting.
REFERENCES

1.

Keith J. Beven, Rainfall-Runoff Modelling, The Primer. John Winley & Sons. LTD,
Chichester, 2001, 324 pp.

2.

2. M.J. Hall & A.w. Minns, Rainfall-runoff modeling as a problem in artificial intelligence:
experience with a neural network, BHS 4th National Hydrology Sym posium , Cardiff, 1998
pp.23-45.

3.

Le Xuan Cau, Applying ANN model to correct meteo-hydrologic data, Journal of
Meteorology and Hydrology, HaNoi, No.7(1999), 1999, pp.23-29.

4.

Nguyen Huu Khai, Research on flash flood in Dinh river basin, Proceedings international
symposium on achiuements oflH P -V in Hydrological research, Hanoi, 2001, pp. 135-145.

5.