You are on page 1of 4

2011 Fourth International Symposium on Computational Intelligence and Design

Application of an improved neural network


to flood forecasting of the Lower Yellow River
Bo Pang, Yuan Liang
College of Water Science, Beijing Normal University
Key Laboratory of Water and Sediment Science, Ministry of Education
Beijing 100875, China
pb@bnu.edu.cn
satisfactory results, especially in large areas exhibiting
marked seasonal behavior (Nash and Brasi, 1983; Liang and
Nash, 1988; Kachroo et al., 1988). However, there were still
remained some limitations in the application of LPM models.
Subtracting of the seasonal means only removed part of the
non-linearity of the flood generation and routing process.
The relationship between the departures was still non-linear.
When the seasonal behavior of the flood procedure is not
significant, linear response function couldnt simulate the
relationship successfully.
The application of the Artificial Neural Network (ANN)
in flood forecasting has become very popular in recent years
(French et al., 1992; Campolo et al., 1999; Khatibi et al,
2011). The advantage of neural networks is capable of
treating complex problems of nonlinear processes. In
particular, it can be an efficient tool for prediction
applications. The paper presents a model named non-linear
perturbation model based on the artificial neural network
(the NLPM-ANN). The model structure is similar to that of
the Linear Perturbation Model (LPM). The deference is that
ANN, instead of linear response function, was used to
simulate the unknown relationship between the input
perturbing terms and the output perturbing terms. The
objective of this study is to investigate if the ANN
approaches in NLPM-ANN can significant improve the
efficiency of flood forecasting, prior to that of linear
approaches in LPM. And considering the seasonal effects,
which also represents the state of the watershed, can
enhance the ANN models efficiency in simulating mode. A
comparison study is conducted involving application of also
the LPM model and the ANN model to the flood forecasting
of lower yellow river in China.

AbstractConsidering seasonal feature of the flood events, a


nonlinear perturbation model based on Artificial Neural
Network is developed. The model structure is similar to that of
the Linear Perturbation Model. The deference is that ANN,
instead of linear response function, was used to simulate the
unknown relationship between the input perturbing terms and
the output perturbing terms. The reach from Huayuankou to
Sunkou, located in the lower yellow river, is selected to test
flood forecasting with this model. The proposed model was also
compared with the LPM model and ANN model. It was found
that the NLPM-ANN model was significantly more efficient
than the original linear perturbation model. The results
demonstrate that the relationship between the perturbations is
high nonlinearity though subtracting the seasonal means and
ANN is capable to simulate the relationship. The results also
indicate that considering the seasonal information can improve
the model efficiency. Subtracting the seasonal means, which
adopted in the LPM, is also a feasible way to reduce the system
complexity and improve the model efficiency of ANN models.
Keywords- flood forcasting; Non-linear perturbation model;
Artificial Neural Networks; Linear Perturbation Model;

I.

INTRODUCTION

Flood is one of the most common natural hazards in the


world, which poses a great threat to human lives and
properties. Flood forecasting plays an important role in the
early warning of flood. Accurate and timely flood
forecasting can aid disaster preparedness, which reduces the
human costs and economical damages induced by flood
events. The mathematical models applied for flood
forecasting are broadly of two types: black-box and
conceptual models. Due to the complexity of the various
physical phenomena involved in the flood generation and
routing, conceptual models require extensive survey and
detailed hydrometric data. It will be costly and sometimes
there are various reasons not to be able to meet data
requirements. Therefore, several black-box models have
been developed and used in practice.
In order to use the information contained in the observed
seasonal variation of the hydrograph, Nash and Brasi (1983)
developed the linear perturbation model (LPM). On
catchments with highly seasonal variation of discharge,
subtraction of the seasonal means from the original series
would remove much of the dependence of linearity. The
LPM was applied in some regions of the world and get
978-0-7695-4500-4/11 $26.00 2011 IEEE
DOI 10.1109/ISCID.2011.112

II.

REVIEWS OF THE COMPARISON MODELS

1) The LPM model


The linear perturbation model (Kachroo et al., 1988),
originally referred to as the hybrid model (Nash and Brasi,
1983), attempted to incorporate and utilize the seasonal
information of actual observed rainfall and discharge time
series in input-output modeling. The LPM is based on two
assumptions. One is that the model output (discharge) is
identical to its inputs (rainfall, discharge in upstream, et al.).
Another is that the input departure series and output
departure series are linearly related. The relation between the
43

rainfall departures xi and the corresponding discharge


departures yi can be expressed in the form

the nodes in each layer. The weights associated with the


connections between the input and the hidden nodes are
denoted by vij , 0 d i d n , 1 d j d l and those between the

yi

G xt  j 1 h j  ei

(1)
where m is the memory length of the region and et
designates the model output error. hi denotes the discrete
series of pulse response ordinates, and
j 1

h
i 1

hidden and the output nodes are denoted by w jk , 0 d j d l ,


1d k d m.
For node Z

in the hidden layer, its effective income

signal, denoted by z c j , is calculated as

(2)

z cj

The input to the LPM consists of two parts. The first is


the series of the seasonal expectations of the input (pd) which
is transformed to the series of the seasonal expectations of
the output (qd) though an undefined relation. The second part,

v 0 j  vij xi , 1 d j d l

(6)

i 1

where x i , 1 d i d n , represents the input to each node in


the input layer.
For node Z j , its corresponding output signal, denoted

which is the remainder of the input ( xi Pi  p d ), is routed


though a linear system. The total output of the LPM is then
the arithmetic sum of the seasonal expectations of the output
(q ) and the output ( y i Qi  q d ) of the linear system.

by

z j , is obtained by using an activation function f (x)


zj

f ( z cj ) , 1 d j d l

(7)

The most widely used activation function is the sigmoid


function (Haykin, 1994). Among several different sigmoid
functions, the one most often used for ANNs is the logistic
function
1
(8)
z j f ( z cj )
1  exp(V z _ in j )
where V is an adjustable parameter used in the activation
function f ( x) .

The seasonal means pd and qd can be get as the seem


means of the seasonal model (SM). That is,
1 L
p d
P
L r 1 d ,r
(3)
q d

1 L
Q
L r 1 d ,r

(4)
where the Pd,r and Qd,r are the observed input and output on
date d in year r and L is the number of years in the
calibration period. As L becomes large, the estimate of pd
and qd are expected to approach a smoothly varying function
of date. To remove small sample fluctuations, the values
obtained from (4) are smoothed by obtaining the Fourier
series representation. For qd, it can be expressed as
p
2S jd
2S jd
(5)
q d a 0  A j cos

 B j sin
n
n
j 1
where a0 is the mean of qd, Aj and Bj are the Fourier
coefficiens, j is the order of the harmonic, p is the maximum
number of harmonics and n is equal to 365 for a daily series.

Among the algorithms used to perform supervised


training, the backpropagation method (Rumelhart et al.,
1986) has emerged as the most successful algorithm for the
design of the multilayer feed forward neural networks
(Haykin, 1994) and in hydrology, and has already been used
widely.
THE NON-LINEAR PERTURBATION MODEL BASED ON
THE ARTIFICIAL NEURAL NETWORK (THE NLPM-ANN))

III.

With the objective of considering both the seasonality


effects and the nonlinearity of the flood generation and
routing process, the nonlinear perturbation model adopting
the artificial neural networks (the NLPM-ANN) is
developed. The influence of seasonal changes is taken into
account by a seasonal mean analysis combined with the
concept of perturbation modelling (Nash and Brasi, 1983).
Without adopting the assumption of the linear relationship
between the input perturbations and the corresponding
output perturbations in LPM, the NLPM-ANN regarded the
relationship as non-linear. Considering the powerful
capability of modelling systems where the explicit form of
the relationship between the variables involved is unknown,
ANN is adopted to simulate the relationship.
In present paper, the procedure for calibration of the
NLPM-ANN can be summarized as follows:

2) The ANN model


Many ANN structures have been proposed and
explored for flood forecasting. Among these different
structures, the multilayer feed forward networks have the
best performance in the context of input-output function
approximation (Haykin, 1994). As a matter of fact, almost
all ANNs explored in flood forecasting are multilayer feed
forward networks (Campolo et al., 1999).
The artificial neural network with a single hidden layer
was used widely. It has input nodes ^X i ( p )`in 1 (and a bias),

hidden nodes ^Z j ( p )`l (and a bias) and output nodes


j 1

^Yk ( p)`mk 1 , where X, Z, and Y represent the input, hidden and

(1) Calculate the smoothed seasonal means id and qd. The


seasonal means are estimated by the same procedure as

output layer respectively, n, l, and m represent the number of

44

that used in the case of the LPM. From (3)~(5), id and qd


can be estimated. By subtracting the seasonal means id
and qd from the original series Pi and Qi, the system
perturbations xi and yi can be estimated.

Sutcliffe, 1970). The second index to assess the model


performance is the index of volumetric fit (IVF), which is
defined as the ratio of the simulated runoff volume to that
observed. Where

(2) Choose initial values of the parameters of the network


(i.e. the connection weights and the neuron thresold
values) and use error back propagation algorithm, find
estimates of the parameter values of the network in such
a way as to minimize the objective function, i.e. the sum
of the squares of the deference between the network
output and the outflow perturbations.

IV.

(Q  Qc)
(Q  Q )
i

Qc
Q
i

(10)

u 100(%)

(11)

Where Qi and Q c are the observed and model estimated


discharges, Q is the mean discharge of the calibration
period. Though these two criteria have some limitations,
they are still the most widely used indices in flood
forecasting for calibrating model parameters and verifying
the model performance (Nash and Sucliffe, 1970).
3) The results
Because there is no local inflow between the two
satations, the discharge of Sunkou station can be expressed
as follow

, where
hence the model simulated outflow Q
i
y i  q d

1

IVF

(3) Calculate the estimate of outflow perturbation y i and


Qi

R 2 and IVF can be expressed as follow

(9)

COMPARATIVE STUDY

1) Study Area
The yellow river is the second largest river in China,
Which has a drainage area of 752,000 km2 and a length of
5464 km. The lower yellow river, downstream of Taohuayu,
has a length of 786 km. Because of the severe sediment
deposition in the river channel, the riverbed is rising
continuously and now up to 10m higher than the surrounding
land surface in some places, which is known as a suspended
river .The plain is an area of great floods because the
riverbed. The river has changed its course across the plain
several times. So the flood forecasting is an important and
difficult mission for the lower yellow river.
Huanyuankou and Sunkou are two main control stations
in the lower yellow river, and the geographical locations are
shown in Fig. 1 (Nicolas, 2008).The discharge time series of
the two stations in flood season (July 1st to October 1st)
from 1998 to 2001 are select to calibrate the NLPM-ANN
model, and the series from 2002 to 2003 are select to verify
the model.

Qs (t )

f (Qh (t  W ), Qh (t  W  1),Qh (t  W  m  1))

(12)

Where Qs(t) is the dischage of Sunkou station, Qh(t) is the


dischage of Huayuankou station, andis the the minmum
travel time of flood wave. By observing the historical
dicharge data of Huayuankou and Sunkou, the minmum
travel time of flood wave between the two stations is 36h. m
represents the influenced time range of input, which is also
called memory length in LPM model.
The results of the LPM, the ANN and the NLPM-ANN
are summized in Tab. 1. The simulated output perturbations
in verification period were plotted in Fig.2. The simulated
flow hydrographs in 2003 were plotted in Fig.3. The results
can be concluded as follows:
(1) The results show that an improvement in model
efficiency values of both calibration and verification periods
were obtained with NLPM-ANN over the LPM. According
to the Nash-Sutcliffe model efficiency index R 2 , the
increases of the NLPM-ANN are 23.58% over the LPM in
calibration, and in verification is 6.31%. Fig. 2 shows the
simulation results of the output perturbations. It is clearly
shown that the NLPM-ANN obtains better simulation
results than the LPM.
(2) A comparison of the NLPM-ANN result with those
of ANN shows that, in terms of model efficiency R2, the
overall performance of NLPM-ANN is better than that of
the ANN. The increases of the NLPM-ANN are 0.03% over
the ANN in calibration period; the corresponding value is
2.6% in verification. It is shown that the performance of
NLPM-ANN is better than that of ANN, though their input
information is similar. Fig. 3 shows the observed and
simulated discharge in 2003. It is shown that the NLPM-

Figure 1. The location of Huayuankou and Sunkou stations.

2) The model efficiency criteria


The main model criterion for assessing the NLPMANN model efficiency is chosen to be the widely used
Nash-Sutcliffe model efficiency index R 2 (Nash and

45

V.

ANN has a better performance in simulating the flow


hydrograph than ANN.
TABLE I.

The non-linear perturbation model adopting artificial


neural network (NLPM-ANN) is presented in this paper.
Some discussions and conclusions of this study are
summarized as follows.
In this study, it is recognized that seasonal hydrological
behavior, as incorporated into the modeling by the LPM
approach is a very important source of information in flood
forecasting. The results showed that the model efficiency
was improved when adding the seasonal information as
model input for ANN model.
The results of NLPM-ANN were compared with those
of the LPM. It is found that the results of the NLPM-ANN
are better than those of the LPM. It can be concluded that
the relationship of the perturbations still remains highly
non-linear though subtracting the seasonal means has
removed much of the nonlinearity of the rainfall-runoff
process. And ANN has a better performance on simulation
the relationship rather than linear response function.
In conclusion, the NLPM-ANN model is a flexible tool
for flood forecasting, especially in the area without detailed
hydrometric data, a not uncommon situation particularly in
developing countries.

SUMMARY OF RESULTS
Calib. period

MODEL

Valid. period

m
R2(%)

IVF

R2(%)

IVF

NLPM-ANN

99.06

1.043

91.13

0.952

LPM

75.48

1.052

84.82

0.940

ANN

99.09

0.999

88.43

0.946

m /s

2500
2000
1500
1000
500
0

-1000

-500

500

1000

1500

2000

2500

-500
3

-1000

m /s

LPM results
Observed data

-1500

REFERENCES:

m /s

(a) LPM model


2500

[1]

D. E. Rumelhart, E. Hinton, J. Williams, Leaning internal


representation by error propagation, in Parallel Distributed
Processing, pp. 318-362, MIT Press, Cambridge, Mass.,USA, 1986.
[2] G.C. Liang, Linear models for river flow routing on large
catchments. J. Hydrol., vol. 103, 1988, pp. 157-188,
doi:10.1016/0022-1694(88)90012-1.
[3] H. Nicolas, Fully coupled 1D model of mobile-bed alluvial
hydraulics: application to silt transport in the LowerYellow River,
PHD report: Bruxellesk kuniversity, Aug 2008, pp. 48.
[4] J. E. Nash and B.I. Brasi, A hybrid model for flood forecasting on
large catchments. J. Hydrol., vol. 65, 1983, pp. 125-137,
doi:10.1016/0022-1694(83)90213-5.
[5] J. E. Nash, J. Sutcliffe, River flow forecasting through conceptual
models. J. Hydrol., vol.10, 1970, pp. 282-290, doi:10.1016/00221694(70)90255-6.
[6] M. N. French, W. F. Krajewski, and R. R. Cuykendall, Rainfall
forecasting in space and time using a neural network. J. Hydrol., vol.
137, 1992, pp. 1-31, doi:10.1016/0022-1694(92)90046-X.
[7] M. Campolo, P.Andreussi, A. Soldati, River flow forecasting with a
neural network model. Water Resour. Res., vol. 35, 1999, pp. 11911197.
[8] R. Khatibi, M. A. Ghorbani, M. H. Kashani, Ozgur Kisi,
Comparison of three artificial intelligence techniques for discharge
routing,
J.
Hydrol.,
vol.403,
2011,
pp.
201212,
doi:10.1016/j.jhydrol.2011.03.007.
[9] R. K. Kachroo, G. C. Liang, K.M. OConnor, Application of the
linear perturbation model (LPM) to flood routing on the Mekong
River. Hydrol. Sci.J., vol.33(2-4) , 1988, pp. 193-214,
doi:10.1080/02626668809491238.
[10] S. Haykin, Neural Networks, MacMillan College Publishing
Company, New York, USA, 1994.

2000
1500
1000
500
0
-1000

-500

500

1000

1500

2000

NLPM-ANN results

-500

2500
3

m /s

Observed data

-1000

(b) NLPM-ANN model


Figure 2 Comparison of the output perturbations simulation results

m /s

3000
2500

Observed
NLPM-ANN

2000

ANN

1500
1000
500

29
20
03
-1
014
20
03
-1
029
20
03
-1
113

14

20
03
-9
-

30

20
03
-9
-

15

20
03
-8
-

20
03
-8
-

31

16

20
03
-7
-

20
03
-7
-

0
20
03
-7
-

DISCUSSIONS AND CONCLUSIONS

Figure 3 The observed and simulated discharge in lower yellow river

46

You might also like