Flood Forecasting Using Neural Networks

A. R. Ghumman, U. Ghani, M. A. Shamim

Civil Engg. Deptt., University of Engg. & Tech. Taxila

Abstract. This paper deals with flood routing in rivers using neural networks.

The unsteady river flow may be formulated in terms of two one-dimensional

partial differential equations. These are the Saint Venant flow continuity and

dynamic equations.

Several methods of solution of these equations are known. These methods are

based upon characteristics of equations, finite difference, finite element and

finite volume. All of these methods have some limitations regarding data

requirements and complications involved in solution of equations. Neural

network techniques have been developed recently. These are easy to use and

need comparatively less data and less labor for solution of the problem. One of

these techniques is used in this research work. The model was applied for flood

routing in River Chenab in Pakistan. Its reach from Marala to Khanki was

selected. Date for various flood events was collected from Meteorological

Department, Lahore and Flood Commission Islamabad. The error between the

observed and simulated values of flood hydrograph ordinates was found to be in

acceptable range.

1 Introduction

Rivers, which provide water to the millions of people settled on their banks and flood

plains, can also bring calamity in the form of floods. Today, all over the world,

hundreds of floods occur annually. The consequences of these floods are the

persistent recurrence of damage and tragedy of death and disease, and a constant

menace to the progress of industry and civilization.

The climatic conditions of Pakistan are hot rainy in summers and dry cold in

winters. There are five large rivers, which are in flood during the monsoon season.

These floods are devastating. For example, during 1988 and 1992, the floods caused

widespread human sufferings, loss of lives and colossal damage of private and public

infrastructure. In 1988, flood resulted in the loss of 370 people and damaged more

than one million hectare of agricultural land. In 1992, flood caused even greater

damage, more than 1000 people died and about 13000 villages with about one million

houses were destroyed. Over two million hectares of agricultural land inundated

resulting in the loss of about 15 percent of both cotton and rice crop. The nation wide

damage was estimated at about $ 2.2 billion (Khan R.A., 2001).

In order to safeguard property and lives, flood control and its dissipation play an

important and vital role. Flood forecasting becomes utmost important in such

Ghumman A., Ghani U. and Shamim M. (2004).

Flood Forecasting Using Neural Networks.

In Proceedings of the First International Workshop on Artiﬁcial Neural Networks: Data Preparation Techniques and Application Development, pages

9-15

DOI: 10.5220/0001148800090015

Copyright

c

SciTePress

conditions. Flood routing techniques have been developed mainly for the study of

floods traveling through channels, reservoirs and lakes. The technique determines the

time and magnitude of flood peaks occurring at various points along a river as the

flood travels downstream. Presently, flood routing is employed for a wide variety of

problems associated with water use. Some of these include; (i) evaluating past floods

for which records are incomplete; (ii) determining hydrograph of channel flow from

hypothetical design floods on tributaries and upstream reaches of the main channels:

(iii) forecasting floods along the main course of a river, by use of predicted

hydrographs at key points in the drainage network; (iv) determining hydrographs

modified by reservoir storage; and (v) studying the effects of water resources

development on the downstream flow conditions.

Flood routing is carried out by solving the unsteady flow equations. These are

partial differential equations and for solution of these equations, many simplifying

assumptions such as uniform roughness, constant channel cross-section with constant

bed slope etc., are made. Without these assumptions, the relationship between

variables becomes more complex and the method of solution needs more data to

describe the channel and wave conditions. Faced with the limitations of time and

economy in the preparation of data, one has to resort to approximate methods of

solution. Along with the labor required for solving the equations, a tremendous work

is required for identification of some of the parameters required, (see for example

Ghumman A.R. 1996).

An effort is made in this study to use a neural networking technique and prepare a

flood-warning scheme for rivers in Pakistan. A reach of River Chenab between

Marala and Khanki is selected for this as it receives flood nearly every year for the

last so many years.

2 Neural Networking Models

Artificial Neural networks are increasingly used in predicting and forecasting water

resource variables (Nash, J.E. and Sutcliffe (1970), French et, M.N. (1992), Zhu, M.L.

and Fujita, M. (1994), Dawson C. W. & Wilby R. L.(2001), Yi-Ming Kuo (2003)).

Hydrologic models can be divided into three broad categories, namely: Physical

distributed models, lumped conceptual models and black box models.

Physical based distributed models require excessive field data whereas in case of

lumped conceptual models, large number of parameters and subsequent difficulty in

calibration is involved. Both of these models are used where detailed understanding of

the hydraulic phenomenon is necessary. Black box models do not contribute much in

enhancing the understanding of hydrological and hydraulic phenomena; nevertheless

in operational hydrology and hydraulic Engineering their usefulness is of utmost

importance. Neural Networking models can be considered as black box models. These

are easy to use and have comparatively less data requirements. This is the reason why

they are becoming popular and are recently being used in the field of Water

10

Resources Engineering also. There are several neural network softwares. EasyNN was

used in this study.

3 Training

The training process estimates the Artificial Neural Networks (ANN) weights and is

similar to the calibration of a mathematical model. The ANNs are trained with a

training set of input and known out put data. The weights are initialized either with a

set of random values, or based upon some previous experience. These weights keep

on changing till the goal is achieved. The goal of learning is to determine a set of

weights that will minimize the error function.

4 Training and Validation Example

The input data of the model were taken as the observed 6 hourly discharges at Marala,

River Chenab, Pakistan. The 6 hourly measured discharge data at Khanki were used

as the target discharges in the EasyNN model calibration and verification. The target

was to forecast flood hydrograph at Khanki. By considering the data of 1973 flood

hydrograph at Marala the training was carried out. The flood data of August 1976 and

1986 was used in the model testing.

5 Data Processing

The model was run both without and with data processing. For processing purpose the

data was normalized by the following formulae

Q

pi

max

Q

Q

i

=

where Q

pi

is processed discharge at ith time step. Q

i

is discharge at ith time step

and

Q

max

is the peak discharge

After normalizing all the input values, the data was again entered into the EasyNN

software. The results were used to calculate the efficiency of the model.

11

6 Network Architecture

6.1 Architecture – I

In the grid, two input columns in one test and four in the second case and one output

column were made. The total number of training example rows were 47. The grid did

not contain any validating rows and had only one querying row. The learning rate and

momentum were set to be 1.0 and 0.6 respectively and were optimized. Growing layer

no. 1 generated the new network with the growth rate changing after every 10 cycles

or 5 seconds.

In the network two input nodes in one test and four in the second case were one

output node were selected. Hidden layer no. 1 was provided with five nodes in first

case and seven in second case whereas hidden layers 2 and 3 were not provided with

any nodes. The first test then started learning and gave the following results.

Hidden layer 1 No. Hidden Nodes 5 Nos

Learning Cycles 517 Learning Rate 1

Learning Momentum 0.6 Minimum error 0.002128

Average error 0.049995 Maximum error 0.228

Target error 0.05

The second test gave the following results after learning.

Hidden layer 1 No. Hidden Nodes 7 Nos

Learning Cycles 273 Learning Rate 1

Learning Momentum 0.8 Minimum error 0.000301

Average error 0.049738 Maximum error 0.323219

Target error 0.05

6.2 Architecture – II

The grid was same as that in architecture-1. The learning rate and momentum were set

to be 1.0 and 0.7 respectively and were optimized. Growing layer no. 2 generated the

new network with the growth rate changing after every 20 cycles or 2 seconds.

In the network two input nodes and one output node were selected. Hidden layer

no 1 was provided with three nodes whereas hidden layers 2 and 3 were not provided

with any nodes. The network then started learning and gave the following results.

Hidden layer 1 No. Hidden Nodes 3 Nos

Learning Cycles 677 Nos Learning Rate 1

Learning Momentum 0.7 Minimum error 0.000951

Average error 0.049995 Maximum error 0.211899

Target error 0.05

12

It was found that the first test in the former network yields a better accuracy than

the other one.

7 Model Performence

The model performance was evaluated both qualitatively by the visual comparison of

the simulated and observed hydrographs and quantitatively using a statistical

parameter namely the Model Efficiency Index (EI) given by the following equations.

ST

SEST

EI

−

=

Where

2

1

)( QQST

Si

N

i

−=

∑

=

2

1

)(

OiSi

N

i

QQSE −=

∑

=

Si

N

i

Q

N

Q

∑

=

=

1

1

, N is the number of data points. Q

si

is simulated discharge

and Q

oi

is the observed discharge at ith time step.

8 Results and Discussion

8.1 Training and Validation Run 1

Architecture 1 was used. Data of 1973 flood was used for learning purpose and that of

1976 & 1986 was used for validation.

The observed and simulated hydrographs are shown in fig. 1 & fig. 2. It is

observed from fig. 1 that observed and simulated hydrographs are similar. For two

input columns, a slight variation in time to peak is observed. The E.I in this case is

nearly 60%. In case of fig. 2 the simulated and observed hydrographs match with each

other. The error between observed and simulated discharges is small. The value of E.I

is 77.3%. In case of four input columns the EI of 1976 flood was found to be 62.32%

and that for 1986 flood was found to be 79.28%.

8.2 Training and Validation Run 2

Architecture 2 was used in this test run. Data of flood hydrographs was taken as same

as that in test run-2. EI in case of 1976 flood was nearly 50% and that in 1986 flood

was estimated to be 30%. This architecture did not give good results

13

8.3 Training and Validation Run 3

Test run 1 was repeated by normalized data. It was observed that the EI was improved

and came out to be 68% for the first figure and 84% for the second. The smaller EI in

case of 1976 hydrograph is due to the difference in shape of hydrographs used for

training and validation.

0

100

200

300

400

500

600

700

800

0 25 50 75 100 125 150 175 200 225 250 275 300

TIME (HOURS)

DISCHARGE (CUSECS)

Q(obs)cusecs

Q (sim)cusecs

Fig. 1. 1976-Flood Hydrographs at Khanki

0

50

100

150

200

250

300

350

400

0 25 50 75 100 125 150 175 200 225 250

TIME (HOURS)

Discharge (cusecs)

Q(Obs) (Cusecs)

Q(Sim) (Cusecs)

Fig. 2. 1986-Flood Hydrographs at Khanki

14

9 Conclusions

A neural network model for discharge forecasting at Khanki, River Chenab, Pakistan

was developed. The model developed was found to perform very well in both training

and validation. The efficiency index of the model was found to be about 84%. The

neural network model is however still dominated by trial and error process in many

aspects. It is important to mention that the selection of the network architecture

significantly influences the output performance of the model and the computational

time. The model requires discharge data but not the topographical data.

10 Future Recommendations

More studies are encouraged to apply neural network models for flood forecasting in

other reaches of this river and other rivers of Pakistan so that the models could be

applied with confidence in operational flood forecasting and warning.

References

1. Dawson, C. W. and Wilby, R. L. “ Hydrological Modeling using Artificial Neural

Networks”, Progress in Physical Geography, Arnold, 2001.

2 French, M.N.; Krajewski, W.F. and Cuykendall, R.R. “Rainfall forecasting in space and

time using a neural network”, Journal of Hydrology, Vol. 137, 1-31, 1992.

3. Ghumman A, R, “Parameter Identification for Sediment Routing in Rivers”, J.o Hydraulic

Research, Vol. 34, 1996.

4. Khan R.A., “Flood Routing for using Hydrodynamic Model” M.Sc Thesis, Civil Engg.

Deptt, Univ. of Engg. & Technology, Taxila, Pakistan, 2001.

5. Nash, J.E. and Sutcliffe, J.V., “River flow forecasting through conceptual models, part I, A

discussion of Principles”, Journal of Hydrology, 10(3), pp. 282-290, 1970.

6. Yi-Ming K. Chen-W. L, Kao-Hung, L., “Evaluation of the ability of an artificial Neural

Network Model to Assess the Variation of Groundwater Quality in an Area of Blackfoot

Disease in Taiwan”, Water Research, ELSEVIER, 2003.

7. Zhu, M.L. and Fujita, M., “Comparison between fuzzy reasoning and neural networks

methods to forecast runoff discharge”, Journal of Hydro science and Hydraulic Engineering,

Vol. 12, No. 2, 131-141, 1994.

15