Simulation of flood flow in a river system using artificial

neural networks
R. R. Shrestha, S. Theobald, F. Nestmann

To cite this version:

R. R. Shrestha, S. Theobald, F. Nestmann. Simulation of flood flow in a river system using artificial
neural networks. Hydrology and Earth System Sciences Discussions, European Geosciences Union,
2005, 9 (4), pp.313-321. <hal-00304835>

HAL Id: hal-00304835

https://hal.archives-ouvertes.fr/hal-00304835
Submitted on 7 Oct 2005

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
Hydrology and Earth System Sciences, 9(4), 313–321Simulation
(2005) of©flood
EGUflow in a river system using ar tificial neural networks

Simulation of flood flow in a river system using artificial neural

networks
Rajesh Raj Shrestha 1, Stephan Theobald 2 and Franz Nestmann 2
1
Department of Hydrological Modelling, UFZ - Centre for Environmental Research Leipzig-Halle, Brückstrasse 3a, 39114 Magdeburg, Germany
2
Institute for Water Resources Management, Hydraulic and Rural Engineering, University of Karlsruhe, D-76128 Karlsruhe, Germany

Email for corresponding author: [email protected]

Abstract
Artificial neural networks (ANNs) provide a quick and flexible means of developing flood flow simulation models. An important criterion for
the wider applicability of the ANNs is the ability to generalise the events outside the range of training data sets. With respect to flood flow
simulation, the ability to extrapolate beyond the range of calibrated data sets is of crucial importance. This study explores methods for
improving generalisation of the ANNs using three different flood events data sets from the Neckar River in Germany. An ANN-based model
is formulated to simulate flows at certain locations in the river reach, based on the flows at upstream locations. Network training data sets
consist of time series of flows from observation stations. Simulated flows from a one-dimensional hydrodynamic numerical model are integrated
for network training and validation, at a river section where no measurements are available. Network structures with different activation
functions are considered for improving generalisation. The training algorithm involved backpropagation with the Levenberg-Marquardt
approximation. The ability of the trained networks to extrapolate is assessed using flow data beyond the range of the training data sets. The
results of this study indicate that the ANN in a suitable configuration can extend forecasting capability to a certain extent beyond the range of
calibrated data sets.

Keywords: artificial neural networks, activation function, backpropagation, hydrodynamic numerical model, multilayer perceptron, Neckar
River.

Introduction al., 2002; Shrestha, 2003). A recent review can be found in

The ability to simulate river flows quickly and accurately is the ASCE Task Committee on Application of Artificial
of crucial importance in flood forecasting operations. Neural Networks in Hydrology (2000). Typical applications
Hydrodynamic models provide a sound physical basis for involve the training of two- to three-layer networks using
this purpose and have the capability to simulate a wide range suitable network architectures like multilayer perceptron,
of flow situations. However, these models require accurate radial basis networks or recurrent networks. The
river geometric data, which may not be available in many performances of the ANNs in river flow prediction have
locations. It is also not possible to integrate observed data been found to be comparable with other data driven
directly at desired locations to improve the model results. modelling approaches (Lekkas et al., 2001; Sivakumar et
In this respect, artificial neural network (ANN) provides a al., 2002).
quick and flexible approach for data integration and model An important criterion in application of ANN for flood
development. flow simulation is to predict flows beyond the range of
The use of ANN-based models for the simulation of flood calibrated data sets. The network may perform very well
flows has been gaining popularity in recent years. Several for the training data set, but may be unable to generalise
researchers have demonstrated the application for rainfall- flows beyond the range of training data sets. Minns (1996)
runoff modelling and streamflow simulation (Thirumalaiah applied ANNs to both real and theoretical catchments, and
and Deo, 1998; Dawson and Wilby, 1999; Imrie et al., 2000; found that the peak flows were considerably underestimated.
Solas et al., 2000; Dolling and Veras, 2002; Shamseldin et Thirumalaiah and Deo (1998) used ANNs for river stage

313
Rajesh Raj Shrestha, Stephan Theobald and Franz Nestmann

forecasting and found that although lower water levels were Bias
x1 wk1
predicted fairly accurately, higher water levels were bk
underestimated. Solas et al. (2000) used average, dry and Activation
x2 wk2 function
wet years’ mean annual precipitation in a rainfall-runoff f(yk)
modelling application of ANNs, and observed that high Ȉ
flows were overestimated for the wet years. yk

The reasons why the ANNs underestimate or overestimate xn-1 wkn-1

Output
Summing
extreme flows may lie in the network structure used and junction
range of training data sets. Thirumalaiah and Deo (1998) xn wkn
suggested that this could be due to a smaller number of
training patterns for higher water levels. Minns (1996) Input Synaptic
signals weights
emphasised the need to ensure that the training data sets
actually contain all conceivable events.
Several methodologies have been proposed to improve Fig. 1. Structure of a neuron of an Artificial Neural Network
the forecasting capability outside the range of training data
sets. In the ANN applications, the inputs data are generally
normalised in the range such as [0 – 1.0] or [–1.0 – 1.0]. To weights providing inter-neuron connection for storage of
accommodate the data beyond the training range, alternative knowledge. A summing junction acts as an adder for
normalisation ranges have been suggested. The scaling of summing input signals weighted by respective weights. An
training data in a range such as [0.1 – 0.9] or [0.2 – 0.8] activation function provides limiting amplitude of the neuron
compared to the range [0.0 – 1.0] has been reported to be an output, typically between [0.0 – 1.0] or [–1.0 – 1.0]. The
effective means of improving generalisation (Imrie et al., structure of a typical neuron of an ANN is shown in Fig. 1.
2000; Dawson et al., 2002). This can accommodate The ANNs may also consist of a number of hidden layers
validation and test data sets in excess of training data sets. between the input and output layers. Such networks are
The upper and the lower limits of activation functions such commonly known as multilayer feedforward networks or
as [0 – 1.0] or [–1.0 – 1.0] also provide limiting amplitude multilayer perceptrons (MLPs). The MLPs are one of the
to the data sets and affect the generalisation capability of most widely used types of neural networks, which can be
the ANNs. Imrie et al. (2000) investigated the effects of trained in a supervised manner to solve highly nonlinear
different output activation functions for improving problems. The capability of the MLPs in modelling dynamic
generalisation using cascade correlation network building systems can be enhanced using time delays. These provide
strategy. Shamseldin et al. (2002) examined the significance a sequence of input vectors that occur in a certain time order.
of different non-linear activation functions for the hidden For the ANNs to predict a wide range of flow situations, it
and the output layers in the context of overall performance is important that the networks are able to generalise different
of the multilayer feedforward networks. An alternative ranges of data sets. A network with too few neurons may
approach is the application of a non-linear activation not approximate different flow situations. A network that is
function at the hidden and the linear function at the output too complex may fit the noise, not just the signal, causing
layers. The application of linear activation in the output over-fitting. This may lead to a network that performs very
layers enables the network to take any range of values well for the training data set, but becomes unable to
(Demuth and Beale, 2000). generalise to a new situation.
This study explores methods for improving generalisation It is customary to use validation and test data sets to assess
with different activation functions at the hidden layers using the capability of the trained ANNs. Validation sets are
multilayer feedforward networks. A case study from the separate sets of data, to be used during the training process
Neckar River in Germany demonstrates the application using to monitor the generalisation capability. Normally, the errors
historical flood data sets. of the validation data sets decrease during the initial training
iterations but begin to rise when the network begins to over-
fit the training data. When the validation errors increase for
Artificial Neural Networks the number of iterations in a criterion specified by ‘early
Artificial neural networks (ANNs) are massively parallel stopping’, the training process is stopped. The test sets are
distributed processors with a natural propensity for storing independent sets of data, not used in training or validation,
knowledge and making it available for use (Haykin, 1994). and are to be used to evaluate the network performance.
Typical neural networks consist of layers of neurons with

314
 Simulation of flood flow in a river system using ar tificial neural networks

ACTIVATION FUNCTIONS the sigmoidal function provides the lowest limiting

The activation function transforms signals yk at a neuron amplitude and the linear function the highest limiting
using some function f (yk). Different activation functions amplitude. The limited amplitude range provided by the
investigated in this study are given below. functional output of the sigmoidal and hyperbolic tangent
functions produces a ‘squashing effect’ to the input signals.
i. Linear function calculates the neuron’s output by simply The sigmoidal and hyperbolic tangent functions are most
summing all the input yk and bias b passed to it. This commonly used only at the hidden layers. If the output layers
may be modified by a factor a to provide different use sigmoidal or hyperbolic tangent functions, the outputs
limiting amplitudes. This function is defined as are restricted to a small range of values. The application of
the linear function at the output layer makes it possible for
f(yk) = ayk for all yk (1) the network to take any value.

ii. Sigmoidal function produces an output in the range 0 Study area and data
to +1. This function has the following form
The Neckar is a major tributary of the Rhine and flows
1 (2) through the region of Stuttgart, Heidelberg and Mannheim
f (y k )
1  exp(-y k ) in south-west Germany. The study area consists of a reach
iii. Hyperbolic tangent function is mathematically of about 100 km from Lauffen to Heidelberg (Fig. 3) with a
equivalent to tanh(yk) and produces an output in the catchment area of 13 787 km2 at the Heidelberg station. Time
range -1 to +1. This function is given by series of flow and water level data at one hour intervals are
available from the gauging stations located at Lauffen,
2 Rockenau and Heidelberg and water level time series are
f (y k ) 1 (3)
1  exp(-2y k ) available from Gundelsheim station for the 1988, 1990 and
iv. Hyperbolic tangent + linear function can be used to 1993 flood events. Flow time series from the major
combine non-linearity of the hyperbolic tangent function tributaries Kocher, Jagst and Elz and the smaller tributaries
with the linear function using the weighing factor ±. Schwarzbach, Elsenz and Itter are also available for the same
This function is of the form years. There is also a functioning one-dimensional
hydrodynamic numeric (HN) model of the study area, with
§ 2 · cross-sections at 100 m interval. Only the simulated flows
f (y k ) ¨¨  1 ¸¸D  (1  D )y k (4)
© 1  exp(-2y k ) ¹ from the HN model at Gundelsheim station, where no flow
Figure 2 shows these four different activation functions for
the data range –2 to +2. From the figures it can be seen that

1.5

0.5
f (yk)

-0.5

-1 Linear function
Linear + Hyperbolic tangent function
Hyperbolic tangent function
-1.5 Sigmoidal function

-2
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
yk

Fig. 2. Activation functions Fig. 3. Study Area

315
Rajesh Raj Shrestha, Stephan Theobald and Franz Nestmann

Table 1. Statistical characteristics of the flow data at Rockenau and Heidelberg gauging stations

Data set   Meanflow Maximum Minimum Standard

(m3 s–1) flow (m3 s–1) flow (m3 s–1) deviation

ROCKENAU
Training set (1988) 958 1930 341 399
Validation set (1990) 583 2225 144 576
Test set (1993) 713 2680 232 579
HEIDELBERG
Training set (1988) 1060 1945 374 401
Validation set (1990) 666 2299 224 604
Test set (1993) 807 2706 275 597

data are available, are included for model training and The forecast horizons at Gundelsheim, Rockenau and
validation. The available flow data were divided in such a Heidelberg with respect to the upstream flows are each 2
way that the maximum flows of validation and test data sets hours. However, since the contribution from the tributaries
exceed the range of training data sets. The flow time series Elz, Schwarzbach, Elsenz and Itter are quite small compared
from the 1988 flood event was used as training set, and to flows in the rivers, their flows can be assumed to be
flood event data from 1990 and 1993 were used as validation constant for the duration of forecast to increase the forecast
and test data sets. The statistical characteristics of the flow horizon. This increases the forecast horizon at Rockenau
data at the prediction stations Rockenau and Heidelberg are station to 5 hours and Heidelberg station to 7 hours. The
summarised in Table 1. forecast horizon can be increased further by integrating with
A cross correlation analysis was performed on the time external models such as the rainfall-runoff models.
series flows to identify a suitable lag time from upstream to
downstream points. The cross correlation analyses of the
time series water level data from Lauffen, Gundelsheim and
River flow prediction
Rockenau stations yielded the suitable lag time for The selection of appropriate input and output data sets is an
Gundelsheim and Rockenau with respect to the upstream important consideration in the ANN modelling. A number
stations. Similarly, the analysis between the flow data from of experiments was performed with the division of river
Rockenau and Heidelberg gave the lag time for the reach into different ANN blocks. In the first experiment the
Heidelberg station. The lag times for the tributaries’ inflows ANN was used to predict flows at the gauging station
were calculated based on their distances. The lag times with Rockenau, based on the upstream flows from Lauffen
respect to forecast stations Gundelsheim, Rockenau and (Neckar), and the tributaries Kocher, Jagst and Elz. However,
Heidelberg for the upstream stations in the Neckar river and although the network functioned quite well for the training
the tributaries are given in Table 2. and validation data sets, it did not perform well for the test
data sets.
As a functioning HN model is available for the study area,
Table 2. Lag time with respect to forecast stations it was decided to integrate the HN simulated results from
Gundelsheim for the ANN training. It was observed that
Forecast station  Upstream stations Lag time (hrs) the integration of HN model results from Gundelsheim
improved the performance of river flow prediction at
Gundelsheim (Neckar) Lauffen (Neckar) 4
Rockenau. Accordingly, the river reach was divided into
Stein (Kocher) 3
three ‘sub-reaches’ represented by independently trained
Untergriesheim (Jagst) 2
ANN blocks. The observed flows at Lauffen, Rockenau and
Rockenau (Neckar) Gundelsheim (Neckar) 3 Heidelberg stations, together with the tributaries and HN
Mosbach (Elz) 2 model results from Gundelsheim, were integrated for the
Heidelberg (Neckar) Rockenau (Neckar) 2 ANN training. The inflows and the desired outflows for each
Eschelbronn (Schwarzbach) 2 of the ANN blocks are summarised in Table 3. Outputs of
Eberbach (Itter) 2 the best performing networks from previous blocks were
Meckesheim (Elsenz) 2 used as inputs to next blocks.

316
 Simulation of flood flow in a river system using ar tificial neural networks

Table 3. Network inputs and desired outputs

Network block River sub-reach Network input Desired output

ANN block 1 Lauffen – Gundelsheim Measured flows from gauging stations at Simulated flows from the HN
Lauffen, and tributaries Jagst and Kocher model at Gundelsheim

ANN block 2 Gundelsheim – Rockenau Simulated flows from the ANN block 1 at Measured flows from gauging
Gundelsheim, and measured flow from stations at Rockenau
gauging station at the tributary Enz

ANN block 3 Rockenau – Heidelberg Simulated flows from the ANN block 2 at Measured flows from gauging
Rockenau, and measured flows from stations at Heidelberg
gauging stations at the tributaries Schwarzbach,
Elsenz, and Itter

The input data sets for the ANN trainings were scaled in performing network blocks in each of the sub-reaches (with
the range [0.2 – 0.8]. Each of the ANN blocks was trained the hyperbolic tangent and linear function in the first hidden
with one input layer, two hidden layers and one output layer. layer for the first and third blocks and hyperbolic tangent
The number of neurons in the networks was kept to a function in the first and second hidden layers for the second
minimum of eight in the first hidden layer, four in the second block) were combined in the Simulink environment and an
hidden layer and one in the output layer. There was no ANN simulation model was formulated. The combined ANN
significant improvement of the model performance with the simulation model is schematised in Fig. 4.
increase in number of neurons. Each of the networks
consisted of a linear activation function in the output layer.
The networks were trained with four different activation Error Measurement
functions in the first hidden layer, namely linear, sigmoidal, Table 4 shows the statistics of the error comparison used in
hyperbolic tangent and a hyperbolic tangent + linear this study. The error measurement process consists of
function. The weighing factor ± for the hyperbolic tangent analysis of errors between observed and calculated values.
+ linear function was varied between 0.4 and 0.8 during The overall performance of trained networks can be judged
training process. with respect to criteria such as the coefficient of efficiency
The networks were trained using the procedures from (CE) and coefficient of determination (R 2 ). These
MATLAB neural network toolbox. This involved network coefficients are independent of the scale of data used and
designing using text files containing MATLAB code (M- are useful in assessing the goodness of fit of the model
files). The training is done using a backpropagation (Dawson et al., 2002; Dawson and Wilby, 1999). The root
algorithm with Bayesian regularisation of the Levenberg- mean square error (RMSE) evaluates the error independent
Marquardt approximation. The early stopping criteria of sample size and can give useful insights into amplitude
provided by the validation data sets were used to prevent errors. The difference in peak flow between observed and
overtraining. The test data sets were used independently for calculated flows was considered to assess the prediction
the evaluation of the model performance. The best capability of the trained networks beyond the calibrated range.

Itter

Kocher
Elsenz
Gundelsheim

Rockenau

Heidelberg

ANN ANN
Neckar(Lauffen) ANN
Block 1 Block 2 Block 3
Jagst
Elz Schwarzbach

Correlated Model Validation Model Validation Model

Inflows & Testing & Testing Outflow

Fig. 4. ANN simulation model

317
Rajesh Raj Shrestha, Stephan Theobald and Franz Nestmann

Table 4. Error measurement formula Rockenau

3000
Observed
Error measurement Name Formula HN Simulated
Hyperbolic tangent + linear function
n 2500 Linear Function

Coefficient of efficiency CE ¦ (Q obs  Qcal ) 2 Hyperbolic tangent function

1 i 1
n
Sigmoidal function

¦(Q
i 1
obs  Qav ) 2
2000

Runoff (m3/s)
2
ª n º
Coefficient of determination R2 «¦ (Q
¬i 1
obs  Qav )(Qcal  Qcal _ av ) »
¼ 1500
ª n ºª n º
«¦ (Q
¬i 1
obs  Qav ) » « ¦ (Q
¼¬ i 1
cal  Qcal _ av ) »
¼

1000
Root mean square errors RMSE 1 n
¦ ( Qobs  Qcal ) 2
n i1

Difference in peak flow DPF Qobs (max)  Qcal (max) 500

90 100 110 120 130 140 150 160 170 180
TIme (hrs)

Where n is the number of observations, Qobs and Qcal are the

Fig. 5. Comparison of model performance of different activation
observed and calculated values respectively, and Qcal_av are the functions at Rockenau using upstream flows from Lauffen and
mean of the observed and calculated values. Qobs (max) and Qcal tributaries
(max) are the maximum of observed and calculated values.

ANN block 1 between Lauffen and Gundelsheim, the

Results network could easily approximate the training data sets. For
training and test data sets, there was a tendency to decreasing
PERFORMANCE OF THE ROCKENAU ANN MODELS error when activation functions with higher limiting
The performance of the ANN models for the test data sets amplitude were used. The application of the sigmoidal and
(1993 flood event) for the prediction of flow at Rockenau hyperbolic tangent functions underestimated the peak flows
station are summarised in Table 5. The ANNs were trained whereas the linear function overestimated the peaks. The
with the upstream flows from Lauffen (Neckar) and the application of the combination of hyperbolic tangent and
tributaries Kocher, Jagst and Elz. All the ANN models using linear function (a = 0.5) at the first hidden layer gave the
different activation functions underestimated the flow at best performance in terms of CE, R 2 and RMSE. The
Rockenau station for the test data sets. The ANN with the difference in peak flow was also found to be least using this
activation function with a lower limiting amplitude such as activation function.
a sigmoidal and hyperbolic tangent function in the hidden In the ANN block 2 between Gundelsheim and Rockenau,
layer had a higher level of underprediction. Figure 5 shows the performance of the networks was similar to the ANN
the partial results of the model corresponding to the 1993 block 1. The application of activation functions with higher
flood event, for the period 21.12.1993 to 25.12.1993. limiting amplitude produced better results. However, there
was an overall trend of underestimation of peaks, even with
the linear activation function. Only with the application of
PERFORMANCE OF THE SUB-REACH MODELS hyperbolic tangent activation functions at the first and
The performances of different activation functions for the second hidden layers was the network able to predict peak
test data sets are summarised in Tables 6, 7 and 8. In the flows with reasonable accuracy.

Table 5. Network performance in the reach between Lauffen and Rockenau

Activation functions Coefficient of efficiency Coefficient of determination Root mean square error Difference in peak flow
CE R2 RMSE (m3 s–1) (m3 s–1)

Sigmoidal 0.9498 0.9771 141 273

Hyperbolic tangent 0.9565 0.9777 132 230
Hyperbolic tangent + linear 0.9534 0.9771 142 176
Linear 0.9535 0.9810 137 169

318
 Simulation of flood flow in a river system using ar tificial neural networks

Table 6. Network performance in the reach between Lauffen and Gundelsheim

Activation functions Coefficient of efficiency Coefficient of determination Root mean square error Difference in peak flow
CE R2 RMSE (m3 s–1) (m3 s–1)

Sigmoidal 0.9847 0.9907 72 118

Hyperbolic tangent 0.9887 0.9922 62 54
Hyperbolic tangent + linear 0.9918 0.9937 56 -36
Linear 0.9895 0.9921 66 -176

Table 7. Network performance in the reach between Gundelsheim and Rockenau

Activation functions Coefficient of efficiency Coefficient of determination Root mean square error) Difference in peak flow
CE R2 RMSE (m3 s–1) (m3 s–1)

Sigmoidal 0.9698 0.9804 116 165

Hyperbolic tangent 0.9634 0.9820 121 246
Hyperbolic tangent + linear 0.9639 0.9801 119 205
Linear 0.9650 0.9787 122 154
Hyperbolic tangent in first 0.9806 0.9877 95 -47
& second hidden layers

Table 8. Network performance in the reach between Rockenau and Heidelberg

Activation functions Coefficient of efficiency Coefficient of determination Root mean square error Difference in peak flow
CE R2 RMSE (m3 s–1) (m3 s–1)

Sigmoidal 0.9563 0.9671 136 215

Hyperbolic tangent 0.9566 0.9817 135 163
Hyperbolic tangent + linear 0.9581 0.9752 142 62
Linear 0.9649 0.9790 109 -126

The statistical performances of the model from Gundelsheim

3000
Gundelsheim to Rockenau are compared with the Rockenau HN-simulated

ANN model, which do not integrate the HN model results Linear function
Hyperbolic tangent+ linear function
from Gundelsheim (Table 5 and 7). The statistical 2500 Hyperbolic tangent function
Sigmoidal function
performance in terms of CE, R2, RMSE and difference in
peak flow of the Gundelsheim to Rockenau ANN models
Runoff (m3/s)

2000
using all activation functions are found to be superior.
The performance ANN block 3 between Rockenau and
Heidelberg also improved with the application of activation 1500

functions with higher limiting amplitude. There was also a

general trend of underestimation of peaks. The difference
1000
in peak flow prediction was also found to be lower using
activation functions with higher limiting amplitude. The
hyperbolic tangent + linear function (a = 0.7) gave the best 500
90 100 110 120 130 140 150 160 170 180
performance in terms of CE, and difference in peak flow. Time (hrs)

Partial results of the model showing peak discharges for

the test data sets in the period between 21.12.1993 and
Fig. 6. Comparison of model performance of different activation
25.12.1993 are shown in Figs. 6, 7 and 8. In Fig. 6, the functions at Gundelsheim (Block 1)

319
Rajesh Raj Shrestha, Stephan Theobald and Franz Nestmann

3000
Rockenau containing the best performing ANN blocks in the previous
HN Simulated
Observed
section was used. Simulations using the HN models were
Hyperbolic tangent function in 2 layers also made by multiplying the upstream flows by 1.5. The
Linear function
2500
Hyperbolic tangent + linear function outputs of the HN model and ANN simulated results were
Hyperbolic tangent function
Sigmoidal function compared with each other. Figures 8, 9 and 10 show the
2000
comparison of the results at Gundelsheim, Rockenau and
Runoff (m3/s)

Heidelberg respectively with 1993 flows multiplied by 1.5.

The comparison of results of the two models indicated a
1500
good match for the Gundelsheim station and an under- or
overestimation for the Rockenau and Heidelberg stations.
1000 The ANN outputs showed underestimation at Gundelsheim
(ANN block 1) and Heidelberg (ANN block 3), and

500
90 100 110 120 130 140 150 160 170 180
Time (hrs) Gundelsheim
4000
NN-Simulated
Fig. 7. Comparison of model performance of different activation HN-Simulated
functions at Rockenau (Block 2) 3500

3000

Heidelberg
Runoff (m3/s)

3000
Observed 2500
HN-Simulated
Linear function
Hyperbolic tangent + linear function 2000
2500 Hyperblic tangent function
Sigmoidal function
1500

2000
Runoff (m3/s)

1000

1500 500
90 100 110 120 130 140 150 160 170 180
Time (hrs)

1000 Fig. 9. Comparison of model performance at Gundelsheim using 1993

flows multiplied by 1.5 (Hyperbolic tangent + linear activation
function)
500
90 100 110 120 130 140 150 160 170 180
Time (hrs)
Rockenau
4000
Fig. 8. Comparison of model performance of different activation NN-Simulated
functions at Heidelberg (Block 3) HN-Simulated
3500

ANNs results are compared with the simulated flows from 3000

hydrodynamic numerical (HN) model. Figures 7 and 8 show

Runoff (m3/s)

2500
the comparison with both HN model results and measured
flow series. 2000
It is to be noted that the statistical performance of the
ANN model for block 1 is not consistent with blocks 2 and 1500

3 as the results of block 1 are compared with the HN model

simulations and blocks 2 and 3 with the measured data. 1000

500
90 100 110 120 130 140 150 160 170 180
ANN MODELS FOR EXTREME FLOWS Time (hrs)

To test the ability of ANNs to predict extreme events,

Fig. 10. Comparison of model performance at Rockenau using 1993
simulations were made with all the upstream inflows flows multiplied by 1.5 (Hyperbolic tangent activation function in
multiplied by a factor 1.5. The ANN simulation model two hidden layers)

320
 Simulation of flood flow in a river system using ar tificial neural networks

4000
Heidelberg from observations and numerical model results. This also
NN-Simulated provides guidance to the network training and enhances the
HN-Simulated
overall model performance. The assessment of results of
3500
the trained networks shows that a combination of hyperbolic
tangent and linear transfer functions at the first hidden layer
3000
generally produced the best performance. This function has
Runoff (m3/s)

higher limiting amplitude and also imparts non-linearity to

2500
the networks. Hence, the ANNs in a suitable configuration
can extend the forecasting capability to a certain extent
2000
beyond the range of calibrated data sets. It is, however,
important to exercise caution in using ANNs for extreme
1500
flood events.

1000
90 100 110 120 130 140
Time (hrs)
150 160 170 180
References
ASCE Task Committee on Application of Artificial Neural
Networks in Hydrology, 2000. Artificial neural networks in
Fig. 11. Comparison of model performance at Heidelberg using 1993
hydrology II: Hydrologic applications. J. Hydrolog. Eng.-ASCE,
flows multiplied by 1.5 (Hyperbolic tangent + linear activation
function)
5, 124–137.
Dawson, C.W. and Wilby, R.B., 1999. A comparison of artificial
neural networks for flow forecasting. Hydrol. Earth Syst. Sci.,
3, 529–540.
overestimation at Rockenau (ANN block 2). It is interesting Dawson, C.W., Harpham C., Wilby, R.B. and Chen, Y., 2002.
Evaluation of artificial neural network techniques for flow
to note here that ANN blocks 1 and 3 have a combination of
forecasting in River Yangtze, China. Hydrol. Earth Syst. Sci.,
linear and hyperbolic transfer functions at the first hidden 6, 619–626.
layers and ANN block 2 has hyperbolic transfer functions Demuth, H. and Beale M., 2000. Neural network toolbox user’s
at the first and second hidden layers. guide. The MathWorks Inc. Online documentation: http://
www.mathworks.com/access/helpdesk/help/toolbox/nnet/
The result highlighted a random behaviour of the ANNs Dolling, O.R. and Veras, E.A., 2002. Artificial neural networks
when used to predict extreme flow events. It might function for streamflow prediction. J. Hydraul. Res., 40, 547–554.
quite well in some cases and not so well in other cases. Haykin, S., 1994. Neural networks a comprehensive foundation
(1st Edition). Macmillan College Publishing Company, Inc.,
Hence caution needs to be exercised in the use of ANNs for New York.
forecasting extreme events. It is also important to specify Imrie, C. E., Durucan, S. and Korre, A., 2000. River flow prediction
the forecasting range of the trained ANNs. using artificial neural networks: generalisation beyond
calibration range. J. Hydrol., 233, 138–153.
Lekkas, D.F., Imrie, C.E. and Lees M.J., 2001. Improved nonlinear
transfer function and neural network methods for flow routing
Conclusions for real-time flood forecasting. J. Hydroinformatics, 3, 153–164.
This paper has presented an ANN-based approach for the Minns, A.W., 1996. Extended rainfall-runoff modelling using
artificial neural networks, A. Müller (Ed.), Proc.
simulation of flood flows in the Neckar River in Germany. Hydroinformatics ’96, Balkema, Rotterdam, the Netherlands.
The river reach was divided into three ‘sub-reaches’ for the 207–213.
ANN trainings. The effects of different activation functions Shamseldin, A.Y., Nasr, A.E. and O’Connor, K.M., 2002.
Comparison of different forms of the multi-layer feed-forward
at the first hidden layer of multilayer perceptron (MLP)
neural network method used for river flow forecasting. Hydrol.
neural networks were evaluated for predicting flows beyond Earth Syst. Sci., 6, 671–684.
a calibrated range. This evaluation was made in terms of Shrestha, R.R., 2003. Flood routing using artificial neural
test data sets with higher peaks, above the range of training networks. Proc. XXX IAHR Congress, JF Kennedy Student
Paper Competition, Thessaloniki, Greece.
data sets. Four different activation functions, the sigmoidal, Sivakumar, B., Jayawardena, A.W. and Fernando, T.M.K.G., 2002.
hyperbolic tangent, linear, and a combination of hyperbolic River flow forecasting: use of phase-space reconstruction and
tangent and linear functions were investigated in this study. artificial neural networks approaches. J. Hydrol., 265, 225–245.
Solas, J.D., Markus, M. and Tokar, A.S., 2000. Streamflow
The results of this study indicate that the ANNs provide forecasting based on artificial neural networks. In: Artificial
an efficient means of flood flow forecasting. Compared to neural networks in hydrology, R.S. Govindaraju and A.
a HN model, which requires a lot of cross-sectional data, Ramachandra Rao (Eds.), Kluwer Academic Publishers,
Dordrecht, The Netherlands. 23–51.
the ANN model can be quickly trained to forecast flows at
Thirumalaiah, K. and Deo, M.C., 1998. River stage forecasting
specific sections in the river reach. The division of river using artificial neural networks. J. Hydrolog. Eng.-ASCE, 3,
reach into ‘sub-reaches’ facilitates the integration of data 26–31.

321