Flow forecast by SWAT model and ANN in Pracana basin, Portugal

Abstract

This study provides a unique opportunity to analyze the issue of flow forecast based on the soil and water assessment tool (SWAT) and artificial neural network (ANN) models. In last two decades, the ANNs have been extensively applied to various water resources system problems. In this study, the ANNs were applied to the daily flow of the Pracana basin in Portugal. The comparison of ANN models and a process-based model SWAT was established based on their prediction accuracy. The ANN model was found to be more successful than the SWAT in relation to better forecast of peak flow. Nevertheless the SWAT model results revealed a better value of mean squared error. The results of this study, in general, showed that ANNs can be powerful tools in daily flow forecasts.

Full text

Flow forecast by SWAT model and ANN in Pracana basin, Portugal
Mehmet C. Demirel
a,1
, Anabela Venancio
b
, Ercan Kahya
c,*
a
Institute of Science and Technology, Istanbul Technical University, Maslak 34469, Istanbul, Turkey
b
HIDROTEC, Escola Superior de Tecnologia, Universidade do Algarve, Campus da Penha 8005-117 Faro, Portugal
c
Civil Engineering Department, Istanbul Technical University, Maslak 34469, Istanbul, Turkey
article info
Article history:
Received 12 May 2008
Received in revised form 31 July 2008
Accepted 4 August 2008
Available online 12 November 2008
Keywords:
Flow
Forecast
Artificial neural networks
SWAT
Portugal
abstract
This study provides a unique opportunity to analyze the issue of flow forecast based on the soil and water
assessment tool (SWAT) and artificial neural network (ANN) models. In last two decades, the ANNs have
been extensively applied to various water resources system problems. In this study, the ANNs were
applied to the daily flow of the Pracana basin in Portugal. The comparison of ANN models and a pro-
cess-based model SWAT was established based on their prediction accuracy. The ANN model was found
to be more successful than the SWAT in relation to better forecast of peak flow. Nevertheless the SWAT
model results revealed a better value of mean squared error. The results of this study, in general, showed
that ANNs can be powerful tools in daily flow forecasts.
Ó2008 Elsevier Ltd. All rights reserved.
1. Introduction
Streamflow, which is known an integrated process of atmo-
spheric and topographic processes, is of prime importance to water
resources planning [19]. In a wide spectrum of engineering appli-
cations, it is critical to have reliable long-term or short-term flow
forecasts. The lead time of day is often used for the flood warning
systems. However, the tools for forecast are not free of error and
usually expensive when they are set in a physical base. Stochastic
and conceptual models have been always common in use [20].Itis
possible to work on a sophisticated model considering both hydro-
logic and climatologic variables, such as precipitation, runoff, tem-
perature and evaporation; however, it is economically preferable
to use a model simulating flow variations on the basis of historical
observations. For this reason, the historical observations will be
used as input to artificial neural networks (ANNs) models to eval-
uate two different flow forecast models. Black-box models are not
physically based models as they tackle with a system in the input–
output manner. Unit hydrograph and autoregressive moving aver-
age models are the types of black-box models used in examining
the rainfall–runoff relation [32]. Burlando et al. [9] applied ARMA
models to hourly rainfall data for forecasting. They made compar-
ison for the point and grid data (average rainfall over the basin)
using autocovariance structure of certain low-order ARMA pro-
cesses. They concluded that the event-based estimation approach
yields better forecasts. The nonlinear approach of ANN can repre-
sent the rainfall–runoff precisely if the input variables are coherent
with output of the system [3]. The ANNs have been extensively
used in hydrology for simulating rainfall–runoff and other hydro-
logical processes [27,23,26,25].
Maier and Dandy [27] presented an extensive literature review
of the ANN model applications and outlined the steps that should
be followed in the development of such models. They examined a
total of 43 papers and 41 out of these papers included the use of
feed forward networks. Majority of these neural networks were
associated with the back propagation algorithm in training part.
Kisi [23] compared two different feed forward neural network
algorithms for the estimation of daily reference evapotranspiration
from available climatic data. He showed that the Levenberg–Mar-
quardt and conjugate gradient algorithms successfully employed
in modeling evapotranspiration process. Lee et al. [26] carried
out three tests to demonstrate the representative elementary wa-
tershed approach could be used for soil moisture predictions. Their
approach was more successful than a distributed model called
CATFLOW. Kisi [25] performed three different ANN techniques,
namely, feed forward neural networks, generalized regression neu-
ral networks and radial basis in monthly flow forecasting. The gen-
eralized regression neural networks were found to be more
successful in forecast of one-month advanced streamflow of the
two stations from the Eastern Black Sea region of Turkey. Zealand
0965-9978/$ - see front matter Ó2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.advengsoft.2008.08.002
*Corresponding author. Present address: Civil Engineering Department, Amer-
ican University of Sharjah, P.O. Box 26666, Sharjah, United Arab Emirates.
E-mail addresses: m.c.demirel@ctw.utwente.nl (M.C. Demirel), kahyae@itu.
edu.tr,ekahya@aus.edu (E. Kahya).
URL: http://atlas.cc.itu.edu.tr/~kahyae (E. Kahya).
1
Present address: Department of Water Engineering and Management, University
of Twente, P.O. Box 217, 7500 AE Enschede, The Netherlands.
Advances in Engineering Software 40 (2009) 467–473
Contents lists available at ScienceDirect
Advances in Engineering Software
journal homepage: www.elsevier.com/locate/advengsoft

et al. [41] carried out a short-term flow forecasting in a part of the
Winnipeg River system in northwest Ontario (Canada) having a
large catchment area of 20.000 km
2
. They then compared the
ANN methods to conventional approaches, indicating that the for-
mer methods resulted in more accurate forecasts. Moreover, a very
good fit to observed flow value were achieved with respect to the
root mean squared error (RMSE) in training and testing parts. Their
results also showed that using specific network architecture for
each forecast lead time was more appropriate in the multi-week
forecasting. Chen and Adams [12] integrated the ANN with semi-
distributed form of conceptual models. They achieved the runoff
generation and water budget among different runoff components
including surface runoff and groundwater by the spatially
distributed model parameters and rainfall inputs for the each sub-
catchment. Baratti et al. [8] used the ANN in the rainfall–runoff
modeling process when different time step durations have to be
considered in the reservoir management. They made numerical
comparisons with observed data that are provided for runoff pre-
diction in the Tirso basin at the S. Chiara section in Sardinia (Italy).
Calvoa and Portelab [10] applied the ANN to forecast daily flow in
the northern Portugal domain. They used Castanheiro and Cida-
delhe stations in Douro watershed to compare the ANN and ARIMA
models. Both models are defined as data-driven approaches which
are based on historical records. Toth et al. [38] compared simple
heuristic short term (1–6 h lead time) prediction results with the
ANN model for the purpose of real time flood forecast. An excellent
collection of studies concerning the ANN are collected in [6,7,16].
The SWAT model is a widely used process-based model that em-
beds most of the hydrologic processes by the principle of water
balance. It can also be used in ungauged catchments [4]. Kaur
et al. [21] used the SWAT model to predict runoff and sediment
loss from Nagwan basin in India. They also developed a decision
support tool to identify the priority areas for soil and water conser-
vation measures.
Few studies made accuracy comparisons between process-
based models and ANNs [37,30,36]. Sivakumar et al. [34] tested
two nonlinear black-box approaches (phase-space reconstruction
and artificial neural networks) for forecasting river flow dynamics.
They used multi-layer perceptron nets and the daily river
flow rates from the Nakhon Sawan station at the Chao Phraya
River basin in Thailand to achieve 1-day and 7-day in advance
forecasts.
Finally, two inspiring works (i.e. [36,30]) provided the perfor-
mance assessment of the SWAT and ANN models in simulating
hydrologic responses for different watersheds. Srivastava et al.
[36] reported that winter months and the models’ inaccurate
base flow simulation both affected the SWAT model performance
on the agricultural watershed located in the south eastern
Pennsylvania. However, the ANN models in their current form
are not spatially distributed watershed modeling systems. Morid
et al. [30] examined the ANN and SWAT models together and
found that the ANN models performed better than the snow
module of SWAT during low flow periods, namely summer, au-
tumn and winter. Conversely the SWAT revealed better results
during high flow period (spring), in particular for peak flow. In
a nutshell, the following question still deserves more scientific
attentions particularly in the regions under flood risk that the
flow changes abruptly: what is the best model to use in flow
forecast?
The main objective of this study is to compare the artificial neu-
ral network model with a process-based semi-distributed model
for the accuracy of flow forecasting scheme. This task is aimed to
be accomplished in the Pracana basin, Portugal. The accuracy of
ANN and SWAT models will be investigated in daily temporal res-
olution. The flow forecast models will be described in detail in the
following section.
2. Description of study area and data
Portugal is located between latitudes 37°N and 42°N and longi-
tudes of 9.5°W and 6.5°W. The country lies in the transition zone
stretching from the subtropical anticyclones (the Azor anticyclone)
to the area of sub poles depressions. The factors that most affect
the climatic conditions in the region are the latitude, the orography
and the influence of the Atlantic Ocean [39].
The Pracana is a sub-basin of Tejo region with an area of
1433 km
2
. It is located in the central east part of Portugal, close
to the border of Spain (Fig. 1).
The climate in this region is, in general, characterized by a tem-
perate climate that features warm and dry summers and winters
with precipitation. Microclimatology varies dramatically with re-
spect to elevation. The precipitation fluctuates from 900 to
1400 mm per year. The average annual temperature is between
9°C and 20 °C. The primary counties are Castelo Branco and Pro-
enca-A-Nova, occupying 80% of the basin. Castelo Branco hosts
most of the industrial facilities that creates a big load of nutrients.
Estimated nitrogen is 213 ton/year and phosphorus is 63 ton/year.
The contamination from agricultural sources is low as the perma-
nent pastures and trees constitute 57% of the utilized agricultural
area. Hence the region cannot be defined as vulnerable [39]. The
dam of Pracana is located between the county of Macau and Vila
Velha de Rodao. It was built for hydroelectric and water regulation
that feeds the Ocreza River. The hydrology of the basin is the main
driving force for the transport of nutrients, sediment, or other
properties, however, the retention of nutrients is not the objective
of this study. The main information required by the SWAT is DEM
(digital elevation model), land use, soil types, and hydrological and
meteorological data. The data used to create the digital terrain
model of the Pracana were obtained from SRTM (Shuttle Radar
Topography Mission, www2.jpl.nasa.gov). The DEM determines
the direction of flow as well as the physical characteristics of the
basin. The hydrographic network can be determined automatically
from the digital terrain model or can be provided via a map. The
land use and soil types are important data which will significantly
influence the water balance. The soil texture is a basic property of
soil physics. The textures were obtained from the soil map devel-
oped by The Commission of the European Communities, Director-
ate General for Agriculture, Coordination of Agricultural Research
in 1985. According to this map; 6.1% of the area is fine texture,
1.2% is coarse texture and 93.7% is medium texture. For other soil
characteristics and vegetation information, readers are referred to
Venancio and Chambel [39]. The SWAT model simulations were
performed over a period of 52-year, 1953–2001. But the compari-
Fig. 1. The location of Pracana basin.
468 M.C. Demirel et al. / Advances in Engineering Software 40 (2009) 467–473

son of the ANN–SWAT was done for the interval 1953–1965 years
in which there is a continuous daily flow and precipitation data
span. The updated data is available at Sistema Nacional de Infor-
macao de Recursos Hidricos (SNIRH). Table 1 provides descriptive
information of the gauging station used in this study.
3. Methodology
Increasing interests and appreciation of the practical manner
and potential efficiency associated with data-driven technologies
(i.e., ANN) and their flow forecasting applications stimulated us
to investigate their effectiveness as a prediction tool using data
in the Pracana basin. In this respect, the principal direction for this
hydrological neural network study is to benchmark neural net-
works and a process-based model (SWAT) for flow forecast by con-
sidering two general cases: single input/single output and multiple
input/single outputs. The SWAT model simulations for the Pracana
basin were detailed in [39] and its outputs were included into the
contents of this comparative study.
3.1. Artificial neural networks
The neural networks is a powerful soft computational tech-
nique for linear and nonlinear approximations in many disci-
plines, inspired by biological cerebral activity called
neuroscience [17,28,11,22,23,1]. The idea of ANN is based on
the estimation of an output by a function of the input as in the
process of biological neuron cell in the brain. This cell network
has the ability in which it can be trained and learned by previous
examples (experience) so that it can recognize the patterns, such
as sounds and faces. The estimation of the model parameters is
called training in ANN terminology. The neural network has three
main layers; input, hidden and output. Each layer may have mul-
tiple units interconnected completely with the adjacent layer and
an adjusted weight is attached to each link in the system (Fig. 2).
The nodes (dendrites) in input layer receive the data then the
nodes in hidden layer (cell body) process and send them to output
layer (axon). The human brain has more than a billion neurons
with many working interconnections; hence, it is able to learn
and then distinguish different human voices. It can also differen-
tiate background noises such as car traffic, ocean wave and
mechanical noise of refrigerators which are very difficult task
for most of the supercomputers today [33]. Wu et al. [40] pro-
posed a practical solution that the optimal number of units in
the hidden layer could be estimated as two thirds of the sum of
the number of input and output neurons. While fewer neurons
could be insufficient to capture intricate relations between pre-
dictors and calculated output, the larger number of hidden nodes
may perform better, but the training time must be increased and
probably the accuracy will then be deficiently affected or the
problem ‘‘over fitted network” will occur [33,40]. Over fitting is
simply the inconsistent behavior of the network. While the net-
work memorizes the data in training part, it fails to work with
new input data in validation part.
The number of nodes in the first layer designated by Min Eq. (1)
equals to the number of input parameters.When the input values
pass to the next layer, they are multiplied by the weight of the con-
nection. Each node jreceives incoming signals from every node iin
the previous layer. Incoming signal (x
i
) is associated with a weight
(w
ji
). The effective signal (E
j
) to node jis the weighted sum of all
incoming signals (Eq. (1)). In the first phase of training, the weights
(i.e., w
ji
) are set to random values:
E
j
¼X
M
i¼1
x
i
w
ji
:ð1Þ
At the next phase, the effective signal (E
j
) passes through a transfer
function (i.e., Eq. (2)) to produce the outgoing signal (y
j
) of node j.
There are transfer functions different than sigmoidal-types (logistic
and hyperbolic tangent function): hard limit transfer function
(bounded to 0 or 1), linear, polynomial, rational function (ratios of
polynomials) and Fourier series (sums of cosines). In the literature,
the most commonly used transfer functions are sigmoidal-type
transfer functions in the hidden layers and linear transfer functions
(y
j
=E
j
) in the output layer due to its advantage in extrapolation be-
yond the range of the training data [17,27,41,10]:
y
j
¼fðE
j
Þ¼ 1
1þexpðE
j
Þ:ð2Þ
The connection weights manifest the importance of input to the
overall estimation process. The fitting error (Eq. (3)) between the
desired and estimated output is used as feedback to enhance the
performance of the network by altering the connection weights:
Error ¼X
N
j¼1
ðy
j
d
j
Þ
2
;ð3Þ
where N= number of output nodes, y
j
= calculated output, and
d
j
= desired data value. This process is repeated until establishing
a successive layer [33]. Therefore, these kinds of networks are called
feed forward back propagation (FF-BP) networks, which are the most
popular supervised algorithm for training networks in prediction,
pattern recognition, and nonlinear function fitting [13,41,40,25,2].
When using a FF-BP network, the sigmoid activation function is of-
ten preferred [34,24,17]. Training (calibrating) is a crucial process,
in which the network is tested by a set of data pairs (input–output)
and changing the initial conditions in each iteration step to achieve
an accurate forecasting. Minimization is performed by calculating
the gradient for each node at the output layer
d
k
¼d
r
k
ðy
k
d
k
Þ;ð4Þ
d
r
k
= the derivative of the sigmoid function applied at y
k
which is
defined for each kth output node. For hidden layer (one layer back),
the gradient function becomes
d
j
¼d
r
j
X
N
i¼1
d
i
w
jk
;ð5Þ
where d
r
j
is the derivative of the sigmoid function and w
jk
= weight
value from hidden node jto output node k. When the input data are
chosen, then the network runs; the weights for each connection are
updated by the procedure in Eq. (6) until the error is minimized to a
predefined error target or the desired number of training periods is
reached:
D
w
jk
¼w
jk

g
d
k
y
j
;ð6Þ
where the notation
g
is the learning rate of each layer back to the
network. Each passes through the training data is called epoch.In
the Matlab routines, the user can define the number of epochs prior
to analysis and manually adjusts until the plausible performance is
achieved in the trial and error period [13]. In this study, we used
Table 1
Geographic information of the gauging station used in this study.
Parameter Gauge identification
PRACANA
Latitude (°N) 39.568
Longitude (°W) 7.816
Altitude (m) 140
River Rio Ocreza
Precipitation station Castelo Branco
Streamflow station Almourao
M.C. Demirel et al. / Advances in Engineering Software 40 (2009) 467–473 469

MSE (Eq. (7)), RMSE for the highly extreme flow and run time for
the performance assessment. Only the run time depends on com-
puter resources:
MSE ¼P
N
i¼1
ðQ
obs:
Q
est:
Þ
2
N;ð7Þ
where Q
obs.
is the observed flow, Q
est.
is the estimated flow, and Nis
the total number of observations of the validation set. In ANN mod-
eling, unlike the SWAT model, a prior knowledge of the underlying
physical processes concerned is not required. Moreover, there is no
need to satisfy preliminary conditions (i.e., normal distribution) as
required in typical statistical models and optimization models.
However, there are few disadvantages of ANNs, such as an exponen-
tial increase in training time with increased data size owing to the
complex relationships used by the network to produce output [33].
3.2. A process-based model: SWAT
Process-based models attempt to formulate the entire physical
process from precipitation to flow in the hydrologic cycle by bal-
ancing the amount of water. These models can provide accurate
estimation of flow on daily, monthly and seasonal time scales.
However, these models require a large number of parameters.
For instance, the Water Balance (Watbal) model has 50 adjustable
parameters as the simple conceptual rainfall–runoff (SCRR) model
consists of seven fitting coefficients [37]. Hence, the success of the
model prediction is dependent on the user’s knowledge about the
region and ability to manipulate the model components. The input
data are usually temperature, humidity, soil moisture, soil texture,
precipitation, evapotranspiration, lateral flow and percolation rate.
The well-calibrated conceptual model presents a reasonable accu-
racy in forecasting flow. However, the number of inputs that are
required to run the model and the difficulties which were dis-
cussed above are limiting the use of the physical based models to
a very small number of river basins [37]. One of the most popular
process-based models is SWAT. It is a mathematical model devel-
oped for the US Department of Agriculture, Agricultural Research
Service. The SWAT is used to analyze the impacts of land use
changes on the runoff and groundwater, production of sediment
and water quality; for example, flow in the tributaries or agricul-
tural issues (e.g., nutrient/pesticide loads) Eq. (8) [37]. The model
simulates the water balance in a watershed and can be formulated
as
SW
final
¼SW
init:
þX
t
i¼1
ðP
p
ðiÞQ
s
ðiÞE
e
ðiÞP
per:
ðiÞQ
r
ðiÞÞ:ð8Þ
In this equation, SW
final
= the final soil water content (mm),
SW
init.
= the soil water content available for plant uptake (initial
water content permanent wilting point water content), t= the
time in days, P
p
(i) = the amount of precipitation on day i(mm),
Q
s
(i) = the amount of surface runoff (mm), E
e
(i) = the amount of
evapotranspiration (mm), P
per
(i) = the amount of percolation
(mm), and Q
r
(i) = the amount of return flow (mm). The SWAT
model uses two phases of hydrologic cycle; one for the land pro-
cesses and the other for the channel processes. The following
phases of hydrologic cycle should be recalled in its realization. Pre-
cipitation may be intercepted and kept in the vegetation canopy or
fall over the soil surface where it will infiltrate into the soil profile
or flow overland as runoff. Runoff arrives relatively quickly to a
stream channel and creates a short-term flow response. Infiltrated
water may be kept in the soil and then evapotranspired to the
atmosphere or it may slowly make its way to the stream water
system via underground paths. The SWAT model has been exten-
sively used and tested since 1993 by mainly hydrologists for soft
engineering related issues [4,35,18,29,14,15]. The digital elevation
model (DEM), a crucial tool for delineating the sub-watersheds
using Arcview GIS software, is integrated to the SWAT model.
The last versions of SWAT (e.g., SWAT 2000) enable users to input
solar radiation, wind speed, relative humidity and evaporation
data from more than one gauge station into the model simulation
structure. Readers are referred to Tokar and Markus [37] and Ar-
nold and Fohrer [5] for further information. Venancio and Chambel
[39] presented adaptation of the parameters and all required pro-
cedures needed by the SWAT model in the Pracana basin as a case
study.
Fig. 2. Conceptual diagram of three-layer neural network model.
470 M.C. Demirel et al. / Advances in Engineering Software 40 (2009) 467–473

4. Results and discussion
Following to Rumelhart et al. [31] Govindaraju and Ramachan-
dra Rao [16] and ASCE [7], we selected the FF-BP ANN model to
employ to daily flow records in the Pracana basin. Three layers
including input, hidden and output nodes were selected as a basis
for our network. The observed historical data (i.e., daily flow
values) is introduced to the model as an input layer. We used
one hidden layer and carefully tested many hidden node options,
but showed here only the results of 6, 22 and 28 node options in
Table 2. The number of units in the output layer is the number of
values to be estimated. In our flow forecast network, we have only
one output node for the flow which will be predicted (see output
column in Table 2). We used a network training function that up-
dates weight and bias values according to gradient descent with
adaptive learning rate. Moreover, the weights were adjusted by
nonlinear sigmoid function (Eq. (2)) for the hidden layer besides
linear function utilized for the output layer. We tried various sce-
narios for the model parameters such as transfer function and
the number of epochs to achieve a better performance; thus, sev-
eral numerical experiments were conducted to the model architec-
ture. Selected model structures and results were given in Tables 2
and 3.
In the first stage the following model was investigated for the
flow simulation. Model 1: S(t)=f(P(t1),S(t1)), where P
t1
and S
t1
are the precipitation and flow recorded in the previous
day, respectively. It should be noted that model 1 is labeled as
Exp-I in Table 2 and evaluated by different criteria in Table 3.
Unfortunately, this model (shown in Fig. 3) failed to generate
acceptable estimations; hence, we decided to incorporate precipi-
tation and flow as an individual input into the model formation.
The modifications that we adapted for a better performance are
as follows:
(i) The data were normalized within a range of 0.1–0.9 as we
used log sigmoid transfer function (Logsig) which only takes
on a value in the interval 0 to +1.
(ii) Since the number of neurons in the hidden layer plays an
important role in the model performance, we tested 6–28
neurons. We noticed that there was no noticeable differ-
ence; hence, we preferred 6 neurons.
(iii) Epochs (iteration) size was adjusted to 100 as a result of the
trail and errors application to the different higher
magnitudes.
Although sometimes arbitrary changes or over fitting may occur
in the large epoch sizes the performance function becomes more
stable after 100 epochs. For example, in the case of the Exp-IV
model, when the epoch size is increased to 300, it was spoiled
and over fitted. An overall evaluation of the applied experiments
leads us to stress that the Exp-IV is the superior structure in terms
of estimating high flow values (Table 3).
There are various criteria to be used in the comparison business
like MSE. In addition we brought forward practical and specific cri-
teria such as ‘‘adequate/poor peak magnitude estimation” and ‘‘run
time”. As concerning a success criteria we set one percentile as
threshold indicating extremely high flow occurrences. In our calcu-
Table 2
ANN model architecture and test scheme.
ID Model Input Output Training Test Network structure Epochs
Exp-I ANN-dS P
p
t1
;S
t1
S
t
410
3
748 2-28-1 1 10
3
Exp-II ANN-dS P
p
t
1S
t
410
3
748 1-22-1 1 10
3
Exp-III ANN-dS P
p
t1
S
t
310
3
1748 1-22-1 4 10
3
Exp-IV ANN-dS S
t1
S
t
410
3
748 1-6-1 100
Exp-V ANN-dS S
t1
;S
t2
S
t
410
3
746 2-6-1 100
Exp-VI ANN-dS S
t1
;S
t2
;S
t3
S
t
410
3
744 3-6-1 300
Exp-VII ANN-dS S
t1
;S
t2
;S
t3
,S
t4
S
t
410
3
742 4-6-1 300
Notations: Exp = experiment, ANN = artificial neural network; S = streamflow; P = precipitation, d = daily mean and p = Pracana station.
Table 3
Model performances.
ID MSE Run time (s) 1% Peak magnitude RMSE Overall evaluation
Exp-I 2576.4 58.187 669, 2952 NP, PPME
Exp-II 2978.3 45.641 705, 3706 NP, PPME
Exp-III 2997.4 133 983, 4421 NP, PPME
Exp-IV 2611.6 5.141 650, 5933 NP, APME
Exp-V 2837.4 4.609 491, 2456 NP, APME
Exp-VI 2861.7 7.203 556, 3769 NP, APME
Exp-VII 2728.5 7.25 515, 2181 NP, APME
SWAT model 2098.30 814.83 711, 5250 NP, PPME
Notations: Exp = experiment, MSE = mean squared error, NP = not precise, PPME = poor peak magnitude estimation and APME = adequate peak magnitude estimation.
Fig. 3. Experiment I: One day in advance flow forecast for Pracana basin.
M.C. Demirel et al. / Advances in Engineering Software 40 (2009) 467–473 471

lations this value correspondence to a specific value of 238 m
3
/s.
According to this threshold value we computed a root mean square
error (RMSE) value in the test part for each model and showed the
results in Table 3. Once again derivatives of Exp-IV model (i.e., Exp-
V, Exp-VI and Exp-VII) appear to be superior to the other models.
We computed significant autocorrelation coefficients in our
flow series from lag-1 to lag-4. Significant correlations were equal
and higher than 0.6 and taken into consideration in the ANN mod-
els. This comparative analysis shows us that it is possible to con-
struct a successful simple and faster model structure based on
the ANN to forecast daily flow in the Pracana basin (Fig. 4).
The SWAT model results were used to make a comparison with
those of the ANN model. The SWAT model appears to have the best
performance against the other ANN models with respect to having
the smallest MSE value (Table 3). On the other hand, Fig. 5 displays
a striking feature that the SWAT model did not perform as good as
the ANN (Exp-IV and it’s derivatives) model in estimating peak
flow values.
5. Conclusion
One of main focuses in this study was to develop the ANN mod-
els for flow forecasting and determine a more accurate architecture
(i.e., number of hidden layers) in the design phase. Comparisons
were made between the ANN model and one of the conventional
forecasting approaches (e.g., SWAT). In order to find the best
ANN architecture for extremely high flows, we tested seven major
alternatives (Table 2) and decided to use the Exp-IV and its deriv-
atives for the further steps.
We found that the process-based model SWAT simulations in
the Pracana basin were not good enough in forecasting peak flow
values. The peak flow inefficiency could be caused by the formula-
tion used in the model. Data preprocessing might be necessary to
get a better performance in SWAT as normalization of the data
yielded a finest accuracy in the ANN for capturing peak magni-
tudes. The lag time of 1 day may not allow simulating phenomena
of high-frequency occurrences. The deficiency of not capturing
peak values becomes an important issue particularly in the studies
of extreme hydrologic events (e.g., floods). However, according to
the criteria of MSE, the SWAT model and first ANN model (i.e.,
Exp-I) which included precipitation and flow into the process, pro-
duced more accurate results than those of the used ANN model.
One of known advantages of the SWAT model is to make reliable
flow simulation when there are available climate and soil data at
ungauged site [30]. The outcomes of this study were in a good
agreement and relation with earlier studies conducted for, in gen-
eral, the SWAT and ANN comparison in daily simulations, specifi-
cally peak flow prediction performance [36]. In Portugal, it is
relatively easier to obtain flow and precipitation records through
the governmental online resources compared to physical charac-
teristics of river basins such as soil moisture, infiltration, soil clas-
ses, groundwater level and evaporation. Hence the black-box
models might emerge as a faster tool to implement on flow fore-
casting business.
Fig. 4. Experiment IV: Lag-1 ANN model training part.
Fig. 5. ANN–SWAT comparison based on Exp-IV forecast.
472 M.C. Demirel et al. / Advances in Engineering Software 40 (2009) 467–473

Acknowledgements
The authors gratefully acknowledge the support of DPT Project
No.: 90187 for this work. Moreover, we thank to Prof. Arthur J.
Mariano (University of Miami) and Mr. Muhittin Tarhan (Istanbul
Technical University) for fruitful comments on the manuscript.
References
[1] Altunkaynak A. Forecasting surface water level fluctuations of lake van by
artificial neural networks. Water Resour Manage 2007;21(2):399–408.
[2] Altunkaynak A, Ozger M, Sen Z. Regional streamflow estimation by standard
regional dependence function approach. J Hydraul Eng – ASCE
2005;131(11):1001–6.
[3] Anctil F, Rat A. Evaluation of neural network streamflow forecasting on 47
watersheds. J Hydrol Eng 2005;10(1):85–8.
[4] Arnold J, Williams J, Srinivasan R, King K. SWAT – soil and water assessment
tool – documentation and users manual. USDA-ARS, Temple, Texas; 1996.
[5] Arnold JG, Fohrer N. SWAT2000: current capabilities and research
opportunities in applied watershed modelling. Hydrol Process
2005;19(3):563–72.
[6] ASCE. Task committee on application of ANNs in hydrology, artificial neural
networks in hydrology. I: Preliminary concepts. J Hydrol Eng
2000;5(2):115–23.
[7] ASCE. Task committee on application of ANNs in hydrology, artificial neural
networks in hydrology. II: Hydrology application. J Hydrol Eng
2000;5(2):124–37.
[8] Baratti R, Cannas B, Fanni A, Pintus M, Sechi G, Toreno N. River flow forecast for
reservoir management through neural networks. Neurocomputing
2003;55(3):421–37.
[9] Burlando P, Rosso R, Cadavid LG, Salas JD. Forecasting of short-term rainfall
using ARMA models. J Hydrol 1993;144(1–4):193–211.
[10] Calvoa I, Portelab M. Application of neural approaches to one-step daily flow
forecasting in Portuguese watersheds. J Hydrol 2007;332(1–2):1–15.
[11] Can I. A new improved na/k geothermometer by artificial neural networks.
Geothermics 2002;31(6):751–60.
[12] Chen J, Adams B. Integration of artificial neural networks with conceptual
models in rainfall–runoff modeling. J Hydrol 2006;318(1–4):232–49.
[13] Demuth H, Beale M. Neural network toolbox for use with MATLAB user’s
guide. Natick (MA): The MathWorks Inc.; 2001.
[14] Di Luzio M, Arnold J, Srinivasan R. Effect of GIS data quality on small watershed
stream flow and sediment simulations. Hydrol Process 2005;19(3):629–50.
[15] Govender M, Everson C. Modelling streamflow from two small South African
experimental catchments using the SWAT model. Hydrol Process
2005;19(3):683–92.
[16] Govindaraju RS, Ramachandra Rao A. Artificial neural networks in
hydrology. Dordrecht, Netherlands: Kluwer; 2000.
[17] Hsu K-L, Gupta H, Sorooshian S. Artificial neural network modeling of the
rainfall runoff process. Water Resour Res 1995;31(10):2517–30.
[18] Jha M, Pan Z, Takle E, Gu R. Impacts of climate change on streamflow in the
upper Mississippi river basin: a regional climate model perspective. J Geophys
Res 2004.
[19] Kahya E, Dracup JA. US streamflow patterns in relation to the El Nino/southern
oscillation. Water Resour Res 1993;28(8):2491–503.
[20] Karabork C, Kahya E. Multivariate stochastic modeling of streamflows in the
Sakarya basin. Turkish J Eng Environ Sci 1999;23(2):133–47 [in Turkish].
[21] Kaur R, Srinivasan R, Mishra K, Dutta D, Prasad D, Bansal G. Assessment of
SWAT model for soil and water management in India. Land use. Water Resour
Res 2003;3:1–7.
[22] Kisi O. Multi-layer perceptrons with Levenberg–Marquardt training algorithm
for suspended sediment concentration prediction and estimation. Hydrol Sci J
2004;49(6):1025–40.
[23] Kisi O. Evapotranspiration estimation using feed-forward neural networks.
Nordic Hydrol 2006;37(3):247–60.
[24] Kisi O. Streamflow forecasting using different artificial neural network
algorithms. J Hydrol Eng 2007;12(5):532–9.
[25] Kisi O. River flow forecasting and estimation using different artificial neural
network techniques. Hydrol Res 2008;39(1):27–40.
[26] Lee H, Zehe E, Sivapalan M. Predictions of rainfall–runoff response and soil
moisture dynamics in a micro-scale catchment using the crew model. Hydrol
Earth Syst Sci 2007;11:819–49.
[27] Maier HR, Dandy GC. Neural networks for the prediction and forecasting of
water resources variables: a review of modelling issues and applications.
Environ Modell Software 2000;15:101–24.
[28] Markus M. Application of neural networks in streamflow forecasting. Ph.D.
thesis, Colorado State University; 1997.
[29] Moon J, Srinivasan R, Jacobs J. Stream flow estimation using spatially
distributed rainfall in the trinity river basin, Texas. Trans ASAE
2004;47(5):1445–51.
[30] Morid S, Gosain AK, Keshari AK. Comparison of the SWAT model and ANN for
daily simulation of runoff in snowbound ungauged catchments. In: Fifth
international conference on hydroinformatics, Cardiff, UK; 2002.
[31] Rumelhart DE, Hinton GE, Williams RJ. Learning representations by back-
propagating errors. Nature 1986;323:533–6.
[32] Salas JD, Markus M, Tokar AS. Streamflow forecasting based on artificial neural
networks. In: Govindaraju R, Ramachandra Rao A, editors. Artificial neural
networks in hydrology. London: Kluwer Publishers; 2000. p. 23–51.
[33] Silverman D, Dracup JA. Artificial neural networks and long-lead precipitation
prediction in California. J Appl Meteorol 2000;31(1):57–66.
[34] Sivakumar B, Jayawardena AW, Fernando TMKG. River flow forecasting: use of
phase-space reconstruction and artificial neural networks approaches. J
Hydrol 2002;265(1–4):225245.
[35] Srinivasan R, Arnold J. Integration of a basin-scale water quality model with
GIS. Water Resour Bull 1994;30(3):453–62.
[36] Srivastava P, McNair JN, Johnson TE. Comparison of process-based and
artificial neural network approaches for streamflow modelling in an
agricultural watershed. J Am Water Resour Assoc 2006;42(3):545–63.
[37] Tokar AS, Markus M. Precipitation–runoff modelling using artificial neural
networks and conceptual models. J Hydrol Eng 2000;4(3):232–9.
[38] Toth E, Brath A, Montanari A. Comparison of short-term rainfall prediction
models for real-time flood forecasting. J Hydrol 2000;239(1–4):132–47.
[39] Venancio A, Martins F, Chambel P, Neves R. Modelacao hidrologica da bacia
drenante da albufeira de pracana Faro: V Congresso Iberico; 4–8 December,
2006.
[40] Wu JS, Han J, Annambhotla S, Bryant S. Artificial neural networks for
forecasting watershed runoff and stream flows. J Hydrol Eng
2005;10(3):216–22.
[41] Zealand CM, Burn DH, Simonovic SP. Short term streamflow forecasting using
artificial neural networks. J Hydrol 1999;214(1–4):3248.
M.C. Demirel et al. / Advances in Engineering Software 40 (2009) 467–473 473