Water Resour Manage (2008) 22:67–82

DOI 10.1007/s11269-006-9144-x

Deterministic Insight into ANN Model Performance

for Storm Runoff Simulation

Kwan Tun Lee & Wei-Chiao Hung & Chung-Chieh Meng

Received: 28 December 2005 / Accepted: 18 December 2006 /

Published online: 24 January 2007
# Springer Science + Business Media B.V. 2007

Abstract The artificial neural network (ANN) theory has been widely applied to practical
applications in hydrology. Since watershed rainfall–runoff processes are nonlinear and
exhibit spatial and temporal variability, the ANN model, which considers watershed
nonlinear characteristics, can usually but not always obtain satisfactory simulation results.
The training of an ANN network is based completely on the reliability of the available
hydrologic records. The objective of this study was to provide deterministic insight into the
limitations of storm runoff simulation when using ANN. Hydrologic records of 42 storm
events from two watersheds in Taiwan were adopted for analysis. A deterministic runoff
model was used to classify the hydrologic records into “usual” and “unusual” storm events.
The analytical results show that the ANN model could provide good simulation results for
“usual” storm events; however, its performance was poor when it was applied to “unusual”
storm events because no consistent hydrologic characteristics could be extracted from the
storm event records using ANN. The success of the ANN model in usual storm discharge
simulations may be mainly due to the input vectors including the previous observed
discharge. Moreover, the number of past periods of rainfall that were set as the input vectors
of the ANN model was found to be highly correlated with the watershed time of
concentration. It can be used to efficiently determine the ANN network structure instead of
using iterative network training.

Key words ANN model . deterministic insight . training data set . rainfall–runoff simulation

1 Introduction

The rainfall–runoff process in a watershed is dominated by the watershed geomorphic and

rainfall characteristics. The process is indeed a highly nonlinear, time and spatial variation

K. T. Lee (*) : W.-C. Hung : C.-C. Meng

Department of River and Harbor Engineering, National Taiwan Ocean University,
Keelung 202, Taiwan, Republic of China
e-mail: [email protected]
68 K.T. Lee, et al.

hydrologic system. For practical application purposes, such a complex hydrologic system
has usually been simplified as a linear, time-invariant, and spatially averaged system (for
examples, see Sherman 1932; Nash 1957; Dooge 1959). Although the linear system model
has been widely applied to short-term rainfall–runoff simulation or long-term flow
prediction with acceptable results, the watershed inherent nonlinear hydrologic character-
istics do induce large simulation errors in some special cases (Minshall 1960; Amorocho
1967; Chiu and Huang 1970; Ding 1974). By taking advantage of high performance
computational techniques, hydraulic routing or simplified kinematic-wave approximation
can handle such nonlinearity, but these approaches are time-consume and laborious while
collecting information about the watershed geomorphology, land cover, and channel
geometry.
The human neuron structure is one of the most complex systems known. With the aid of
high performance computers, the artificial neural network theory has been widely studied
and applied during the past 20 years (Hopfield 1982; Rumelhart et al. 1986; Caudill 1987;
French et al. 1992; Hsu et al. 1995; Tokar and Johnson 1999; ASCE Task committee 2000a,b;
Tayfur 2002; Rao et al. 2003; Muleta and Nicklow 2004; Kumar et al. 2004). An artificial
neural network system is a massively parallel-distributed information processing system that
has certain performance characteristics resembling the biological neural networks of the
human brain (Haykin 1994). The characteristics of a neural network depend on the pattern of
connections between layers. The general structure of a neural network includes an input
layer, several hidden layers, and an output layer. Each layer contains several neurons, and the
inputs are weighted and transferred through a transfer function to generate outputs. The ANN
has been proven to provide better solutions when applied to: (1) complex systems that are
poorly described or understood; (2) problems which deal with noise or involve pattern
recognition; diagnosis, abstraction, and generalization; and (3) situations where the input is
incomplete or ambiguous by nature (Tokar and Johnson 1999).
Since no generally accepted rule can be used to determine an optimal ANN structure, an
ANN system is usually designed through trial and error processes. A training process,
which includes iterative adjustment and optimization of the weight matrices as well as bias
vectors, is applied to find the best ANN structure to minimize the errors between observed
records and predicted outputs. Back-propagation, conjugate gradient algorithms, radial
basis functions, cascade correlation algorithms, recurrent ANN, and self-organizing feature
maps are usually adopted to determine the ANN structure. Significant growth in interest in
this computational mechanism has occurred since Rumelhart et al. (1986) developed a
mathematically rigorous theoretical framework for neural networks. Consequently, ANNs
have been extensively applied to hydrology engineering. As reviewed by the ASCE Task
Committee (2000b), ANNs have been applied to the rainfall–runoff process, streamflow
prediction, water quality, ground water, precipitation, and other water resources problems.
A good physical understanding of the hydrologic process being modeled can help one
select the input vector and design a more efficient network. Previous researches have tried
to compare simulation results obtained using ANN, empirical, and physically based
approaches (Dawson and Wilby 1998; Zealand et al. 1999; Anmala et al. 2000;
Thirumalaiah and Deo 2000; Tayfur 2002; Jain and Indurthy 2003), but no detailed
investigation has been reported to show in which cases the ANN can provide good
simulation results, and which cases the physically based approach can provide better
simulation results than ANN can. In this study, we sought to gain deterministic insight into
the limitations of ANN modeling for rainfall–runoff simulation. A validated deterministic
runoff model was adopted to identify the hydrologic characteristics of selected storm events
Deterministic insight into ANN model performance for storm runoff simulation 69

and then linked to the simulation results obtained using the ANN models. Moreover, the
relationship between the number of past periods of rainfall input vectors of the ANN model
and the watershed time of concentration was studied. We hope the deterministic insight into
the watershed rainfall–runoff process provided here can aid researchers in selecting
adequate ANN model structures.

2 Artificial Neural Networks

The ANN is an information processing system for emulating neuron information transfer in
biological systems. In a feed forward network, there are an input layer, several hidden
layers, and an output layer. The information is passed from the input layer through the
hidden layers to the output layer. In this study, we focused on a feed forward and back
propagation ANN system. To avoid redundancy, only a brief description of the ANN theory
will be given here.

2.1 Structure of a Feed Forward System

The ANN is composed of multi-unit processors or artificial neurons. The outputs of the
system are functions of the inputs, the weighted coefficients, and the threshold values. The
sigmoid function, which is a continuous and differentiable function, has usually been
adopted by previous researchers to simulate the nonlinearity of the system. Consequently,
the outputs of the system can be expressed as:

X  1
Yj ¼ f Wij Xi
q j ¼ P ð1Þ

ð Wij Xi
q j Þ
1þe

in which Yj is the output from the jth neuron in a specified layer; Xi is the input of the ith
neuron in the previous layer; Wij is the weight of the connection joining the jth neuron
in a layer with the ith neuron in the previous layer; and θj is the threshold value, also
called the bias, of the jth neuron in a specified layer. The output of the sigmoid function
is bounded between 0 and 1, and the input to the function can be negative or positive
infinite.
A three-layer ANN system was examined in this study. The system contains an input
layer, a hidden layer, and an output layer. The nodes in the same layer are independent, but
they are linked to those in the subsequent layer. The input layer receives information from
outside of the system, and then this information is transferred by using Eq. 1 to the hidden
layer. Data processing is also applied to transfer the information from the hidden layer to
the output layer. It is, therefore, called a feed forward process because the information is
only transmitted from the previous layer to the next layer, with no recurrent process in this
data transfer process.

2.2 Back-Propagation Algorithm

When constructing a feed forward ANN, one needs to define the weight coefficients and
threshold values in the system. Hydrologic data is, therefore, required to calibrate the
coefficients; this process is called network training. The back-propagation algorithm is
widely used for network training. A recurrent process based on the gradient steepest descent
70 K.T. Lee, et al.

method is used to transfer the estimation error back to the original system to modify the
coefficients in order to arrive at the minimum estimation error. The error function used in
this study is:

1X 2
E¼ ðTt
Yt Þ ð2Þ
2 t

where E is the estimated error; Tt is the desired output at time t; and Yt is the corresponding
ANN output at time t. Since the back-propagation algorithm is a first-order approximation
based on the steepest gradient descent, sometimes, the solution is trapped in a local
minimum, which should be avoided during network training.

3 ANN Model Development

Selecting of training data that represent the characteristics of a watershed and

meteorological patterns is extremely important in ANN modeling (Yapo et al. 1996).
Although insufficient training data cannot enable the network to generalize the patterns of
the physical phenomena, a model trained using high noise data may result in poor
forecasting. Noise in the hydrologic records may arise from errors in the measurement of
rainfall and discharge quantity, or unusual hydrologic phenomena that should not be used to
represent the general watershed rainfall–runoff relationships. As reported by the ASCE
Task Committee (2000a), routine procedures such as plotting and examining the statistics
are sometimes effective judging the reliability of the data and possibly removing outliers.
However, it is difficult to use a plotting method to examine the adequateness of the hourly
rainfall and discharge records because temporal fluctuations of rainfall intensity and
discharge are comparatively significant. In this study, a deterministic runoff model was used
to identify representative storm event records for the watersheds. Hydrologic records from
general storm events were identified as “usual storm events,” and those corresponding to
exceptional conditions, for example, very dry or very wet antecedent conditions, or
significant spatial variation of rainfall, were identified as “unusual storm events” for
developing different kinds of ANN models.
We focused on a feed forward and back propagation ANN system. A three-layer ANN
system, which contains an input layer, a hidden layer, and an output layer, was adopted. In
conventional rainfall–runoff models, the model inputs are the past several periods of
precipitation, and the model output is the present discharge at the watershed outlet. In ANN
approaches, the past several periods of both precipitation and discharge are usually adopted
as input vectors. The general function of the ANN model output can, therefore, be
expressed as:

b ðtÞ ¼ f ½Pðt
1Þ; Pðt
2Þ;    ; Pðt
mÞ; Qðt
1Þ; Qðt
2Þ;    ; Qðt
nÞ
Q ð3Þ

in which Qb ðtÞ is the predicted discharge at time t; P(t−m) is the observed precipitation at
time t−m; and Q(t−n) is the observed discharge at time t−n. Since no general rule is
available for determining the number of input vectors, the prior values of m and n may be
based on a correlation matrix of training data from the study watersheds (Tokar and
Johnson 1999).
Deterministic insight into ANN model performance for storm runoff simulation 71

4 Deterministic Runoff Model

The deterministic model that we adopted was the geomorphologic instantaneous unit
hydrograph model (GIUH) proposed by Rodriguez-Iturbe and Valdes (1979). The rainfall–
runoff process can be represented by integrating all the rainfall excess particles moving
along different paths towards the watershed outlet to form the outflow hydrograph. The
resultant IUH can be expressed as (Rodriguez-Iturbe and Valdes 1979):
X 
uðtÞ ¼ fxoi ðt Þ*fxi ðtÞ*fxj ðtÞ*    *fxΩ ðt Þ  PðwÞ ð4Þ
w2W w

in which u(t) is the instantaneous unit hydrograph; fxj ðtÞ is the travel-time probability
density function in state xj with a mean value of Txj ; W is the runoff travel path space; P(w)
is the probability of a specified runoff path w; and * denotes a convolution integral. Lee and
Yen (1997) applied kinematic-wave theory to estimate the mean travel times on overland
areas and channels of each order. The runoff travel time in the ith-order overland areas and
channels can be expressed, respectively, as follows (Lee and Yen 1997)

0 11
m
n L
o o
Txoi ¼ @ 1=2 i A ð5Þ
S oi iem
1

20 11 3
m
Bi 6@ m 2ie nc Loi Lci A 7
Txi ¼ 4 hcoi þ
hcoi 5 ð6Þ
2ie Loi 1=2
S ci Bi

in which no and nc are the roughness coefficients for the overland areas and channels,
respectively; Loi and Lci are the mean overland and channel lengths of the ith-order sub-
basin, respectively; S oi and S ci are the the mean ith-order overland and channel slopes,
respectively; Bi is the width of the ith-order channel; ie is the rainfall excess intensity; m is
an exponent; and hcoi is the inflow depth of the ith-order channel due to water transported
from upstream reaches, which can be determined from the watershed channel network
structure. The kinematic wave based GIUH model is then denoted as KW-GIUH model in
the following sections.

5 Model Application

5.1 Description of the Study Watersheds

The Wu-Tu and Heng-Chi watersheds in northern Taiwan were selected for application of
the ANN model in order to gain deterministic insight. Both of the stream networks of the
two watersheds are fourth order, and the sizes of the watersheds are 204 and 54 km2,
respectively. The watershed geomorphic factors of these two watersheds for application of
the KW-GIUH model are listed in Table 1. Both of the watersheds are upland forest
watersheds with heavy brush, and the channels consist of large boulders with weeds and
72 K.T. Lee, et al.

Table 1 Geomorphic factors of the Wu-Tu and Heng-Chi watersheds

Watershed Stream Number of Mean length of Mean of ith- Mean slope of Mean slope of
(1) order i ith-order ith-order channel order drainage ith-order ith-order
(2) channel Ni Lci (km) area Ai (km2) overland S oi channel S ci
(3) (4) (5) (6) (7)

Wu-Tu 1 89 1.14 1.43 0.4133 0.0825

2 20 2.35 5.63 0.3640 0.0514
3 4 3.49 18.00 0.3909 0.0386
4 1 39.42 203.89 0.3454 0.0181
Heng-Chi 1 30 0.78 1.05 0.4786 0.1312
2 6 3.13 6.91 0.4750 0.0701
3 2 1.79 19.81 0.4001 0.0524
4 1 5.14 53.92 0.3222 0.0246

brush banks. The averaged annual rainfall is 4,847 mm for Wu-Tu and 3,353 mm for Heng-
Chi. Hydrologic records show that high rainfall intensity caused by typhoons and
thunderstorms occurs mainly between June and September in these watersheds. The
proportion of the base flow compared to the total runoff of the storm is relatively small. The
major component of the discharge is due to rapid surface runoff, which results not only
from the high rainfall intensity, but also from the steep slopes of the watersheds.

5.2 Data Screening and ANN Models Training

Hydrologic records from these two watersheds from 1975 to 2000 were collected for
analysis. Storm records with major measurement errors have been excluded in this analysis.
Since the success of an ANN model depends on the reliability of the hydrologic records,
systematic examination of the quality of the input data is required. The KW-GIUH model
was used to investigate the reliability and hydrologic characteristics of the storm event data.
The model has been tested in Taiwan (Lee and Yen 1997; Lee 1998; Lee et al. 2001; Lee et
al. 2006a,b) and the U. S. (Yen and Lee 1997). Hourly rainfall and discharge data from 42
storm events were collected for analysis. In performing the runoff model, the rainfall-excess
hyetograph was determined by deducing the abstractions from the rainfall using the Horton
infiltration equation.
Storm event records from the study watersheds were firstly compared with the simulated
results obtained using the KW-GIUH model. Two criteria were used to evaluate the
goodness of fit of the model. The estimate error of the peak discharge can be expressed as:

b p ðtÞ
Qp ðtÞ
Q
EQp ð%Þ ¼  100 ð7Þ
Qp ðt Þ
b p ðtÞ is the predicted peak discharge;
where EQp is the estimate error of the peak discharge; Q
and Qp(t) is the observed peak discharge. The estimate error of time to peak discharge is
given by:

ETp ¼ Tbp
Tp ð8Þ

bp is the predicted time to peak

where ETp is the estimate error of time to peak discharge; T
discharge; and Tp is the observed time to peak discharge.
Deterministic insight into ANN model performance for storm runoff simulation 73

Table 2 Storm events analyzed in this study and the simulated results of KW-GIUH model

Watershed Storm events nc no Observed KW-GIUH simulation Remark

(1) (2) (3) (4) Qp(m3 s−1) (8)
(5) EQp(%) ETp(h)
(6) (7)

Wu-Tu 09/17/1986 0.05 12.0 603 5.93 2 Unusual

07/27/1987 0.05 6.0 257 6.49 1 Unusual
06/23/1990 0.05 2.0 235 4.26 1 Usual
08/19/1990 0.05 3.0 532 10.21 1 Usual
09/20/1992 0.05 0.1 389 −4.02 −4 Unusual
06/18/1994 0.05 18.0 532 10.21 1 Unusual
08/071994 0.08 1.0 344 8.38 1 Usual
10/09/1994 0.05 1.0 481 1.06 2 Usual
07/30/1996 0.05 1.0 1,090 10.21 2 Usual
09/27/1996 0.05 5.0 678 9.34 −1 Unusual
08/17/1997 0.05 1.2 1,040 8.37 1 Usual
08/29/1997 0.05 2.0 960 6.30 1 Usual
10/15/1998 0.05 10.0 1,050 5.92 −2 Unusual
10/25/1998 0.05 3.0 1,050 9.52 1 Usual
04/24/2000 0.05 1.0 323 4.33 0 Usual
06/12/2000 0.09 8.0 211 0.66 1 Unusual
07/09/2000 0.03 1.0 325 6.76 0 Usual
08/22/2000 0.05 10.0 350 8.35 1 Unusual
09/09/2000 0.08 2.0 182 4.33 2 Usual
11/08/2000 0.05 1.0 632 2.80 1 Usual
12/13/2000 0.05 2.0 365 4.06 1 Usual
Heng-Chi 08/02/1975 0.05 0.6 164 3.38 −1 Usual
08/07/1984 0.05 5.5 251 6.52 1 Unusual
02/08/1984 0.05 0.8 137 4.47 −1 Usual
09/16/1985 0.05 1.2 592 1.27 −1 Usual
10/03/1985 0.05 0.05 245 3.93 1 Unusual
08/22/1986 0.05 0.08 336 1.23 0 Unusual
09/17/1986 0.05 1.2 457 5.06 −1 Usual
09/08/1987 0.10 2.0 329 6.69 0 Unusual
10/23/1987 0.05 0.5 363 0.13 −2 Usual
08/18/1990 0.05 0.6 492 6.34 −1 Usual
08/30/1990 0.05 0.5 668 0.21 −1 Usual
09/07/1990 0.05 0.01 290 −1.45 0 Unusual
06/19/1991 0.05 0.01 233 2.85 0 Unusual
08/27/1992 0.15 2.0 149 2.20 0 Unusual
06/05/1993 0.05 6.0 179 3.40 0 Unusual
08/07/1994 0.05 12.0 120 4.59 −1 Unusual
08/31/1994 0.05 15.0 154 7.45 −1 Unusual
07/30/1996 0.05 1.5 243 6.25 −1 Usual
08/17/1997 0.05 0.8 195 −2.48 −3 Usual
08/28/1997 0.05 0.5 132 5.70 0 Usual
10/31/2000 0.05 2.5 317 5.30 0 Unusual

As shown in Table 2, runoff simulation performed using KW-GIUH model could usually
produce acceptable results while comparing with the hydrological records. The estimate
error of peak discharge was almost always limited to 10%, and the estimate error of time to
peak was usually less than ±2 h. Columns 3 and 4 show the values of channel-flow
74 K.T. Lee, et al.

roughness coefficient nc and overland-flow roughness coefficient no. The channel-flow

roughness reflects the channel bottom roughness, riverbank vegetation, and channel
irregularity and obstruction (Chow 1959). The overland-flow roughness reflects the surface
grain roughness, microtopographic features, ground surface cover, vegetation, infiltration
capacity, and antecedent soil moisture conditions (Engman 1986; Weltz et al. 1992; Singh
1996). In this study, the calibrated channel-flow coefficient nc was found to be around 0.05
for both the Wu-Tu and Heng-Chi watersheds. For usual storm events, the limited range of
the calibrated coefficient no for the Wu-Tu watershed was 1.0∼3.0, and it was 0.5∼1.5 for
the Heng-Chi watershed because the land cover conditions in these two watersheds vary,
and the soil in Wu-Tu showed a higher infiltration capacity than did that in the Heng-Chi
watershed. The ranges of the above mentioned values for the overland-flow roughness
coefficient were considered reasonable when compared with the values provided by the
Hydrologic Engineering Center (2000). Nevertheless, the calibrated roughness values for
some exceptional storm events records were found to be outside the reasonable range,
which may be due to significant variation of the rainfall spatial distribution or extreme
antecedent hydrologic conditions because the seasonal effect on vegetation cover in Taiwan
is negligible. Since the KW-GIUH model is a semi-distributed model and cannot account
for soil moisture variation, these special hydrologic conditions will result in the values of
the model parameters outside the reasonable range. Such storm events were then designated
as “unusual” storm events in this study (Column 8 in Table 2).
Careful investigations were conducted to determine possible reasons for the exceptional
range of the value of the coefficient. For example, the 3 Oct. 1985 storm and the 22 Aug.
1986 storm in the Heng-Chi watershed were found to be associated with very wet
antecedent conditions of short duration that included intense rainfall; consequently, lower
values of no were required to reflect the fast, concentrated surface-flow processes. On the
other hand, the 31 Aug. 1994 storm was a mild storm with very dry antecedent conditions;
a high no value was, therefore, used to reflect the large portion of slowing flow moving
toward the watershed outlet.
By using different sets of training data, we developed three types of ANN models,
namely, ANN1, ANN2, and ANN3. The ANN1 model structure was trained using the whole
set of hydrologic records collected in this study, which included “usual” and “unusual”
storm events shown in Table 2. The ANN2 model was trained using only the records from
the “usual” storm events, and the ANN3 model was trained using only the records from the
“unusual” storm events. The data sets used for model training, testing, and verification are
listed in Table 3. The performance of the models and input vectors characteristics are
discussed detail in the following section.

6 Results and Discussion

6.1 Model Performance Achieved Using Different Training Data Sets

Training processes, which included iterative adjustment and optimization of the weight
matrices as well as bias vectors, were applied to find the best ANN structure that minimized
the deviation between the observed records and predict outputs. Additional input vectors
added into the model were trained in different steps. The goodness-of-fit statistics for the
present model were computed for the training and testing procedures, and compared with
those for the best-fit model in the previous step. Iterative training for the 1-h ahead
Deterministic insight into ANN model performance for storm runoff simulation 75

Table 3 Selected storm events in ANN models’ training, testing, and verification

Watershed Models Data process Storm events

(1) (2) (3) (4)

Wu-Tu ANN1 Training 06/23/1990, 06/18/1994*, 07/30/1996, 07/27/1987*,

08/29/1997, 07/09/2000, 08/22/2000*, 09/09/2000,
11/08/2000
Testing 08/07/1994, 10/09/1994, 09/17/1986*, 10/25/1998,
04/24/2000, 06/17/2000*
Verification 08/19/1990, 09/20/1992*, 09/27/1996*, 08/17/1997,
10/15/1998*, 12/13/2000
ANN2 Training 06/23/1990, 07/30/1996, 08/29/1997, 07/09/2000,
09/09/2000, 11/08/2000
Testing 08/07/1994, 10/09/1994, 10/25/1998, 04/24/2000
Verification 08/19/1990, 09/20/1992*, 09/27/1996*, 08/17/1997,
10/15/1998*, 12/13/2000
ANN3 Training 07/27/1987*, 06/18/1994*, 08/22/2000*
Testing 09/17/1986*, 06/17/2000*
Verification 08/19/1990, 09/20/1992*, 09/27/1996*, 08/17/1997,
10/15/1998*, 12/13/2000
Heng-Chi ANN1 Training 08/02/1975, 02/08/1985, 09/16/1985, 10/03/1985*,
09/17/1986, 10/23/1987, 08/07/1994*, 08/31/1994*,
10/31/2000*
Testing 08/18/1990, 08/30/1990, 09/07//1990*, 08/27/1992*,
06/05/1993*
Verification 08/07/1984*, 08/22/1986*, 06/19/1991*, 07/30/1996,
08/17/1997, 08/28/1997
ANN2 Training 08/02/1975, 02/08/1985, 09/16/1985, 09/17/1986,
10/23/1987
Testing 08/18/1990, 08/30/1990
Verification 08/07/1984*, 08/22/1986*, 06/19/1991*, 07/30/1996,
08/17/1997, 08/28/1997
ANN3 Training 10/03/1985*, 08/07/1994*, 08/31/1994*, 10/31/2000*,
06/19/1991*
Testing 09/07/1990*, 08/27/1992*, 06/05/1993*
Verification 08/07/1984*, 08/22/1986*, 06/19/1991*, 07/30/1996,
08/17/1997, 08/28/1997

prediction discharge, Qb ðtÞ, showed that the input vectors of the optimal ANN model should
consist of the past 3 h of precipitation, P(t−1), P(t−2), and P(t−3), and the past hour of
discharge, Q(t−1), for the Wu-Tu watershed; on the other hand, the input vectors should
consist of the past 2 h of precipitation, P(t−1) and P(t−2), and the past hour of discharge, Q
(t−1), for the Heng-Chi watershed.
The estimate error of the peak discharge (EQp) and root mean square error (RMSE) of
the verification storm events obtained using the ANN1, ANN2 and ANN3 models are shown
in Table 4. In these verification events, half of the hydrologic records were from “usual”
storm events and the other half records were from “unusual” storm events. We found that
the simulation performance of the ANN1 model was, in general, slightly better than that of
the ANN2 model, and that the model performance of the ANN2 model was better than that
of the ANN3 model. This was because the ANN1 model was trained using complete
 76

Table 4 Simulation results of ANNs and KW-GIUH models

Watershed Verification Observed ANN1 ANN2 ANN3 ANNNF KW-GIUH

(1) events Qp(m3 s−1)
(2) (3) EQp (%) RMSE EQp (%) (6) RMSE EQp (%) RMSE EQp (%) RMSE EQp (%) RMSE
(4) (m3 s−1) (m3 s−1) (8) (m3 s−1) (10) (m3 s−1) (12) (m3 s−1)
(5) (7) (9) (11) (13)

Wu-Tu 08/19/1990 825 5.44 38.25 5.92 38.78 −20.57 56.86 −6.03 85.79 10.21 59.87
09/20/1992* 389 −15.01 19.11 −12.20 16.23 −12.93 20.27 −15.01 24.61 −4.02 21.69
09/27/1996* 678 17.90 20.46 17.63 19.42 10.17 20.81 25.37 56.31 9.34 27.34
08/17/1997 1,040 0.79 24.10 −0.95 24.32 −33.61 79.68 12.10 53.82 8.37 120.39
10/15/1998* 1,050 1.55 47.01 −2.08 42.93 −33.96 126.81 2.28 47.02 4.92 58.60
12/13/2000 364 3.81 8.82 2.81 12.35 4.37 11.92 6.56 38.72 4.06 18.43
Heng-Chi 08/07/1984* 251 49.48 34.12 59.32 37.21 41.92 39.26 51.47 33.55 6.52 16.10
08/22/1986* 336 −0.26 29.11 12.56 33.39 −21.31 24.81 65.42 44.56 1.23 34.13
06/19/1991* 233 −9.69 22.31 44.58 29.70 15.96 37.63 10.93 29.31 2.85 39.81
07/30/1996 260 8.13 14.02 7.16 15.06 5.13 12.17 36.87 24.80 6.25 19.33
08/17/1997 195 −6.94 10.18 −4.09 12.39 −0.02 8.06 −8.83 12.47 −2.48 37.78
08/28/1997 132 7.11 11.66 2.55 11.85 17.42 10.16 117.61 47.91 5.70 32.87

*unusual storm events

K.T. Lee, et al.
Deterministic insight into ANN model performance for storm runoff simulation 77

hydrologic data sets, which included both the “usual” and “unusual” storm events; thus the
ANN structure could account for various hydrologic conditions. The ANN2 model was
trained using only the “usual” storm event data. Although it could represent the major
runoff process of the watershed, the model could not include some exceptional hydrologic
conditions of the watershed. Consequently, the performance of ANN2 was sometimes
poorer than that of ANN1. As mentioned previously, the “unusual” events were the storm
events in the watershed that occurred under exceptional hydrologic conditions, which could
result from very dry or very wet antecedent conditions, or from significant variation of the
spatial distribution of the rainstorms. Therefore, only some of the hydrologic characteristics
of the watershed could be extracted from the training data sets. The ANN3 model could not
provide a complete structure reflecting various hydrologic conditions.
In short term rainfall–runoff simulations, if the estimate error of peak discharge is less
than 10%, the simulation can be considered acceptable. Consequently, we found that the
ANN1 and ANN2 models could generally produce good simulation results for the “usual”
storm events, but not for the “unusual” storm events. Nevertheless, when the KW-GIUH
model parameters values (the overland- and channel-flow roughness coefficients) were
adjusted, the hydrologic response of the deterministic runoff model could adapt to
accommodate the exceptional hydrologic conditions, thus producing good simulation
results for the “unusual” storm events.
The records and simulated hydrographs obtained using ANN1, ANN2, ANN3, and KW-
GIUH for two example storms in the Wu-Tu watershed are shown in Figs. 1 and 2. Figure 1
shows the verification of a storm event on 17 Aug. 1997 in the Wu-Tu watershed, which
was classified as a “usual” storm event in this study. The figure shows that the deviation

0
10

20
30
Wu-Tu 08/17/1997
1200 recorded
ANN1
1000 ANN2
ANN3
KW-GIUH
800

600

400

200

0 12 24 36 48 60 72 84 96
Time (h)
Fig. 1 Models verification of a usual storm event in the Wu-Tu watershed
78 K.T. Lee, et al.

0
10
20
30
40

Wu-Tu 09/27/1996
800 recorded
ANN1
ANN2
600 ANN3
KW-GIUH

400

200

0 10 20 30 40 50 60
Time (h)
Fig. 2 Models verification of an unusual storm event in the Wu-Tu watershed

between the records and the simulated hydrographs is minor for the ANN1 and ANN2
models but is large for the ANN3 model. Figure 2 shows the verification of an “unusual”
storm event on 27 Sept. 1996, which demonstrates that the KW-GIUH model can more
accurately predict the storm hydrograph peak while the model parameters are adequately
adjusted.

7 Input Vectors Characteristics

As mentioned previously, a trial and error procedure was used to find the best-of-fit ANN
model in this study. Tokar and Johnson (1999) reported that correlation matrixes could be

Table 5 Correlation matrix of the input and output vectors in the Wu-Tu watershed

Vectors P(t−5) P(t−4) P(t−3) P(t−2) P(t−1) Q(t−1) Q(t)

P(t−5) 1.000
P(t−4) 0.640 1.000
P(t−3) 0.440 0.640 1.000
P(t−2) 0.396 0.440 0.640 1.000
P(t−1) 0.309 0.396 0.440 0.640 1.000
Q(t−1) 0.620 0.655 0.661 0.677 0.566 1.000
Q(t) 0.588 0.620 0.655 0.661 0.677 0.982 1.000
Deterministic insight into ANN model performance for storm runoff simulation 79

Table 6 Correlation matrix of the input and output vectors in the Heng-Chi watershed

Vectors P(t−5) P(t−4) P(t−3) P(t−2) P(t−1) Q(t−1) Q(t)

P(t−5) 1.000
P(t−4) 0.682 1.000
P(t−3) 0.544 0.682 1.000
P(t−2) 0.418 0.544 0.682 1.000
P(t−1) 0.409 0.418 0.544 0.682 1.000
Q(t−1) 0.602 0.640 0.691 0.731 0.604 1.000
Q(t) 0.560 0.602 0.640 0.691 0.731 0.915 1.000

used a prior rule for selecting the input vectors of ANN models. Therefore, we examined
the correlation matrixes of the input vectors to the system output. Tables 5 and 6 show the
correlation matrixes of the input and output vectors in the Wu-Tu and Heng-Chi
watersheds. If a significant drop in the correlation coefficient values between Q(t) and
P(t−1), P(t−2), ..., P(t−m) can be used as a criteria for precipitation input vector selection,
then the adequate precipitation input vectors should be P(t−3), P(t−2), and P(t−1) for the
Wu-Tu watershed, and P(t−2) and P(t−1) for the Heng-Chi watershed (see the bold
numerals in Tables 5 and 6).
In deterministic hydrology, precipitation during different periods is the only input used
for watershed rainfall–runoff modeling. The number of input vectors (the length of the
system-input memory) represents the persistence of the input vectors within the hydrologic

0
20

40
60 Wu-Tu 08/19/1990
recorded
feedback
1000
non-feedback

800

600

400

200

0 12 24 36 48
Time (h)
Fig. 3 Real-time feedback and non-feedback ANN models simulations in the Wu-Tu watershed
80 K.T. Lee, et al.

response system. Therefore, the watershed time of concentration, which is defined as the
time of flow from the hydraulic remotest point on the watershed to the outlet (Chow 1964),
can be used to estimate the past periods of precipitation that should be considered in the
ANN system. Applying a storm averaged rainfall excess intensity, ie, of 30∼50 mm h−1 in
Eqs. 5 and 6, we found that the values of the watershed time of concentration were around
3.2 h for the Wu-Tu watershed, and around 2.2 h for the Heng-Chi watershed. These values
of the time of concentration correspond to the best-of-fit ANN models with precipitation
input vectors P(t−3), P(t−2) and P(t−1) for the Wu-Tu watershed, and with P(t−2) and P
(t−1) for the Heng-Chi watershed. This demonstrates that the number of precipitation input
vectors can be determined either from the correlation matrix or from the watershed time of
concentration estimate. However, the later may be more convenient to be used in practice.

7.1 Real-Time Feedback and Non-Feedback ANN Systems

In conventional watershed rainfall–runoff modeling, GIUH model for example, rainfall

during different periods is the only input of the hydrological model. However, the ANN
model shown in Eq. 3 is a real-time feedback system because the real-time observation
quantities, Qðt
1Þ; Qðt
2Þ;    ; Qðt
nÞ; are treated as input vectors used to update
the latest hydrologic response of the watershed. Since the ANN model considers watershed
latest hydrologic information, it captures the evolution of the hydrodynamic system.
Simulation results obtained using the ANN model are, therefore, usually in good agreement
with records.

0
10

20
30 Heng-Chi 08/28/1997
Recorded
Feedback
300
Non-feedback

250

200

150

100

50

0 12 24 36 48 60 72 84 96
Time (hr)
Fig. 4 Real-time feedback and non-feedback ANN models simulations in the Heng-Chi watershed
Deterministic insight into ANN model performance for storm runoff simulation 81

However, real-time discharge information may not be available for rainfall–runoff

simulation in the field. It is crucial to know whether the ANN models would preserve the
same accuracy for runoff prediction when the past n periods observation quantities,
Qðt
1Þ; Qðt
2Þ;    ; Qðt
nÞ; are replaced by previous model predicted quantities,
b ðt
1Þ; Q
Q b ðt
2Þ;    ; Q
b ðt
nÞ. Two example storms in the study watersheds were
adopted to examine the difference between using the input vector of the past period
observed quantity, Qðt
1Þ and the past period ANN model predicted quantity, Q b ðt
1Þ
As shown in Figs. 3 and 4, smaller deviations between the records and simulated results
were obtained when the past 1 h observed quantity (real-time feedback model) was used
rather than using the past 1 h model predicted quantity (non-feedback model). For these two
storms the root mean square errors (RMSEs) of the real-time feedback ANN model were
38.3 and 11.7 m3 s−1 for the Wu-Tu and Heng-Chi watershed, respectively. However, the
RMSEs of the non-feedback model increased to 85.8 and 47.9 m3 s−1 for these two
watersheds. Simulation results of the non-feedback ANN model (ANNNF) for other storms
are shown in Table 4. It can be concluded that the good performance of the ANN model
simulation may be mainly due to the fact that the past period observation quantity was
included as one of the input vectors to reflect the evolution of the hydrologic response
system. It has shown that simulation results of the non-feedback ANN model were even
worse than the results generated by using the KW-GIUH model.

8 Conclusions

Three types of ANN models were developed in this study, based on different groups of
training data sets, which were classified into “usual” and “unusual” storm events using a
deterministic runoff model. Although the ANN1 model has difficulty to obtain acceptable
simulation results for the “unusual” storm events, it performs better for runoff simulation
than the ANN2 and ANN3. Thus, the ANN model developed using complete data set is
suggested. We also found that the accuracy of the ANN models was mainly due to
including the real-time observation quantities into analyses. So the real-time feedback
information is required to assure good quality of simulation using ANN.

Acknowledgements This study is part of a research work supported by the National Science Council,
Taiwan, ROC, under grant NSC 89-TPC-7-019-011. Financial support from the National Science Council is
gratefully acknowledged.

References

Amorocho J (1967) The nonlinear prediction problem in the study of runoff cycle. Water Resour Res 3
(3):861–880
Anmala J, Zhang B, Govindaraju RS (2000) Comparison of ANNs and empirical approaches for predicting
watershed runoff. J Water Resour Plan Manage, ASCE 126(3):156–166
ASCE Task Committee (2000a) Artificial neural networks in hydrology. I: Preliminary concepts. J Hydrol
Eng ASCE 5(2):115–123
ASCE Task Committee (2000b) Artificial neural networks in hydrology. II:Hydrologic applications. J Hydrol
Eng ASCE 5(2):124–137
Caudill M (1987) Neural networks primer. Part I. AI expert:46–52
Chiu CL, Huang JT (1970) Nonlinear time-varying model of rainfall–runoff relation. Water Resour Res 6
(1):1277–1286
Chow VT (1959) Open-channel hydraulics. McGraw-Hill, New York
82 K.T. Lee, et al.

Chow VT (1964) Handbook of applied hydrology. McGraw-Hill, New York

Dawson CW, Wilby R (1998) An artificial neural network approach to rainfall–runoff modeling. Hydrol Sci J
43(1):47–66
Ding JY (1974) Variable unit hydrograph. J Hydrol 22:53–69
Dooge JCI (1959) A general theory of the unit hydrograph. J Geophys Res 64(1):241–256
Engman ET (1986) Roughness coefficients for routing surface runoff. J Irrig Drain Eng ASCE 112(1):39–53
French MN, Krajewski WF, Cuykendall RR (1992) Rainfall forecasting in space and time using a neural
network. J Hydrol 137:1–31
Haykin S (1994) Neural Networks: a comprehensive foundation. MacMillan, New York
Hopfield JJ (1982) Neural networks and physical system with emergent collective computational abilities.
Proc Natl Acad Sci 79:2554–2558
Hsu K, Gupta HV, Sorooshian S (1995) Artificial neural network modeling of the rainfall–runoff process.
Water Resour Res 31(10):2517–2530
Hydrologic Engineering Center (2000) Hydrologic modeling system HEC-HMS, Technical reference
manual, U. S. Army Corps of Engineers.
Jain A, Indurthy SKVP (2003) Comparative analysis of event-based rainfall–runoff modeling techniques –
deterministic, statistical, and artificial neural networks. J Hydrol Eng ASCE 8(2):93–98
Kumar DN, Raju KS, Sathish T (2004) River flow forecasting using recurrent neural networks. Water Resour
Manag 18:143–161
Lee KT (1998) Generating design hydrographs by DEM assisted geomorphic runoff simulation: a case study.
J Am Water Resour Assoc 34(2):375–384
Lee KT, Yen BC (1997) Geomorphology and kinematic-wave based hydrograph deviation. J Hydraul Eng
ASCE 123(1):73–80
Lee KT, Chang C-H, Yang M-S, Yu W-S (2001) Reservoir attenuation of floods from ungauged basins.
Hydrol Sci J 46(3):349–362
Lee KT, Chung Y-R, Lau C-C, Meng C-C, Chiang S (2006a) A windows-based inquiry system for design
discharge based on geomorphic runoff modeling. Comput Geosci 32(2):203–211
Lee KT, Ho Y-H, Chyan Y-J (2006b) Bridge blockage and overbank flow simulations using HEC-RAS in the
Keelung River during the 2001 Nari typhoon. J. Hydraul Eng, ASCE 132(3):319–323
Muleta MK, Nicklow JW (2004) Joint application of artificial neural networks and evolutionary algorithms
to watershed management. Water Resour Manag 18:459–482
Minshall NE (1960) Predicting storm runoff on small experimental watersheds. J Hydraul Eng ASCE
HY8:17–38
Nash JE (1957) The form of instantaneous unit hydrograph. Int’l Assoc Sci Hydrol, Pub 45(3):114–121
Rao SVN, Thandaveswara BS, Bhallamudi M, Srinivasulu V (2003) Optimal groundwater management in
deltaic regions using simulated annealing and neural networks. Water Resour Manag 17:409–428
Rodriguez-Iturbe I, Valdes JB (1979) The geomorphologic structure of hydrologic response. Water Resour
Res 15(6):1409–1420
Rumelhart DE, Hinton GE, Williams RJ (1986) Learning internal representations by error propagation.
Parallel Distributed Processing. MIT, Cambridge, MA, pp 318–362
Sherman LK (1932) Streamflow from rainfall by unit-graph method. Eng News-Rec 108(7):501–505
Singh VP (1996) Kinematic wave modeling in water resources – surface water hydrology. Wiley, USA.
Tayfur G (2002) Artificial neural networks for sheet sediment transport. Hydrol Sci J 47(6):879–892
Thirumalaiah K, Deo MC (2000) Hydrological forecasting using neural networks. J Hydrol Eng ASCE 5
(2):180–189
Tokar AS, Johnson PA (1999) Rainfall–runoff modeling using artificial neural networks. J Hydrol Eng ASCE
4(3):232–239
Weltz MA, Arslan AB, Lane LJ (1992) Hydraulic roughness coefficients for native rangelands. J Irrig Drain
Eng ASCE 118(5):776–790
Yapo PO, Gupta VK, Sorooshian S (1996) Automatic calibration of conceptual rainfall–runoff models:
sensitivity to calibration data. J Hydrol 181:23–48
Yen BC, Lee KT (1997) Unit hydrograph derivation for ungaged watersheds by stream order laws. J Hydrol
Eng ASCE 2(1):1–9
Zealand CM, Burn DH, Simonovic SP (1999) Short term streamflow forecasting using artificial neural
networks. J Hydrol 214:32–48