Journal of Hydrology 535 (2016) 211–225

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

A comparison between wavelet based static and dynamic neural

network approaches for runoff prediction
Muhammad Shoaib a,⇑, Asaad Y. Shamseldin a, Bruce W. Melville a, Mudasser Muneer Khan b
a
Department of Civil and Environmental Engineering, The University of Auckland, Private Bag 92019, Auckland, New Zealand
b
Department of Civil Engineering, Bahauddin Zakariya University, Multan, Pakistan

a r t i c l e i n f o s u m m a r y

Article history: In order to predict runoff accurately from a rainfall event, the multilayer perceptron type of neural net-
Received 3 August 2015 work models are commonly used in hydrology. Furthermore, the wavelet coupled multilayer perceptron
Received in revised form 26 January 2016 neural network (MLPNN) models has also been found superior relative to the simple neural network
Accepted 28 January 2016
models which are not coupled with wavelet. However, the MLPNN models are considered as static and
Available online 6 February 2016
This manuscript was handled by Corrado
memory less networks and lack the ability to examine the temporal dimension of data. Recurrent neural
Corradini, Editor-in-Chief, with the network models, on the other hand, have the ability to learn from the preceding conditions of the system
assistance of Gokmen Tayfur, Associate and hence considered as dynamic models. This study for the first time explores the potential of wavelet
Editor coupled time lagged recurrent neural network (TLRNN) models for runoff prediction using rainfall data.
The Discrete Wavelet Transformation (DWT) is employed in this study to decompose the input rainfall
Keywords: data using six of the most commonly used wavelet functions. The performance of the simple and the
Rainfall runoff modelling wavelet coupled static MLPNN models is compared with their counterpart dynamic TLRNN models.
Dynamic recurrent models The study found that the dynamic wavelet coupled TLRNN models can be considered as alternative to
Discrete wavelet transformation the static wavelet MLPNN models. The study also investigated the effect of memory depth on the perfor-
mance of static and dynamic neural network models. The memory depth refers to how much past infor-
mation (lagged data) is required as it is not known a priori. The db8 wavelet function is found to yield the
best results with the static MLPNN models and with the TLRNN models having small memory depths. The
performance of the wavelet coupled TLRNN models with large memory depths is found insensitive to the
selection of the wavelet function as all wavelet functions have similar performance.
Ó 2016 Elsevier B.V. All rights reserved.

1. Introduction without taking into consideration the spatial variability and

Prediction of runoff produced from a watershed as a result of
rainfall event is a key area of research in hydrology. It is considered
as one of the most complex hydrological process to be modelled
because of the involvement of number of variables in the mod-
elling process and the enormous spatial and temporal variability
of watershed characteristics. Since the establishment of rational
method in 1850 (Mulvany, 1850) for calculation of the peak dis-
charge, numerous hydrological models have been proposed. These
models include two main categories: the theory driven (conceptual
and physically-based) models and the data driven (empirical and
black-box) models. Conceptual models describe the general sub-
processes and the physical mechanisms of the hydrological cycle
⇑ Corresponding author.
E-mail addresses: msho127@aucklanduni.ac.nz (M. Shoaib), a.shamseldin@
auckland.ac.nz (A.Y. Shamseldin), b.melville@auckland.ac.nz (B.W. Melville),
mkha222@aucklanduni.ac.nz (M.M. Khan).
stochastic characteristics of the rainfall–runoff process. Physically
based models involve the solution of a system of partial differential
equations in order to simulate various constituent processes of the
hydrological cycle. Data-driven models consider the hydrological
system as a black-box and try to establish a relationship between
historical inputs (such as rainfall, evaporation etc.) and outputs
(such as runoff).
Among data-driven models, the artificial neural network (ANN)
models has emerged as powerful black-box models and received a
great attention during last two decades. The idea of ANN is inspired
by the operation of the biological neural networks of the central
nervous system of human brain. Mathematically, an ANN is a com-
pound nonlinear function with numerous factors that are adjusted
in such a way that the ANN output becomes comparable to the
observed output. The ANN approach has been successfully used
for different modelling problems in various branches of science
and engineering. In the field of hydrology, French et al. (1992) were
the first to use ANN for forecasting rainfall. Shamseldin (1997)

http://dx.doi.org/10.1016/j.jhydrol.2016.01.076
0022-1694/Ó 2016 Elsevier B.V. All rights reserved.
212 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225
pioneered the use of ANN in modelling rainfall–runoff relationship.
The ANN has been successfully applied in many hydrological stud-
ies (e.g. Akiner and Akkoyunlu, 2012; Antar et al., 2006; Arsenault
et al., 2015; Aziz et al., 2014; Jain et al., 2004; Lallahem and Mania,
2003; Mekanik et al., 2013; Nourani et al., 2009a; Piotrowski et al.,
2015; Senthil Kumar et al., 2005). An extensive review of ANN in
hydrological applications can be found in ASCE Task Committee
(2000a, 2000b), Abrahart et al. (2012) and Tayfur (2012).
However, Cannas et al. (2006) pointed out that the ANN based
models may not be able to deal with non-stationary data until
pre-processing of the input and/or output data is performed. Appli-
cation of wavelet transformation (WT) on time series data has been
found effective in addressing this issue of non-stationary data
(Nason and Sachs, 1999). The WT decomposes the time series data
into its sub-constituents and these sub-constituents are used as
external inputs to the ANN. The resulting model is known as the
hybrid wavelet model. These hybrid models improve the perfor-
mance of ANN by capturing the important temporal and the spec-
tral information embedded in the time series data. Various studies
used WT in order to improve the results of the ANN based
hydrological models (e.g., Wang and Ding, 2003; Cannas et al.,
2006; Nourani et al., 2009a, 2009b; Tiwari and Chatterjee, 2010;
Adamowski and Sun, 2010; Kisi, 2011; Singh, 2012; Shoaib et al.,
2014a, 2014b; Altunkaynak and Nigussie, 2015). Wang and Ding
(2003) used a three layered hybrid wavelet feed forward neural
network (FFNN) model with back propagation (BP) training algo-
rithm for forecasting shallow groundwater levels and river dis-
charges. Cannas et al. (2006) applied hybrid wavelet MLPNN
models for forecasting river flows. Nourani et al. (2009a) presented
a hybrid wavelet MLPNN model for prediction of precipitation
while Nourani et al. (2009b) utilized a FFNN model with BP train-
ing algorithm for modelling rainfall–runoff process. Tiwari and
Chatterjee (2010) employed a wavelet coupled MLPNN model for
flood forecasting purposes. Wavelet coupled flow forecasting
MLPNN model for non-perennial rivers was presented by
Adamowski and Sun (2010). Likewise, Singh (2012) presented
wavelet-MLPNN conjunction models for prediction of flood events.
Hybrid wavelet MLPNN and radial basis function neural network
(RBFNN) models were used by Shoaib et al. (2014a) for comparing
the performance of various wavelet coupled models.
Most of the ANN models including the simple and the wavelet
coupled models used in hydrology are of static nature relying on
the MLPNN model to learn the relationship between the observed
input and the observed output. MLPNN is a static network as it
allows only one-way information flow from the input layer to
the output layer. Moreover, it is also considered as memory less
network because of absence of any memory or recursion compo-
nent to store the past information at any given time step. Further-
more, the MLPNN models lack the capability to examine the
temporal dimension of data and cannot instinctively learn from
the preceding conditions of the system (Saharia and
Bhattacharjya, 2012). This is very vital in case of the hydrological
systems since the current response of a hydrologic system can be
very reliant on their preceding states. An implicit method of encod-
ing temporal characteristics in static ANN is to use a sliding win-
dow of input sequences (e.g. Coulibaly et al., 2000a, 2000b; Kisi
et al., 2013; Lohani et al., 2012; Tayfur and Guldal, 2006; Tayfur
et al., 2014). In this method, a form of static memory is implicitly
provided to the MLPNN by selecting an input vector comprising
of the fixed number of past events relevant to the current system
response. But incapability of this method to encode temporal pat-
terns with randomly selected time intervals makes it unsuited for
conditions that require high forecasting efficiency (Saharia and
Bhattacharjya, 2012). The concept of signal delays play an impera-
tive role in the biological neural network system of human brain.
This concept has prompted the development of dynamic recurrent
neural network (RNN) models. RNN models have the capability to
learn from the preceding conditions of the system as they facilitate
time delay units through feedback connections and thus have
attracted much attention recently. The application of RNN can be
found in many studies (e.g. Anmala et al., 2000; Assaad et al.,
2005; Badjate and Dudul, 2009; Chang et al., 2012; Chiang et al.,
2004; Coulibaly and Baldwin, 2005; Coulibaly and Evora, 2007;
Güldal and Tongal, 2010; Kale and Dudul, 2009; Kote and
Jothiprakash, 2008; Ma et al., 2008; Muluye, 2011; Serpen and
Xu, 2003).
It is evident from the literature reviewed and cited in this paper
that the use of static MLPNN and dynamic RNN models is increas-
ing in hydrological studies, but most of the hybrid wavelet ANN
models are relying only on the static MLPNN models. However,
to our present knowledge, no study has yet been conducted to
evaluate the potential of wavelet coupled dynamic neural network
models. This study, is therefore, conducted to compare the perfor-
mance of hybrid wavelet static MLPNN models and dynamic time
lagged recurrent neural network models for runoff prediction using
rainfall data. The performance of hybrid wavelet models is sensi-
tive to the selection of a particular mother wavelet function, the
choice of decomposition level and the preference of appropriate
input variables. This study will, therefore, investigate effect of var-
ious most commonly used wavelet functions, the choice of suitable
decomposition level and the selection of suitable delay signal for
the hybrid wavelet RNN models. The paper is arranged in the fol-
lowing manner. Section 1 gives the introduction and the review
of literature. Section 2 is the methodology section which also
describes data used in the study. In this section, the theoretical
background of MLPNN, time-lagged neural network (TLNN) recur-
rent models, the development of simple and the hybrid wavelet
static and dynamic models are discussed. Also in this section, the
performance indices used to evaluate the developed models are
presented. The results of the different developed models are dis-
cussed in Section 3. The conclusions of the paper are presented
in Section 4.
2. Methodology
2.1. Artificial neural networks (ANN)
2.1.1. Multilayer perceptron neural network (MLPNN)
The MLPNN consists of a number of neurons arranged in a series
of consecutive layers. Typically, it consists of an input layer, a hid-
den layer and an output layer. Each neuron receives an array of
inputs and produces a single output. The output of a neuron in
the input layer will be input for the neuron in the hidden layer.
Similarly, the output of the neuron in the hidden layer will be input
for the output layer. Each neuron in all the layers processes its
input by a mathematical function known as the neuron transfer
function. The neurons in the input layer have connection with
the neuron in the hidden layer while neuron in the output layer
is only connected to the neuron in the hidden layer. There is no
direct connection between the neuron in the input layer with the
neurons in the output layer. MLPNN is the most widely used neural
network type in various application of hydrology (Dawson et al.,
2002; Maier and Dandy, 2000). More theoretical background of
ANN and its various applications in water resources engineering
can be found in Tayfur (2012).
2.1.2. Time-lagged recurrent neural network (TLNN)
Conventionally, MLPNN, where neurons in one layer are only
connected to neurons in the next layer, have been used for predic-
tion and forecasting applications. Nevertheless, recurrent net-
works, where neurons in one layer can be connected to neurons
 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225 213

in the next layer, the previous layer, the same layer and even to
themselves, have been proposed as alternatives to MLPNN
(Warner and Misra, 1996). Time is an important aspect of learning.
We may include time into neural network design implicitly or
explicitly (French et al., 1992). A time lagged recurrent neural net-
work (TLNN) is used in the present study. The static processing ele-
ments/neurons of the MLPNN are replaced by the neurons with
short-term memories in the TLRNN models. These short term
memories may be attached to any layer in the network, producing
very sophisticated neural topologies which may be very useful for
time series prediction and system identification (Charaniya and
Dudul, 2013). The addition of a short-term memory structure only
in the input layer of the static MLPNN is a straightforward method
of implicit representation of time. The resulting neural network
type is called as focused time-lagged recurrent neural network
(FTLRNN). This memory is used to store past information, which
can be used to analyse the temporal variations in the dataset in a
more effective manner.
There are two methods to implement the short term memory
structure in the TLRNN models. These methods include the Tapped
delay line (TDL) memory and the Gamma memory. The TDL mem-
ory may be viewed as single input-multiple output network as it
consists of p unit delays with (p + 1) terminals as shown in Fig. 1
(a). In the figure, the Z
1 represents the unit delay. The memory
depth (D) is fixed as k and its memory resolution (R) is fixed as
unity. The memory depth (D) refers to how far into past the mem-
ory stores information while the memory resolution (R) refers to
the degree to which information regarding input is stored. The
Gamma memory introduced by Hopfield (1982) provides control
over the memory depth by building a feedback loop around each
unit delay as represented in Fig. 1(b). The unit delay Z
1 of TDL is
replaced by the transfer function G(z) given below;
lZ
1
GðzÞ ¼ ð1Þ
1
ð1
lÞZ
1
and the output of each tap of a discrete Gamma memory structure is
given by the following two equations (Motter and Principe, 1994):
X 0 ðtÞ ¼ XðtÞ ð2Þ
X k ðtÞ ¼ ð1
lÞX k ðt
1Þ þ lX k
1 ðt
1Þ; k ¼ 1; . . . ; k ð3Þ
where l is an adjustable parameter used to find an optimal compro-
mise between memory depth and resolution during training. The
order of Gamma memory (k) is product of (D) and (R). For a kth
order Gamma memory, the memory depth is approximated by
(k/l) and memory resolutions l. Lallahem and Mania (2003),
Coulibaly et al. (2000a, 2000b), Sajikumar and Thandaveswara
(1999) and Rummelhart et al. (1986) have successfully demon-
strated application of time-lagged recurrent models in hydrology.
2.1.3. Wavelet transformation (WT)
Mathematical transformations are often used to extract infor-
mation from a time series data which is not readily available in
its raw form. Wavelets are mathematical functions which give a
time-scale representation of the time series data and their relation-
ships. They are suitable for analyzing time series data that contain
non-stationaries. Wavelet transform (WT) is regarded as being
capable of revealing aspects of the original raw time series data
such as trends, breakdown points, and discontinuities that other
time series analysis techniques might miss (Adamowski and Sun,
2010; Singh, 2012). There are two types of the wavelet transforma-
tions; Continuous Wavelet Transformation (CWT) and Discrete
Wavelet Transformation (DWT). The Continuous Wavelet
Transform (CWT) of a time series f(t) is defined as follows:

x(t) x0(t)

G(z)

Memory
Hidden Output
Layer
X(n)
Layer Layer 1-

Z-1

x1(t)

X(n-1)
G(z)

1-

X(n-2)
Z-1

x2(t)

X(n-k) G(z)
(a)
1-

(b)
Z-1

xn(t)

Gamma Memory

Fig. 1. Schematic diagram of TLRNN with (a) TDL memory, (b) Gamma memory.
214 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225
Z þ1  
1 t
b
W a;b ðtÞ ¼ f ðtÞ pffiffiffi w dt

1 a a
where ⁄ refers to the complex conjugate and w(t) which is called the
wavelet function or the mother wavelet. The entire range of the sig-
nal/time series is analyzed by the wavelet function by using the
parameters ‘a’ and ‘b’. The parameter ‘a’ and ‘b’ are the dilation
(scale) and translation (position) parameters, respectively. The cal-
culation of CWT coefficients at each scale ‘a’ and translation ‘b’
results in a large amount of data. This problem was resolved in
the DWT which operates scaling (low pass filter) and wavelet (High
pass filter) functions. The DWT scales and positions are based on
power of two (dyadic scales and positions) and can be defined for
a discrete time series f(t);
Z j¼J
Wða; bÞD ¼ 2
j=2 w ð2
j=2
kÞf ðtÞdt
j¼1
where the real numbers, j and k are the integers which control the
wavelet dilation and translation, respectively. In WT, a time series
passes through a low pass filter and a high pass filter and produces
detail information d(n) and the approximation a(n), respectively.
The high pass filters are used to analyse the high frequencies while
the low pass filters are used to analyse the low frequency content of
the time series. The approximations are the high scale, low fre-
quency components of the signal while detail represents the low
scale, high frequency components. The low frequency content of
the time series data is the most significant part and it gives the sig-
nal its identity while on the other hand, the high frequency content
imparts flavour or nuance. The detail signals can catch trivial attri-
butes of interpretational value in data while the approximation
shows the background information of data (Nourani et al., 2009a;
Tiwari and Chatterjee, 2010). This decomposition process continues
until the desired level is achieved. The maximum possible number
of levels depends on the length of the signal/data. At each decompo-
sition level, the low pass and the high pass filters produces the sig-
nals spanning only half the frequency band. This doubles the
frequency resolution double as the uncertainty in the frequency is
reduced by half. The decomposition of temporal data by DWT satis-
fies the following conditions; f(t) = d1 + a1 = d1 + d2 + a2 = d1 + d2 +
d3 + a3 and so on. Further details on the DWT can be found in many
text books including Daubechies (1992) and Addison (2002).
2.2. Development of hybrid wavelet models
The static MLPNN and the dynamic TLNN models are integrated
with DWT in this study to develop hybrid wavelet models. DWT is
used to decompose the rainfall time series into sub-series of
approximation and details in order to reveal its temporal and the
spectral information. The hybrid wavelet multilayer perceptron
neural network (WMLPNN) and the hybrid wavelet time lagged
neural network (WTLNN) models are then developed using the
decomposed data as input and the observed runoff as the output.
The performance of the hybrid wavelet models are compared with
their respective simple multilayer perceptron neural network
(SMLPNN) and the simple time lagged neural network (STLNN)
models. A schematic diagram of hybrid wavelet model is presented
in Fig. 2. The SMLPNN and the STLNN models comprises of three
layers; input, hidden and output layer. The number of neurons in
the input and output layers are fixed as one for the simple models.
For the wavelet coupled models, the number of neurons in the
input layer is determined on the basis of the selected decomposi-
tion level. The selection of the number of neurons in hidden layer
is important for maximizing the ANN performance. The selection of
appropriate number of neurons in hidden layer in the present
study is determined by trial and error procedure. This is accom-
plished by calibrating the network and evaluating its performance
using an increasing number of hidden neurons. In order to obtain
ð4Þ
near maximum efficiency with as few neurons as necessary
(Hammerstrom, 1993). The sigmoid activation function is used as
the transfer function for the neurons of the hidden and the output
layer for the simple and wavelet coupled models. Calibration of the
simple and hybrid models in the present study is done by using the
Levenberg–Marquardt algorithm (LMA) because of its simplicity
(Adamowski and Sun, 2010). LMA is an iterative technique which
can be used of an objective function which in this study is
expressed as sum of squares of nonlinear functions. The stopping
criteria for the training of the developed models in the present
study is either a maximum of 100 epochs or training is set to ter-
minate when the mean squared error (MSE) of the cross validation
testing data set begins to increase. This is an indication that the
network has begun to over-train. Overtraining is when the network
ð5Þ simply memorizes the training set and is unable to generalize the
problem.
The calibrated network is then tested by presenting different set
of validation data which has not been used in calibration. The net-
work is re-calibrated by changing the number of neurons in the
hidden layer, if it failed to perform satisfactory during validation
phase. Testing of the network ensures that it has learned the gen-
eral patterns of the system and has not simply memorized a given
set of data.
2.3. Performance indices
In the present study, the performance of the developed models
is determined by using various statistical tests that define errors
associated with the models. The performance of the developed
models is assessed in terms of statistical measures of goodness of
fit. In the current study, two statistical indices, namely, the Root
Mean Squared Error (RMSE) and the Nash–Sutcliffe Efficiency
(NSE) Nash and Sutcliffe (1970) are used. More details on these sta-
tistical indices along with their equations can be found in Tayfur
(2012). The RMSE is used to measure the estimated output accu-
racy. The RMSE values range between zero and 1. An RMSE value
of zero indicates perfect match between the estimated and the
observed discharges. Similarly, an RMSE of 1 suggests that there
is no match between the estimated and the observed discharges.
Karunanithi et al. (1994) suggested that the RMSE value is a good
measure for indicating the goodness of fit for high flows. The NSE
is one of the most widely used criteria for assessment of the hydro-
logical models performance. The NSE provides a measure of ability
of the model to predict observed values. In general, high values of
NSE (up to 100%) and small values for RMSE indicate a good model.
These two statistical measures can be used to evaluate perfor-
mance of hydrological models satisfactorily (Legates and McCabe,
1999).
3. Data
The daily rainfall runoff data of two catchments located in dif-
ferent hydro-climatic conditions in the world is used in the present
study. The two catchments are the Baihe catchment located in
north eastern China and the Brosna catchment located in Ireland.
The Baihe catchment with an area of 61,780 km2 is a sub-basin
of Upper Hanjian River basin as shown in Fig. 3(a). The Hanjian
River is the largest tributary of the Yangtze River in China, covers
a total drainage area of approximately 151,000 km2 (34°300 –30°4
90 N, 106°140 –114°560 E) with a total length of 1577 km. The altitude
of the basin decreases from 3500 m in the northwest to 88 m at the
Danjiangkou reservoir in the southeast (Sun et al., 2014). The Han-
jian River is about 737 km long from northwest to Baihe station in
southeast. The average width of the main stream of Hanjiang is
 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225 215

Wavelet
an
Transformation
d1

Rainfall
MLP/ Runoff
Data d2 TLNN

dn

Fig. 2. Schematic diagram of hybrid wavelet models.

Fig. 3. Location of the study area (a) Baihe, (b) Brosna.

about 200–300 m with an average slope of 6% and surrounded by
high and steep mountains, narrow valleys and swift torrents. This
mountainous basin is semi-arid in nature, lies in a typical subtrop-
ical monsoon climate region with rainfall mainly during the sum-
mer and autumn seasons. Normally, the evaporation is greater
than rainfall from November to April and in August, whereas rain-
fall is less in the other months. The daily rainfall data for a period of
eight years starting from 1st January 1972 onwards from nineteen
rain gauge stations and three hydrological stations are used to
obtain daily average rainfall. The concurrent discharge data used
in the study is the daily averaged data measured at the Baihe sta-
tion. The data for a period of six years from 1972 to 1979 (2190
data points) is used for training the models, whereas the remaining
data for a period of two years from 1978 to 1979 (730 data points)
is used for validation purpose.
The Brosna catchment located in Ireland having drainage area of
1207 km2 up to the Ferbane flow gauging station is used in this
study. The catchment has a very flat topography except for
some undulations caused by glacial deposits. There is no noticeable
evidence of substantial groundwater movement across the
216 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225
topographical boundary of the catchment. Daily rainfall data for a
period for ten years (1st January 1969 to 31st December 1978) are
lumped by averaging the data from four rain gauge stations. The
concurrent discharge data used in the study is the daily averaged
data measured at the Ferbane station. The first eight years data
from 1969 to 1976 (2920 data points) is used for the training/
calibration purpose while remaining two years data from 1977 to
1978 (730 data points) is used for the testing purpose. The location
of the catchments along with rainfall gauging sites is shown in
Fig. 3(b).
The summary of statistical characteristics of the training and
the testing data are given in Table 1. Furthermore, the results of
cross correlation analysis between lagged rainfall data and dis-
charge is shown in Fig. 4 for both selected catchments of Baihe
and Brosna. It can be seen from Fig. 4 that the lag-3 day rainfall
time series r(t
3) has the maximum correlation with the
observed discharge.
4. Results and discussion
4.1. Selection of suitable wavelet function
In order to decompose the raw time series data into approxima-
tion and details using WT, a wavelet function is required. There are
different wavelet families available each containing different
wavelet functions such as Daubechies wavelet families contains
db1, db2, db3 and so on as the wavelet functions. These different
wavelet functions are categorized by their unique features includ-
ing the region of support and the number of vanishing moments.
The region of support of wavelet is related with the span length
of the wavelet which affects its feature localization possessions
of a signal while the vanishing moment confines the wavelet’s abil-
ity to signify polynomial behaviour or information in a signal. The
details of different wavelet families can be found in many text
books such as Daubechies (1992) and Addison (2002). This study
uses six most commonly used wavelet functions in hydrology are
used in this study. These wavelet functions includes the simple
wavelet functions Haar wavelet, the db4 and db8 wavelet functions
of most commonly wavelet family Daubeches, sym4 and sym8
wavelet functions with sharp peak of Symlet wavelet family and
discrete approximation of Meyer (dmey) wavelet function.
4.2. Selection of decomposition level
A DWT involves the decomposition of the raw time series data
into approximation and details. A DWT decomposition consist of
log2 (N) levels/stages at most where N is the total number of data
points. Selection of suitable decomposition level for the develop-
ment of hybrid wavelet models is considered as one of the impor-
tant factors to be considered. This decomposition associated with
the seasonal and periodic variations embedded in the hydrologic
data under consideration. Level one decomposition yield
Table 1
Statistical summary of data of test catchments.
Catchment Data type Calibration period
Maximum Average
(mm/day) (mm/day)
Baihe Rainfall 47.08 2.59
Evaporation 12.80 2.89
Discharge 28.25 1.04
Brosna Rainfall 32.67 2.20
Evaporation 9.80 1.31
Discharge 6.94 0.98
approximation at level one (a1) and one detail (d1), level two
decomposition gives approximation (a2) and two details (d1 and
d2), level three decomposition results in approximation (a3) and
three details (d1, d2 and d3) and so on. Level one detail sub-series
d1 describes the features of original data detectable at a scale of
up to two days. The d2 sub-series represent features of the raw data
measurable at a scale of 2–4 days. Likewise, d3 can detect features
on approximately weekly basis in the original data. Various meth-
ods have been applied in different studies to select suitable decom-
position level. Adamowski and Sun (2010), Partal and Kisßi (2007)
and Kisi and Shiri (2011) used trial and error method to select
decomposition level and employed level eight, level ten and level
three decomposition, respectively. Aussem et al. (1998), Nourani
et al. (2009a, 2009b), Tiwari and Chatterjee (2010) and
Adamowski and Chan (2011) used the following formula to select
suitable decomposition level:
L ¼ intðlogðNÞÞ ð6Þ
where L is the level and N is the total data points in data. Neverthe-
less, Nourani et al. (2011) and Moosavi et al. (2013) argued that this
equation was derived only for fully autoregressive data by consider-
ing the length of data only without giving consideration to the sea-
sonal signature of the hydrologic process. Shoaib et al. (2014a, b)
and Shoaib et al. (2015) favoured the use of level nine decomposi-
tion based on the results of regression analysis between wavelet
transformed rainfall data at each decomposition level and runoff
data. Therefore, a regression analysis between wavelet transformed
rainfall data and observed runoff is carried out in the present study
for the two selected catchments to determine the appropriate
decomposition level. For this purpose, lagged-1 rainfall data is
transformed for all possible decomposition levels (i.e. from level
one to level nine at most) using six selected wavelet functions. A
total of 108 decomposition levels are obtained for both selected
catchments. The results of the regression analysis are presented in
Fig. 5 for both selected catchments. It is found from Fig. 5 that coef-
ficient of determination R2 has the lowest value at the first decom-
position level. The value of R2 is found to increase with the increase
in decomposition level and the R2 value is found maximum at
decomposition level nine. Therefore, level nine decomposition is
also selected in the current study as it is giving best results for both
selected catchments. Level nine decomposition contains one large
scale sub-signal approximation (a9) and nine small scale sub-
signals detail (d1, d2, d3, d4, d5, d6, d7, d8 and d9). The detail sub-
series d4 corresponds to variation up to 16-day, d5 up to 32-day
(about month), d6 to 64-day, d7 to 128-day (about four month), d8
to 256-day (about eight and half month) and d9 to 512-day mode
(about seventeen month). Therefore, selected level nine decomposi-
tion holds d8 and d9 sub-signals which are liable for detecting sea-
sonal variations in the input rainfall data on nearly annual basis.
This annual periodicity is considered as a very important and
foremost seasonal cycle in the hydrologic time series data.
Validation period
Coefficient of variation Maximum Average Coefficient of variation
(mm/day) (mm/day)
2.17 79.98 2.48 2.53
0.80 8.10 2.53 0.71
1.89 22.66 0.78 2.23
1.63 27.56 2.47 1.51
1.04 6.90 1.32 1.04
0.83 6.62 1.22 0.86
 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225 217

0.80 0.80
Training
0.70 (a) Testing (b)
Training
0.70 Testing

Correlation Coefficient

Correlation Coefficient
0.60 0.60
0.50 0.50
0.40 0.40
0.30 0.30
0.20 0.20
0.10 0.10
0.00 0.00
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Lag (days) Lag (days)

Fig. 4. Cross-correlation results of rainfall–runoff data for (a) Baihe, (b) Brosna.

Fig. 5. Variation of regression R2 with decomposition level (a) Baihe, (b) Brosna.

4.3. Selection of memory depth
4.3.1. Static MLPNN model
In order to use temporal information contained in the time ser-
ies data with the static MLPNN model, a straight forward method
of implicit representation of time is to add a short-term memory
structure in the input layer of a static MLPNN in the form of lagged
rainfall data. The performance of the simple and hybrid wavelet
static MLPNN models is very much dependent on the selection of
appropriate number of lagged rainfall data series/memory depth
in the input vector. This is because the response of a hydrological
system is inherently dependent on their previous states, so the
use of past (time lagged) data is necessarily required in order to
encode temporal features of the input data. There are several
approaches for the selection of appropriate input vector for the
data driven models. One of the most common approach for the
selection of appropriate input vector is considering a sliding win-
dow of input sequences in the input vector. With this approach,
the input vector starts from containing only lagged-1 day time ser-
ies data in the input vector. The input vector is then modified by
successively adding one more lagged time series into it and this
continues up to a specific lag time. This approach has been succes-
sively applied in many hydrological studies including Furundzic
(1998), Tokar and Markus (2000), Riad et al. (2004), Chua et al.
(2008), Moosavi et al. (2013), Shoaib et al. (2015), etc. This specific
lagged time can be estimated either by trial and error or may be
selected from present time to the time where the lagged time data
is most correlated with the observed discharge/runoff. The other
common approach of selection of appropriate input vector is the
selection of most correlated lagged time series data in the input
vector only and neglecting the poorly lagged rainfall data series.
Nayak et al. (2005a, 2005b, 2007) has successfully applied this
approach of selection of appropriate input vectors for data driven
models. Another method of selection of appropriate input vector
for the ANN models was suggested by Sudheer et al. (2002). This
method involves the selection of input vector on the basis of statis-
tical properties of cross-correlation, auto-correlation and partial
correlation of the lagged input time series with the output time
series. However, some hydrological studies including Nayak et al.
(2004) and Senthil Kumar et al. (2005) criticizes the application
of statistical properties of cross-correlation, auto-correlation and
partial correlation for the selection of appropriate input vector
for ANN models as these properties considered a linear relation-
ship between the input and output variables and additional vari-
able effect to take account any non-linear residual dependencies
is not considered in this method. Moreover, Shoaib et al. (2014a,
2014b) compared in detail various approaches for the selection
of appropriate input vector and favoured the use of most common
approach for the wavelet coupled ANN models. Hence, the follow-
ing five input vectors are considered in the present study to predict
runoff at the present time using SMLPNN and WMLPNN models:
I1 r(t
1)
I2 r(t
1), r(t
2)
I3 r(t
1), t(
2), r(t
3)
I4 r(t
1), t(
2), r(t
3), r(t
4)
I5 r(t
1), r(t
2), r(t
3), r(t
4), t(t
5)
The first input vector I1 contains only lagged-1 day rainfall data
series r(t
1). The second input vector I2 is obtained by adding
218 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225

lagged-2 day rainfall data series r(t
2) in the first input vector and
then the external input vector is further modified by successively
adding one more lagged time rainfall data series into input vector
and this continues up to a 5-day lagged rainfall data series to
predict runoff at the current time step.
The performance of the static simple and the wavelet MLPNN
models in terms of performance indices of NSE (%) and the RMSE
for both selected catchments and for all selected six wavelet func-
tions are shown in Figs. 6 and 7. Fig. 6(a) and (b) illustrates the NSE
(%) and RMSE values of WMLPNN models for Baihe catchment with
input vector I1. It is evident from figure that WMLPNN models out-
performed their counterpart simple MLPNN models for all six
selected wavelet functions. The NSE and RMSE values are found
to be 4.5% and 1369 cumecs, respectively with input vector Ii for
the simple models. The WMLPNN models developed with wavelet
functions db8 and sym8 are found best among all the wavelet func-
tions and yielded NSE and RMSE values of about 50% and
900 cumecs, respectively with I1 input vector. For I2 input vector
containing both lagged one and two day rainfall data series in
the input vector, the NSE is increased to about 20% and RMSE value
is decreased to about to 1200 cumecs in case of the simple models
(not shown here).
The WMLPNN model developed in this study with the wavelet
function db8 is found to be the best among all WMLPNN models
developed and gives the NSE and RMSE values of about 82% and
400 cumecs, respectively. As shown in Fig. 6(c) and (d), the NSE
(%) value is further increased from 20% with I2 input vector to
about 55% with I3 input vector while RMSE value is found to be
decreased from 1200 cumecs with I2 to 800 cumecs with I3 input
vector. For input vector I3, the performance of WMLPNN models
developed with all selected wavelet functions except Haar wavelet
is almost found equal with NSE and RMSE values of about 80% and
400 cumecs, respectively. For input vector I4 and I5, the NSE values
are further increased to 72% and 75%, respectively for the simple
MLPNN models while RMSE values are reduced to 700 cumecs with
I4 and 600 cumecs with I5 input vector (not shown here). However,
no significant improvement can be seen for WMLPNN models with
input vector I4 and I5 relative to input vector I3. For input vector I3,
I4 and I5, the NSE and RMSE are found about 80% and 400 cumecs,
respectively as shown in Fig. 6(c)–(f).
The NSE and RMSE values of simple and wavelet coupled mod-
els for the Brosna catchment with all five input vectors is repre-
sented in Fig. 7. It is evident from the figure that the simple
MLPNN model yielded NSE value of about 6% with input vector
I1, 12% with I2, 19% with I3, 25% with I4 and 30% with I5 input vector
containing one to five lagged rainfall data series to predict runoff at
the current time step as shown in Fig. 7(a), (c) and (e), respectively.
A RMSE value of about 10 cumecs is found for simple models with
all five input vectors considered without showing any significant
improvement as shown in Fig. 7(b), (d) and (f). In the case of the
WMLPNN models, the wavelet coupled MLPNN model with db8
wavelet function is found the best among all the wavelet coupled
models for input vector I1 with NSE value of about 40% and RMSE
value of about 8 cumecs as shown in Fig. 7(a) and (b), respectively.
For input vector I2, the WMLPNN models developed with wavelet
functions db8 and sym4 are found best and yielded the NSE and
RMSE values of about 52% and 8 cumecs, respectively (not shown
here). For input vector I3 and I4, the wavelet models developed
with wavelet functions db8, sym4, sym8 and dmey wavelet func-
tions are the best models with NSE and RMSE values of about
60% and 7 cumecs, respectively. However, the wavelet coupled
model developed with wavelet function db8 is found to be best
among all wavelet coupled models for input vector I5 with NSE
and RMSE values of about 60% and 7 cumecs, respectively as shown
in Fig. 7(e) and (f). It can be concluded on the basis of above
analysis that the db8 wavelet function is the only function which

100 (a) I1 Training 1400 (b) I1 Training Testing

Testing 1200
80
1000
NSE (%)

60
RMSE

800
40 600
400
20
200
0 0
Simple Haar db4 db8 sym4 sym8 dmey Simple Haar db4 db8 sym4 sym8 dmey

100 (c) I3 Training Testing 1400 (d) I3 Training

Testing
1200
80
1000
NSE (%)

RMSE

60 800
600
40
400
20 200
0
0 Simple Haar db4 db8 sym4 sym8 dmey
Simple Haar db4 db8 sym4 sym8 dmey

100 (e) I5 1400 (f) I5 Training

Training Testing
1200 Testing
80
1000
NSE (%)

60
RMSE

800
40 600
400
20
200
0 0
Simple Haar db4 db8 sym4 sym8 dmey Simple Haar db4 db8 sym4 sym8 dmey

Fig. 6. Performance of static MLPNN models for the Baihe catchment.

 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225 219

100 (a) I1 Training 15 (b) I1 Training Testing

80 Testing
10

RMSE
NSE (%)
60
40
5
20
0 0
Simple Haar db4 db8 sym4 sym8 dmey Simple Haar db4 db8 sym4 sym8 dmey

100 16 (d) I3
(c) I3 Training Testing Training Testing
14
80 12
NSE (%)

10

RMSE
60
8
40 6
4
20
2
0 0
Simple Haar db4 db8 sym4 sym8 dmey Simple Haar db4 db8 sym4 sym8 dmey

100 (e) I5 Training

Training Testing
15 (f) I5
Testing
80
10

RMSE
NSE (%)

60

40
5
20

0 0
Simple Haar db4 db8 sym4 sym8 dmey Simple Haar db4 db8 sym4 sym8 dmey

Fig. 7. Performance of static MLPNN models for the Brosna catchment.

performed well for all the selected input vectors considered and for
both selected catchments of Baihe and Brosna in this study during
training as well as testing phases.
4.3.2. Dynamic TLRNN model
Time lagged recurrent neural networks (TLRNN) are the
extended form of static MLPNN models with short term memory
structure and recurrent connections. The input layer used the
inputs lagged by multiple time steps before presented to the net-
work and thus forms the short term memory. There are several
memory structures to choose such as TDL, Gamma and Laguarre.
As stated earlier in Section 2.1.2, the Gamma memory is selected
for the present study. It is very critical to determine how much
short memory is required to describe the temporal pattern of the
data in case of TLRNN. This is because as it is not known a priori
where, in time, the relevant information is. It is the depth (D)
parameter which determines the depth of the memory. The depth
parameter (D) of the Gamma memory is used to compute the num-
ber of taps to be used for storing past information. As the short
term memory of one to five day is considered in the development
of static MLPNN models where input vector I1 and I5 corresponds
to the memory of one day and five days, respectively. Ideally, the
same memory should be used for the TLRNN network models.
However, the same memory cannot be employed for the TLRNN
models developed using Gamma memory, as the number of taps
in case of Gamma memory is calculated by the formula T = 2D/3.
This formula yields a memory of 0.67, 1.34, 2, 2.67, 3.34, and 4 days
with D taken as one, two, three, four, five and six, respectively.
Therefore, the value of D in the present study is varied from one
to six in the development of simple and hybrid wavelet TLRNN
models, so that its performance can be compared with the static
MLPNN models with a memory of one to five days.
Typically, the TLRNN has three layers, namely, the input layer,
the hidden layer and the output layer with feedback connection
from the hidden layer back to the input layer. Training of the
TLRNN is accomplished with the back-propagation through time
with trajectory learning factors. The simple time lagged recurrent
neural network (STLRNN) models are developed with rainfall r(t)
and runoff q(t) at the current time as the input and desired target,
respectively. The hybrid wavelet time lagged recurrent neural net-
work (WTLRNN) are developed using discrete wavelet transformed
input rainfall data at current time as input and observed runoff q(t)
as the desired target. The level nine decomposition is also
employed for wavelet transformation of the input rainfall data r
(t) which contains one approximation and nine details sub-signal
of the data.
Results of the STLRNN and WTLRNN models for Baihe and
Brosna catchments are presented in Figs. 8 and 9, respectively for
different values of D ranging from 1 to 6. It was found that the
STLRNN models are the least performing models. The STLRNN
model yielded NSE and RMSE values of about 2% and 1200, respec-
tively for D = 1 and D = 2 which refers to the tap of about 0.67–
1.34 days. Furthermore, the WTLRNN models developed with
wavelet functions db8 and sym4 are the best performing models
for D equals to one and two. The WTLRNN models yielded the
NSE and RMSE values of about 50% and 1000 cumecs, respectively.
The results for D = 1 are presented in Fig. 8(a) and (b) while results
for D = 2 are not shown here. It is further found that the perfor-
mance of simple models is increased exponentially when D is set
as three which corresponds to the tap of two days. It yielded the
NSE and RMSE values of about 60% and 800 cumecs, respectively
against the 75% NSE value and 600 cumecs RMSE values of best
WTLRNN model with db8 wavelet function. The NSE value of
TLRNN is further increased to about 69% and 77% with D set as
three and four, respectively. The corresponding RMSE values of
the simple models are found as about 750 and 600 cumecs. The
WTLRNN model developed with db8 wavelet function is again
found to be best performing wavelet models by yielding NSE and
RMSE values of about 80% and 550 cumecs, respectively with D
set as four. The results for D = 3 are presented in Fig. 8(c) and (d)
while results for D = 4 are not presented here. The performance
of all simple and wavelet models is found to be similar for D is
220 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225

100 1400 (b) D = 1 Training

(a) D = 1 Testing
1200
80

NSE (%)
Training Testing 1000

RMSE
60 800
40 600
400
20
200
0 0
Simple Haar db4 db8 sym4 sym8 dmey Simple Haar db4 db8 sym4 sym8 dmey

100 Training 1400 (d) D = 3 Training

(c) D = 3
Testing 1200 Testing
80
1000
NSE (%)

RMSE
60 800
40 600
400
20
200
0 0
Simple Haar db4 db8 sym4 sym8 dmey Simple Haar db4 db8 sym4 sym8 dmey

100 Training (e) D = 6 1400 (f) D = 6 Training

90 Testing 1200 Testing
80
70 1000
NSE (%)

RMSE
60 800
50
40 600
30 400
20
200
10
0 0
Simple Haar db4 db8 sym4 sym8 dmey Simple Haar db4 db8 sym4 sym8 dmey

Fig. 8. Performance of TLRNN models of Baihe catchment.

100 Training
(a) D = 1 16 (b) D = 1 Training
Testing 14 Testing
80 12
NSE (%)

RMSE

60 10
8
40 6
4
20
2
0 0
Simple Haar db4 db8 sym4 sym8 dmey Simple Haar db4 db8 sym4 sym8 dmey

100 Training 16 (d) D = 3 Training

Testing (c) D = 3 14 Testing
80 12
NSE (%)

RMSE

60 10
8
40 6
4
20
2
0 0
Simple Haar db4 db8 sym4 sym8 dmey Simple Haar db4 db8 sym4 sym8 dmey

100 Training (e) D = 6 16

Testing (f) D = 6 Training
80 14 Testing
12
NSE (%)

RMSE

60 10
8
40 6
20 4
2
0 0
Simple Haar db4 db8 sym4 sym8 dmey Simple Haar db4 db8 sym4 sym8 dmey

Fig. 9. Performance of TLRNN models of Brosna catchment.

taken as five and six. The results for D = 6 are presented in Fig. 8
(e) and (f) while results for D = 5 are not presented here. It is evi-
dent from the above analysis for the Baihe catchment that D has
an important role in the performance of TLRNN models. Further-
more, the performance of simple models is found least with D
ranging from one to two compared to their counterpart wavelet
coupled models. However, the difference between performance
of simple and wavelet models in terms of NSE (%) and RMSE values
is very small for D ranging from three to six.
In addition, the performance of wavelet coupled models is
found to be insensitive of wavelet function with higher values
of D (five and six) as all the wavelet functions are found to be
 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225
performing equally well. This may be due to the fact that the larger
memory depths contained maximum temporal information and
therefore be retrieved well by the all the wavelet functions includ-
ing the simpler Haar wavelet function as well. Fig. 9(a)–(f) shows
the performance of simple and wavelet coupled TLRNN models
for the Brosna catchment with different values of D ranging from
one to six. The performance of simple TLRNN models for D from
one to two almost remains same in terms of NSE (%) and RMSE val-
ues. The simple TLRNN models yields NSE and RMSE values of
about 3% and 11 cumecs, respectively. The WMLPNN models
developed with the db8 wavelet function is found as the best
wavelet coupled models for D values one and two. The best wave-
let coupled models yields the NSE and RMSE values of about 45%
and 9 cumecs, respectively. The results for D = 1 are shown in
Fig. 9(a) and (b) while no results are presented here for D = 2. Sim-
ilarly, the best WMLPNN model yields about 50% NSE value and
8 cumecs, respectively. The results for D = 1 are shown in Fig. 9
(a) and (b) while no results are presented here for D = 2. Similar
to the Baihe catchment, the NSE values of simple model is expo-
nentially increase to about 40% with D as three and four, to about
45% with D as five and to about 48% with D set as six as shown in
Fig. 9(c) and (e). Similarly, the RMSE value decreases from
10 cumecs with D set as two to about 8 cumecs with D set as six
for the simple models. The wavelet coupled TLRNN models devel-
oped with wavelet functions db4 and db8 are found the best mod-
els among WTLNN models developed with other four wavelet
functions with D set as three and four. The best WTLRNN models
yields NSE value of about 65% and 70% for D value set as three
and four respectively. Likewise, the corresponding RMSE values
of about 7 cumecs is found for the best WTLRNN models. The
results for D = 3 are shown in Fig. 9(c) and (d) while no results
are presented here for D = 4. Similar to Baihe catchment, the
WTLRNN models developed with selected six wavelet functions
are found performing equally good for D value set as five and six
as shown in Fig. 9(e) and (f) for D = 6.
4.4. Discussion
The performance of the best wavelet coupled static MLPNN and
the dynamic TLRNN models are compared in this section with dif-
ferent memory depths. From the above analysis shown, the db8
wavelet function is found to produce the beset results for small
as well as for large memory depths, so wavelet coupled models
developed with db8 wavelet function are considered for this com-
parison. The best performance of the db8 wavelet function may be
due to the point that by decomposing a time series data up to max-
imum level, the information embedded in the data can be revealed
well by a composite wavelet function such as db8.
Moreover, time series data mostly have high frequencies occurs
only for small period, while low frequencies spans almost whole
range of the signal/temporal data. Consequently, the wavelets hav-
ing higher support length are able of capturing the low frequencies
more precisely. The db8 wavelet performed well compared to
other wavelets used in the present study as it is having reasonable
support width and also contains good time–frequency localization
property. All of these characteristics together allow the hybrid
wavelet models developed with the db8 wavelet function to cap-
ture both the underlying trend as well as the short term variability
in the time series data.
Figs. 10 and 11 represent performance of different models for
the Baihe and the Brosna catchments, respectively considering dif-
ferent memory depths. Fig. 10(a) and (b) shows the performance of
static wavelet coupled MLPNN models while Fig. 10(c) and (d) rep-
resents the performance of TLRNN models with different memory
information. It is evident from Fig. 10(a) and (b) that performance
in terms of NSE (%) and RMSE of the wavelet coupled static MLPNN
221
model increases drastically (increase in NSE (%) value and decrease
in RMSE value) from I1 with only one day lagged rainfall data series
to I2 input vector with lagged one and lagged two rainfall data ser-
ies and thereafter it almost remains constant with I3, I4 and I5 input
vector. The addition of one more lagged data series in I2 and there-
after did not improve the performance significantly. This may be
attributed to the fact that maximum temporal and spectral infor-
mation contained in the data has been presented to the network
by the lagged one and lagged two wavelet transformed rainfall
data series. The addition of lagged three, four and five day input
rainfall data did not provided any further information to the net-
work and thus resulted in no significant improvement. However,
for the simple static MLPNN model which uses lagged rainfall data
in its raw form in the input vector behaved entirely different than
its counterpart wavelet coupled model. It is evident from Fig. 10
(a) and (b) that the performance of the simple model kept on
increasing from I1 to I2 then I3 and up to I4 and then it remains
almost constant. This may be due to the fact that the rainfall data
in its raw form did not contain too much temporal information.
That is why with the addition of one more lagged time data series
in the input vector, the performance of the simple models
improved significantly and this continues up to I4 input vector
which contains lagged one to lagged four rainfall data series in
the input vector. With further addition of one more lagged rainfall
data series in the input vector I4, no significant improvement can
be seen. This means maximum temporal information contained
in the rainfall data has been presented to the network. As for as
the performance of wavelet coupled TLRNN models is concerned
with different memory depths, the results are presented in
Fig. 10(c) and (d). It is evident from the figures that the memory
depth has affected the performance significantly. The trend is
almost similar for both simple and wavelet coupled models. It
can be seen from the figures that the performance of both the sim-
ple and wavelet coupled models increases drastically from mem-
ory depth two to memory depth three. The memory depth two
and three refers to the memory of about 1.5 and 2 days, respec-
tively. The performance of the wavelet coupled TLRNN model is
consistent with the static MLPNN model as both are giving maxi-
mum performance at the memory depth of two days. Both the sim-
ple and wavelet coupled models are found to perform best with
memory depth of five which corresponds to the about three days
and thereafter it starts decreasing. It can be further seen from
Fig. 10 that the performance of wavelet coupled MLPNN models
outperformed their counterpart simple MLPNN models with input
vector I1, I2 and I3. However, with input vectors I4 and I5 which con-
tains up to lagged four and lagged five rainfall data series, respec-
tively, the performance of simple and wavelet coupled MLPNN
models are not too much significantly different. But wavelet couple
TLRNN models only outperformed the simple TLRNN models only
with D set as one and two. With D taken as three up to six, the both
the models are almost following the similar trend.
Fig. 11 shows the impact of memory depth on the performance
of the static simple and wavelet coupled MLPNN and dynamic
TLRNN models. Fig. 11(a) and (b) presents the performance of sta-
tic simple and wavelet coupled models while Fig. 11(c) and (d)
shows the results for the TLRNN models. It is evident from
Fig. 11(a) and (b) that the performance of the wavelet coupled
models outperformed their counterpart simple MLPNN models
for all the input vectors considered in the study. However, the
wavelet coupled TLRNN models only outperformed their counter-
part simple models with memory depth of one and two only. With
D taken as three up to six the performance of simple TLRNN models
is not significantly different with their counterpart wavelet cou-
pled models. It can, therefore, be concluded on the basis of above
analysis that wavelet coupled static MLPNN and dynamic TLRNN
models performed better relative to their counterpart simple
222 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225

100 1400
(a) (b)
90
1200
80

RMSE (cumecs)
Simple
70 1000
Wavelet

NSE (%)
60 800
50
Simple 600
40
Wavelet
30 400
20
200
10
0 0
I1 I2 I3 I4 I5 I1 I2 I3 I4 I5

100 1400
(c) (d) Simple
90
1200
80 Wavelet
70 1000

NSE (%)
NSE (%)

60 800
Simple
50
Wavelet 600
40
30 400
20
200
10
0 0
D1 D2 D3 D4 D5 D6 D1 D2 D3 D4 D5 D6

Fig. 10. Impact of memory depth on the performance of models for Baihe catchment.

100 20
90
(a) Simple
18
(b) Simple
Wavelet Wavelet
RMSE (cumecs)

80 16
70 14
NSE (%)

60 12
50 10
40 8
30 6
20 4
10 2
0 0
I1 I2 I3 I4 I5 I1 I2 I3 I4 I5

100 16
90
(c) (d)
14
80
12
70
NSE (%)

NSE (%)

60 10
50 8
40 6
Simple
30 Simple
Wavelet 4
20 Wavelet
10 2
0 0
D1 D2 D3 D4 D5 D6 D1 D2 D3 D4 D5 D6

Fig. 11. Impact of memory depth on the performance of models for Brosna catchment.

models. However, the performance of the simple TLRNN models
with memory depth of about 1.5 days is not significantly lesser
than their counterpart wavelet coupled models in contrast with
the simple MLPNN models. Furthermore, the performance of the
wavelet coupled MLPNN and TLRNN models is almost similar for
both the selected catchments in contrast to their simple counter-
part models. This may be due to the fact that the wavelet trans-
formed data contains temporal and spectral information which
when presented to any network type either static or recurrent
enhances the performance significantly which is not the case of
simple models which make use of rainfall input data in its raw
form.
Finally, the flow duration curves (FDCs) of the observed and
best models are presented in Fig. 12 in order to check the ability
of the best models to simulate the high, low and medium trails
of the observed hydrograph for the both selected catchments. From
the FDC, the 10 percentile flow (the flow that is equalled or
exceeded 10% of the period of record) can be considered as high
flow percentile threshold. Similarly, 11 to 89 percentile flow is con-
sidered as medium flow range while 90 percentile flows are con-
sidered as the low flow threshold. The medium flow percentile
can be further divided in to high medium flow percentile and
low medium flow percentile from 11 to 49 percentile and 50 to
89 percentile flows respectively. Fig. 12(a) and (b) shows the FDCs
of static MLPNN and dynamic TLRNN models, respectively for the
Baihe catchment while 12(c) and (d) presents the FDCs for the cor-
responding models of Brosna catchment. It is evident from Fig. 12
(a) that the simple model with input vector I2 containing lagged
one and lagged two rainfall data series) can only captures the high
flow trails of the observed hydrograph while it over estimates dur-
ing the medium and low flow trails of the observed hydrograph.
The performance of the corresponding wavelet model is much bet-
ter than its counterpart simple model as it well trails the high,
medium and low flow trails of the observed hydrograph. It can
be further seen that the performance of the simple and wavelet
coupled models with I5 input vector (containing from lagged
 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225 223

100000 100000
(a) Observed (b)
Observed
Discharge (cumecs) 10000 Simple I2

Discharge (cumecs)
Simple_D2
Simple I5 10000 Simple_D5
Wavelet I2
Wavelet_D2
1000 Wavelet I5
Wavelet_D5

1000
100

10 100
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Percent of time indicated flow was equalled or exceeded Percent of time indicated flow was equalled or exceeded

100 100
(c) Observed (d) Observed
Simple I2 Simple_D6
Wavelet I2 Wavelet_D6
Discahrge (cumecs)

Discahrge (cumecs)
Simple I5 Simple_D2
Wavelet I5 Wavelet_D2

10 10

1 1
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Percent of time indicated flow was equalled or exceeded Percent of time indicated flow was equalled or exceeded

Fig. 12. Flow duration curves.

rainfall data series to lagged five day rainfall data series) in trailing  The dynamic TLRNN models are found to perform equally well
low, medium and high flows of the observed hydrograph is almost compared to the commonly used static MLPNN models and
similar and these modes are equally good as the wavelet coupled hence can be considered as an alternative to the MLPNN
models with input vector I2. Almost similar trend can be found models.
for the dynamic TLRRNN models for the Baihe catchment as shown  The performance of the wavelet coupled MLPNN models is
in Fig. 12(b). Wavelet coupled TLRNN models with input vector I2 found sensitive to the selection of wavelet function. The db8
input vector can be considered as the best model. It is apparent is found best among the six selected wavelet functions. How-
from Fig. 12(c) and (d) for the Brosna catchment that the simple ever, performance of the dynamic TLRNN models is only found
and wavelet models (both static and dynamic models) with input sensitive to the selection of the wavelet function with less
vector I2 and I5 captures the features of the observed hydrograph memory depth and again the db8 wavelet function is found to
well only for the high flow and medium high flows. The wavelet yield the best results. With large memory depth, the perfor-
coupled models outperformed their counterpart simple models mance of the all wavelet coupled TLRNN models is found almost
for the medium low and low flow trails of the observed hydro- similar with all six wavelet functions even with the simple Haar
graphs shown in Fig. 12(c). For the TLRNN models, the simple mod- wavelet function. It can, therefore, be concluded to use the db8
els with D take as 2 only describes the average behaviour of the wavelet function with static MLPNN models and also for
hydrograph without giving any information rearing low, medium dynamic TLRNN models with smaller memory depth. However,
and high flow trails of the observed hydrograph. The performance any wavelet function can be used with dynamic TLRNN models
of the wavelet coupled TLRNN model with D taken as 2 is much with larger memory depth.
better than its counterpart simple mode in trailing low, medium  The performance of the wavelet coupled MLPNN and TLRNN
and high flows of the observed hydrographs. It is further evident models is equally well for all input vector/memory depths.
from Fig. 12(d) that the simple TLRNN model with D taken as 6 However, the simple TLRNN models outperformed their coun-
is unable to trail the medium low and low flow periods of the terpart simple MLPNN models with small memory depth.
observed hydrograph. The counterpart wavelet coupled TLRNN
model is found to be the best in trailing all the features of the
observed hydrograph. It can be concluded on the basis of above
analysis of FDCs that the wavelet coupled models can well captures References
the features of the observed hydrograph compared with their out-
Abrahart, R.J., Anctil, F., Coulibaly, P., Dawson, C.W., Mount, N.J., See, L.M.,
performed simple models and this is true for both static MLPNN Shamseldin, A.Y., Solomatine, D.P., Toth, E., Wilby, R.L., 2012. Two decades of
and dynamic TLRNN models. anarchy? Emerging themes and outstanding challenges for neural network river
forecasting. Prog. Phys. Geogr. 36 (4), 480–513.
Adamowski, J., Chan, H.F., 2011. A wavelet neural network conjunction model for
groundwater level forecasting. J. Hydrol. 407 (1–4), 28–40.
5. Conclusions Adamowski, J., Sun, K., 2010. Development of a coupled wavelet transform and
neural network method for flow forecasting of non-perennial rivers in semi-arid
watersheds. J. Hydrol. 390 (1–2), 85–91.
The following conclusions can be drawn from the present Addison, P.S., 2002. The Illustrated Wavelet Transform Handbook. Institute of
study: Physics Publishing, London.
224 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225
Akiner, M.E., Akkoyunlu, A., 2012. Modeling and forecasting river flow rate from the
Melen Watershed, Turkey. J. Hydrol. 456, 121–129.
Altunkaynak, A., Nigussie, T.A., 2015. Prediction of daily rainfall by a hybrid
wavelet-season-neuro technique. J. Hydrol. 529, 287–301.
Anmala, J., Zhang, B., Govindaraju, R.S., 2000. Comparison of ANNs and empirical
approaches for predicting watershed runoff. J. Water Resour. Plann. Manage.
126 (3), 156–166.
Antar, M.A., Elassiouti, I., Allam, M.N., 2006. Rainfall-runoff modelling using
artificial neural networks technique: a Blue Nile catchment case study.
Hydrol. Process. 20 (5), 1201–1216.
Arsenault, R., Gatien, P., Renaud, B., Brissette, F., Martel, J.-L., 2015. A comparative
analysis of 9 multi-model averaging approaches in hydrological continuous
streamflow simulation. J. Hydrol. 529, 754–767.
ASCE Task Committee on Application of Artificial Neural Networks in Hydrology,
2000a. Artificial neural networks in hydrology. I: preliminary concepts. J.
Hydrol. Eng., 115–123 http://dx.doi.org/10.1061/(ASCE)1084-0699(2000)5:2
(115).
ASCE Task Committee on Application of Artificial Neural Networks in Hydrology,
2000b. Artificial neural networks in hydrology II: hydrologic applications. J.
Hydrol. Eng., 124–137 http://dx.doi.org/10.1061/(ASCE)1084-0699(2000)5:2
(124).
Assaad, M., Boné, R., Cardot, H., 2005. Study of the behavior of a new boosting
algorithm for recurrent neural networks. Artificial Neural Networks: Formal
Models and their Applications – ICANN 2005. Springer, pp. 169–174.
Aussem, A., Campbell, J., Murtagh, F., 1998. Wavelet-based feature extraction and
decomposition strategies for financial forecasting. J. Comput. Intell. Finance 6
(2), 5–12.
Aziz, K., Rahman, A., Fang, G., Shrestha, S., 2014. Application of artificial neural
networks in regional flood frequency analysis: a case study for Australia. Stoch.
Env. Res. Risk A. 28 (3), 541–554.
Badjate, S.L., Dudul, S.V., 2009. Multi step ahead prediction of north and south
hemisphere sun spots chaotic time series using focused time lagged recurrent
neural network model. WSEAS Trans. Inform. Sci. Appl. 6 (4), 684–693.
Cannas, B., Fanni, A., See, L., Sias, G., 2006. Data preprocessing for river flow
forecasting using neural networks: wavelet transforms and data partitioning.
Phys. Chem. Earth, Parts A/B/C 31 (18), 1164–1171.
Chang, L.-C., Chen, P.-A., Chang, F.-J., 2012. Reinforced two-step-ahead weight
adjustment technique for online training of recurrent neural networks. IEEE
Trans. Neural Netw. Learn. Syst. 23 (8), 1269–1278.
Charaniya, N.A., Dudul, S.V., 2013. Time Lag recurrent Neural Network model for
Rainfall prediction using El Niño indices. Int. J. Sci. Res. Publ. (IJSRP) 3 (1), 367.
Chiang, Y.-M., Chang, L.-C., Chang, F.-J., 2004. Comparison of static-feedforward and
dynamic-feedback neural networks for rainfall–runoff modeling. J. Hydrol. 290
(3), 297–311.
Chua, L.H.C., Wong, T.S.W., Sriramula, L.K., 2008. Comparison between kinematic
wave and artificial neural network models in event-based runoff simulation for
an overland plane. J. Hydrol. 357 (3), 337–348.
Coulibaly, P., Anctil, F., Bobee, B., 2000a. Daily reservoir inflow forecasting using
artificial neural networks with stopped training approach. J. Hydrol. 230 (3),
244–257.
Coulibaly, P., Anctil, F., Bobée, B., 2000b. Neural network-based long-term
hydropower forecasting system. Comput.-Aided Civil Infrastruct. Eng. 15 (5),
355–364.
Coulibaly, P., Baldwin, C.K., 2005. Nonstationary hydrological time series forecasting
using nonlinear dynamic methods. J. Hydrol. 307 (1), 164–174.
Coulibaly, P., Evora, N.D., 2007. Comparison of neural network methods for infilling
missing daily weather records. J. Hydrol. 341 (1), 27–41.
Daubechies, I., 1992. Ten Lectures on Wavelets (CBMS-NSF Regional Conference
Series in Applied Mathematics). Society for Industrial and Applied Mathematics.
Dawson, C.W., Harpham, C., Wilby, R.L., Chen, Y., 2002. Evaluation of artificial neural
network techniques for flow forecasting in the River Yangtze, China. Hydrol.
Earth Syst. Sci. Discuss. 6 (4), 619–626.
French, M.N., Krajewski, W.F., Cuykendall, R.R., 1992. Rainfall forecasting in space
and time using a neural network. J. Hydrol. 137 (1), 1–31.
Furundzic, D., 1998. Application example of neural networks for time series
analysis: rainfall–runoff modeling. Signal Process. 64, 383–396.
Güldal, V., Tongal, H., 2010. Comparison of recurrent neural network, adaptive
neuro-fuzzy inference system and stochastic models in Eğirdir Lake level
forecasting. Water Resour. Manage. 24 (1), 105–128.
Hammerstrom, D., 1993. Working with neural networks. IEEE Spectrum 30 (7), 46–
53.
Hopfield, J.J., 1982. Neural networks and physical systems with emergent collective
computational abilities. Proc. Natl. Acad. Sci. USA 79 (8), 2554–2558.
Jain, A., Sudheer, K.P., Srinivasulu, S., 2004. Identification of physical processes
inherent in artificial neural network rainfall runoff models. Hydrol. Process. 18
(3), 571–581.
Kale, S.N., Dudul, S.V., 2009. Intelligent noise removal from EMG signal using
focused time-lagged recurrent neural network. Appl. Comput. Intell. Soft
Comput., 1.
Karunanithi, N., Grenney, W.J., Whitley, D., Bovee, K., 1994. Neural networks for
river flow prediction. J. Comput. Civil Eng. 8 (2), 201–220.
Kisi, O., 2011. Wavelet regression model as an alternative to neural networks for
river stage forecasting. Water Resour. Manage. 25 (2), 579–600.
Kisi, O., Shiri, J., 2011. Precipitation forecasting using wavelet-genetic programming
and wavelet-neuro-fuzzy conjunction models. Water Resour. Manage. 25 (13),
3135–3152.
Kisi, O., Shiri, J., Tombul, M., 2013. Modeling rainfall–runoff process using soft
computing techniques. Comput. Geosci. 51, 108–117.
Kote, A.S., Jothiprakash, V., 2008. Reservoir inflow prediction using time lagged
recurrent neural networks. In: First International Conference on Emerging
Trends in Engineering and Technology, 2008. ICETET’08. IEEE.
Lallahem, S., Mania, J., 2003. A nonlinear rainfall–runoff model using neural
network technique: example in fractured porous media. Math. Comput. Modell.
37 (9), 1047–1061.
Legates, D.R., McCabe, G.J., 1999. Evaluating the use of ‘‘goodness-of-fit” measures
in hydrologic and hydroclimatic model validation. Water Resour. Res. 35 (1),
233–241.
Lohani, A.K., Kumar, R., Singh, R.D., 2012. Hydrological time series modeling: a
comparison between adaptive neuro-fuzzy, neural network and autoregressive
techniques. J. Hydrol. 442, 23–35.
Ma, X.-X., Chen, X., Guo, H.-F., 2008. Reservoir annual runoff ANFIS forecast model
based on wavelet analysis. China Rural Water Hydropower 7, 005.
Maier, H.R., Dandy, G.C., 2000. Neural networks for the prediction and forecasting of
water resources variables: a review of modelling issues and applications.
Environ. Modell. Softw. 15 (1), 101–124.
Mekanik, F., Imteaz, M.A., Gato-Trinidad, S., Elmahdi, A., 2013. Multiple regression
and Artificial Neural Network for long-term rainfall forecasting using large scale
climate modes. J. Hydrol. 503, 11–21.
Moosavi, V., Vafakhah, M., Shirmohammadi, B., Behnia, N., 2013. A wavelet-ANFIS
hybrid model for groundwater level forecasting for different prediction periods.
Water Resour. Manage., 1–21
Motter, M., Principe, J.C., 1994. A gamma memory neural network for system
identification. In: Neural Networks. IEEE International Conference on IEEE
World Congress on Computational Intelligence. IEEE.
Muluye, G.Y., 2011. Improving long-range hydrological forecasts with extended
Kalman filters. Hydrol. Sci. J. 56 (7), 1118–1128.
Mulvany, T.J., 1850. On the use of self-registering rain and flood gauges. Making
Observations of the Relations of Rain Fall and Flood Discharges in a Given
Catchment. Transactions and Minutes of the Proceedings of the Institute of Civil
Engineers of Ireland, Dublin, Ireland, Session 1.
Nash, J.E., Sutcliffe, J.V., 1970. River flow forecasting through conceptual models.
Part 1: A discussion of principles. J. Hydrol. 10 (3), 282–290.
Nason, G.P., Von Sachs, R., 1999. Wavelets in time-series analysis. Philos. T. Roy. Soc.
A 357 (1760), 2511–2526.
Nayak, P.C., Sudheer, K.P., Ramasastri, K.S., 2005. Fuzzy computing based rainfall–
runoff model for real time flood forecasting. Hydrol. Process. 19 (4), 955–968.
Nayak, P.C., Sudheer, K.P., Rangan, D.M., Ramasastri, K.S., 2004. A neuro-fuzzy
computing technique for modeling hydrological time series. J. Hydrol. 291 (1–
2), 52–66.
Nayak, P.C., Sudheer, K.P., Rangan, D.M., Ramasastri, K.S., 2005. Short-term flood
forecasting with a neurofuzzy model. Water Resour. Res. 41 (4).
Nayak, P.C., Sudheer, K.P., Jain, S.K., 2007. Rainfall–runoff modeling through hybrid
intelligent system. Water Resour. Res. 43.
Nourani, V., Alami, M.T., Aminfar, M.H., 2009a. A combined neural-wavelet model
for prediction of Ligvanchai watershed precipitation. Eng. Appl. Artif. Intell. 22
(3), 466–472.
Nourani, V., Kisi, Ö., Komasi, M., 2011. Two hybrid artificial intelligence approaches
for modeling rainfall–runoff process. J. Hydrol. 402 (1), 41–59.
Nourani, V., Komasi, M., Mano, A., 2009b. A multivariate ANN-wavelet approach for
rainfall–runoff modeling. Water Resour. Manage. 23 (14), 2877–2894.
Partal, T., Kisßi, Ö., 2007. Wavelet and neuro-fuzzy conjunction model for
precipitation forecasting. J. Hydrol. 342 (1), 199–212.
Piotrowski, A.P., Napiorkowski, M.J., Napiorkowski, J.J., Osuch, M., 2015. Comparing
various artificial neural network types for water temperature prediction in
rivers. J. Hydrol. 529, 302–315.
Riad, S., Mania, J., Bouchaou, L., Najjar, Y., 2004. Predicting catchment flow in a
semi-arid region via an artificial neural network technique. Hydrol. Process. 18
(13), 2387–2393.
Rummelhart, D.E, McClelland, J.L, P.R. Group, 1986. Parallel Distributed Processing.
Explorations in the Microstructure of Cognition. Foundations, vol. 1. The MIT
Press.
Saharia, M., Bhattacharjya, R.K., 2012. Geomorphology-based time-lagged recurrent
neural networks for runoff forecasting. KSCE J. Civil Eng. 16 (5), 862–869.
Sajikumar, N., Thandaveswara, B.S., 1999. A non-linear rainfall–runoff model using
an artificial neural network. J. Hydrol. 216 (1), 32–55.
Senthil Kumar, A.R., Sudheer, K.P., Jain, S.K., Agarwal, P.K., 2005. Rainfall–runoff
modelling using artificial neural networks: comparison of network types.
Hydrol. Process. 19 (6), 1277–1291.
Serpen, G., Xu, Y., 2003. Simultaneous recurrent neural network trained with non-
recurrent backpropagation algorithm for static optimisation. Neural Comput.
Appl. 12 (1), 1–9.
Shamseldin, A.Y., 1997. Application of a neural network technique to rainfall–runoff
modelling. J. Hydrol. 199 (3–4), 272–294.
Shoaib, M., Shamseldin, A.Y., Melville, B.W., 2014a. Comparative study of different
wavelet based neural network models for rainfall–runoff modeling. J. Hydrol.
515, 47–58.
Shoaib, M., Shamseldin, A.Y., Melville, B.W., Khan, M.M., 2014b. Hybrid wavelet
neuro-fuzzy approach for rainfall–runoff modeling. J. Comput. Civil Eng. 30 (1),
04014125.
Shoaib, M., Shamseldin, A.Y., Melville, B.W., Khan, M.M., 2015. Runoff forecasting
using hybrid Wavelet Gene Expression Programming (WGEP) approach. J.
Hydrol. 527, 326–344.
 M. Shoaib et al. / Journal of Hydrology 535 (2016) 211–225
Singh, R., 2012. Wavelet-ANN model for flood events. In: Deep, K., Nagar, A., Pant,
M., Bansal, J.C. (Eds.), Proceedings of the International Conference on Soft
Computing for Problem Solving (SocProS 2011) December 20–22, 2011, vol.
131. Springer, Berlin/Heidelberg, pp. 165–175.
Sudheer, K., Gosain, A., Ramasastri, K., 2002. A data-driven algorithm for
constructing artificial neural network rainfall–runoff models. Hydrol. Process.
16 (6), 1325–1330.
Sun, Y., Tian, F., Yang, L., Hu, H., 2014. Exploring the spatial variability of
contributions from climate variation and change in catchment properties to
streamflow decrease in a mesoscale basin by three different methods. J. Hydrol.
508, 170–180.
Tayfur, G., 2012. Soft Computing in Water Resources Engineering: Artificial Neural
Networks, Fuzzy Logic and Genetic Algorithms. WIT Press.
Tayfur, G., Guldal, W., 2006. Artificial neural networks for estimating daily total
suspended sediment in natural streams. Nord. Hydrol. 37, 69–79.
225
Tayfur, G., Zucco, G., Brocca, L., Moramarco, T., 2014. Coupling soil moisture and
precipitation observations for predicting hourly runoff at small catchment
scale. J. Hydrol. 510, 363–371.
Tiwari, M.K., Chatterjee, C., 2010. Development of an accurate and reliable hourly
flood forecasting model using wavelet–bootstrap–ANN (WBANN) hybrid
approach. J. Hydrol. 394 (3–4), 458–470.
Tokar, A.S., Markus, M., 2000. Precipitation-runoff modeling using artificial neural
networks and conceptual models. J. Hydrol. Eng. 5 (2), 156–161.
Wang, W., Ding, J., 2003. Wavelet network model and its application to the
prediction of hydrology. Nat. Sci. 1 (1), 67–71.
Warner, B., Misra, M., 1996. Understanding neural networks as statistical tools. Am.
Stat. 50 (4), 284–293.