August 2012, P.P. 673 685
www.cafetinnova.org
! "!
$
*
)+ , $,
)
%& ' (
+-
.
!
! "#
#
)
,
/0+
%
)0 ,
$
%
& )*)+,)
( -#
.
12*
&# '
.(
,
(
(
3
4 This paper deals with the prediction of hydrologic behavior of the runoff for the one of the largest
discharge carrier International River, Brahmaputra, located in Assam (India) at the Pandu station, by using daily
time unit. The flow regime dominated by high data non stationary and seasonal irregularity due to Himalayan
climate fallout. The influence of data preprocessing through wavelet transforms has been investigated. For this, the
main time series of flow data were decomposed to multiresolution time series using discrete wavelet
transformations. Then these decomposed data were used as input to Artificial Neural Network (ANN) for multiple
lead time flow forecasting. Various types of wavelets were used to evaluate the optimal performance of models
developed. The forecasting accuracy of the models has been tested for multiple lead time upto 4 days using different
decomposition levels. The performance of the proposed hybrid model has been evaluated based on the performance
indices such as root mean square error (RMSE), coefficient of efficiency (CE) and mean relative error (MRE).The
results shows the better forecasting accuracy by the proposed combined hybrid model over the single ANN model
in hydrological time series forecasting.
/ "#
5
%
& %
0
%
4
Hydrology system is influenced by many factors such as
weather, land cover, infiltration, evapotranspiration, so
it includes the good deal of stochastic dependent
component, multi time scale and highly non linear
characteristics. Hydrologic time series are often non
linear and non stationary. Inspite of high flexibility of
Artificial Neural Network (ANN) in modelling
hydrologic time series, sometimes signals are highly
non stationary and exhibits seasonal irregularity. In
such situation, ANN may not be able to cope with non
stationary data if pre processing of input and/or output
data is not performed (Cannas et al.,2006). Pre
processing data referes to analysing and transforming
input and output variables in order to detect trends,
minimise noise, underline important relationship and
flatten the variables distribution in a time series. These
analysis and transformations help the model learn
relevant patterns. Data Pre processing techniques,
which facilitates stabilisation of the mean and variance,
and seasonality removal, are often applied to remove
non stationarity in data used to build soft computing
models.
Most of the researchers using ANN prefer using Feed
Forward Back Propagation (FFBP) algorithm. But
major disadvantages of FFBP are slow convergence,
often trapped in local minima, weak extrapolator and
$$ 1
2
.
unable to cope up with high non linearity and non
stationary data series. To minimize these limitations,
Wavelet transformation attached with ANN as data
preprocessing technique can be adopted to achieve
higher forecasting accuracy and consistency in multi
lead time ahead (Deka and Prahlada,2012). non
stationary time series can be decomposed into certain
number of stationary time series by wavelet transform
(WT). Then different single prediction methods are
combined with wavelet transform to improve the
prediction accuracy. In most of the hybrid models, WT
is used as data preprocessing technique. The wavelet
transformed data aid in improving the model
performance by capturing helpful information on
various resolution levels. Due the above mentioned
advantages of WT, it has been found that the
hybridization of wavelet transformation with other
models like ANN, FL, ANFIS, linear models, etc.,
improved the results significantly than the single regular
model (Prahlada and Deka, 2011).
Wavelet theory (Mallat, 1989) is first developed in the
end of 1980s of last century. Now a days, it has been
applied in many fields, such as signal process, image
compression, voice code, pattern recognition,
hydrology, earthquake investigation and many other
non linear science fields.
#02050405 Copyright ©2012 CAFET INNOVA TECHNICAL SOCIETY. All rights reserved.
PARESH CHANDRA DEKA, LATIFA HAQUE and ANIRUDDHA GOPAL BANHATTI
Wensheng and Ding (2003), carried out a multi time
scale prediction of ground water level at Beijing and
daily discharge of Yangte River Basin at China using
Hybrid Model of Wavelet Neural Network. The
proposed WLNN model focused on improving the
precision and prolonging the forecasting time period.
Kim and Valdes (2003) developed nonlinear model for
drought forecasting based on a conjunction of wavelet
transforms and neural networks in the Conchos river
basin in Maxico. The results indicate that the
conjunction model using dyadic wavelet transform
significantly improves the ability of neural network in
forecasting.B. Cannas et al. (2005) modelled the river
flow forecasting one month ahead with Neural
Networks and Wavelet Analysis using monthly runoff
data for the Tirso Basin, Italy. For the non stationary
and seasonal irregularity of runoff time series, the best
results were obtained using data clustering and discrete
Wavelet Transform combination. Tests showed that
neural networks trained with pre processed data showed
better performance.
Zhou et al(2008)developed monthly discharge
predictor corrector
model
based
on
wavelet
decomposition using 52 years records of monthly
discharge at Yichang station of Yangtse river The
decomposed times series data were used as input to
ARMA model for prediction which improves the
prediction accuracy. Y.R. Satyaji Rao et al. (2009)
carried out modelling using Hydrological Time Series
data adopting Wavelet Neural Network for four west
flowing rivers in India .The results of daily Streamflow
and monthly Groundwater level series modelling
indicated that the performances of WNN Models are
more effective than ANN Models. A rainfall runoff
modelling using Wavelet ANN approach was carried
out by Nourani et al., (2009), for predictions of runoff
discharge one day ahead of the Ligvanchai watershed at
Tabriz, Iran. The daily rainfall and runoff time series for
21 years were used. The time series were decomposed
upto four levels by using Haar, Daubechies (db2),
Symlet (sym3) and Coiflet (coif1). The Study showed
that both short and long term runoff discharges could be
predicted considerably with higher merit of Haar
wavelet in comparison with the others. They also
recommended that Wavelet Transform could be used for
trend analysis in watersheds.
Another Intermittant streamflow forecasting model was
developed by Kisi,O. (2009)called neurowavelet (NW)
model combining two methods discrete wavelet
transform (DWT) and artificial neural network (ANN),
for 1 day ahead forecasting. The new subtime series
having high correlation coefficient were used as input to
the ANN model. The NW model was found to be much
better than the ANN in high flow estimation. The test
results showed that the DWT could significantly
6
increase the accuracy of the ANN model in modelling
intermittent streamflows.
Many works related to sediment load estimation were
alsoreported in the literature using wavelet
transformations (T. Rajaee et al.,2010, Kisi,O.2010,
Rajaee, T., and et al. 2011) Shiri and Kisi, (2010),
examined the performance of short term and long term
streamflow forecasting using a wavelet and neuro fuzzy
conjunction model to investigate the daily, monthly and
yearly streamflow of Derecikviran station on Filyos
River in the Western Black Sea region of Turkey using
31 years of streamflow data. The results obtained
showed that the neuro fuzzy (NF) and wavelet neuro
fuzzy (WNF) models increased the accuracy of the
single NF models especially in forecasting yearly
streamflow.
The effects of various decomposition level of wavelet in
flow forecasting performance was investigated by
Adamowski, J., and Sun, K. (2010) in non perennial
rivers in semi arid watersheds at lead times of 1 and 3
days for two different rivers in Cyprus. The discrete
trous wavelet transform was used to decompose flow
time series data into 8 levels wavelet coefficients which
are used as inputs to Levenberg Marquardt artificial
neural network models to forecast flow. WA ANN
model provided more accurate results than regular
ANN. Further,wavelet combined with Genetic
programming and ANFIS were also reported in very
recent literature in streamflow forecasting and rainfall
runoff modeling(Kisi and Shiri,2011; Nourani
et
al.,2011).
In fact, in general, and in Brahmaputra basin in
particular, runoff time series consists of high non
linearity and non stationary, and neural network models
may not be able to cope with these two different aspects
if no pre processing of the input and/or output data is
performed. In this study, wavelet transforms and neural
networks have been applied to predict the hydrologic
behavior of the runoff for the Brahmaputra basin located
in Assam (India) at Pandu (Guwahati) station by using
daily flow data. Wavelet analysis is used to pre process
the data to be fed to a traditional multilayer perception
(MLP) neural network.
7
%&
%
&
4
The Brahmaputra is the fourth largest river in the world
in terms of average discharge at mouth, with a flow of
19,830 cumec. The hydrologic regime of the river
responds to the seasonal rhythm of the monsoons and to
the freeze thaw cycle of the Himalayan snow. The rainy
season (May to October) accounts for 82% of the mean
annual flow at Pandu. The station Pandu is located near
Guwahati city of Assam, India. The discharge is highly
fluctuating in nature. Discharge per unit drainage area in
International Journal of Earth Sciences and Engineering
ISSN 0974 5904, Vol. 05, No. 04, August 2012, pp. 673 685
Discrete Wavelet Ann Approach in Time Series Flow Forecasting
A Case Study of Brahmaputra River
6
the Brahmaputra Basin River is among the highest of
major rivers of the world. At Pandu, the Brahmaputra
yields 0.0306 cumecs per sq.km and the mean annual
flood discharge is 51,156 Cumecs. The maximum
discharge being 61,000 Cumecs and the minimum being
2432 Cumecs for Pandu Station.The study area (station
Pandu) is shown in figure 1 below.
8 9
%
#& %
%4
2. Root Mean Square Error,
$
∑ (4 − 3 )
2
=1
=
$
3. Mean relative error, (%)
Most of the previous studies were carried out using
selective type of wavelet attached with one wavelet for
the same data point. Also the optimal level of
decomposition in a particular type of wavelet is the
difficulty in proper combination selection. Again the
forecasting accuracy for higher lead time is not well
documented so far. Keeping all this point of view, in
this paper an attempt has been made to investigate the
potential and applicability of Hybrid Model by
combining Wavelet and ANN with objectives to address
the above mentioned scenarios using time series data.
Development of various combined models for selective
input scenarios for Multistep Lead Time and their
performance evaluation has also been carried out.
=
1
$
$
∑
4 −3
4
=1
× 100
Where, X=observed values, Y=predicted values, N =
total number of values, and x 5 4 4
The objectives of this paper are to examine how
successfully Wavelet Transform have been used in
hydrologic problem. The focus of this work for
hydrologists in practicing field.
In this study, Discrete Wavelet Transform and Neural
Networks has been applied for multistep lead time
predictions. Daily discharge data is collected for 20
years for the period from 1980 1999. Statistical analysis
and Data Normalization of data has been done to
analyze the pattern and the type of data. Out of total
available data,2/3 of data set used for training and 1/3 of
data set used for testing in MATLAB PACKAGE using
ANN and Wavelet Tool.
The data statistics for both training and testing are given
below which contains minimum, maximum, mean,
standard deviation (Sd) and skewness coefficient. It is
obvious from the table that the extreme values of the
available data are in the training set. When classifying
the data into training and testing subsets, it is essential
to check that the data represents the same statistical
population. It was earlier observed that high skewness
coefficient has a considerable negative effect on ANN
performance. The table 1 shows the statistical behavior
of the observed available time series flow data which
reveals the large variation in data at the station Pandu.
Prediction is made using 1 day, 2 days, 3 days and four
days lead time. The Performance of ANN and Hybrid
Wavelet ANN Model is analyzed by the performance
indices like Coefficient of Efficiency (CE), Root Mean
Squared Error (RMSE), and Mean Relative Error
(MRE).
* '
%
4
1. Coefficient of Efficiency,
#
85
'
67
& &
8
" :4
Artificial Neural Networks are mathematical inventions
inspired by observations made in the biological systems.
ANN has gained popularity among Hydrologist in
recent decades due to its large array of application in the
field of Engineering and research. The purpose of ANN
is mapping function i.e., mapping an input space to an
output space.
ANN has excellent flexibility and high efficiency in
dealing with nonlinear and noisy data in Hydrological
modeling. Some of the advantages of using ANN Tool
are Input Output mapping, Self adaptive, Real Time
Operation, Fault Tolerance and Pattern Recognition.
ANN is a massively parallel distributed processor made
International Journal of Earth Sciences and Engineering
ISSN 0974 5904, Vol. 05, No. 04, August 2012, pp. 673 685
PARESH CHANDRA DEKA, LATIFA HAQUE and ANIRUDDHA GOPAL BANHATTI
up of simple processing units, which has a natural
prosperity for strong experimental knowledge and
making it available for use. A typical ANN consists of a
number of nodes that are organized according to a
particular arrangement. It consists of “Neurons” which
are interconnected computational elements that are
arranged in a number of layers which can be single or
multiple. Each pair of neurons is linked and is
associated with weights.ANN are trained by adjusting
the values of these connection weights between network
elements. The weighted inputs in each layer are
processed from neurons in the previous layer and
transmit its output to neurons in the next layer.
A transfer function/activation function is used to
convert a weighted function of input to get the output.
Usually non linear sigmoidal activation functions are
2
*
9
%
$==
9
9 =
2
87
4
The term wavelet means small wave. The smallness
refers to the condition that the function is of finite
length. The wave refers to the condition that it is
oscillatory. The term mother implies that the functions
used in the transformation process are derived from one
main function, the mother wavelet. A mother wavelet
that is continuously differentiable with compactly
supported scaling function and high vanishing moments.
875
'
used as reported in the literature which was also adopted
in this study. The inputs to the ANN model were
normalized and kept within the range of 0.1 to 0.9.
The learning rate and momentum coefficient are
influential parameters which controls the convergence
rate but to optimize them for best output .Here,both the
parameters were kept with constant 0.4 and 0.6
respectively throughout the network structure for
various number of hidden neurons. The basic structure
of ANN is shown in figure 2 below.
In this study,various ANN models were developed using
various input combinations for multiple leadtime.The
input combinations were consists of lagged data upto
past four days.The output was kept as single flow
variable for multiple leadtime.
%
#
;$
<
17520.11
11011.62
0.59
2432
61015
4
More, technically, a wavelet is a mathematical function
used to divide a given function or continuous time
signal into different scale components. The wavelet
6 6
#
#
;$
<
18897.3
10306.13
0.64
5539
54100
;$
<
17904.94
10836.02
0.59
2432
61015
$$
transform of a signal is capable of providing time and
frequency information simultaneously, hence providing
a time frequency representation of the signal. Usually
one can assign a frequency range to each scale
component. Each scale component can then be studied
with a resolution that matches its scale. To do this, the
data series is broken down by the transformation into its
“wavelets”, that are “scaled” and “shifted” version of
the mother wavelet (Nason and Von Sachs, 1999).
A Wavelet analysis is a set of building blocks to build
or represent a signal or function. Wavelet analysis has
become a common tool for analyzing localized
variations of power within a time series. Wavelet
analysis gives information both time and frequency
International Journal of Earth Sciences and Engineering
ISSN 0974 5904, Vol. 05, No. 04, August 2012, pp. 673 685
Discrete Wavelet Ann Approach in Time Series Flow Forecasting
A Case Study of Brahmaputra River
6
domain of the signal. Wavelet transformed data of
original time series improves the ability of a predicting
model by capturing useful information on various
resolution levels (Kim and Valdes 2003).The signal and
corresponding wavelet is shown in figure 3 for better
understanding of concept.
&
1
'
;$
<4
The Continuous Wavelet Transform (CWT) of a signal
x(t) is defined as follows:
CWTxψ (τ,s) =
(1)
Where‘s’ is the scale parameter, ‘Ƭ’ is the translation
parameter and the ‘*’ denote the complex conjugate.
Here, the concept of frequency is replaced by that of
scale, determined by the factor ‘s’
ψ
(2)
is the transforming function and it is called mother
wavelet.
'
;
<4
The Discrete Wavelet Transform (DWT) allows one to
reduce the computation time and it is considerably
simpler to implement than CWT. High pass and low
pass filters of different cutoff frequencies are used to
analyze the signal separating the signal at different
scales. The signal is passed through a series of high pass
filters to analyze the high frequencies, and it is passed
through a series of low pass filters to analyze the low
frequencies.
y[n] = (x*g) [n] =
ylow[n] =
yhigh[n] =
..(4)
..(5)
The time series after wavelet decomposition allows one
to have a look at the signal frequency at different scales.
The discrete wavelet transform allows reducing
computation time than CWT. High pass and low pass
filters of different cutoff frequencies are used to
separate the signal at different scales. The scale is
changed by upscaling and downscaling operations
(Cannas et.al, 2005).
877$
878
The DWT of a signal x is calculated by passing it
through a series of filters. First the samples are passed
through a low pass filter with impulse response g
resulting in a convolution of the two.
(3)
0 /
The time series is decomposed into one containing its
trend (the approximation) and one containing the high
frequencies and the fast events (the detail). The filtering
procedure is repeated every time some portion of the
signal corresponding to some frequencies is removed
and the outputs giving the detail coefficients (from the
high pass filter) and approximation coefficients (from
the low pass) are obtained depending on the chosen
decomposition level. It is important that the two filters
are related to each other and they are known as a
quadrature mirror filter. Figure 4 shows the low scale
and high scale concept of signal and wavelet.
The resolution of the signal, which is a measure of the
amount of detail information in the signal, is changed by
the filtering operations, and the scale is changed by
upsampling
and
downsampling
(subsampling)
operations. Subsampling a signal corresponds to
reducing the sampling rate, or removing some of the
samples of the signal. Upsampling a signal corresponds
to increasing the sampling rate of a signal by adding
new samples to the signal.
The decomposition process can be iterated, with
successive approximations being decomposed in turn,
so that one signal is broken down into many lower
resolution components. This is called the wavelet
decomposition tree (Figure 5).
"&
&
International Journal of Earth Sciences and Engineering
ISSN 0974 5904, Vol. 05, No. 04, August 2012, pp. 673 685
PARESH CHANDRA DEKA, LATIFA HAQUE and ANIRUDDHA GOPAL BANHATTI
1
Most of the wavelets used in discrete wavelet transform
are fractal in nature. They are expressed in terms of a
recurrence relation so that to see them we must do
several iterations. Wavelets are mathematical functions
that were developed by scientists working in several
different fields for the purpose of sorting data by
frequency. Translated data can then be sorted at a
resolution which matches its scale. Studying data at
different levels allows for the development of a more
complete picture. Both small features and large features
are discernable because they are studied separately.
Unlike the discrete cosine transform, the wavelet
transform is not Fourier based and therefore wavelets
do a better job of handling discontinuities in data.The
Different Wavelets used are Haar, Daubechies wavelets,
Coiflet which were described briefly in the following
section4
3
4
The Daubechies wavelets are a family of orthogonal
wavelets defining a discrete wavelet transform and
characterized by a maximal number of vanishing
moments for some given support. With each wavelet
type of this class, there is a scaling function (also called
father wavelet) which generates an orthogonal
multiresolution analysis. The Daubechies wavelets have
surprising features such as intimate connections with
the theory of fractals. Daubechies wavelets are widely
used in solving a broad range of problems, e.g. self
similarity properties of a signal or fractal problems,
signal discontinuities, etc. The Daubechies wavelets are
not defined in terms of the resulting scaling and wavelet
functions; in fact, they are not possible to write down in
closed form. This wavelet type has balanced frequency
responses but non linear phase responses. Daubechies
wavelets use overlapping windows, so the high
frequency coefficient spectrum reflects all high
6 >
%
frequency changes. Therefore Daubechies wavelets are
useful in compression and noise removal of audio signal
processing.
,
4
The Haar wavelet operates on data by calculating the
sums and differences of adjacent elements. The Haar
wavelet operates first on adjacent horizontal elements
and then on adjacent vertical elements. After each
transform is performed the size of the square which
contains the most important information is reduced by a
factor of 4. The next step in the image compression
process is quantization. In Haar wavelet, the basic
functions are scaled and translated versions of a "mother
wavelet” ψ(t).
The Haar wavelet transform has a number of advantages
such as it is conceptually simple, fast, memory efficient,
since it can be calculated in place without a temporary
Array and it is exactly reversible without the edge
effects that are a problem with other.
$ '
4
Coiflets are discrete wavelets designed by Ingrid
Daubechies, at the request of Ronald Coifman, to have
scaling functions with vanishing moments. The wavelet
is near symmetric, their wavelet functions have N / 3
vanishing moments and scaling functions N / 3 − 1, and
has been used in many applications using Calderón
Zygmund Operators. Both the scaling function (low
pass filter) and the wavelet function (High Pass Filter)
must be normalized by a factor 1/ .The wavelet
coefficients are derived by reversing the order of the
scaling function coefficients and then reversing the sign
of every second one.The various type of wavelet used in
the study are shown in figure 6.
International Journal of Earth Sciences and Engineering
ISSN 0974 5904, Vol. 05, No. 04, August 2012, pp. 673 685
Discrete Wavelet Ann Approach in Time Series Flow Forecasting
A Case Study of Brahmaputra River
6
%#
,&3 %
" :4
The hybrid model consists of a three layer feed forward
perceptron structure so that the first layer is the wavelet
neurons unit with the inputs runoff time series sub
signals obtained via a wavelet transform. In the
proposed approach, the runoff signals are firstly
decomposed into sub signals with different scales .i.e. a
large scale sub signal and several small scale sub
signals in order to obtain temporal characteristics of the
input time series. For a given time series, the time series
corresponding to a(t)(i.e. Qa(t)) is approximation sub
signal(large scale) of the original signal and jth detailed
sub signal(small scale) is identified by j(i.e. Qj(t))
where j are decomposition levels of the runoff(Q(t))
time series.
The wavelet decomposition of non stationary time
series into different scales provides an interpretation of
the series structure and extracts the significant
information about its history, using few coefficients.
%
$
Lag time flow data
#
These reasons make this technique attractive to
researchers for time series analysis of non stationary
signals (Nason and Von Sachs, 1999).
This study deals with some irregular mother wavelets
such as Haar,db 2(Daubechies wavelet of order2), and
coif1 which are illustrated in figure 6.The network
architecture that yielded the best results in terms of
determination coefficient and root mean square error on
the training and verifying steps may be determined
through trial and error process. The time series data
before going through the network are usually
normalized between 0 and 1 as sigmoidal activation
function was used in the study.
)
%
4
In the first stage, a multilayer perceptron (MLP) feed
forward ANN model with raw data was used to model
the daily flow at multistep lead time. Here, testing
results are shown in table 2.
&
2
23
3
1
0
'
%
1 day ahead
2 days ahead
3 days ahead
4 days ahead
Each MLP was trained with 2 to 15 hidden neurons in a
single hidden layer and scheme of Levenberg
Marquardt back propagation was used as the training
algorithm. No significant improvements in the model
performance were observed when the number of hidden
neurons was increased from a threshold. The training
was stopped at the stage where the error in the
validation data set begins to rise. Hence, overfitting the
training data is avoided and the model generalization
capacity was kept intact. At this stage, model efficiency
[CE] reveals unsatisfactory performance. This may be
12
13
14
13
%
)9 +
; 8? <
$+
9)+ @
1584
2213
2691
3460
0.87
0.85
0.83
0.78
4.34
6.39
8.16
8.62
due to significant fluctuations of the data around the
mean value.
In the second step, preprocessed data were fed to the
ANN model to improve the model accuracy using the
discrete wavelet transform. Here, the influence of used
mother wavelet type as well as decomposition level on
the model efficiency has been investigated. The time
series were decomposed to various levels for three kinds
of wavelet transforms such as Haar, db 2 and Coif1
.The results of this study are presented in table 3 for all
the cases.
International Journal of Earth Sciences and Engineering
ISSN 0974 5904, Vol. 05, No. 04, August 2012, pp. 673 685
6>
PARESH CHANDRA DEKA, LATIFA HAQUE and ANIRUDDHA GOPAL BANHATTI
%
&
9
1
$
0
;) 9
. %
&
1 day ahead
2day ahead
3 day ahead
4 day ahead
5
899.88
503.49
8 >>
8>> >6
867.43
726.87
869.88
860.23
8
876.09
>68
1135.47
DB 2
COIF1
HAAR
DB 2
COIF1
HAAR
DB 2
COIF1
HAAR
DB 2
COIF1
HAAR
7
853.85
895.45
> > 5
861.04
907.09
>77 86
896.28
870.52
8 5 6
1062.21
914.01
>7 68
For multilevel sub signals, which are used as input
neurons, the attached weights by ANN will be different
at various decomposition levels. Usually higher weights
are applied to the worthy level of the data or signals.
The validation time series of level 6 decomposition by
db 2 wavelet are shown in figure 6 after reconstructed
via trained ANN. Analyzing the table 3 results, it is
clearly observed that in validation phase, level 1,2,3,4,5
and 6 decomposition give relatively similar performance
but difficult to draw a conclusive inference. Increasing
the lead time, model efficiency slightly decreased due to
higher level lead to large number of parameters in
complex non linearity for ANN. Each parameters
produces error in forecasting data and net errors leads to
decreasing performance. This is not too much
2
9 %
ANN
WLNN
&
NON LINEAR
HYBRID
7
5% &
%
1584.42
309.88
The combined model performing better in lead time
than single ANN model as observed from the table
4.Figure 7 below shows how the original signal is
reconstructed and the formation of Daubacheis wavelet
tree. Also, various decomposition levels of db 2 is
shown in figure 8 with details and approximations.
(Here x axis is time in days, y axis is flow in cumecs)
Where s is full signal(flow);a1,a2,a3,a4,a5,a6 are
approximations;d1,d2,d3,d4,d5,d6 are details
The testing performance results of MRE for both ANN
and combined model are also shown in figure 9 to
visualise the error variations for different leadtime
8
861.74
55
527.13
883.14
956.50
>6
861.13
843.85
890.91
914.09
928.27
> 5 8
+<
906.52
891.62
5 55
>58
934.71
408.16
1406.98
866.58
>6 5
8> 8
570.24
863.26
871.53
505.62
867.02
895.75
>5 5
8 5
466.56
862.91
886.96
881.81
>
5
6
880.56
87
494.98
895.75
938.74
>> >>
895.18
845.93
> 8
457.04
65 7
836.78
significant in all the levels which may be considered as
optimal decomposition level for the data set. The
decomposition level more than 6 leads to low efficiency
as not included in the result table.
As observed in the data, there are sharp upward and
downward jumps in the time series because of several
factors such as rainfall, landslides, hydraulic failure.
The structure of DB 2 wavelet, which is similar to
signal, could capture the signal features such as peaks
can be able to deliver high efficiency. This phenomenon
may guide the selection of wavelet type.
To evaluate the ability of proposed model, various
comparisons have been made as presented in table 4.
9
0
%
# )9 +;
<
7% &
%
8% &
%
2213
2691
408.16
441.52
% &
%
3460
461.92
forecast. It can be confirmed that WLNN model
outperformed ANN model for all the leadtime forecast
considered in the study.
The performance of hybrid model for high peak flow as
well as low flow are presented for one day ahead
forecasting in figure 10 for the year 1999.The proposed
model output is in close agreement with observed flow
.The ANN model also performs in a similar way with
hybrid model for one day ahead flow forecasting. But
for the higher lead time such as 4th day lead time,
hybrid model performance is observed better than ANN
model results as shown in figure 11.
International Journal of Earth Sciences and Engineering
ISSN 0974 5904, Vol. 05, No. 04, August 2012, pp. 673 685
6>5
Discrete Wavelet Ann Approach in Time Series Flow Forecasting
A Case Study of Brahmaputra River
&
1
%
International Journal of Earth Sciences and Engineering
ISSN 0974 5904, Vol. 05, No. 04, August 2012, pp. 673 685
6>7
PARESH CHANDRA DEKA, LATIFA HAQUE and ANIRUDDHA GOPAL BANHATTI
,
6"
&
/
6"
;
:
# #
<
,
/
0
=
>
)
:
8
International Journal of Earth Sciences and Engineering
ISSN 0974 5904, Vol. 05, No. 04, August 2012, pp. 673 685
:
8
6>8
Discrete Wavelet Ann Approach in Time Series Flow Forecasting
A Case Study of Brahmaputra River
7
9
efficiency in validation set because of the time series
may have different long, intermediate and short trends.
The hybrid model involved in effective training of time
series data but also considers the influence of each sub
signal by magnifying its weights relatively (Noorani
et.al, 2009).
From the results, it is observed that the accuracy for
hybrid model is improved and the better estimation of
peaks reflects in the figure 10 and figure 11.It is to be
mentioned that capturing the flood peaks along with the
transition is very important in any river flood modeling.
The major difference of these two methods is related to
training procedure. The process leads to decrease model
7
0
%
&
<
#
6#
<???8
International Journal of Earth Sciences and Engineering
ISSN 0974 5904, Vol. 05, No. 04, August 2012, pp. 673 685
PARESH CHANDRA DEKA, LATIFA HAQUE and ANIRUDDHA GOPAL BANHATTI
7
6 $
4
In this study, wavelet transform was used to decompose
the runoff time series of Brahmaputra river basin so as
to provide multiscale features of data or signal. These
decomposed signals (sub signals) were used as input to
ANN model to forecast runoff for multistep lead time.
This proposed model has shown much improvement in
forecasting than single ANN model which used
undecomposed data as inputs. The proposed model has
shown the capacity to simulate non linear behavior of
the phenomenon more accurately than single ANN
model.
Analyzing the data preprocessing using wavelet
transformation showed that the best results were
obtained with DWT, at level 2 3 to predict runoff values
as output. For multiple lead time forecasting, the
proposed model outperformed single ANN model as
lead time increases. The trend, fluctuations, seasonality,
jumps components in the time series understood in a
better way by the model due to the wavelet
transformation. Overall, these results reveal the
potentiality of combining data clustering and use of
DWT in runoff forecasting.
Further, the effect of wavelet type on model
performance was investigated using three different types
of wavelet transforms. The results show the high quality
performance of db 2 in comparing with others.
Normally, runoff time series peaks which is similar to
single peaked db 2 wavelet provides better
>
#
@
& 6#
6>
<???8
approximation of sharp events exists in daily discharge
records.
Also, the increase in decomposition level may not be
efficient in forecasting ability. Hence; an optimum level
may be selected considering length of the data.
In order to complete current study, it is suggested to use
present strategy for other stations in the basin using
weekly and monthly flow data as different signal
characteristics and also including other time series
variables like rainfall, temperature, evapotranspiration.
Due to wavelet capabilities, the trend of basin
hydrological behavior can be analyzed more effectively
for which long term quality and continuous data will be
required.
:
" %#
4
The authors greatly acknowledged the Water resources
deptt. Govt.of Assam, India for providing necessary
data for the analysis. Authors also thanked the deptt. Of
Applied mechanics &Hydraulics for infrastructural
support to this work.
) '
4
[1] Adamowski, J., and Sun, K. (2010). “Development
of a coupled wavelet transform and neural network
method for flow forecasting of non perennial rivers
in semi arid watersheds.” Journal of Hydrology,
390, 85 91.
[2] Cannas, B., Fanni, A, Sias, G, Tronchi, S, Zedda,
M.K. (2005). “River flow forecasting using Neural
International Journal of Earth Sciences and Engineering
ISSN 0974 5904, Vol. 05, No. 04, August 2012, pp. 673 685
6>
Discrete Wavelet Ann Approach in Time Series Flow Forecasting
A Case Study of Brahmaputra River
Networks and Wavelet Analysis.” EUG (2005),
European Geosciences Union, Vienna, Ausrtia,
vol.7, 24 29.
[3] Cannas.B, Fanni.A, See.L., Sias.G., (2006), “Data
preprocessing for river flow forecasting using
neural networks: wavelet transforms and data
partitioning”. Physics and Chemistry of the Earth
31 (18), 1164–1171.
[4] Deka, P.C.and Prahlada, R(2012) ‘Discrete
wavelet neural network approach in significant
waveheight
forecasting
for
multistep
leadtime’.Ocean Engg.43,32 42.
[5] Kim, T.and Valdes, J.B. (2003). “Nonlinear model
for drought forecasting based on a conjunction of
wavelet transform and neural networks.” Journal of
Hydrologic Engineering, l.8 (6), 319 328.
[6] Kisi, O. (2009). “Neural Networks and Wavelet
Conjunction Model for Intermittent Stream flow
Forecasting.” Journal of Hydrologic Engineering,
14(8), 773–782.
[7] Kisi, O. (2010). “Daily Suspended Sediment
Estimation Using Neuro Wavelet Models.” Int. J.
Earth Sci., 99:1471 1482.
[8] Kisi,O. and Shiri,J(2011).“Precipitation forecasting
using wavelet genetic programming and wavelet
neuro fuzzy conjunction models”.Water Resources
Management, 25:3135 3152.
[9] Mallat S.G. (1989). “A theory for multiresolution
signal
decomposition:
The
wavelet
representation.”IEEE Trans. Pattern Anal. Mach.
Intell., 11(7), 674 693.
[10] Nourani, V., Komasi, M., Mano.A. (2009). “A
Multivariate ANN –Wavelet Approach for Rainfall
Runoff Modeling.” Water Resources Manage
23:2877 2894.
[11] Nason G P,Von Sachs R.1999.Wavelet in time
series analysis. Phil Trans Roy Soc 357; 2511
2526.
[12] Nourani, V., Alami, T Mohammad, Aminfar,
Mohammad.H (2009). “A combined neural wavelet
model for prediction of Ligvanchai watershed
precipitation.” Elsevier, Engineering Applications
of Artificial Intelligence, 22, 466–472.
[13] Prahlada,R and Deka,P .C(2011).“ Hybrid wavelet
neural network model for improving forecasting
accuracy of time series significant wave height”.Int.
J.Earth Science and Engg.vol.4 (5)., Oct.857
866.(ISSN0974 5904).
[14] Rajaee, T., Mirbagheri, S.A, Nourani, V., Alikhani,
A. (2010). “Prediction of daily suspended sediment
load using wavelet and neuro fuzzy combined
model.” Int. J. Environ. Sci. Tech., 7 (1), 93 110.
[15] Rajaee, T., Nourani, V., Kermani, M.Z., Kisi, O.
(2011). “River suspended sediment load
prediction:Application of ANN and wavelet
conjunction model”.,ASCE j. hydrologic Engg.,
16(8),August,613 627.
[16] Rao,Y.R. Satyaji and B. Krishna (2009),
“Modelling Hydrological Time Series data using
Wavelet Neural Network Analysis”, IAHS
Publication 333,101 110.
[17] Shiri, J., and Kisi, O. (2010). “Short term and long
term stream flow forecasting using a wavelet and
neuro fuzzy conjunction model.” Journal of
Hydrology, 394, 486 493.
[18] Nourani, V.,Kisi,O.,Komasi.M. (2011). “Two
hybrid artificial intelligence approaches for
modeling rainfall – runoff process.” Journal of
Hydrology, 402, 41 49.
[19] Wensheng Wang and Jing Ding, (2003) “Wavelet
Network Model and Its Application to the
Prediction of Hydrology”, Nature and Science
1(1):67 71.
[20] Wang.W.,Hu.S.,Li. Y.(2011). “Wavelet transform
method for synthetic generation of daily
streamflow”.Water Resources Management, 25:41
57.
[21] Zhou H.C., Peng, Y., Liang G H. (2008). “The
research of monthly discharge predictor corrector
model based on wavelet decomposition”Water
Resources Management, 22; 217 227.
International Journal of Earth Sciences and Engineering
ISSN 0974 5904, Vol. 05, No. 04, August 2012, pp. 673 685