Escolar Documentos
Profissional Documentos
Cultura Documentos
197
Department of Civil Engineering, Indian Institute of Technology, Guwahati 781 039, India
chand@iitg.ernet.in
Abstract The measurement of discharge in major rivers is very important and serves
as the base information for hydrological analysis. The rating curve is used to assess
the discharge from the measured stage values in the gauging sites. The rating curve
has important bearing on the correct assessment of discharge. The usefulness of the
fuzzy neural network modelling approach in deriving the stagedischarge relationship
is discussed. The performances of a neural network model, a modularized neural network model, a conventional curve-fitting approach and a fuzzy neural network model
for deriving the rating curve are compared using a case study. Overall, the fuzzy
neural network model gives the best results.
Key words artificial neural network model; fuzzy neural network model; modularized neural
network model; stagedischarge relationship; Brahmaputra River basin, India
198
A fuzzy system adaptively infers and modifies its fuzzy associations from representative numerical samples. Fuzzy sets aid in providing information in a comprehensible or natural form, and can handle uncertainties at various levels (Hundecha et
al., 2001). The knowledge contained in fuzzy systems is transparent to the user. See &
Openshaw (1999) proposed a combined approach of fuzzy logic, neural network and
genetic algorithm optimization for flood forecasting.
The concept of neuro-fuzzy systems has emerged as researchers have tried to
combine the transparent, linguistic representation of a fuzzy system with the learning
ability of an ANN (Brown & Harris, 1994). A neuro-fuzzy system uses an ANN
learning algorithm to determine its parameters (i.e. fuzzy sets and fuzzy rules) by
processing data samples. Therefore, it can be trained to perform an input/output
mapping, just as with an ANN, but with the additional benefit of being able to provide
the set of rules on which the model is based (Mitra & Hayashi, 2000). There have been
many attempts to synthesize fuzzy neural network (FNN) models and, according to
their integration methodologies, two major categories of FNN can be identified. One is
based on the fuzzification of conventional neural network model and the other is based
on the implementation of a conventional fuzzy systems using neural networks (Chung
& Duan, 2000).
STUDY AREA
The system considered for the analysis is the Brahmaputra River, in Assam State,
India. It originates as Tsan-po in Tibet, to the east of the Mansarovar Lake and is
2898 km long (Fig. 1). The basin lies between latitudes 2413 and 3130N and
longitudes 82 and 964E. Three gauging stations, namely Bessamara, Bhurbandha
and Pandu (hereafter referred to as gauging stations I, II and III), are considered and
daily stagedischarge values collected for these stations are used for this study (see
Table 1). The upstream gauging stations I and II are located at about 340 and 150 km
upstream of gauging station III, respectively. For this study, four models, namely a
conventional curve-fitting model, a single neural network model, a modularized neural
network model and a fuzzy neural network model, are considered for deriving the
stagedischarge relationship.
CONVENTIONAL MODEL
Streamflow measurements are normally derived using the stagedischarge curve for
transforming the record of stage into a record of discharge (Maidment, 1993). A rating
curve of the following form is commonly used:
Q = a (G G0 ) b
(1)
where Q is the discharge (m3 s-1), G is the river stage (m), G0 is the river stage (m) at
which the discharge is nil, and a and b are constants.
This type of relationship is established by fitting a smooth curve between the stage
and discharge records by either an ordinary or a logarithmic scale. The stage for zero
discharge is an unknown and its determination possesses some difficulties for major
199
A fuzzy neural network model for deriving the river stagedischarge relationship
rivers. A major limitation of this approach is that it is not able to consider the
hysteresis effect. Also in this study, G0 has been estimated from the observed data and
the best values of a and b for the given range of stage are obtained by the least-squares
error method (Table 1).
SINGLE ARTIFICIAL NEURAL NETWORK MODEL
The single ANN model developed for this study is a feed-forward neural network
model with error back-propagation algorithm (Rumelhart & McClelland, 1987). After
a detailed check of different combinations, two experiments on the ANN model are
developed herein: a single input, single output model (Experiment 1) and a four input,
200
Table 1 Data availability, values of a and b for the conventional model and neural network architecture
for different gauging stations considered.
Gauging station
Data availability:
Gauging station I:
Experiment 1
Experiment 2
Gauging station II:
Experiment 1
Experiment 2
Gauging station III:
Experiment 1
Experiment 2
Conventional
model:
a
b
NN architecture
(number of neurons):
Input
Hidden
layer
layer
Output
layer
Training
Testing
1600
1600
480
257
148.00
2.38
1
4
3
4
1
1
1000
800
358
346
106.55
2.55
1
4
2
8
1
1
2500
2500
891
889
127.68
2.45
1
4
3
8
1
1
(a)
(b)
Fig. 2 The two experiments (a) experiment 1, and (b) experiment 2, considered for
analysis.
single output model (Experiment 2) (Fig. 2). The training of the data is carried out by
monitoring the indices, mean square error (MSE) and mean relative error (MRE),
defined as follows:
A fuzzy neural network model for deriving the river stagedischarge relationship
p
1
MSE =
y (jt ) y j
pq q j =1
(2)
p
y (jt ) y j
1
MRE =
y (t ) 100
pq q j =1
j
(3)
in which p is the number of patterns used for training, q is the number of nodes in the
output layer, yj(t) is the target output pattern value (used for training) and yj is the
output from the neural network model.
The MSE value indicates the goodness of fit at high output values and the MRE
value indicates the goodness of fit for moderate output values (Karunanithi et al.,
1994). The number of neurons in the hidden layer, the number of hidden layers and the
normalization factor for the data considered are decided after many trials by studying
the performance of the ANN training indices. The details of different neural network
models developed for the two experiments considered are listed in Table 1.
MODULARIZED NEURAL NETWORK MODEL
For improving the performance of the neural network model, Zhang & Govindaraju
(2000) suggested a modularized neural network approach for rainfallrunoff
modelling. They showed the advantages of segregating the patterns considered for
training and organized the neural networks in a modular architecture to handle
complex sets of rainfallrunoff data. A modularized neural network modelling
approach is attempted herein for developing the stagedischarge relationship (Fig. 3).
Discharge
Trained ANNs
Very low
Rule base
using stage value
Fuzzy
inferencing
Low
Medium
High
Discharge
201
202
Description
Gauging station I:
Modularized modelling
Module I
Module II
Module III
Module IV
Average performance
Single NN modelling
Fuzzy NN modelling
Very Low
Low
Medium
High
Average performance
Conventional modelling
Gauging station II:
Modularized modelling
Module I
Module II
Module III
Module IV
Average Performance
Single NN modelling
Fuzzy NN modelling
Very Low
Low
Medium
High
Average performance
Conventional modelling
Gauging station III:
Modularized modelling
Module I
Module II
Module III
Module IV
Average performance
Single NN modelling
Fuzzy NN modelling
Very Low
Low
Medium
High
Average performance
Conventional modelling
MRE
80.03
81.10
83.28
85.37
81.09
83.27
85.37
87.80
400
400
400
400
80.03
87.80
1600
0.66
6.58
7.06
8.00
5.50
6.05
28.93
21.06
15.83
23.69
22.38
26.74
80.03
80.71
82.11
84.23
82.11
84.23
86.14
87.80
600
800
800
600
80.03
87.80
1600
1.41
5.08
6.91
9.77
5.82
7.28
35.01
38.35
24.31
16.78
29.00
31.59
58.13
60.07
61.761
64.261
60.06
61.76
64.26
66.94
250
250
250
250
58.13
66.94
1000
0.09
3.34
1.36
17.83
8.73
8.97
8.60
31.52
27.50
15.67
20.82
23.22
58.13
59.52
60.93
63.18
60.92
63.15
65.71
66.94
375
500
500
375
58.13
66.94
1000
1.08
6.41
12.73
17.96
9.55
10.08
20.08
44.03
25.39
18.49
28.09
30.34
40.79
41.92
43.75
46.96
41.91
43.74
46.95
49.95
625
625
625
625
40.79
49.95
2500
3.65
18.10
1.01
6.53
7.32
7.99
32.60
19.26
4.14
5.72
15.43
17.34
40.79
41.42
42.70
45.40
42.70
45.39
46.95
49.95
937
1250
1250
938
40.79
49.95
2500
13.56
10.51
2.10
4.53
7.48
28.24
31.42
20.53
6.51
4.70
15.46
39.04
A fuzzy neural network model for deriving the river stagedischarge relationship
203
crisp set. In this approach, instead of developing a neural network model for the whole
set of data, four neural network models are developed, one for each group of data. The
data range is fixed after examining the available patterns. For Experiment 1, the four
groups and the respective range of stage level (in m) are given in Table 2 for three
gauging stations. After completion of the training, with the help of a rule base, which
works on the criteria fixed for segregation, the appropriate neural network will be fired
to obtain the final result.
A ( x) = 0 m > x or x > l
(4)
(5)
where m is the lower limit of a fuzzy set, l is the maximum limit for a fuzzy set and
A(x) is the membership function of a fuzzy set.
For each fuzzified grouping, namely Very Low, Low, Medium and High, a
separate neural network model is developed to establish the membership function of
each group (Fig. 3), i.e. the fuzzy surface of each fuzzy set is developed using the
trained neural network model. The developed fuzzy surface for each fuzzy set is not a
standard shape. This arbitrary shape is decided using the neural network. These fuzzy
shapes are wavy matching behaviour patterns that undulate over the domain. The
membership function for each fuzzy set is established using ANN based on the
observed stagedischarge data using a neural network. Figure 4 shows the membership
functions for the Medium and High fuzzy sets developed using the neural network.
For training the neural network with sigmoidal activation function, the output data
have to be normalized using a normalization factor so that the results are comparable for
error back-propagation training. The membership function for the discharge data is derived
based on the normalization factor used in the respective neural network model. This
function is a straight line, increasing set with a slope equal to the normalization factor. The
fuzzy associate memory developed for the fuzzy model has the following rules:
204
Rule 1: If the stage is Very Low then the discharge is Very Low.
Rule 2: If the stage is Low then the discharge is Low.
Rule 3: If the stage is Medium then the discharge is Medium.
Rule 4: If the stage is High then the discharge is Large.
The fuzzy reasoning used here is the simple monotonic method, a basic fuzzy
implication technique. Mathematically, for a simple proportional implication function:
If x is Y then z is W
(6)
z = f (( x, Y ),W )
(7)
where x is the stage value, z is the discharge value, and Y and W are the fuzzy sets
corresponding to the stage and discharge values. The value of the output is estimated
directly from a corresponding truth membership grade in the antecedent fuzzy region
(Cox, 1994). This type of reasoning is followed when only one rule is fired and
suitable for the stage level considered. On the other hand, if two rules are fired for a
particular stage value, the min-max rule of implication is executed (Fig. 4). The
consequent fuzzy sets resulting from two rules will be producing an output fuzzy
region by aggregation, which will be defuzzified to give the expected discharge value
based on near edge of the support set defuzzification procedure (Fig. 4).
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0
0
82
83
84
85
86
87
11
16
3 -1
0.8
3 -1
( x 10 m s )
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0
84
Membership function
1
Discharge ( x 10 m s )
0.8
0.6
0.4
0.2
0
1
6 11 16 21 26
Membership function (LARGE)
Discharge
0
85
86
87
88
11
16
21
26
3
3 -1
Discharge ( x 10 m s )
The performances of different models considered are assessed using MSE and MRE
indices for training and testing series. In Experiment 1, the conventional model is
compared with other models, as the ANN, MNN and FNN models have the stage value
A fuzzy neural network model for deriving the river stagedischarge relationship
205
II
III
Model description
MNN
Single NN
FNN
MNN
Single NN
FNN
MNN
Single NN
FNN
MRE
5.66
5.83
5.68
8.54
9.03
8.22
4.55
5.56
4.88
Table 4 Module-wise training results (MSE 106 (m3 s-1)2) for gauging station II.
Module I
Module II
Module III
Module IV
Experiment 1:
Single NN model
0.31
3.45
13.88
18.23
MNN model
0.09
3.34
13.67
17.83
Experiment 2:
Single NN model
0.31
2.14
4.43
7.93
MNN model
0.29
0.96
2.87
7.77
as the only input. However, the conventional model is excluded from the discussions
related to Experiment 2, in which ANN, MNN and FNN models have four inputs.
Further, the training and testing data set lengths are different for Experiment 1 and
Experiment 2. This is due to the fact that, in different gauging sites, observations for
some days in-between were not available for technical reasons. Hence some data are
not considered for Experiment 2, which requires details of previous days.
The single ANN model, the modularized ANN model (MNN) and the fuzzy neural
network model (FNN) training results are given in Tables 2 and 3 for both experiments. Both the testing and the model training results of the conventional model are
poorer than those of the other models.
The MNN model shows good improvement in the total MSE and MRE indices in
the training data set relative to the single ANN model and theconventional model
(about 220% in the case of MSE in both experiments). The performances of single
ANN and MNN models in training are studied in detail by considering different
modules for gauging station II data using experiments 1 and 2. When the MSE index is
considered in each module, the MNN model performs better than the ANN model in
all the modules (Table 4). Hence, the overall MSE index also shows that the MNN
model is better than the single ANN model.
However, the MNN model performs poorly in the testing. The MSE and MRE
indices are high for the MNN model (Table 5). For testing data, the MSE results for
different modules for both the experiments are given in Table 6. The single ANN
model performs better in three modules and one module respectively in experiments 1
and 2. However, the overall performance of the MNN model is inferior. The MNN
model suffers in the split zones. For example, consider the modules I and II split
regions, and the modules III and IV split regions (59.8560.10 m and 64.00 64.50 m,
respectively). The MNN model shows poorer performance than other models in these
206
Experiment 1:
MSE 106 (m3 s-1)2
MRE
Experiment 2:
MSE 106 (m3 s-1)2
2.20
2.65
2.64
4.07
12.07
18.00
15.15
28.77
2.86
3.84
2.26
-
5.87
8.65
5.83
29.29
32.19
27.87
38.45
58.62
58.66
49.84
64.62
1.32
4.50
0.86
-
9.81
14.69
6.79
5.94
20.36
6.35
20.17
4.92
15.33
11.34
24.86
6
3 -1 2
PMSE 10 (m s )
16.86
19.03
16.32
1.83
2.23
1.75
-
4.75
5.37
3.83
MRE
1.04
4.66
0.61
Table 6 Module-wise testing results (MSE 106 (m3 s-1)2) for gauging station II.
Model
Module I
Module II
Module III
Module IV
Experiment 1:
Single NN
MNN model FNN model
model
0.09
0.09
0.22
0.98
0.87
0.72
50.66
56.14
37.07
77.35
84.93
85.74
Experiment 2:
Single NN
MNN model FNN model
model
0.31
0.10
0.01
2.14
0.10
0.08
4.43
1.09
0.81
7.93
12.88
1.85
regions (Figs 5 and 6). Due to this fact, the overall performance of the MNN model
suffers. Hence, it can be concluded that, in the case of MNN, when the data are
classified into four groups crisply, the knowledge captured from the four groups is not
generalized properly. The discontinuity results from the absence of knowledge about
its neighbouring regions. When all the four modules are used together, in testing, the
performance of the MNN model suffers more in the split zones. The nature of the data
is also very crucial for this performance. The modularized approach is very suitable for
studies involving clustering of available information.
In the case of the FNN model, segregation of patterns is done by having
overlapping in the neighbouring regions. Since the training is done by overlapping for
each region, the discontinuity is avoided in the data. Hence for the FNN model, the
knowledge gathering during training is not affected as in the case of the MNN model.
Thus, the FNN model performs better due to the training given with the overlapping
data, which avoids discontinuity and also improves the performance of the FNN
model.
A fuzzy neural network model for deriving the river stagedischarge relationship
207
4500
Observed
ANN
FNN
MNN
Discharge in m3s-1
4000
3500
2500
59.4
59.6
59.8
60
60.2
60.4
60.6
60.8
61
61.2
61.4
61.6
Stage in m
Fig. 5 Performance of different models in the module I and module II split region.
Observed
ANN
FNN
MNN
21000
Discharge in m3 s-1
3000
19000
17000
15000
13000
64
64.1
64.2
64.3
64.4
64.5
64.6
64.7
64.8
64.9
65
Stage in m
Fig. 6 Performance of different models in the module III and module IV split region.
Based on testing results, when the performances of single ANN and FNN models
are compared, the FNN performs better than the ANN in all cases except one (about
620% better MSE value than the single ANN). Further, the FNN model performs
better in all modules for Experiment 2 and in two modules in Experiment 1 (Table 6).
When the testing results of different models are compared for Experiment 2
(gauging station II), the hysteresis effect might be better represented by the FNN
model (Fig. 7). The MNN model shows poor performance in the split zone (64.0
65.0 m stage value) and does not properly represent the hysteresis effect in this study.
However, the FNN model gives better performance in this zone also (Fig. 7).
208
Observed
ANN
FNN
MNN
Discharge in m3 s-1
25000
20000
15000
10000
5000
0
62
63
63
64
64
65
65
66
66
67
67
Stage in m
It is observed that in many research studies, using the MRE to rank different models
according to their performance is different from the ranking using MSE. To overcome
the difficulties in using MSE and MRE, a new statistical measure called the pooled
mean square error (PMSE) has been used, which combines the effects of MSE and
MRE (Elshorbagy et al., 2000). For the calculation of the PMSE, a threshold value of
5% of the actual value is considered in this analysis. The PMSE is estimated as:
N /2
PMSE =
(y
(t )
j
y j .k j
j =1
(8)
j =1
where N is the set size and kj is the rank of the residual error.
For this analysis, the gauging station II data are considered (Table 5). In all the
cases, a better performance of the FNN model is noted. The complexity of the ranking
process for different model performances using two indices, namely the MSE and
MRE, is eased by the PMSE index.
It is worth noting that for the gauging station II, the FNN model gives 34.45 and
30.79% improvement over the ANN model when considering the MSE and MRE
indices, respectively. For the same case, the FNN model shows 40.89% improvement
over the ANN model when the PMSE index is considered. This shows the better ability
of the FNN model compared to the other models in estimating a discharge value that is
closer to the observed value within the threshold of the 5% limit.
CONCLUSION
This study shows the ability of the fuzzy neural network model in developing the
stagedischarge relationship. The FNN model has better predictive ability than the
single ANN, the modularized ANN and conventional models. The MNN model gives
A fuzzy neural network model for deriving the river stagedischarge relationship
209
inferior results relative to the single ANN model and suffers due to data discontinuity
even though the training of the MNN shows better values for the performance
indicators. The FNN model is more flexible than the other models considered with
more options of incorporating the fuzzy nature of the real-world system.
Acknowledgement The authors wish to acknowledge the support given by Assam
State Flood Control Department, India, by providing the necessary data for the
analysis.
REFERENCES
Brown, M. & Harris, C. (1994) Neuro-Fuzzy Adaptive Modeling and Control. Prentice-Hall, Englewood Cliffs, New
Jersey, USA.
Chandramouli, V. & Raman, H. (2001) Multi-reservoir modeling with dynamic programming and neural networks.
J. Water Resour. Plan. Manage. ASCE 127(2), 8998.
Chung, F. L. & Duan, J. C. (2000) Multistage fuzzy neural network modeling. IEEE Trans. Fuzzy Systems 8(2), 125142.
Cox, E. (1994) The Fuzzy Systems Handbook. Academic Press Professional, Boston, Massachusetts, USA.
Elshorbagy, A., Simonovic S. P. & Panu, U. S. (2000) Performance evaluation of artificial neural networks for runoff
prediction. J. Hydrol. Engng ASCE 5(4), 424427.
Hu, T. S., Lam, K. C. & Ng, S. T. (2001) River flow time series prediction with a range dependent neural network. Hydrol.
Sci. J. 46(5), 729745.
Hundecha, Y., Bardossy, A. & Theisen, H. W. (2001) Development of a fuzzy logic-based rainfallrunoff model. Hydrol.
Sci. J. 46(3), 363376.
Jain, S. K. & Chalisgaonkar, D. (2000) Setting up stagedischarge relations using ANN. J. Hydrol. Engng ASCE 5(4),
428433.
Karunanithi, N., Grenney, W. J., Whitley, D. & Bovee, K. (1994) Neural networks for river flow prediction. J. Computers
Civ. Engng ASCE 8(2), 201220.
Maidment, D. R. (1993) Handbook of Hydrology. McGraw-Hill, New York, USA.
Minns, A. W. & Hall, M. J. (1996) Artificial neural networks as rainfallrunoff models. Hydrol. Sci. J. 41(3), 399418.
Mitra, S. & Hayashi. Y. (2000) Neuro-fuzzy rule generation: survey in soft computing framework. IEEE Trans. Neural
Networks 11(3), 748768.
Rumelhart, D. & McClelland, J. L. (1987) Parallel Distributed Processing: Explorations in the Microstructure of
Cognition, vols I and II, MIT Press, Cambridge, Massachusetts, USA.
Rajurkar, M. P., Kothyari, U. C. & Chaube, U. C. (2002) Artificial neural networks for daily rainfallrunoff modelling.
Hydrol. Sci. J. 47(6), 865877.
See, L. & Openshaw, S. (1999) Applying soft computing approaches to river level forecasting, Hydrol. Sci. J. 44(5),
763778.
See, L. & Openshaw, S. (2000) A hybrid multi-model approach to river level forecasting, Hydrol. Sci. J. 45(4), 523536.
Tawfik, M., Ibrahim, A. & Fahmy, H. (1997) Hysteresis sensitivity neural network for modeling rating curves.
J. Computers in Civ. Engng ASCE 11(3), 206211.
Xiong, L. & OConnor, K. M. (2002) Comparison of four updating models for real-time river flow forecasting, Hydrol.
Sci. J. 47(4), 621639.
Zhang, B. & Govindaraju, R. S. (2000) Prediction of watershed runoff using Bayesian concepts and modular neural
networks. Water Resour. Res. 36(3), 753762.