J.

Water Resource and Protection, 2010, 2, 880-887

doi:10.4236/jwarp.2010.210105 Published Online October 2010 (http://www.SciRP.org/journal/jwarp)

Development of Flood Forecasting System Using Statistical

and ANN Techniques in the Downstream Catchment of
Mahanadi Basin, India
Anil Kumar Kar1, Anil Kumar Lohani2, Narendra Kumar Goel1, Gopal Prasad Roy3
1
Department of Hydrology, Indian Institute of Technology, Roorkee, India
2
National Institute of Hydrology, Roorkee, India
3
Department of Water Resources, Government of Orissa, Bhubaneswar, India
E-mail: [email protected], [email protected]
Received May 6, 2010; revised July 27, 2010; accepted August 12, 2010

Abstract

The floods in river Mahanadi delta are due to either dam release of Hirakud or due to contribution of inter-
cepted catchment between Hirakud dam and delta. It is seen from post-Hirakud periods (1958) that out of 19
floods 14 are due to intercepted catchment contribution. The existing flood forecasting systems are mostly
for upstream catchment, forecasting the inflow to reservoir, whereas the downstream catchment is devoid of
a sound flood forecasting system. Therefore, in this study an attempt has been made to develop a workable
forecasting system for downstream catchment. Instead of taking the flow time series concurrent flood peaks
of 12 years of base and forecasting stations with its corresponding travel time are considered for analysis.
Both statistical method and ANN based approach are considered for finding the peak to reach at delta head
with its corresponding travel time. The travel time has been finalized adopting clustering techniques, there by
differentiating high, medium and low peaks. The method is simple and it does not take into consideration the
rainfall and other factors in the intercepted catchment. A comparison between both methods are tested and it
is found that the ANN methods are better beyond the calibration range over statistical method and the effi-
ciency of either methods reduces as the prediction reach is extended. However, it is able to give the peak
discharge at delta head before 24 hour to 37 hour for high to low peaks.

Keywords: Flood Forecasting, Mahanadi Basin, Hirakud Dam, Statistical Method, ANN Architecture,
Clustering

1. Introduction order to establish a Rainfall-Runoff model, rainfall data

of all station is highly essential. For further strengthening
Flood is a regular phenomenon in the delta of Mahanadi of model evaporation, soil moisture, temperature, dam
basin. The discharge of the intercepted catchment of be- release etc. are needed. Sometimes all information is not
low Hirakud dam has contributed largely to the flood at available on real time basis making it difficult for estab-
delta. The delta is highly fertile and thickly populated lishment of a physical based model. In that case statisti-
with population density 400-450 persons per sq.km. The cal methods are more useful.
drainage capacity of deltaic rivers is 25500 m3/s only. Any The statistical methods are generally graphical or in
discharge above this can create a flood like situation. the form of mathematical relationship, developed with
Establishment of a workable flood forecasting method the help of historical data, using statistical analysis.
for the downstream catchment of Mahanadi basin is al- These include simple gauge to gauge relationships,
ways under scanner. A lot of hydro-meteorological in- gauge to gauge relationship with additional parameters
formation is needed for establishment of a model. The and rainfall-peak stage relationship. These relationships
dam at upstream is operated by a rule curve. The inter- can be easily developed and are most commonly used in
cepted catchment is very big around 40000 sq.km. In India as well as other countries of the world [1]. Spok-

Copyright © 2010 SciRes. JWARP

 A. K. KAR ET AL. 881

kerreff [2] has mentioned about application of statistical Mahanadi River.

model till 1995 on river Rhine.
Consequently, there is need for models capable of ef- 2. Study Area
ficiently forecasting water levels and discharge rates. In
this regard application of ANN is more effective. Earlier The Mahanadi basin lies between 800-30` to 860-50` of
the works of Bruen and Yang [3], Campolo et al. [4,5], East Longitude and 190-20` to 230-35` of North Latitude.
Coulibaly et al. [6], Dawson and Wilby [7], Imrie et al. The river Mahanadi is an interstate river. The catchment
[8], Lekkas et al. [9], Lohani [10], Minns and Hall [11], area of the basin is 141569 sq. km. The major part of this
Muhamad and Hassan [12], Mukerji et al. [13], Solo- catchment lies in Chhatisgarh (75136 sq. km.) (Figure 1).
mantine and Xue [14], Solomantine and Price [15], Zea- As this part is the upstream part, 143 channels in this
land et al. [16] emphasized the application of artificial zone generally feeds the Hirakud reservoir having a cat-
neural networks over other methods. In flood forecasting chment of 83400 sq.km. The downstream catchment has
both the peak and travel time has equal significance. The three main tributaries like Jeera, Ong and Tel with cat-
chments 2383, 5128 and 25045 sq.km. respectively. So
travel time is also of great importance in prediction of
the contributions from the Tel catchment always remain
the stage. It varies with stages and channel condition.
predominant. Even the flood of 2008 is mainly due to the
Travel time generally reduces when water approaches the
contribution of this tributary. It has produced a peak dis-
top of the bank. As the river overflows flooding over the
charge of 33762 cumecs during 2008. So establishment
flood plain the travel time may begin to increase again of a flood forecasting model below the joining of these
due to the relatively rough surfaces lying in the overbank tributaries reduces the ambiguity. The river Tel joins at
stages. In order to give justification for different types of Patharla to the main river Mahanadi and our base station
peaks with respect to travel time clustering method is Khairmal is at downstream of Patharla station. The other
adopted in this study. Earlier, Zhang and Hall [17] have tributaries Ong and Jeera join also at the upstream of
applied clustering methods in regional flood frequency base station (Khairmal). The whole river from Khairmal
analysis and Xiongrui et al. [18] in establishing rainfall- is taken as one unit ignoring the contribution of further
runoff relationships. In view of the above an attempt has small tributaries. A schematic presentation shows the
been made in this study to apply both statistical method distance and travel time from Hirakud to Mundali pres-
and ANN based approach for forecasting the peak dis- ently being used for the official purposes of Department
charge and travel time in the downstream reach of the of Water Resources, Government of Orissa [19].

Hirakud 115 km Khairmal 106 km Barmul 95 km Mundali (delta head)

(12-18 hours) B.S (12-16 hrs.) I.S (12-16 hrs.) F.S

Figure 1. Showing catchment details with different zones of Mahanadi basin.

Copyright © 2010 SciRes. JWARP

882 A. K. KAR ET AL.

In our study the station at Khairmal is taken as the network is application oriented. However, the multi layer
base station (B.S) which is 115 km away and with an feed forward networks are most commonly used for hy-
average lead time of 12-16 hour for Hirakud release to drological applications [7].
reach here. Barmul is the intermediate station (I.S) and Generally four distinct steps applied to any ANN –
Mundali is the forecasting station (F.S) just upstream of based solution including flood forecasting problems.
Cuttack city. First of all data transformation is the initial step. The
input always varies in a larger range. So in order to les-
3. Data Availability sen the range the data is put in logarithmic scale. Then
normalization and scaling is done. Data sets with many
Adequate data is needed for the formulation of the fore- variations make training difficult. In order to prevent it
casting services. The development of the river forecast- the data are usually scaled using statistical, min-max,
ing procedure requires historical hydrological data and sigmoidal or principal component transformations [21].
for the preparation of operational forecasts sufficient Also the absolute input values are scaled to avoid as-
current information is required. Discharge data of (3 ymptotic issues [22].
hourly) is available at base, intermediate and forecasting The second step is the fixing the network architecture
station for a period of 1996-2007. Discharge of 2830 in which for a particular problem the number of hidden
cumecs and above at forecasting stations are considered layers, neurons in each layer and the connectivity be-
as peaks and its corresponding peaks at intermediate and tween neurons are set. Many experimental results say
base station are collected with their related travel times that one hidden layer may be enough for most forecast-
after drawing the corresponding time series. Out of this ing problems [6,17]. The studies of Cybenko [23], Hor-
80 peaks are considered for our analysis, taking 60 peaks nik et al. [24] revealed that a single hidden layer is suffi-
for calibration and 20 peaks for testing of models. Gen- cient for ANNs to approximate any non-linear transfer
erally, it can be said that it is necessary to have a mini- function. The number of neurons in each layer depends
mum of 10 years of basic hydraulic data, available to upon the problem being studied. Less number of neurons
develop adequate river forecasting procedures, the pri- in hidden layer will make the network with less degree of
mary requirement being that the period of record should freedom for learning and more number of neurons will
contain a representative range of peak flows [20]. lead towards more time and over fitting [25]. Validation
set error is often used to determine the optimal number
4. Methodology of hidden neurons for a given study.
In the third step is the finalization of a learning algo-
The selection of an appropriate flood forecasting model rithm for training the network. The parameters are final-
depends on the availability of the data, output desired etc. ized for the training data set to be applicable for any kind
On the basis of the analytical approach for the develop- of testing data. ANN architecture is considered to be
ment of flood forecasting method can be classified as: trained when the difference between ANN output and
1) Methods based on statistical approach. observed output is very small.
2) Methods based on mechanism of formation and Finally the validation step is applied to get the per-
propagation of floods. formance of the network. The optimal number of hidden
neurons and training iterations can also be determined
4.1. Statistical Approach through validation. The selection of acceptable model is
finalized on the basis of RMSE, R2 and efficiency.
Finding correlation between stage and discharges be-
tween upstream and downstream gauging stations is one 4.3. Clustering Approach
of the simplest methods. This gives better result when
there is less influence of tributaries joining the main As the data set varies in a wide range, it requires some
stream in between or the intercepted catchment is not distinction and in particular when different peaks are
influenced by heavy rainfall. available for the same travel time. In order to make
proper justice for different peaks with respective travel
4.2. ANN Approach time, it requires clustering. Both K-mean and Fuzzy
C-mean clustering methods are attempted.
There are many ANN architectures and algorithms de-
veloped. Out of them most common are Multi layer feed 4.3.1. K-Mean
forward, Hoppfield networks, Radial basis function net- This method was developed by MacQueen [26]. It is best
work, Recurrent network, Self organization feature maps, described as a partitioning method. It partitions the data
Counter propagation networks. Selection of a particular into K mutually exclusive clusters and returns a vector of

Copyright © 2010 SciRes. JWARP

 A. K. KAR ET AL. 883

indices indicating to which of the K-clusters it has as- for finding the travel time. But in order to further distin-
signed each observation. The algorithm to clusters N guish K-mean and Fuzzy C mean clustering methods are
objects based on attributes into K partitions where K < N. adopted to find a better answer for travel time of differ-
The optimization function ent peaks.

  x j   i
k 2
V  (1) 5. Results and Discussion
i 1 x j S i


It tries to achieve minimum intra cluster variance or First of all statistical approach is applied into calibration
the squared error function. Where there are K clusters Si dataset (60 nos.). The relationships between the dischar-
= 1,2……K, and Vi is the centroid or mean point of all ges between Khairmal-Barmul, Barmul-Mundali, Khair-
the points Xj Є Si. [27]. mal-Mundali and Khairmal-Barmul-Mundali are devel-
oped. These relationships are put on the testing data sets
4.3.2. Fuzzy C-Mean (FC) (20 nos.) and the results are recorded in Table 1.
In this method the affinity of a site to undergo either two Where, QM = discharge at Mundali (Forecasting Sta-
or more clusters are visualized. Earlier developed by tion), QB = discharge at Barmul (Intermediate Station)
Dunn [28] and improved by Bezdeck [29] is basically and QK = discharge at Khairmal (Base Station).
used for pattern recognition. Here the data are bound to The same dataset is again put into ANN architecture
each cluster by means of a membership function which using MATLAB codes. The trial has been taken with a
represents the Fuzzy behavior of this algorithm. It shows 3-layer feed forward network. Different combinations of
how to group data points that populate some multidi- feed forward network with changing transfer function,
mensional space into a specific number of different clus- number of neurons and epochs varying at an increment
ters. The objective function of 50 are trailed. The combinations which are mostly as
2 per performance criteria fixed are noted. The Cascade
Jm  i 1  j 1 uijm xi  c j , 1  m  
N C
(2) feed forward network has been most successful for
m = any real number, uij = degree of membership of Khairmal-Barmul and Khairmal-Barmul-Mundali and
other two cases are with Feed forward network. Numbers
xi in cluster j , xi = ith of d-dimensional measured
of neurons in input and output layers are mentioned in
data, c j = d-dimension center of cluster. column (2) of Table 2. In all cases ‘tansig’ neurons are
Similarly, different ANN architectures are attempted used in first layer, ’purelin’ in second layer and ‘trainbr’
over the discharge and corresponding travel times. The remains the training function. Both ‘tansig’ and ‘purelin’
calibration and validation data sets remain same for both are better compatible with feed forward networks. Again
the methodology. The same methodology is also adopted the two layer sigmoid/linear network can represent any
Table 1. Relationship between discharges using simple statistical method (peak to peak).

RMSE (m3/s) R2 Efficiency

Between stations Equation
Training Testing Training Testing Training Testing
Khairmal-Barmul QB = 1.122*QK −289.4 2382.62 1287.4 0.94 0.974 0.952 0.9678
Barmul-Mundali QM = 1.009*QB + 678.17 1740.26 1275.16 0.968 0.989 0.948 0.9735
Khairmal-Mundali QM = 1.144QK + 225.04 2623.76 1769 0.929 0.9826 0.929 0.9490

QM = 0.84197QB + 0.1996QK
Khairmal-Barmul-Mundali 1705.85 1272.12 0.970 0.990 0.934 0.9736
+ 468.72

Table 2. Relationship between discharges using ANN architecture.

RMSE (m3/s) R2 Efficiency
Between stations ANN architecture Epochs
Training Testing Training Testing Training Testing
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Khairmal-Barmul CFF 1,7,1 3000 2249.1 1132.2 0.9720 0.9891 0.9446 0.9752
Barmul-Mundali FF 1,7,1 3000 1681.5 1251.9 0.9852 0.9962 0.9705 0.9745
0.9506
Khairmal-Mundali FF 1,5,1 2000 2441.5 1741.7 0.9684 0.9789 0.9375

Khairmal-Barmul-Mundali CFF 2,15,1 4000 1697 1324.1 0.9850 0.9922 0.9700 0.9714

Copyright © 2010 SciRes. JWARP

884 A. K. KAR ET AL.

functional relationship between inputs and outputs if the with a poor correlation. The lack of correlation does not
sigmoid layer has enough neurons [30]. The ‘trainbr’ imply lack of association. Because ‘r’ measures only
function updates the weight and bias values as per Le- linear association, a strict curvilinear relationship would
venberg-Marquardt optimization. It minimizes the com- not necessarily be reflected in a high ‘r’ value. Con-
bination of squared errors and weights and then deter- versely, correlation between two variables does not
mines the correct combination so as to produce a net- guarantee that they are causatively connected [31].
work that generalizes well as per Bayesian regularization However, to have a better answer for efficiency and
process. It reduces the difficulty of determining the op- RMSE values clustering methods are adopted for the
timum network architecture also over fitting of training discharge value at base station and its corresponding
dataset is prevented, whereas in other networks gener- travel time for the calibration dataset. At least 4 clusters
alization and early stopping is required for reducing over are attempted applying K-mean and Fuzzy c-mean me-
fitting. The performance functions are set as RMSE, R2 thod using the same MATLAB code. The datasets are
and efficiency. The details of the network applied for divided into 4 clusters and their varying ranges are re-
different combinations are shown in Table 2. corded. For each range the average travel time has been
The plots of the statistical method and ANN network calculated. In both methods low peak (2660-8256 m3/s)
over the testing value of 20 numbers of observation as has large travel time (37.77 hours) to reach at delta head
obtained from MATLAB is also produced in Figures 2(a) where as high peaks has taken nearly 25 hours to reach at
to (d) respectively. delta head. As each peak is categorized the respective
For travel time the corresponding discharge and tra- travel time for different discharge values are fixed.
vel times between the stations are trailed with the same However when the result is put over the test dataset
calibration and testing data range applying both Statisti- K-mean clustering has given 80% efficiency with RMSE
cal (Table 3) and ANN architectures (Table 4). The value of 2.5 hour (Table 5). One value of testing dataset
ANN is also applied in the same way with many num- has been removed because it seems that it has some re-
bers of trials. cording errors. So travel time test series has been carried
It has been revealed that none of the case has a good out with 19 numbers of data.
efficiency, RMSE and R2. All the relations are ended A plot (Figure 3) has been drawn with observed test-

(a)

(b)

Copyright © 2010 SciRes. JWARP

 A. K. KAR ET AL. 885

(c)

(d)
Figure 2. (a) Comparison between observed and computed discharge (Khairmal-Barmul); (b) Comparison between observed
and computed discharge (Barmul-Mundali); (c) Comparison between observed and computed discharge (Khairmal-Mundali);
(d) Comparison between observed and computed discharge (Khairmal-Barmul-Mundali).

Table 3. Relationship between discharges and travel time using simple statistical method.

RMSE(hours) R2 Efficiency
Between stations Equation
Training Testing Training Testing Training Testing
-0.32
Khairmal-Barmul TB = 270.7*(QK) 3.42 2.93 0.409 0.261 0.465 0.2973

Barmul-Mundali TM = 64.96*(QB)-0.15 4.34 2.14 0.134 0.607 0.1366 0.507

Khairmal-Mundali TM = 239.1*(QK)-0.22 5.67 3.95 0.394 0.568 0.4191 0.508

where, TM = time (hour) to peak at Mundali, TB = time (hour) to peak at Barmul.

Table 4. Relationship between discharges and travel time using ANN.

RMSE (hours) R2 Efficiency

Between stations Equation
Training Testing Training Testing Training Testing

Khairmal-Barmul CFF,15,1 (2000) 3.449 2.998 0.6826 0.6259 0.4546 0.2634

Barmul-Mundali CFF,10,1 (3000) 4.350 2.173 0.3933 0.7961 0.1352 0.4906
Khairmal-Mundali FF,15,1 (2000) 5.683 3.043 0.6506 0.8306 0.4170 0.6965

Copyright © 2010 SciRes. JWARP

886 A. K. KAR ET AL.

Table 5. Peak values with corresponding travel time obtained by using different clusteringmethods

K-mean Fuzzy-C-Mean

Travel Testing Testing

Discharge range Testing Discharge range Travel time Testing
time RMSE RMSE
(m3/s) Efficiency (m3/s) (hours) Efficiency
(hours) (hours) (hours)

2660-8256 37.77 2660-8256 37.77

10329-15508 30.53 10329-16300 30
16102-22810 26.06 16499-25470 27.07
2.52 0.8011 4.27 0.4003
25470-36300 25 28102-36300 25.2

Figure 3. Plot between testing travel time and estimated by different clustering methods.

ing values and validation result of different clustering Khairmal-Mundali reach (efficiency = 80.11% and
values. The plot shows the result of K-mean clustering is RMSE = 2.55 hour). As the data recording is of 3 hour
almost close with that of observed, whereas result of interval this kind of result is still encouraging. The travel
Fuzzy C mean for some low peak values. For high peaks time between base and forecasting stations varies be-
both K-mean and Fuzzy have same results. As K-mean tween 24-37 hour. By applying clustering method justice
has able to model both high and low peaks with an effi- has been made for different magnitude of peaks. The
ciency of 80.1%,which is higher than Fuzzy C-mean results are handy to operate. The moment one gets the
methods, the discharge range values and its correspond- discharge at Khairmal (Base Station), he can immedi-
ing travel time fixed by this method may be taken as the ately calculate the discharge to be obtained at Mundali
forecast travel time between base station and forecasted (Forecasting Station) with its corresponding travel time.
station. The travel time has not calculated for other in- The result at Mundali (Forecasting Station) with respect
termediate stations as our concentration is with the peak to Khairmal (Base Station) basing on the ANN network
at base station and travel time of at least 24 hour. is best for taking the flood forecasting effectively be-
cause in taking Khairmal-Barmul and Barmul-Mundali
6. Conclusions efficiency is above 97% and RMSE value is 500 m3/s
less than Khairmal-Mundali but lead time of only 12
Both statistical and ANN methods are applied for down- hours remains after Barmul. So considering the factor of
stream Hirakud catchment between stations Khairmal, higher lead time for more relief measures we are making
Barmul and Mundali. The forecasting obtained by both a tradeoff with efficiency. Thus calculating peak dis-
the methods is encouraging. Although the ANN method charge and travel time at Mundali on the basis of Khair-
has a better performance, the existence of statistical me- mal discharge and time of occurrence will be a best
thod cannot be denied as far as discharge measurement is workable method for practicing hydrologists and field
considered. However, the adoption of clustering method engineers. Still the results are running with the limitation
in order to find travel time with respect to different peaks of data insufficiency, large recording intervals, non-con-
is another finding. The best result for travel time is for sideration of influence of other climatic variation.

Copyright © 2010 SciRes. JWARP

 A. K. KAR ET AL. 887

7. References [15] D. P. Solomatine and R. K. Price, “Innovative Approaches

to Flood Forecasting Using Data Driven and Hybrid
[1] Central Water Commission of India, “Manual on Flood Modeling,” 6th International Conference on Hydroin-
Forecasting,” 1980. formatics – Liong, World Scientific Publishing Company,
[2] E. Spokkereff, “Flood Forecasting for the River Rhine in New Jersey, 2004.
the Netherlands,” Proceeding of the Symposium the Ex- [16] C. M. Zealand, D. H. Burn and S. P. Simonovic, “Short
treme of Extremes: Extraordinary Floods, Reykjavic, Term Streamflow Forecasting Using Artificial Neural
Ireland, July 2000, pp. 347-352. Network,” Journal of Hydrology, Vol. 21, No. 4, 1999, pp.
[3] M. Bruen and J. Yang, “Functional Networks in Real- 32-48.
Time Flood Forecasting - A Novel Application,” Journal [17] J. Zhang and M. J. Hall, “Regional Flood Frequency
of Advances in Water Resources, Vol. 28, No. 9, 2005, pp. Analysis for the Gan-Ming River Basin in China,” Jour-
899-909. nal of Hydrology, Vol. 296, No. 1-4, 2004, pp. 98-117.
[4] M. Campolo, P. Andreeussi and A. Soldati, “River Flood [18] Y. Xiongrui, X. Jun and Z. Xiang, “Flood Forecasting by
Forecasting with a Neural Network Model,” Water Re- Coupling Clustering Methods and Artificial Neural Net-
sources Research, Vol. 35, No. 4, 1999, pp. 1191-1197. works,” Proceeding of Symposium HS 3006, IUGG, Pe-
[5] M. Campolo, A. Soldati and P. Andreeussi, “Artificial rugia, 2007.
Neural Network Approach to Flood Forecasting in the [19] Water Resources Department, Government of Orissa,
River Arno,” Hydrological Sciences Journal, Vol. 48, No. “Time of Flow in Mahanadi River between Hirakud &
4, 2003, pp. 381-398. Mundali,” 2009. http://www.dowrorissa.gov.in/Flood/
[6] P. Coulibaly, F. Anctil and B. Bobee, “Daily Reservoir DailyFloodBulletin.html
Forecasting Using Artificial Neural Network with Stopped [20] S. N. Ghosh, “Flood Control and Drainage Engineering,”
Training Approach,” Journal of Hydrology, Vol. 230, No. A. A. Balkema Publishers, Rotterdam, 1997.
3-4, 2000, pp. 244-257. [21] K. L. Priddy and P. E. Keller, “Artificial Neural Networks,
[7] C. W. Dawson and R. L. Wilby, “Hydrological Modeling an Introduction,” SPIE Press, Bellingham, 2005.
Using ANN,” Progress in Physical Geography, Vol. 25, [22] S. Haykin, “Neural Networks,” Macmillan College Pub-
No. 1, 2001, pp. 80-108. lishing Company, Macmillan, 1994.
[8] C. E. Imrie, S. Durucan and A. Korre, “River Flow Pre- [23] G. Cybenko, “Approximation by Superposition of a Sig-
diction Using Artificial Neural Network: Generalization moidal Function,” Journal of Mathematics Control Signal
beyond the Calibration Range,” Journal of Hydrology, System, Vol. 2, No. 4, 1989, pp. 303-314.
Vol. 233, No. 1-4, 2000, pp. 138-153. [24] K. Hornick, M. Stinchcombe and H. White, “Multilayer
[9] D. F. Lekkas, C. Onof, M. J. Lee and E. A. Baltas, “Ap- Feed forward Networks are Universal Approximators,”
plication of Neural Networks for Flood Forecasting,” Neural Networks, Vol. 2, No. 5, 1989, pp. 359-366.
Global Nest: The International Journal, Vol. 6, No. 3, [25] N. Karunanithi, W. J. Granny, D. Whitely and K. Bovee,
2004, pp. 205-211. “Neural Networks for River Flow Prediction,” Journal of
[10] A. K. Lohani, “ANN and Fuzzy Logic in Hydrological Computing in Civil Engineering, Vol. 8, No. 2, 1994, pp.
Modeling and Flow Forecasting,” Ph.D Dissertation, In- 201-220.
dian Institute of Technology, Roorkee, 2007. [26] J. B. MacQueen, “Some Methods for Classification of
[11] A. W. Minns and M. J. Hall, “Artificial Neural Networks Multivariate Observation,” Proceeding of 5th Berkely
as Rainfall-Runoff Models,” Hydrological Sciences Symposium on Probability and Statistics, University of
Journal, Vol. 41, No. 3, 1996, pp. 399-417 California Press, Berkely, 1967, pp. 281-297.
[12] J. R. Muhamad and J. N. Hassan, “Khabur River Flow [27] K-Means Algorithm, 2009. http://biocomp.bioen.uiuc.
Modeling Using Artificial Neural Network,” Al_Rafidain edu/oscar/tools/kmeans.html
Engineering, Vol. 13, No. 2, 2005, pp. 33-42. [28] J. C. Dunn, “A Fuzzy Relative of the ISODATA Process
[13] A. Mukerjee, C. Chatterji, N. Singh and N. S. Raghuvan- and its Use in Detecting Compact Well-Separated Clusters,”
shi, “Flood Forecasting Using ANN, Neuro-Fuzzy and Journal of Cybernetics, Vol. 3, No. 3, 1973, pp. 32-57.
Neuro-GA Models,” Journal of Hydrologic Engineering, [29] J. C. Bezdeck, “Pattern Recognition with Fuzzy Objective
Vol. 14, No. 6, 2009, pp. 647-652. Function Algorithm,” Plenum Press, New York, 1967.
[14] D. P. Solomatine and Y. Xue, “M5 Model Trees and [30] Neural Network Toolbox, Matlab, 2004.
Neural Networks: Application to Flood Forecasting in the [31] World Meteorological Organization (WMO), “Guide for
Upper Reach of The Huai River in China,” Journal of Hydrological Practices,” Geneva, 1994.
Hydrologic Engineering, Vol. 9, No. 6, 2004, pp. 491-501.

Copyright © 2010 SciRes. JWARP