Ahmed 2014

JKSUES 145 No.
of Pages 8
26 May 2014
Journal of King Saud University Engineering Sciences (2014) xxx, xxxxxx
1
King Saud University
Journal of King Saud University Engineering Sciences

www.ksu.edu.sa
www.sciencedirect.com
2 ORIGINAL ARTICLE
4 Prediction of dissolved oxygen in Surma River

5 inuenced by biochemical oxygen demand
6 Q1 and chemical oxygen demand using the articial
7 neural networks (ANNs)

8 Q2 A.A. Masrur Ahmed *
9 Department of Civil Engineering, Leading University, Sylhet, Bangladesh
10 Received 23 December 2013; accepted 13 May 2014

11
13 KEYWORDS Abstract The objective of this study is to develop a feed forward neural network (FFNN) model
14
15 Radial basis function neural and a radial basis function neural network (RBFNN) model to predict the dissolved oxygen from
16 Q3 network; biochemical oxygen demand (BOD) and chemical oxygen demand (COD) in the Surma River, Ban-
17 Feed forward neural net- gladesh. The neural network model was developed using experimental data which were collected
18 work; during a three year long study. The input combinations were prepared based on the correlation
19 Dissolved oxygen; coefcient with dissolved oxygen. Performance of the ANN models was evaluated using correlation
20 Surma River coefcient (R), mean squared error (MSE) and coefcient of efciency (E). It was found that the
ANN model could be employed successfully in estimating the dissolved oxygen of the Surma River.
Comparative indices of the optimized RBFNN with input values of biochemical oxygen demand
(BOD) and chemical oxygen demand (COD) for prediction of DO for testing array were
MSE = 0.465, E = 0.905 and R = 0.904 and for validation array were MSE = 1.009, E = 0.966
and R = 0.963. Comparing the modeled values by RBFNN and FFNN with the experimental data
indicates that neural network model provides reasonable results.
21 2014 Production and hosting by Elsevier B.V. on behalf of King Saud University.
1. Introduction 22
The concentration of dissolved oxygen (DO) reects 23

equilibrium between oxygen-producing processes (e.g. photo- 24
* Tel.: +880 1912140065; fax: +880 821720307. synthesis) and oxygen-consuming processes (e.g. aerobic 25
E-mail addresses: engr.masrur@outlook.com, engr.masrur@gmail. respiration, nitrication, and chemical oxidation) and depends 26
com. on many factors such as temperature, salinity, oxygen deple- 27
Peer review under responsibility of King Saud University. tion, sources of oxygen and other water quality parameters. 28
Therefore it is very desirable to create a DO model of the 29
Surma River so that water quality can be optimized through- 30
Production and hosting by Elsevier out a time period. 31
1018-3639 2014 Production and hosting by Elsevier B.V. on behalf of King Saud University.
http://dx.doi.org/10.1016/j.jksues.2014.05.001
Please cite this article in press as: Ahmed, A.A.M. Prediction of dissolved oxygen in Surma River inuenced by biochemical oxygen demand and
Q1 chemical oxygen demand using the articial neural networks (ANNs). Journal of King Saud University Engineering Sciences (2014), http://
dx.doi.org/10.1016/j.jksues.2014.05.001
JKSUES 145 No. of Pages 8
26 May 2014
2 A.A. Masrur Ahmed
32 The water quality models such as River and Stream Water the three data sets were 1.50 for training, 1.44 for validation, 84
33 Quality Model (QUAL2K) (Chapra and Pellettier, 2003) and and 1.23 for testing. 85
34 Water Quality Analysis Simulation Program (WASP) (Wool The aim of this paper is to construct a feed forward neural 86
35 et al., 2006) are very complicated as they require more informa- network (FFNN) model and a radial basis function neural net- 87
36 tion of the river system. It makes sense to utilize articial intel- work (RBFNN) model to predict the dissolved oxygen in the 88
37 ligence (AI) techniques to derive the crucial information about Surma River, Bangladesh and demonstrate its application to 89
38 the river water quality. Articial neural networks have been identifying complex nonlinear relationships between input 90
39 increasingly used in the prediction of water quality variables and output variables. The Surma River has been preferred 91
40 (Nash and Sutcliffe, 1970; French et al., 1992; Zhu and due to its importance in water supply as well as its high ecolog- 92
41 Fujita, 1994; Yi-Ming et al., 2003). An ANN learns to solve a ical importance. The proposed model may contribute to more 93
42 problem by developing a memory capable of correlating a large efcient management as well as to preventive activities. 94
43 number of input patterns with a resulting set of yields. ANN
44 models mimic somewhat the learning process of a human brain.
2. Materials and methods 95
45 They operate like a black box model, requiring no detailed
46 information about the system (Ahmed et al., 2013). Another
47 advantage of using ANNs is their ability to handle large and 2.1. Study area and water quality data 96
48 complex systems with many interrelated parameters.

49 Articial neural networks (ANNs) have been successfully The Surma River in Sylhet was selected as the study area for 97
50 applied in a number of studies focusing on water quality pre- ANN applications. The Surma River is located in the north- 98
51 diction in rivers (Niroobakhsh et al., 2012; Singh et al., eastern region of Bangladesh within the administrative districts 99
52 2009), lakes (Stefan et al., 1995), reservoirs (Rankovic et al., of Sylhet and Sunamganj. The Surma River has many tributar- 100
53 2010; Kuo et al., 2007) and waste water treatment plants ies and hilly streams joining it from the KhasiJaintia Hills of 101
54 (Abyaneh, 2014). Elhatip and Komur (2008) modeled the the Indian state of Meghalaya. The Surma River ows through 102
55 water quality parameters in the reservoir of Mamasin Dam. the central part of the Sylhet city, plays a very important role 103
56 The results illustrated the ability of ANNs to predict the dis- in the economic development of the city. The natural canals 104
57 solved oxygen of the recharge and discharge areas at the passing through the city are responsible for surface runoff con- 105
58 Mamasin Dam site. Li et al. (2007) used the ANN technique veyance from its urban catchments to the receiving Surma 106
59 to predict the Microcystis spp. population of Lake Dianchi River. However, the water quality of the Surma River is being 107
60 in China. They obtained a correlation of determination (R2) deteriorated day by day due to human activities and industrial 108
61 of 0.911 between the measured and predicted data of Micro- efuents, which are built up on its bank (Ahmed et al., 2010; 109
62 cystis spp. However, Karul (1999) used total phosphates Alam et al., 2007). 110
63 (TP), NO3N, NH3N, water temperature, electrical conduc- The data set utilized in this study was produced through 111
64 tivity, pH, turbidity, secchi depth and suspended solids as monitoring of the water quality of the Surma River. Monthly 112
65 input variables to predict the chlorophyll-a concentration. In sampling was carried out during the period of three years 113
66 his study, he found correlation coefcient of 0.92. (20102012). Four sampling sites are identied in Fig. 1 and 114
67 Niroobakhsh et al. (2012) used two ANN networks, multi- Table 1. One was at the center of the city, where the depth var- 115
68 layer perceptron (MLP) and radial basis function (RBF) to ied from 2 to 10 m, depending on the water level. The second 116
69 compute the total dissolved solid (TDS) concentrations for was in the downstream of the central part with a depth ranging 117
70 the Jajrood River of Iran. In their study, they found that from 2.3 to 15 m. The third was near the Shahjalal Bridge, with 118
71 MLP and RBF are able to simulate water quality variables a depth ranging from 3 to 9 m, about 1 km from the central 119
72 of Jajrood River with more than 90% accuracy. Soyupak part. The fourth was in the upstream part, with a depth rang- 120
73 et al. (2003) used the neural network model to predict the ing from 2 to 8 m, 5 km away from the city. For the analysis, 121
74 pseudo steady state time dependent and space dependent dis- 160 samples were selected for the study. The water quality 122
75 solved oxygen concentrations in three different reservoirs. parameters measured were: chemical oxygen demand (COD), 123
76 The correlation coefcient was more than 0.95 both for the biochemical oxygen demand (BOD) and dissolved oxygen 124
77 measured and model computed output variables. Singh et al. (DO). The chemical analyses were performed using standard 125
78 (2009) computed DO and BOD levels in the Gomti River in methods (APHA, 1995). The basic statistics of the measured 126
79 India using three-layer feed forward neural networks with water quality variables in the Surma River and correlation 127
80 back propagation learning. The coefcient of determination coefcient with DO are presented in Table 2. As there are only 128
81 for modeled values and observed DO values were 0.70, 0.74, two independent variables and BOD has a very signicant 129
82 and 0.76 for the training, validation and test sets, respectively. inuence over DO, BOD has been selected as an individual 130
83 The respective values of root means squared error (RMSE) for input variable for RBFN-II and FFNN-II models. Moreover 131
Table 1 Global Positioning of the sampling points.

Point Location Latitude Longitude
a Kajir Bazar 24530 16.3200 N 91510 30.4400 E
b Keane Bridge 24530 17.1600 N 91520 03.3000 E
c Shahjalal Upashahar 24520 57.2600 N 91520 43.6200 E
d Burhanudins Mazar 24520 37.4000 N 91530 36.7000 E
26 May 2014
Prediction of dissolved oxygen in Surma River 3
Figure 1 Sampling locations of water samples of the Surma River.
Table 2 Basic Statistics of the measured water quality variables in the Surma River.
Variable Unit Minimum Maximum Mean SD CDO
BOD mg/l 0.6 17.3 3.79 2.86 0.727
COD mg/l 2.2 19.43 6.65 2.11 0.557
DO mg/l 1.9 17.30 5.40 2.45 1.000
SD: standard deviation; CDO: correlation with DO.
Q7 Table 3 Data Combination using correlation co-efcient between DO other variables.

Input name Data combination Network architecture
FFNN-I BOD + COD 2-30-1
FFNN-II BOD 1-20-1
RBFN-I BOD + COD 2-55-1
RBFN-II BOD 1-65-1
143
132 the RBFN-I and FFNN-I models were constructed with both The output of a neuron can be expressed as : out fn 1 145
133 Q4 variables (Table 3). 146
X
R
Where n xj xj b; 2
148
134 2.2. Articial neural network model j1
149
135 The articial neural networks are composed of a set of arti- x1, x2, . . .. . ., xR are the input signals; 150
136 cial neurons which are inspired by biological systems. The x1, x2, . . .. . ., xR are the weights of the neuron; 151
137 model of a neuron is represented in Fig. 2. Back propagation b is bias value; and 152
138 (BP) is a gradient descent algorithm in which the gradient is f (n) is the activation function. 153
139 computed for nonlinear multilayer networks. The ANN 154
140 parameters (weights and biases) can be adjusted to minimize The linear and sigmoid are the most common used activa- 155
141 the sum of the squares of the differences between the actual tion functions in the construction of articial neural networks 156
142 values and network output values. (Rankcovic et al., 2010). 157
26 May 2014
4 A.A. Masrur Ahmed
Figure 2 A typical multilayer perceptron ANN architecture (Ahmed et al., 2013).
158 The linear function is written as the hidden unit output function. The RBF meta model is based 200
159
161 fn n 3 on radial basis functions using cones (circular hyperboloid), 201
and it is mathematically represented as follows: 202
203
162 and the logistic sigmoid function is dened as X
n
163
1 fx wi ;jjx xi jj 7
fn 4 i1 205
165 1 en
where n is the number of sampling points; x is the vector of 206
166 However, Haykin (1999) identied a sigmoid function that
input variables; x is the center of basis function /, |||| is any 207
167 can be used as a hyperbolic tangent function
168 lp norm (typically is Euclidean norm, this kind of norm is used 208
1 en in this study) and w is the unknown weighting coefcient. 209
fn 5
170 1 en Therefore, an RBF is actually a linear combination of n basis 210
171 The output y at linear output node can be calculated as: functions with weighted coefcients. 211
172 RBF can be expressed as a matrix format:
0 P 1 212
R 213
xjx
XB
z b
i;j1 i1
C f Ak 8 215
Bx1;i2 1 e
j1
y @ PR C
A b12 6 216
x
174
i0
1e j1 jxi;j1 bi1 where f fx1; fx2; . . . ::; fxmT ; Aij Ujjxi xj jji
175 where R is the number of inputs, z is the number of hidden 1; 2; . . . :; m; j 1; 2; . . . ; n 9 218
176 neurons, xi.j(1) is the rst layer weight between the input j The coefcient vector k is obtained by solving Eq. (9). An 219
177 and the ith hidden neuron, x1,i(2) is the second layer weight RBF using the aforementioned highly nonlinear functions does 220
178 between the ith hidden neuron and output neuron, bi(1) is a not work well for linear responses. To solve this problem, we 221
179 biased weight for the ith hidden neuron and b1(2) is a biased can augment an RBF by including a polynomial function such 222
180 weight for the output neuron. Feed forward neural networks that. 223
181 propagate data linearly from input to output and they are 224
X
n X
m
182 the most popular and most widely used models in many fx wi ;jjx xi jj cj pj x 10
183 practical applications (Rankovic et al., 2010). In this paper, i1 j1 226
184 LevenbergMarquardt algorithm was used as the training
185 algorithm and log-sigmoidal (logsig) was chosen for the activa- where m is a total number of terms in the polynomial, and c 227
186 tion function. (j = 1,2, . . ., m) is the corresponding coefcient. 228
187 A radial basis function network has a feed-forward struc- In this paper, 70% of the total data sets are assigned for 229
188 ture consisting of a single hidden layer for a given number of training, 15% data sets are allocated for testing and last 230
189 locally tuned units which are fully interconnected to an output 15% data sets are assigned for the validation array. The over- 231
190 layer of linear units (Dibike et al., 1999; Mason et al., 1996). all performance of the model was evaluated based on the per- 232
191 Learning in RBF network is carried out in two phases: rst, formances of all three modes. 233
192 for the hidden layer, and then for the output layer. The hidden
193 layer is self-organizing; its parameters depend on the distribu- 2.3. Performance determination parameters 234
194 tion of the inputs, not on the mapping from the input to the
195 output. The output layer, on the other hand, uses supervised In the research training of ANN models of different architec- Q5 235
196 learning to set its parameters. A RBF hidden unit has one tures applied an automatic performance analysis of the net- 236
197 parameter associated with each input unit. These parameters works based on the correlation coefcient (R), mean squared 237
198 w are not weights placed on the input; rather they are the error (MSE) and coefcient of efciency (E) was performed. 238
199 co-ordinates in input space of a point, that is, the center of The R value indicates the strength and direction of a linear 239
26 May 2014
1. Correlation coefcient is dened as the degree of 247

Table 4 Performance parameters of the articial neural
correlation between the experimental and modeled values 248
network models.
(Rankovic et al., 2010). 249
Model Mode E MSE R
RBFN-I Testing 0.905 0.465 0.904 250
v 251
Validation 0.966 1.009 0.963 u N
X
N uX X N
Whole 0.936 0.654 0.944
R yk ytk z=t yk y2 tk z2 11
RBFN-II Testing 0.846 1.181 0.861
k1 k1 k1 253
Validation 0.855 0.942 0.897
Whole 0.892 1.190 0.892 254
FFNN-I Testing 0.997 2.199 0.918 2. Mean squared error (MSE) measures the average of the 255
Validation 0.862 0.693 0.898 squares of the errors. The smaller values of MSE ensure 256
Whole 0.936 0.709 0.936
the better performance. The MSE is calculated as: 257
FFNN-II Testing 0.938 0.550 0.939
Validation 0.855 2.925 0.854
258
Whole 0.880 1.478 0.870 259
1 XN
MSE tk yk 2 12
N k1 261
240 relationship between two variables. Therefore, an initial
241 screening of different ANN models was conducted. Then, the 262
242 ANN models were further examined to decide which one is 3. The coefcient of efciency (E) has been widely used to 263
243 the superlative. For this reason, visual inspection of time-series evaluate the performance of hydrologic models. Nash and 264
244 plots of measured and predicted DO was performed. For the Sutcliffe (1970) dened the coefcient of efciency which 265
245 performance analysis, the following parameters were ranges from minus innity to 1.0, with higher values indi- 266
246 calculated for each ANN model. cating better agreement. It is calculated as: 267
18 Observed Value
16 FFNN-I
Dissolved Oxygen (mg/l)
14 FFNN-II
12
10
8
6
4
2
1
6
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
101
106
111
116
121
126
131
136
141
146
151
156
Obserbaon
Figure 3 The observed and modeled dissolved oxygen values in whole array using feed forward neural network.
(a) y = 0.9156x + 0.5166 (b) y = 0.8153x + 1.2949

20 R = 0.8757 18 R = 0.7561
18 16
16 14
14 12
12
Outputs
Outputs
10
10
8
8
6
6
4 4
2 2
0 0
0 10 20 0 10 20
Targets Targets
Figure 4 Scatter plot of observed versus modeled dissolved oxygen concentration for feed forward neural network (a) FFNN-I; (b)
FFNN-II.
26 May 2014
6 A.A. Masrur Ahmed
268 forward neural network and radial basis function neural 289
269
PN 2 network have been illustrated in Figs. 3 and 5. However, the 290
tk yk
E 1 Pk1
N 2
13 scatter plot of observed versus modeled dissolved oxygen 291
271 k1 tk z concentration for feed forward neural network and radial basis 292
function neural network is shown in Figs. 4 and 6. 293
272 where, yk and tk denote the network output and measured The model RBFN-I with all the input parameters is found 294
273 value from the kth element; y and z denote their average as the most appropriate model for dissolved oxygen prediction 295
274 respectively, and N represents the number of observations. with high correlation coefcient (R), high coefcient of 296
efciency (E) and low mean squared error (MSE) value. The 297
275 3. Results and discussion respective correlation coefcient for RBFN-I model is 0.904, 298
0.963 and 0.944 for testing, validation and whole array. How- 299
276 The ANN models were trained utilizing different numbers of ever, the MSE and E for the three data sets are 0.465 and 0.905 300
277 neuron in the hidden layer, for training, testing, validation for testing, 1.009 and 0.966 for validation, and 0.654 and 0.936 301
278 and whole array (training + testing + validation) for an input for whole set. The prediction ability of RBFN-I is almost sim- 302
279 combination. The ANN models that had a higher R value were ilar to Ying et al. (2007) who found the correlation coefcient 303
280 tabulated and three performance determination parameters as 0.94 between the measured and modeled DO values and 304
281 were calculated to check the DO prediction capability of two Soyupak et al. (2003) where they found a correlation coef- 305
282 selected ANN models with different network architecture. cient of 0.950 between the measured and modeled variables. 306
283 Among the selected two ANN models, the developed ANN The FFNN-II and RBFN-II models constructed with only 307
284 model with RBFNN simulated the DO concentrations of the BOD showed similar performance for dissolved oxygen predic- 308
285 Surma River more accurately when compared to that with tion with high correlation coefcient and coefcient of 309
286 FFNN architecture. The performance parameters for the best efciency for all arrays (training, testing, validation and whole 310
287 DO forecast ANN models are shown in Table 4. The observed array). The respective correlation coefcient for RBFN-II 311
288 and modeled dissolved oxygen values in whole array using feed model is 0.861, 0.897 and 0.892 for testing, validation and 312
18 Observed Value
RBFN-I
16
RBFN-II
Dissolved Oxygen (mg/l)
14
12
10
2
1
6
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
101
106
111
116
121
126
131
136
141
146
151
156
Observaon
Figure 5 The observed and modeled dissolved oxygen values in whole array using radial basis neural network.
(a) y = 0.9526x + 0.39 (b) y = 0.8518x + 0.9724

20 R = 0.8906 18 R = 0.7964
18 16
16 14
14
12
12
Outputs
Outputs
10
10
8
8
6 6
4 4
2 2
0 0
0 10 20 0 10 20
Targets Targets
Figure 6 Scatter plot of observed versus modeled dissolved oxygen concentration for radial basis function neural network (a) RBFN-I;
(b) RBFN-II.
26 May 2014
313 whole array accordingly. However, the MSE and E for the two dissolved oxygen concentration for Surma River, Bangladesh. J. 371
314 data sets are 1.181 and 0.846 for testing, 0.942 and 0.855 for Appl. Technol. Environ. Sanitation 3 (3), 135140. 372
315 validation, and 1.19 and 0.892 for whole set. However, another Alam, M.J.B., Muyen, Z., Islam, M.R., Islam, S., Mamun, M., 2007. 373
316 ANN model of feed forward neural network, FFNN-II has Water quality parameters along rivers. Int. J. Environ. Sci. 374
Technol. 4 (1), 159167. 375
317 coefcient of efciency (E) values of 0.938, 0.855 and 0.880
APHA, 1995. Standard methods for the examination of water and 376
318 for testing, validation and whole sets respectively. Moreover, wastewater, 19th ed., American Public Health Association, Wash- 377
319 a high value of MSE has been found for the validation array. ington, DC. 378
320 The results of this study can be compared with Kuo et al. Chapra, S., Pellettier, G., 2003. QUAL2K: A modeling framework for 379
321 (2007) who applied an ANN model for predicting the dissolved simulating river and stream water quality. Documentation and 380
322 oxygen in the Te-Chi reservoir. The correlation coefcients for users manual. Civil and Environmental Engineering Dept., Tufts 381
323 modeled values and observed DO values were 0.75 and 0.72 for University, Medford, MA. Available from: <http://www.epa.gov/ 382
324 training and test data sets, accordingly. Similarly, Singh et al. athens/wwqtsc/html/qual2k.html>. 383
325 (2009) found coefcient of determination for modeled values Dibike, Y.B., Solomatine, D.P., Abbott, M.B., 1999. On the encap- 384
326 and observed DO values to be 0.85, 0.85, and 0.77, for the sulation of numerical-hydraulic models in articial neural network. 385
J. Hydraulic Res. 37 (2), 147161. 386
327 training, validation and test sets, respectively. They utilized a
Elhatip, H., Komur, M.A., 2008. Evaluation of water quality 387
328 three-layer feed forward neural network with back propaga- parameters for the Mamasin Dam in Aksaray city in the central 388
329 tion learning to predict DO and BOD levels for Gomti River Anatolian part of Turkey by means of articial neural networks. 389
330 (India). Rankovic et al. (2010) applied articial neural network Environ. Geol. 53, 11571164. 390
331 to develop a feed forward neural network (FFNN) model to French, M.N., Krajewski, W.F., Cuykendall, R.R., 1992. Rainfall 391
332 predict the dissolved oxygen in the Gruza Reservoir, Serbia. forecasting in space and time using a neural network. J. Hydrol. 392
333 They used the LevenbergMarquardt algorithm to train the 137, 131. 393
334 FFNN. In their study they found the coefcient of correlation Haykin, S., 1999. Neural Networks: A Comprehensive Foundation. 394
335 for the test set to be 0.848 and 0.840 for the whole data set. The Prentice-Hall, New Jersey, p. 842. 395
336 respective values of root means squared error and bias for the Karul, C., 1999. Development of an articial neural network model for 396
the estimation of chlorophyll-a in lakes, Doctor of Philosophy, and 397
337 three data sets are 2.25 and 0.14 for training, 1.84 and 0.87 for
thesis submitted to the Department of Environmental Engineering, 398
338 validation, and 1.38 and 0.22 for testing. METU, Ankara. 399
339 The paper revealed that the proposed ANN models with Kuo, J.T., Hsieh, M.H., Lung, W.S., She, N., 2007. Using articial 400
340 minimum input parameters such as biochemical oxygen neural network for reservoir eutrophication prediction. Ecol. 401
341 demand and chemical oxygen demand could be successfully Model. 200, 171177. 402
342 used for predicting BOD and COD concentrations. Li, H., Hou, G., Dakui, F., Xiao, B., Song, L., Liu, Y., 2007. 403
Prediction and elucidation of the population dynamics of Mycro- 404
343 4. Conclusions cystis spp. in Lake Dianchi (China) by means of articial neural 405
networks. Ecol. Inform. 2, 184192. 406
Mason, J.C., Price, R.K., Temme, A., 1996. A neural network model 407
344 In this paper, ANN models were developed to predict dis- of rainfall-runoff using radial basis functions. J. Hydraulic Res. 34 408
345 solved oxygen in the Surma River. The proposed model shows (4), 537548. 409
346 efciency in forecasting the dissolved oxygen concentration in Nash, J.E., Sutcliffe, J.V., 1970. River ow forecasting through 410
347 water bodies. The results showed that the radial basis function conceptual models, part I, a discussion of principles. J. Hydrol. 10 411
348 neural network model prepared by BOD and COD provided (3), 282290. 412
349 high correlation coefcient (R), high coefcient of efciency Niroobakhsh, M.S.H., Musavi-Jahromi, S.H., Manshouri, M., Sedghi, 413
S., 2012. Prediction of water quality parameter in Jajrood River 414
350 (E) and low mean squared errors (MSE) value for all three
basin: Application of multi-layer perceptron (MLP) perceptron and 415
351 modes. However, the feed forward neural networks with two
radial basis function networks of articial neural networks (ANNs) 416
352 variables provided convincing results. It has been conrmed in African Journal of Agricultural Research, 7 (29), 41314139. 417
353 that DO in the Surma River can be predicted with acceptable Rankovic, V., Radulovi, J., Radojevic, I., Ostojic, A., Comic, L., 2010. 418
354 accuracy from a small set of variables using feed forward Neural network modeling of dissolved oxygen in the Gruza 419
355 neural network (FFNN) and radial basis function neural net- reservoir, Serbia. Ecol. Model. 221, 12391244. 420
356 work (RBFNN). Dissolved oxygen is an important parameter Singh, K.P., Basant, A., Malik, A., Jain, G., 2009. Articial neural 421
357 for the water bodies. The obtained result may be applied to network modeling of the river water qualitya case study. Ecol. 422
358 power DO estimations to be utilized in water management Model. 220, 888895. 423
359 and treatment systems. Soyupak, S., Karaer, F., Gurbuz, H., Kivrak, E., Senturk, E., Yazici, A., 424
2003. A neural network-based approach for calculating dissolved 425
oxygen proles in reservoirs. Neural Comput. Appl. 12, 166172. 426
360 References Stefan, H.G., Fang, X., Wright, D., Eaton, J.G., McCormick, J.H., 427
1995. Simulation of dissolved oxygen proles in a transparent, 428
361 Q6 Abyaneh, H.Z., 2014. Evaluation of multivariate linear regression and Dimictic Lake. Limnol. Oceanogr. 40, 105118. 429
362 articial neural networks in prediction of water quality parameters. Wool, T.A., Ambrose, R.B., Martin, J.L., Comer, E.A., 2006. Water 430
363 J. Environ. Health Sci. Eng. 12 (40), 18. Quality Analysis Simulation Program (WASP) Version 6.0 431
364 Ahmed, M., Ahmed, A.A.M., Mazumder, R.K., 2010. Deterioration DRAFT: Users Manual. U.S. Environmental Protection Agency, 432
365 of water quality of Surma River inuenced by Natural Canals Athens, GA. Available from: <http://www.epa.gov/athens/ 433
366 Passing through Sylhet City of Bangladesh, Proceedings Interna- wwqtsc/html/wasp.html>. 434
367 tional Conference on Environmental Aspects of Bangladesh (Sept. Yi-Ming, K., Chen, W.L., Kao-Hung, L., 2003. Evaluation of the 435
368 4) pp. 182185. ability of an articial neural network model to assess the variation 436
369 Ahmed, A.A.M., Hossain, M.I., Rahman, M.T., Chowdhury, M.A.I., of groundwater quality in an area of Blackfoot disease in Taiwan. 437
370 2013. Application of articial neural network models for predicting Water Res., 18. 438
26 May 2014
8 A.A. Masrur Ahmed
439 Ying, Z., Jun, N., Fuyi, C., Liang, G., 2007. Water quality forecast Zhu, M.L., Fujita, M., 1994. Comparison between fuzzy reasoning and 442
440 through application of BP neural network at Yuquio reservoir. J. neural networks methods to forecast runoff discharge. J. Hydro 443
441 Zhejiang Univ. Sci. 8, 14821487. Sci. Hydraulic Eng. 12 (2), 131141. 444
445
446

Ahmed 2014

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Ahmed 2014

Hochgeladen von

Copyright:

Verfügbare Formate

JKSUES 145 No.

King Saud University

Journal of King Saud University Engineering Sciences

4 Prediction of dissolved oxygen in Surma River

7 neural networks (ANNs)

9 Department of Civil Engineering, Leading University, Sylhet, Bangladesh

10 Received 23 December 2013; accepted 13 May 2014

The concentration of dissolved oxygen (DO) reects 23

Production and hosting by Elsevier out a time period. 31

48 complex systems with many interrelated parameters.

Table 1 Global Positioning of the sampling points.

Figure 1 Sampling locations of water samples of the Surma River.

Q7 Table 3 Data Combination using correlation co-efcient between DO other variables.

Figure 2 A typical multilayer perceptron ANN architecture (Ahmed et al., 2013).

175 where R is the number of inputs, z is the number of hidden 1; 2; . . . :; m; j 1; 2; . . . ; n 9 218

186 tion function. (j = 1,2, . . ., m) is the corresponding coefcient. 228

1. Correlation coefcient is dened as the degree of 247

(a) y = 0.9156x + 0.5166 (b) y = 0.8153x + 1.2949

(a) y = 0.9526x + 0.39 (b) y = 0.8518x + 0.9724

Das könnte Ihnen auch gefallen