River Water Quality Modelling Using Artificial Neural ...

Available online at www.sciencedirect.com

ScienceDirect Aquatic Procedia 4 (2015) 1070 – 1077

INTERNATIONAL CONFERENCE ON WATER RESOURCES, COASTAL AND OCEAN ENGINEERING (ICWRCOE 2015)

River Water Quality Modelling using Artificial Neural Network Technique Archana Sarkara*, Prashant Pandeyb a

National Institute of Hydrology, Roorkee-247667, Uttarakhand India b Doon University, Dehradun, Uttarakhand, India

Abstract

Dissolved oxygen (DO) concentrations have been used as primary indicator of stream water quality. A problem of great social importance is determining how to best retain the quality of stream water and maintain DO concentrations using various pollution control activities. This paper presents the use of artificial neural network (ANN) technique to estimate the DO concentrations at the downstream of Mathura city, India, located at the bank of River Yamuna in the state of Uttar Pradesh, India. In the analysis, the most commonly used feed forward error back propagation neural network technique has been applied. Monthly data sets on flow discharge, temperature, pH, biochemical oxygen demand (BOD) and dissolved oxygen (DO) at three locations, namely, Mathura (upstream), Mathura (central) and Mathura (downstream) have been used for the analysis. Feed forward error back propagation algorithm, the most commonly used ANN technique, was used to develop three types of ANN models using different combinations of input variables and input stations, namely: (a) All the data sets for stations Mathura (upstream), Mathura (central) and Mathura (downstream) except DO values at Mathura (downstream) (b) All data sets for the stations Mathura (upstream), and Mathura (central), and (c) All the data sets for the stations Mathura (upstream). The performance of the ANN technique has been evaluated using statistical tools (in terms of root mean square error and coefficient of correlation). The predicted values of DO showed prominent accuracy by producing high correlations (upto 0.9) between measured and predicted values. © by Elsevier B.V.by This is an open © 2015 2015Published The Authors. Published Elsevier B.V.access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of organizing committee of ICWRCOE 2015. Peer-review under responsibility of organizing committee of ICWRCOE 2015 Keywords: Artificial Neural Network,,Water Quality, Dissolved Oxygen, River Yamuna

* Corresponding author. E-mail address: [email protected]

2214-241X © 2015 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of organizing committee of ICWRCOE 2015 doi:10.1016/j.aqpro.2015.02.135

Archana Sarkar and Prashant Pandey / Aquatic Procedia 4 (2015) 1070 – 1077

1071

1. Introduction The problem of water quality management plays an important role in water pollution control and river basin planning. The possibility of a pollutant being discharged to the river as municipal and industrial waste is a constant concern to those diverting and using water from rivers. Dissolved oxygen (DO) is a primary indicator of the river water quality and has received lot of attention in the literature in recent past. The resultant DO concentration at any location in a river is the result of many processes occurring at the upstream location, which include, de-oxygenation, re-aeration, photosynthesis, respiration, sediment oxygen demand, water temperature and the discharge. The two cities namely, Mathura and New Delhi, located at the at the bank of River Yamuna in Uttar Pradesh and Delhi, India, discharge significant amount of municipal and industrial waste to the River Yamuna without prior treatment in most of the cases. The quality of the River Yamuna has been sinking in recent past. As greater demands are placed on River Yamuna, the ability to simulate potential pollution build up in river becomes increasingly important. In order to enforce water quality standards (that is to ensure that the maximum allowable concentration of a substance in a given water body is not exceeded), river water quality models are used extensively in research as well as in the design. However, many models are based on the assumption of linearity functions. For prediction of dissolved oxygen in a stream under scenarios of interest, different deterministic models have been attempted in the past. However, in practice the statistical accuracy of the models commonly is poor because natural systems tend to be too complex for state of the art deterministic modelling methods. ANNs provide a quick and flexible means of creating models for estimation of stream water quality. In recent years ANNs have shown exceptional performance as regression tools, especially when used for pattern recognition and function estimation. An Artificial Neural Network (ANN) is a computational method inspired by the studies of the brain and nervous system in biological organisms. ANNs represent highly idealized mathematical models of our present understanding of such complex systems. One of the characteristics of the neural networks is their ability to learn. A neural network is not programmed like a conventional computer program, but is presented with examples of the patterns, observations and concepts, or any type of data, which it is supposed to learn. Through the process of learning (also called training) the neural network organizes itself to develop an internal set of features that it uses to classify information or data. In comparison to the conventional methods, ANNs tolerate imprecise or incomplete data, approximate results, and are less vulnerable to outliers. They are highly parallel, i.e., their numerous independent operations can be executed simultaneously. Due to its massively parallel processing architecture the ANN is capable of efficiently handling complex computations, thus making it the most preferred technique today for high speed processing of huge data. In addition, there are many advantageous characteristics of ANN approach to problem solving viz.: (1) application of a neural network does not require a priori knowledge the underlying process; (2) one may not recognize all the existing complex relationships between various aspects of the process under investigation; (3) a standard optimisation approach or statistical model provides a solution only when allowed to run to completion whereas a neural network always converges to an optimal (sub-optimal) solution condition and; (4) neither constraints nor an a priori solution structure is necessarily assumed or strictly enforced in the ANN development. These characteristics render ANNs to be very suitable tools for handling various hydrological modelling problems). However, the number of uses for ANNs is increasing rapidly and in recent years they have been successfully used for the prediction of economic, water resources, water quality and hydrologic time series (Chakraborty et al., 1992; Windsor and Harker, 1990; Daniell, 1991; DeSilets et al., 1992; Lachtennacher and Fuller, 1994; Karunanithi et al., 1994; Schizas et al., 1994). The problem of water quality management plays an important role in water pollution control and river basin planning. In this paper, ANNs are applied for simulating the DO values in River Yamuna at the downstream of Mathura city, which is receiving tremendous amount of municipal and industrial wastes from the urban area and non-point source pollutants from its surroundings. The architecture of ANN used in the present work are discussed in the next section.

1072


2. Background of Artificial Neural Networks (ANNs) 2.1 Architecture of ANNs An ANN is a computing system made up of a highly interconnected set of simple information processing elements, analogous to a neuron, called units. The neuron collects inputs from both a single and multiple sources and produces output in accordance with a predetermined non-linear function. An ANN model is created by interconnection of many of the neurons in a known configuration. The primary elements characterizing the neural network are the distributed representation of information, local operations and non-linear processing. The learning process or training forms the interconnection between neurons and is accomplished by using known inputs and outputs, and presenting these to the ANN in some ordered manner. The strength of these interconnections is adjusted using an error convergence technique so that a desired output will be produced for a known input pattern. Several ANN architectures have been reported in the literature, namely, Rosenblatt's perceptron (Rosenblatt, 1961), ADALINE (Widrow and Hoff, 1960), Error-back propagation (Rumelhart et al., 1986), Hopfield's network ( Hopfield and Tank, 1985), Self organizing network (Kohonen, 1988), ART (Grossberg, 1982) and a few other ANN architectures. In the present study, the feed forward error back propagation algorithm (Rumelhart et al., 1986) is used for ANN training. In most of the ANN applications as seen in the literature, the feed forward error back propagation neural network is the most popular and widely used ANN architecture. The ANN network used for the present study is shown in Figure 1. There are three basic layers or levels of data processing units viz., the input layer, the hidden layer and the output layer. Each of these layers consists of processing units called nodes of the neural network. The interconnections between nodes of different layers are called weights of the neural network. These weights are updated or modified iteratively using the generalized delta rule or the steepest gradient descent principle (ASCE Task Committee, 2000).

Inputs

Output

Input layer

Hidden layer

Output layer

Fig. 1: Structure of a multi-layer feed forward Artificial Neural Network model

2.2 Training of ANN An ANN stores the knowledge about the problem in terms of weights of inter-connections. The process of determining ANN weights is called learning or training. The ANNs are trained with a training set of input and known output data. At the beginning of training, the initial value of weights can be assigned randomly or based on experience. The weights are systematically changed by the learning algorithm such that for a given input, the difference between the ANN output and the actual output is small. Many learning examples are repeatedly presented


1073

to the network, and the process is terminated when this difference is less than a specified value. At this stage, the ANN is considered trained. An ANN is better trained as more input data are used. The number of input nodes, output nodes and the nodes in the hidden layer depend upon the problem being studied. If the number of nodes in the hidden layer is small, the network may not have sufficient degrees of freedom to learn the process correctly. If the number is too high, the training will take a long time and the network may sometimes over-fit the data (Karunanithi et al., 1994). After training is over, the ANN performance is validated. Depending on the outcome, either the ANN has to be re-trained or it can be implemented for its intended use. ANN theory has been described in many books such as Vemuri (1992) and Yegnanarayana (1999). 2.3 Performance evaluation A large number of statistical criteria are available to compare the goodness/adequacy of any given model. The performance evaluation statistics used for ANN training in the present work are root mean square error (RMSE), coefficient of correlation (R) and determination coefficient (DC). These parameters have been determined using the following equations (Neural Power, 2003): n

¦ Qi qi i 1

RMSE

(1)

n

¦ Q

Q qi q

¦ Q

q

n

i

R

2

i 1 n

i

Q

2

q

i

(2)

2

i 1

¦ Q Q ¦ Q q n

n

2

2

i

DC

i

i 1

i 1

¦ Q n

i

Q

(3)

2

i 1

where

Q

1 n ¦ Qi , q ni1

1 n ¦ qi , Q= observed , q=calculated ni1

3. The study area and data collection The Yamuna is the largest tributary of Ganga and its catchment area comprises about 42% of the Ganga basin area in the Indian Territory. Total catchment area of Yamuna in 3,66,223 sq.km. Its drainage area comprises parts of the States of Himanchal Pradesh, Uttaranchal, Haryana, Rajasthan, Madhya Pradesh, Uttar Pradesh, and entire Union Territory of Delhi. Of the total catchment, about 3% comes under hilly area and the remaining is almost equally distributed between plains and plateau regions. The river Yamuna originates from the Yamunotri glacier near Bandarpunch at an elevation of about 6320 metres above mean sea level in the Tehri Garhwal district of Uttaranchal. Figure 2 illustrates the synoptic view of the Yamuna River basin in India and locations of various cities along its banks including the Mathura city which is downstream of New Delhi. Water quality data for the years 1990 to 1996 used in this study were collected from the Central Water Commission. Monthly data sets on flow discharge, temperature, pH, biochemical oxygen demand (BOD) and

1074


dissolved oxygen (DO) were available for Mathura (upstream), Mathura (central) and Mathura (downstream) for Mathura City on River Yamuna. No instance of white-water or shallow flow depth was noticed in the datasets and so the conditions were representative of typical river flow conditions.

Fig.2: The Yamuna River Basin in India and location of some major cities along its banks


1075

4. ANN application and results Standardization of input data is an important step of data processing before ANN can be applied. In the present study, the input data for a variable x were standardized by the ANN model (Neural Power, 2003) In estimation of parameters of a hydrologic model, the available data are divided in two parts. The first part is used to calibrate the model and the second, to validate it. This practice is known as ‘split-sample’ test. The length of calibration data depends upon the number of parameters to be estimated. The general practice is to use half to twothird of the data for calibration and the remaining for validation. The data of River Yamuna consisted of 72 patterns where the measurements of flow discharge, travel time, temperature, pH, electrical conductivity, BOD and DO were available. Out of the data sets, 48 patterns selected randomly were used for training and 24 patterns for testing. To simulate the Dissolved oxygen, ANNs based on feed forward error back propagation algorithm were used. For ANN development, a number of combinations of input variables and input stations were tried namely: (a) All the data sets for stations Mathura (upstream), Mathura (central) and Mathura (downstream) except DO values at Mathura (downstream) (b) All data sets for the stations Mathura (upstream), and Mathura (central), and (c) All the data sets for the station Mathura (upstream). For the analysis, the ANN consisted 14 input nodes, 10 input nodes and 5 input nodes for the cases (a), (b), and (c) respectively. The output layer had a single node corresponding to dissolved oxygen at downstream of Mathura. The ANN training epoch consisted of 5000 cycles. According to Hsu et al. (1995), three-layer feed forward ANNs can be used to model real-world functional relationships that may be of unknown or poorly defined form and complexity. Therefore, only three-layer networks were tried in this study. During training, the number of nodes in the hidden layer was considered to be less than or equal to 2*input layers+1 (Swingler, 1996), which really provided the best results. It may also be added that as a result of training, a set of weights that represents the ‘knowledge’ of ANN is obtained and one does not get an explicit equation to work with. The comparative performance of various ANN models based on RMSE, Coefficient of correlation and determination coefficient for DO estimation are given in Table 1. It can be seen from Table 7.6 that during calibration, the values of RMSE, R and DC for all the developed models vary in the range of 1.71 to 2.89, 0.852 to 0.907 and 0.722 to 0.8226 respectively. However, the values of RMSE, R and DC vary in the range of 1.52 to 6.91 cumec, 0.654 to 0.928 and 0.283 to 0.856 respectively during the model validation. As explained in the earlier sections, three types of ANN models have been developed with different combinations of data and stations, i.e., ANN1: All the data sets for stations Mathura (upstream), Mathura (central) and Mathura (downstream) except DO values at Mathura (downstream); ANN2: All data sets for the stations Mathura (upstream), and Mathura (central); and ANN3: All the data sets for the station Mathura (upstream). Performance of the three ANN models is different. The best performing ANN model is the ANN2 model with values of RMSE, R, DC as 1.71, 0.907, 0.822 respectively during calibration and 1.52, 0.928, 0.856 respectively during validation. ANNR2 model has a total of 10 input variables consisting of monthly flow discharge, temperature, pH, biochemical oxygen demand (BOD) and dissolved oxygen (DO) at Mathura (upstream) and Mathura(central). The results indicate that the ANN model performance is best with optimum number of input variables. In case of ANN1 with 14 input variables, the performance is lower than ANN2 due to more number of input variables which increase model complexity. An increase in the complexity of the models might mislead the modeler to overfit the training data and lead to poor forecasts (Tokar and Johnson, 1999). In case of ANN3, the performance reduces drastically compared to ANN2 because of less number of input variables which are unable to explain the underlying physical process. Therefore, in ANN model development, it is very important to use optimum number of input variables and for the present study, the results indicate that for DO simulation at Mathura (downstream), 14 input variables as used in ANN2 are optimum. The graphical results of the best performing ANN, i.e., ANN2 are shown in Figure 3 in the form of time series plots and scatter plots which depict a good match between the observed and simulated DO by ANN method.

1076

Archana Sarkar and Prashant Pandey / Aquatic Procedia 4 (2015) 1070 – 1077 Table 1: Comparative performance of various ANN models ANN Model

Training (Calibration) RMSE 2.89

ANN-I ANN-II

1.71 2.35

ANN-III

Observed DO

Testing (Validation)

R

DC

RMSE

R

DC

.879

.726

3.05

.794

.519

.907

.822

1.52

.928

.856

.852

.722

6.91

.654

.283

30

Simulated DO by ANN

30

Simulated DO by ANN

25

Dissolved Oxygen

25 20 15 10

20

15 10

5

5

0

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47

0 0

Time (months)

30

Calibration 25

Simulated DO by ANN

Simulated DO by ANN

25 20 15 10

5 0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

Dissolved Oxygen

20

Observed DO

(a) Observed DO

10

Time (months)

20 15 10 5 0 0

10

20

Observed DO

(b)

Validation

Fig.3: Observed and computed DO values for ANN2

30


1077

5. Conclusions Very limited efforts have been made in the past specifically in India to utilise ANN for water quality modelling purposes. It is understood that the estimation of water quality of a river at any location is a tedious work due to the non-linear behaviour of different water quality variables. The ANNs have been successfully used in many hydrological studies and this was a motivating factor for its application to the present study. The performance of ANN was tested using RMSE, R and DC. It was found that the ANN approach turned out to be an efficient approach for water quality modelling. Even more accurate predictions of DO could be obtained with a shorter time base input data, e.g., 10-daily or daily data. An added advantage of ANN is that it does not require any assumption about the range of flow discharge, temperature, BOD and DO. However, the input data should be consistent and the controlling factors should be the same for training and test data. References ASCE Task Committee, (2000). Artificial neural network in hydrology. I: Preliminary concepts. Journal of Hydrologic Engineering, ASCE, 5(2), 115-123. Chakraborty, K., Mehrotra, K., Mohan, C.K. and Ranka, S. (1992). Forecasting the behaviour of multivariate time series using neural networks. Neural Networks, 5, 961-970. Daniell, T.M. (1991). Neural networks- Applications in hydrology and water resources engineering. Paper presented at International Hydrology and Water Resources Symposium. Institutions of Engineers, Perth, Australia. DeSilets, L., Golden, B., Wang, Q. and Kumar, R. (1992). Predicting salinity in the Chesapeake Bay using backpropogation. Computer Operation Research, 19(3/4), 277-285. Grossberg, S. (1982). Studies of mind and brain: Neural principals of learing perception, Development, Cognition and Motor Control. Reidell Press Boston. Haykin, S., Neural Networks - a Comprehensive Foundation. Macmillan, 1994, New York. Hopfield, J. and Tank, D. (1985). Neural computations of decisions in optimization problems. Biological Cybernatics, 52, 141-152. Hsu, K-L., Gupta, H.V., and Sorooshian, S. (1995). “Artificial neural network modeling of the rainfall-runoff process.” Water Resources Res., 31(10), 2517-2530. Karunanithi, N., Grenney, W.J., Whitley, D. and Bovee, K. (1994). Neural networks for river flow prediction. J. Comput. Civ. Eng., 1(2), 201220 Kohonen, T. (1988). Self organisation and associative memory. 2nd Ed., Springer-Verlag, New York. Lachtermacher, G. and Fuller, J.D. (1994). Backpropogation in hydrological time series forecasting. In Stochastic and Statistical Methods in Hydrology and Environmental Engineering, Vol.3, Time Series Analysis in Hydrology and Environmental Engineering, pp. 229-242, edited by k.W.Hipel, A.I. Mcleod, U.S.Panu and V.P.Singh, Kluwer Acd., Norwell, Mass.. Neural Networks, (2003). Neural networks professional version 2.0. CPC-X Software, Copyright: 1997-2003. A Demo version downloaded from the Internet. Rumelhart, De.E., Hinton, G. E. and Williams, R.J. (1986). Learning representation by back propagating errors. Nature, 323(9), 533-536. Rosenblatt, F. (1961). Principals of neurodynamics: Perceptrons and the theory of brain mechanisms. Spartan, New York. Schizas, C.N., Pattichis, C.S. and Michaelides, S.C. (1994). Forecasting minimum temperature with short time length data using artificial neural networks. Neural Network World. 4(2), 219-230. Swingler, K. (1996). Applying neural networks- A Practical guide. Academic Press Ltd., London. Widrow, B. and Hoff, M. (1960). Adaptive switching circuits. In Western Electronic Show and Convention, Convention Records, IEEE, 96-104. Windsor, C.G. and Harker, A.H. (1990). Multi-variate financial index prediction- A neural network study. In International Neural Network Conference, vol.1, Kluwer, Acad., Norwell, Mass. Vemuri, V.R. (1992). Artificial Neural Networks: Concepts and Control Applications. IEEE Computer Society Press, California. Yegnanarayana, B. (1999). Artificial Neural Networks. Prentice-Hall of India Pvt. Ltd., New Delhi.