Prediction of Water Quality Index of Pavna River using ANN Model

Prediction of Water Quality Index of Pavna River using ANN Model

Sujana Prajithkumar1 ,

M E Student( Env. Engg),

DY Patil College of Engineering ,Akurdi , Pune-411044,India

Dr. Shweta Verma2,

.Assistant Professor,

DY Patil College of Engineering, Akurdi, Pune-411044,India

S. J. Mane

3.Assistant Professor,

DY Patil College of Engineering, Akurdi, Pune-411044,India

Abstract – Rapid industrialization and urbanization witnessed the deterioration of most of the river bodies. Interpretation of complex water quality data is difficult to understand also it is difficult to communicate during decision making process. Assembling the various parameters of the water quality data into one single number is essential in for easy interpretation of data. The present study is concentrated on the prediction of WQI using ANN model. Pavna river originates south of Lonavala from western ghatts and flows a total of nearly 60Km to meet the Mula river in Pune. Pavna river flows through Pune city covering Pimpri Chinchwad Area. Pimpri Chinchwad area is developing area and due to industrialization and high population growth the river is getting polluted. The data for the prediction of water quality index were collected from the water treatment Plant, Nigidi sector 21.The collected data were for the point Ravet Intake.(2001-2014)(monthly).In the present study it is found that the modular neural network give better results as compared to the radial basis function neural network. R value for the modular neural net work is 0.99 and that for radial basis function is 0.98.

Key wordsWQI, Artificial neural network, radial basis neural network, modular neural network

1. INTRODUCTION :-

   Water is the prime requirement for the existence of life and thus it has been mans endeavour for the immemorial to utilize the available resources. The growing population exerts a great pressure on this resource. The increased population growth and ill planned exploitation of the water resources created a situation, where the very survival of man has become endangered[1].

   Interpretation of complex water quality data is difficult to understand. Also it is difficult to communicate during decision making process. Assembling the various parameters of the water quality data into one single number is essential for decision making purpose. The purpose of an index is to transform the large quantity of data into information that is easily understandable by the general public. Water quality index exhibits the overall water quality at a specific location and specific time based on several water quality parameters. The index result represents the level of water quality in a

   given water basin, such as lake, river or stream.[2] WQI is a set of standards used to measure changes in water quality in a particular river reach over time and make comparisons from different reaches of a river. A WQI also allows for comparisons to be made between different rivers. This index allows for a general analysis of water quality. It also helps for the analysis of water many levels that affect a streams ability to host life and to check whether the overall quality of water bodies poses any potential threat to various uses of water. A computer programming using MATLab had been used for training and testing of the ANN .After a number of training trials ,the best neural network model was generated. The maximum number of epochs was chosen by a trial and error approach. Trial and error method was used to find the most suitable network model for the WQI analysis using the different ten parameters. The result of WQI was valid in respect of experimental and ANN model.

   In recent years the trend has been to use statistical method instead of traditional and domestic methods to forecasting WQI. Lee Yoot Khan et al conclude that the modular neural network was found to be the most suitable model for the termination of the WQI[2].A A Mansur et al predicted the dissolved oxygen in Surma river using fed forward neural network and radial basis function neural network[3]. Both the methods provided better results. Sundarambal et al concluded that the ANN model can be used even for sea water quality forecasting[4]. In this study the modular neural network and radial basis function models of ANN are used to forecast the WQI of pavna river at ravet intake.

2. STUDY AREA

   The Pavna river in Pune is selected as the study area for ANN application. Pavna river originates south of Lonavala from western ghatts and flows a total of nearly 60Km to meet the Mula river in Pune.Pavna river flows through Pune city covering Pimpri Chinchwad Area. Pimpri Chinchwad area is developing area and due to industrialization and high population growth in thease area the river is getting polluted. The study area is Ravet intake from where water is collected and treated in water treatment plant in Nigidi.

3. MATERIALS AND METHODOLOGY

   There are different methods for the determination of water quality Index. In this particular study the methods used are given below:

   1. Weightage Rating method

   2. Prediction of WQI by using ANN

      1. Weigtage Rating Method:

         Factors which have higher permissible limits are less harmful because they can harm quality of river water when they are present in very high quantity. So weightage of factor has an inverse relationship with its permissible limits.

         Therefore Wi (1/ Sn)Or (1) Wi = K/Sn

         Where, K = constant of proportionality Wi = unit weight of all chemical factor Vs or Sn = Standard alue of ith parameter

         71

         Values of K were calculated as: K = 1

         Sl No	Parameters	CPCB/IS	IS (10500)
         1	PH	6.5-8.5	6.5-8.5
         2	EC	<300**	—
         3	Turbidity	–	5
         4	TDS	<500	<500
         5	Total Alkalinity	200	200
         6	Total hardness	300	300
         7	DO	6	–
         8	Chloride	250	250
         9	Nitrate	20	45
         10	Calcium	75	75

         Wi – total unit weight of all chemical factors. Using the water quality index, all the samples were categorized into the following five classes: excellent (0 – 25), good (26 – 50),

         moderately polluted (51 – 75), severely polluted (76-100) and unfit for human consumption(above 100) based on their suitability.[5

         Table 1 the permissible values of various pollutants for drinking water(expressed in mg/l except ph and ec recommended by indian standards and cpcb standard[2]

         1

         7 1 =

         1

         ( 1 )+( 1

         )+ 1

         +( 1

         )+( 1

         )+(

      2. Pre-processing of data

         ()

         ( )

         ( )

         At the initial stages of the project, real environmental data on

         1 )+ 1 +

         ( )

         . +( 1 ) + ( 1 ) (2)

         the condition of Pavna river over a period of time 10 years beginning from 2001 to 2014 wereacquired from water

         treatment plant sector 21 Nigidi, Pune. These make a total of

         97. Out of these data 70 datas are set for training and

         The weightage of all the factors were calculated on the basis of the above equation.

         WQI = Wi x Qi/Wi (3)

         Wi x Qi = Wi(pH) x Qi(pH) + Wi(EC) + Qi(EC)+ Wi(TDS) xQi(TDS) + Wi(DO) x Qi(DO) + .

         …………+ Wi(Calcium) x Qi (Calcium)

         (4)

         Qi = 100[(Va-Vi)/(Vs-Vi)]

         (5)

         = Rating scale

         Va =average measured values of water sample at study area. Vi =standard value of ith parameter

         Vs =ideal value for pure water (0 for all parameters except pH and DO)

         remaining 27 data are set for testing. Trial and error method is tried for the different combinations of the data.

      3. Training data

      The network architecture for WQI consists of ten input nodes and one output node. The parameters of choices as input must have an influence on the desired output. For the current project the ten water pollution parameters are selected to be as input nodes. These input and output nodes are used to produce ANN model. The input variables are DO concentration, concentration of Nitrates, concentration of Calcium, Total Suspended Solids concentration, concentration of Total hardness, concentration of Alkalinity, concentration of Chlorides, Total Hardness concentration. All the eight parameters are measured in milligram per litre. The remaining three parameters are pH, turbidity and conductivity.

4. ARTIFICIAL NEURAL NETWORK Artificial Neural networks are non-linear mapping structure which is inspired by the observed process of natural networks of biological neurons in brain. It consists of simple computational units called neurons. These are highly inter connected. ANNs become very popular nowadays because of their wide range of applicability and the ease with which it can treat complicated problems.[6]

   Fig 1 The schematic representation of a artificial neural network

   1. Development of ANN Model

      ANNs are constructed with layers of units, and hence it is termed as multilayer ANNs. The first layer of multilayer ANN is known as input units. In statistical nomenclature input units are known as independent variables. Last layer is called output units. In statistical nomenclature these are known as dependent variables or response variables. All the other units in the model is known as hidden units and it includes hidden layers. There are two function to achieve the behaviour of a unit in a particular layer, which normally are the same for all units within the whole ANN.

      Input function and Output function

      Input into a node is a weigted sum of outputs from nodes connected to it. The input function is given by the equation:

      Neti = Wij Xj + µi

   2. Modular Neural Network Function

      Modular feed forward is the special class of Multilayer perceptron (MLP). modular neural network process their inputs using several parallel MLPs and recombine the results. Less number of training examplers are required and training time will also be reduced due to the smaller number of weights. In this study the transfer function used is TanhAxon and the learning rule is LevenbergMarqua.

   3. Radial Basis Neural Network Function

   Radial basis function network are nonlinear hybrid networks typically containing a single hidden layer of processing elements.(PE).It consists of three layers ,the input layer

   ,hidden layer and out put layer as shown in fig3. The input variables x1,x2 …xn are input assigned layers and passes through the hidden layers. The hidden layers consists of the array of n numbers. Each nodes contains the Gaussian Radial function network. It can have any number of outputs with any activation function. It based on the squared Euclidean distance between the input vector and the weight vector. These are also known as transfer function and are similar to the sigmoid function. In this study transfer function is sigmoid and learning rule is LevenbergMarqua.

   (6)

   Fig 2 Mathematical representation of neural network

   Where Neti describes the result of the net inputs xi (weigted by the weights wij)impacting on unit i.Also wij are weights connecting j to neuron i ,x j is output from unit j and µi is the threshold for neuron i, Threshold term is baseline input into to a node in absence of any other input. If a weight wij is negative it is termed inhibitory because it decreases net input

   ,otherwise it is called excitatory. Each unit takes its net input and applies an activation function to it. A number of nonlinear functions have been used in the literature as activation function. The threshold function is useful in situation where the input and outputs are binary encoded.most common choice in activation function is sigmoid function[6] ,such as

   g(netinput) = [1+ e-netinput]-1 (7)

   g(netinput) = tanh(netinput) (8)

   Fig 3 A typical radial basis function

   Radial basis function is non linear hybrid networks typically containing a single hidden layer of processing element(PEs).Gaussian function is used as a transfer function

   ,rather than the standard sigmoid functions used by MLPs. Learning rules applied are supervised and unsupervised applied to the outputlayers. These network tends to learn more faster than MLPs.

5. RESULT AND DISCUSSION

The ANN models were trained using number of neurons in the hidden layers and for training and testing. After number of trials by changing the length of data set for training ,testing and also changing the number of neurons in the hidden layer the maximum value of R was obtained. The modular neural network consists of TanhAxon as transfer function and LevenbergMarqua as learning rule. In case of radial basis neural network sigmoid function as transfer function and LevenbergMarqua is used as a learning rule. The number of epochs was set to 1000 to 5000 throughout the trial and error process. The number of ANN models were developed using modular neural network and radial basis function. The hidden

Desired Output and Actual Network Output(using radial basis function)

150

100

layer is taken to be between 1-2.The training and testing data set an epochs were varied as shown in table 2 in order to achieve better accuracy.

Exemplar

WQI

Output

1 3 5 7 9 1113151719212325

50

0

WQI

Output

Among the selected two ANN models the developed ANN model with Modular neural network model stimulated the WQI of pavna river with great accuracy when compared with RBFNN architecture.The maximum value of R was 0.990, MSE=9.542 and MAE=2.400 as shown in table. In this case the value of R obtained is quite high (0.990) and reliable for short term prediction. The observed and modelled WQI values in whole array using modular neural network and radial basis Function neural network is illustrated in fig 3a and 3b and fig.4a and fig 4b which indicates that the modelling has been

quite successful. Fig4.a The observed and modelled WQI values in radial basis network

Desired Output of WQI and Actual Network Output of WQI (using modular neural network)

120

100

Output

80

60

40

20

Output WQI vs. DesiredWQI

y = 0.970x + 0.445

R² = 0.961

100

80

60

40

WQI Output

WQI

0

1 3 5 7 9111315171921232527

Exemplar

WQI

40 60 80 100

WQI Desired

Output

WQI Output vs. WQIDesired

WQI Output

Fig3.a The observed and modelled WQI values in modular neural network

y = 0.907x + 6.550

R² = 0.982

100

80

60

40

40

60

80

100

WQI Desired

Fig3.b Scatter plot of observed verses modelled WQI for modular neural network

Fig4.b Scater plot of observed verses modelled WQI for radial basis neural network

TABLE 2 SUMMARY OF ANN MODEL

The modular basis function with all input parameters are found to be the most appropriate model for WQI prediction with high correlation co-efficient of 0.990and a mean square error (MSE) value of 9.542 and that for RBFNN is found to be 0.980 and 18.490respectively.

6. CONCLUSION

In this model ANN models were developed to predict the WQI in Pavna river at Ravet intake. The proposed model shows efficiency in forecasting the WQI in water bodies. The result showed that the modular network model prepared by different ten water quality parameter provided high R(0.990) value. It has been observed that the WQI of Pavna river can be predicted using both with acceptable accuracy using Modular Neural Network as well as Radial basis function network models.

REFFERENCES:

1. Maheshkumar .Akkaraboyina 1,prof B.S.N.Raju2,A Comparative study of water quality indices of river Godavari, Journal IJRD- eISSN : 2278-067X, pISSN : 2278-800 Volume 2, Issue 3 (July 2012), PP. 29-34

2. Lee Yoot Khauan,Noraliza Hamzah ,and Rozita Jailani, Prediction of Water quality Index (WQI)based on Artificial Neural Network.(ANN) Journal 2002 IEEE 157-161

3. A A Masrur Ahemed ,Prediction of dissolved oxygen in Surma River influenced by biochemical oxygen demand and chemical oxygen demand using Artificial Neural Network,Jounal of King Saud UNIVERSITY,-

   ENG.SCI(2014)xxx,xxx-xxx

4. Sunarambal Palani, Shie-Yui Liong,Pavel Tkalich,An ANN application for water quality forecasting, Marine Pollution Bulletin56(2008)1586-1597

5. S. Kalavathy1, T. Rakesh Sharma1, P. Sureshkumar2* Water Quality Index of River Cauvery in Tiruchirappalli district, Tamilnadu, Journal ARCH. ENVIRON. SCI. (2011), 5, 55-61

6. Girish kumar Jha ,Artificial Neural Networks ,Indian Agricultural Research Institute .New Delhi.

7. Papita Das Saha ,R .Sengupta ,Jhuma Saha ,P K Banerjee Assessment of water quality characteristics of River GAnga At Kolkata Region ,India using Water quality Index and ANN simulation method. Journal ARCH…ENVIRON.SCI.(2012),6,34-41

8. Rene E R ,Saidutta M B Prediction of Water Quality Indices by Regression Analysis and Artificial Neural Networks,INT.J.ENVIRON.RES,2(2):183-188,SPRING 2008

9. K.V.Nagaraj, Book ,Practical Handbook on Environmental Engineering II

10. APHA , Book, Standard methods for the examination of water and waste water