Sie sind auf Seite 1von 10

1

Keywords: Neural Network, Sewage treatment, Biological Oxygen Demand, Chemical


Oxygen demand, Suspended solids.

Introduction:
Artificial Neural Networks (ANN) has been used extensively for solving numerous problems
in different fields of engineering. ANN has proven to be an important alternative for system
analysis when compared to multivariate regression technique or other such statistical tools.
Apart from linear or multiple regression, ANN may be used as a substitute for Autoregressive
Integrated Moving Average (ARIMA) method. Also, this technique can be exploited to
extract the patterns and subsequently detect the trends in data that are very complex to be
analysed by using other available computational techniques.
Oliveira-Esquerre et al. (2002) presented a study on the effluent BOD prediction using ANN
in biological waste water treatment for a pulp and paper mill. Gontarski et al. (2002)
demonstrated an approach to have quantitative prediction of parameters in the output stream
of an activated sludge system treating the wastewater in the production of purified terephtalic
acid. Cinar et al. (2006) developed an ANN model to predict the performance of a membrane
bioreactor in the treatment of cheese whey wastewater. Maier and Dandy (2000) modelled a
water resource system using the feed forward Back Propagation Neural Network (BPNN) to
examine the accuracy level in BOD prediction. Gondarski et al. (2000) simulated an
industrial waste water treatment plant using ANN.
Pu and Hung (1995) used ANN to predict the performance of a trickling filter in municipal
waste water treatment plant. Hamoda et al. (1999) determined the correlation coefficient in
effluent BOD prediction using ANN in the treatment of waste water. Belanche et al. (1999)
predicted the effluent COD in waste water treatment system with the application of ANN.
Cote et al. (1995) developed a dynamic model for an improved prediction of the effluent
parameters in activated sludge process with the application of neural network. Conlin et al.
(1997) presented the application of an artificial neural network model to predict the pressure
drop in the operation of a water treatment plant. Hack and Kohru (1996) estimated the waste
water process parameters with the application of ANN to predict the scope for elimination of
Total Organic Carbon (TOC) in the treatment plant and subsequently determine the
correlation index between the observed data and predicted results. Neural network has been
applied in waste water treatment to model the Sludge Volume Index (SVI) for the purpose of
2

forecasting sludge bulking (Capodaglio et al., 1991). Also neural networks have been used to
predict the effect of heavy metals on the performance of activated sludge process (Tyagi and
Du, 1992). Some more studies on the modelling of biological waste water treatment
processes have been carried out using ANN (Hamoda et al., 1999; Lee and Park, 1999;
Wilcox et al., 1995; Zho et al., 1999).
Although a plethora of neural networks has been applied successfully in different fields, very
few applications have been done to study the anaerobic biological waste water treatment
processes. In fact, the prediction of effluent sewage quality in a municipal waste water
treatment plant pertaining to developing countries is yet to be found in the body of published
literature. An approach has been made in this paper to determine the correlation coefficient
and root mean square error between the predicted and actual effluent data in the effluent
stream of a sewage treatment plant. This study has aimed to develop an ANN model to
predict the BOD, COD and SS in the outflow stream of a sewage treatment plant located in
western part of India.
In order to predict the process upsets and enable the process plant operator to work in a
proactive manner for minimizing the fluctuations in output stream, ANN can be exploited and
subsequently a comparison can be done between the predicted effluent quality and the actual
operational data. Timely control in the operation of sewage treatment plant often appears to
be difficult due to variation in the composition of raw sewage, strength and flow rates. Thus,
an effecting control on plant operation often becomes a complicated issue. Moreover, there is
a five-day delay period in determining the effluent BOD. In practical situation, such delay is
coupled with the Hydraulic Retention Time (HRT) and it may be too-late to make proper
adjustments in the sewage treatment process. Furthermore, it has become necessary to finalise
strict and appropriate control strategies in sewage treatment plants due to the increasingly
stringent environmental regulations for the treated effluents that would be discharged to the
adjacent river water bodies.
ANN model development
In neural network, the nodes are regarded as processing elements in each layer. The nodes of
one layer are connected linearly through weights to the adjacent nodes in the succeeding
layer. The learning process in neural network adjusts the weights from time to time as the
network proceeds. Information enters the neural network at the input layer. All the layers
process these signals through the network until the output layer is reached. The connection
3

weightage of the neural network is liable to change in response to inputs and states of the
network. In the present study, the parameter learning process has been considered to analyse
the network. The process of parameter learning relates to updating the connection weights in
the network.
With the application of neural network, an accurate mapping can be obtained from the input-
output pairs of data without having any prior knowledge on their functional relationship
(Emmanouilides and Petron, 1997). Gardner and Dorling (1999) presented a study on the
Multilayer Perceptron (MLP) neural network to model any highly complex relationship. MLP
neural network learns to model the relationship during supervised training process, when
presented with input output data series and can be trained accurately for new set of data. In
this study, MLP neural network model has been applied to train the input data.
In a neural network analysis, it is necessary to divide the data into three sets namely training,
validation and testing. However, the training data set presents the major portion of data. The
goal of training the neural network is to extract the underlying patterns in the data set. The
training of data should be adequately extensive as well as representative in character
(Gardner and Dorling, 1999). The training set is used to train the network and adjust the
weight. The training process of neural network involves adjustment of the weight and bias so
that the error between target (or observed output) and the model predicted results attain the
minimum value. The validation data set is used during training to check the general
performance of the model. The process of training would be stopped when the best
performance on validation data set has been reached. Finally, the testing of model is done on
the basis of testing data set.
Fig. 1 presents a 3-layered feed forward neural network model with a single hidden layer.
The feed forward ANN model is interconnected by a number of neurons. Generally, such
network comprises of three layers namely input, hidden and output. Initially, the input
parameters are fed to the first layer (i.e. input layer). Subsequently, these input parameters are
multiplied by the weights to obtain the weighted inputs in respective cases. Each neuron in
the hidden layer receives the weighted inputs along with bias from each of the neurons of the
input layer. The values of weight and bias are set during the process of network training. At
the outset, weights are set in random manner. The network is trained by using the input and
target (or observed output) so that the errors between the target and model predicted output
values attain the minimum.
4


N
i
i i ij
a x w z
1
1

When,
x
i
= i
th
input node parameter
i = 1, 2, 3........... N
N = no. of nodes in the input layer
w
ij
= weight for i
th
input node to j
th
hidden layer node
j = 1, 2, 3........... M
M = no. of nodes in the hidden layer
a
i
= bias parameter for the i
th
input node
The output of hidden layer (z
h
) representing the input to output layer can be obtained by using
the sigmoidal transfer function as below:
z
h
= f(z
1
)
Similarly, each neuron in the output layer receives the output from hidden layer as below:


M
i
j j jk
b x w z
1
2

When,
x
j
= output parameter from the j
th
hidden node
b
j
= bias parameter for j
th
hidden node
k = 1, 2, 3............L
L = no. of nodes in the output layer
w
jk
= weights for j
th
hidden node to k
th
output node
The outcome of the output layer (z
0
) can be calculated by linear transfer function:
z
o
= f(z
2
)
5

The BP algorithm can be written as
X
w+1
=X
w
-L
w
.G
w

Where,
X
w
= vector of weights and biases
L
w
= learning rate
G
w
= gradient
The Back Propagation (BP) algorithm can be used to calculate the weights. The back
propagation approach denotes the process of finding a gradient vector in the structure of
neural network. In this process the gradient vector is calculated in the direction opposite to
the flow of output of each node resulting to back propagation learning rule. Thus, the basic
learning rule for the network is based on gradient descent method. During the training
process, the algorithm can revise the weights and bias terms iteratively to achieve the
condition of minimal error for the total network. This procedure employs a chain rule to
guide the weight and bias values for movement along the negative gradient of the control
function.
The back propagation MLP is an adaptive network, whose nodes (or neurons) perform the
same function on the incoming signals. The node function is a composite of the weighted sum
and differentiable non-linear activation function known as transfer function. The sigmoidal
function used for activation is given below:
x
x
e
f

1
1

This function provides smooth non-zero derivatives w.r.t input signals. The function can also
be termed as squashing function, while the input to this function is squashed to the range (0-
1). Also, this function is called as sigmoidal function and its s- shaped curve exhibit
smoothness and asymptotic properties.
The Levenberg Marquardt (LM) algorithm can be used in the training process to speed up the
convergence to a network capable of estimating the minimal global error. The squared error
measure for the input-output pair of data can be obtained to reduce the error between the
calculated and the target values by adjusting the interconnection between the layers. In this
6

process, the weights are adjusted using the gradient decent BP algorithm. In the present
study, the ANN Tool Box in MATLAB 7.0 has enabled to import the raw sewage treatment
data for developing the basic network. The network structure has been created based on the
choice of network type, no. of neurons in each layer, training process and transfer function.
Consequently, the trained network is adapted to minimise the errors.
Now, E =

P
p
p
e
1

e
p
=


o
N
k
k k
h o
1
2

When
E = squared error measure for the total data set
e
p
= error for p
th
training pattern
k = 1, 2, 3.................... N
o

N
o
= total number of output nodes
o
k
= predicted output of network at the k
th
node
h
k
= target (observed output) at the k
th
node


Wastewater Treatment System:
Data collection
The influent and effluent data were obtained for a period of two years from the historical data
base of a municipal waste water treatment plant located at the western part of India. The
sewage waste water treatment plant has been operating with full scale UASB reactors. Based
on the influent and effluent parameters e.g. inflow rate, BOD, COD and SS concentration, an
attempt has been made in this paper to predict the effluent BOD, COD and SS level of the
treatment systems. The maximum and minimum values of influent and effluent parameters
along with mean and standard deviation in respective cases have been presented in Table 1.
Results and Discussion
7

All data sets were normalised and divided into three groups namely training, validation and
testing. Since the different parameters have different units, the data was normalised to lie
between 0 and 1, so as to avoid any numerical overflow and underflow. Thus, the data
pertaining to each variable were scaled corresponding to the range of data. MATLAB
(Version 7.0) was used to frame the three layered feed forward network consisting of an input
layer with four neurons, one hidden layer with six neurons and one output layer with one
neuron. In this study, the MLP neural network has been applied to train the input data. The
nodes in hidden layer were varied commensurate with the no. of inputs to understand the
effect of such changes in the network structures. At the outset, training was done with the
same no. of input and hidden layer nodes. This was followed by network training with a
hidden layer having more nodes compared to those in the input layer. The no. of neurons in
the hidden layer were finalised from a series of trial runs considering 1-10 neurons to obtain
the minimal error. The goal of error reduction was achieved by adjusting the interconnection
between the layers; whereas the weights in the process were adjusted with the use of gradient
descent BP algorithm.
Thus, a weight based neural network model was developed to estimate the key effluent
parameters e.g. BOD, COD, SS for the sewage treatment process.
In this context, three alternative models were formulated on the basis of data distribution ratio
i.e. training: validation: testing as below for the prediction of output results:
a. Model 1 (M-I) - 60 : 20 : 20
b. Model 2 (M-II) - 70 : 15 : 15
c. Model 3 (M-III) - 80 : 10 : 10
Each of the above models represented the percentage of training, validation and testing data
points out of the total data set. With the increase in percentage of training data points from
60% to 70% corresponding to the models M-I and M-II, the model performance and
prediction ability was found to improve on an overall basis. However, the performance of
model M-III did not appear to be satisfactory. Further, the variation of no. of neurons in the
hidden layer helped to obtain the optimum results for each of the effluent parameters. A
number of ANN networks having different architecture were trained to arrive at the optimised
model for each of the effluent parameters. This was done on the basis of compatibility of
correlation coefficient and RMSE values.
8

Finally, the ANN model output results were compared with the observed data to validate the
neural network model. Further, the performance of neural network models was judged on the
basis of correlation coefficient (R) and root mean square error (RMSE) as presented in Table
2. The RMSE (%) was reduced with the use of model M-II. The predicted results obtained
from the application of ANN models corresponding to BOD, COD and SS have been
summarised in Table 1. In this analysis, an excellent compatibility of results was obtained
with the incorporation of data distribution ratio of 70:15:15 in the prediction of effluent
parameters e.g. BOD, COD and SS. The correlation coefficients and RMSE vales for
training, validation and testing data sets corresponding to each effluent quality parameter
have been shown in Table 2. The mean values and standard deviations obtained in the
prediction of effluent parameters are highly satisfactory when compared to the respective
observed data sets (Table 3). The average error obtained in the prediction of effluent
parameters wrt observed (or target) data were within the range of + 4.6 to 8.3. Also, the ANN
models with the optimised structure showed standard deviation of errors in the range of + 5.3
to 9.6 when compared to the observed data sets (Table 4) in the prediction of effluent
parameters.
Consequently, the ANN models developed in this exercise have been found to be highly
satisfactory in the prediction of effluent parameter values based on the given input data. This
approach has indicated an excellent performance of the ANN models in the prediction of
effluent parameters. Furthermore, the results of optimised ANN models have demonstrated
the acceptability of correlation coefficients in respective cases for the prediction of effluent
parameters in quantitative terms.

Conclusion:

References:



9










Table 1. Features of ANN Model for Effluent Quality Prediction
Sl.No. Description
Effluent Parameters
BOD COD SS
1. Data distribution ratio 70:15:15 70:15:15 70:15:15
2. No. of iterations 14 6 14
3. No. of hidden neurons 8 6 6
4. R (training) 0.99 0.97 0.99
5. R (validation) 0.89 0.86 0.98
6. R (testing) 0.92 0.84 0.79
7. RMSE (training) 0.011 0.066 0.08
8. RMSE (validation) 0.155 0.08 0.08
9. RMSE (testing) 0.167 0.27 0.14

Table 2. ANN Predicted and Observed Effluent Quality
Sl.No. Description
Effluent Parameters
BOD
(mg/l)
COD
(mg/l)
SS
(mg/l)
1. Predicted mean 32.75
103.11 39.01
2. Predicted Standard Deviation + 5.06 +12.36 + 4.83
10

3. Observed Mean 32.57 102.21 38.96
4. Observed Standard Deviation + 4.78 + 12.35 + 5.32
5. Error % for Mean
6. Error % for Standard Deviation














Error back propagation algorithm is commonly used as the learning algorithm. These
technique applies the input values of a sample data set to be learned as the inputs of neural
network. The process compares the outputs of network with the observed output values to
compute the errors. The errors so obtained can be analysed to identify these synaptic weight
that had to be modified by the network for reducing the error of observed data set. The
algorithm iterates these steps for each data set until the average error reaches below the
predermined threshold.

Das könnte Ihnen auch gefallen