Artificial Neural Network - Water Quality Management: Assessment and Interpretation - page 77

studied an improved back propagation of the ANN model for eutrophication of
eastern china, the model was applied to four eastern lakes and results described the
ANN is suitable for predicting eutrophication. Kuo et al. ( 2007 ) used back-propa-
gation neural network to relate the number of water quality indicators such as DO, TP,
Chl-a and secchi dish depth in a reservoir in central Taiwan and they concluded that
the ANN is able to predict these indicators with reasonable accuracy. Kanani et al.
( 2008 ) predicted the salinity levels for 1 month in advance in a Talkheh Rud River
(Iran) by applying the MLP and IDNN models, the result was reasonable The Illinois
state water survey conducted a study to assess the potential of the ANN in forecasting
weekly nitrate-nitrogen concentration. Three ANN models were applied to predict
weekly Nitrate N concentration in the Sangamon River near Decatar, Illinois, based

on the previous week

—

s precipitation, air temperature, and discharge and past nitrate
concentration. The result model was more accurate than the linear regression model
having the same input and output (Momcilo et al. 2003 ). Asadollahfardi et al. ( 2010 )
applied the MLP for predication of eutrophication in Anzalily Wetland using TN, TP
as input in the MLP and predicted BOD parameters, the results indicated reasonable
accuracy. Asadollahfardi et al. ( 2011 ) studied static and dynamic neural network to
TDS of Talkheh-Rud River, Iran, and predicted the future of salinity. Asadollahfardi
et al. ( 2013 ) predicted sodium adsorbtion ratio (SAR) of Chelghazy River in Kurd-
istan (Iran) using MLP neural network, and the results was acceptable.

'

5.3 Arti cial Neural Network Theory

5.3.1 Theory of ANN

Bearing in mind natural neural and its components, scientists developed an arti cial

neural system. This is the smallest unit of an ANN. An arti cial neural system

consists of three components including weighting (W), bias (b) and transfer func-
tion (f). These three components are unique to each neural system. In Fig. 5.1 ,

“ p ”

and “ n ” are input and output while “ a ” is net output. The junctions 1 and 2 in the

figure indicate the schematic of an arti cial neural system. Function of an arti cial

neural network would be called “ p ” .

n = wp þ b ð 5 : 1 Þ

a = f ðÞ = fwp þ b ð Þ ð 5 : 2 Þ

Generally, the ANN is divided into two groups, static and dynamic. Time is not a
key parameter in the static of the ANN but it is one of the main parameters in
dynamic networks.
Hornik et al. ( 1989 ) proved the

“ ”

which expressed
that a feedforward neural network with a hidden layer of sigmoid tangent and linear
output layer would be able to estimate each complex function (Cybenko 1989 ;

universal approximator theory