EGION 
. This paper 
eteorological 
' wind speed, 
hips with the 
forecast the 
ir di-oxide in 
es are within 
rs render it 
which will 
n dispersion 
e. Therefore, 
it is directly 
n to use a 
. It is in this 
ificance as it 
1 the system. 
to study the 
ration of the 
tal Chennai 
h as petro- 
ower plants. 
se emissions 
the region. 
| parameters 
ind ambient 
‘or the years 
ured for this 
bd from the 
| were also 
e Regional 
Ictuations as 
ms such as 
atmosphere. 
33 f : O à summation 
Input Layer 
(1k. xlk) 
   
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring", Hyderabad, India, 2002 
activation ; © 
E utput layer 
(yl k. yOR 
  
  
0; 
  
  
  
. + 
~ + 
  
  
Learning/Training layer 
  
  
Figure 1. Basic Components of simple Artificial Neural Network 
Initially the meteorological parameters viz., wind speed, wind 
direction and ambient temperatures which primarily affect air 
flow are well analyzed. For this purpose, hourly data for the 
month of January corresponding to winter season for the years 
1999 and 2000 are chosen. This data is segregated on an hourly, 
daily and monthly basis. Similarly to account for diurnal 
fluctuations, the day is divided into six slots of four hourly 
duration and the wind and temperature patterns were studied. 
As wind plays a vital role in transport of pollutants, the data for 
wind direction and speed is analyzed slot wise. The 
dependence of temporal changes in pollutant concentration is 
thus observed and its sensitivity to changes in meteorological 
factors was analyzed. The relationship that exist between 
meteorological parameters and pollutant species is observed 
and the non-linearities among the different groups of variables 
such as wind speed, temperature and concentration are noted. 
Each data line has information about the year, month, day, hour 
as well as the corresponding measured values of meteorological 
parameters such as wind speed, wind direction and ambient 
temperature. The experimental data for concentration 
corresponding to SO; is also included. The data set numbering 
816 is split into the training and test set. 
ARTIFICIAL NEURAL NETWORKS 
Neural networks can be effectively used for a variety of 
applications, which have relationships that are dynamic, non- 
deterministic and non-linear among its parameters. So it is an 
alternative to conventional techniques which are often limited 
by strict assumptions of normality, linearity etc. Hence ANN 
can be used as a forecasting tool for the prediction of weather, 
air quality etc. (Boznar, 1993). 
An ANN is a model that emulates a biological neural network. 
It has a parallel-distributed architecture with a large number of 
nodes and connections to which are each associated a weight. 
Each neuron receives several inputs through these connections 
called synapses. The inputs are the activations of the incoming 
neurons multiplied by the weight of the synapses. The 
activation of the neuron is computed by applying a threshold 
function to this product. There are three components of any 
neural net model and these are input layer, output layer and 
learning / training layer. In addition functions like threshold, 
summation and activation also form part of any ANN model 
(Beale, 1998). This is shown in Fig.l. Here input segment 
consists of one set of input given in vector form as 
(x1k,x2k..xIk) with I members for set k and similarly the output 
in (y1k,y2k..yOk) for O members of k™ set. The main aim is to 
map input layer to output layer for different sets of inputs, k = 
1to N. The summation and activation function rely on weights, 
thresholds and transfer types —linear or non-linear, so that the 
ANN adjusts for different input to get desired outputs within a 
range of prescribed errors. The learning / training layer is what 
makes ANN applications different from routine mathematical 
methods. Even though initial ANN's were more focused on 
minimizing error between input and output layers through any 
standard optimization method like least square (LSM), the 
realization that the input-output relation need not always be 
confined to mathematical basis but more on variety of random 
parameters which do not permit mathematical simulation. Here 
is where adaptive ANN is needed where besides learning from 
sets of data, premise to introduce AI (Artificial Intelligence) 
concepts in the form of knowledge becomes necessary. The 
knowledge part of this adaptive ANN (White, 1992) can be in 
two segments one from training and another from data- 
dependent ** if , then" or simple predicate logic type knowledge. 
Now the ANN can be modified as shown in Fig.2. Here 
another layer of knowledge base is added which gives inputs 
for knowledge-based :