ther the input 
it-the-time-of- 
al errors of 
nts of output. 
eflect overall 
based on one 
sets of data. 
/in Fig. 3. 
wn in Table 1. 
in an hourly 
onth and year. 
sis are 24 per 
th. Now the 
d sets of data 
if inputs are 
  
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring", Hyderabad, India,2002 
  
  
  
  
  
  
  
Year | Month | Day Hour Wind Wind Temperature SO, 
speed direction °C concn. 
m/s (degrees) pg/m’ 
99 01 05 01 0.2 298 20.9 > 171,29 
99 01 05 02 0.3 299 20.3 0.83 
05 03 0.8 296 18.0 0.83 
99 01 05 24 0.9 33 23.3 171.29 
  
  
  
  
  
Table 1. Sample of hourly input and output data 
The purpose of training is to enable the network to get trained 
in such a way so as to get closer to the desired output. The 
training pattern in the present study uses the measured hourly 
values of meteorological parameters such as wind speed, wind 
direction and ambient temperature as inputs to the neurons. 
Their values are normalized. The weight vectors are first 
initialized with random values. Such an input layer serves as a 
buffer which communicate the inputs to the hidden layers. This 
model makes use of an adaptive neural network approach using 
a set of ‘if and then’ rules. The assignment of the weight 
fractions for day is different from that of night. For wind speed 
there exists an inverse relation with concentration while the 
converse is true for temperature. Hence in accordance with this 
knowledge gained from the analysis of the observed data set, 
weights are suitably assigned. A look up table is constructed, 
(a sample shown in Table-2) which assigns weights for the 
various meteorological inputs depending on their respective 
values. If the percentage error is large then the weights are 
further modified using suitable knowledge until optimization 
occurs. In this manner the neural network was trained using 
408 data lines for the month of January 1999. After numerous 
iterations the network gets trained and gives an output of 
acceptable error rate. This trained network is then used to test 
the data for the year 2000. 
RESULTS AND DISCUSSION 
The results obtained using the adaptive neural network during 
the training period is presented graphically in figures 4 to 6 for 
hourly, daily and monthly basis. In figure 4, the hourly values 
of measured and predicted concentration for a sample day of 
the month of January are presented. The average error is less 
than 15% for 75% of the values for the specified day. 
However there are a few exceptions with error shooting up to as 
high as 6796 and 9096 for the 1* and 14” hour respectively. 
This could be attributed to the errors that have crept in during 
measurements or due to other allied phenomena quite typical 
for the day, which has not been accounted for in the model. 
The same logic is extended for all the other days of the month 
for which observations are available. While considering daily 
values, either 24hour mean values, median, maximum or 
minimum, or even rate of change values could be chosen. 
However the best representative value for the day with 
appropriate weight fractions would yield an output very close to 
the desired output. Similar is the case for monthly values. The 
measured and predicted values of SO, concentration on a daily 
basis are shown in figure 5, for those values with average error 
less than 10%. 
  
   
   
  
  
  
   
  
  
   
  
  
  
   
  
  
  
   
  
   
   
   
   
  
  
   
  
   
  
  
   
  
  
   
  
  
  
  
  
  
Wind Speed Weights for Temperature Weights for 
m/s Wind speed SC temperature 
<0.9 50 «21 35 
0.91 — 1.15 40 - 45 21.1 - 25 40 
  
Table 2. Sample of the look up table for weights