OO
IAPRS & SIS, Vol.34, Part 7, "Resource and Environmental Monitoring", Hyderabad, India,2002
Knowledge Base
007/655
NA
Input * Y» E
layer Learning/Training layer output
layer
Figure 2. Modified Artificial Neural Network
weights, accumulation and learning rule modification.
Typically for the problem chosen the inputs and output
measurements, have so many factors influencing them like for
example
data may be fuzzy i.e. imprecise and uncertain
data may be measurements, which have inherent
uncertainties or errors-- for input
Similarly the set of data used for mapping may be fuzzy or
imprecise or if they are also measurements, may contain errors
and uncertainties. So additional rules and logic were added to
ANN in the form of "if....then" rules or using simple predicate
logic and later tested for training.
In the present study the inputs have been chosen as measured
values of temperature, wind speed and wind direction and the
outputs are measurements of pollutant concentration of SO,
NOx etc. As mentioned earlier both input and output values are
measurements with associated inadequacies. Further the input
measurements are location, season and weather-at-the-time-of-
measurement dependent along with the usual errors of
measurement and so are the pollutant measurements of output.
Now, some a-priori knowledge is needed to reflect overall
nature of inputs as well as to modify the weights based on one
or two cycles of training the ANN for some sets of data.
Present architecture used for ANN is shown below in Fig. 3.
Details of one set of input data and output are shown in Table 1.
This is typical as it pertains to a certain hour, in an hourly
measurements on a particular date, of a certain month and year.
The total sets of data if one groups on hourly basis are 24 per
day and on daily basis it is 30 or 31 for a month. Now the
ANN is ready for application and using this and sets of data
typical training weights were obtained even if inputs are
scattered on a weekly or monthly basis.
Inpuis
Knowledge for weights, threshold, learning
Outputs
-Lemp.
-wind
speed
-wind
directn.
* fuzzy
“errors
Tocation
"Seas onal
Wweather-on-
Concn. of
pollutants
A such as
S302, NOx
eic.
the day Learning and Training layer
Figure 3. Artificial Neural Network for the problem