Predicting Frost Using Artificial Neural Network

Predicting Frost Using
Artificial Neural Network

AUTHORS

Mr. Abhishek Jain

Ms. Ramyaa

Dr. Ronald W. McClendon

Dr. Gerrit Hoogenboom



Presenters

Mr. Abhishek Jain

Ms. Ramyaa

Overview


Introduction


Prediction Approach



Objectives



Methodology



Results



Future Work


Classification Approach




Objectives



Methodology



Results



Future Work


Questions




Introduction


Need


Georgia Automated
Environmental
Monitoring Network.


Why ANN?


Forecasting
–

Two
formats




Prediction




Classification


Predicting Approach


The overall goal of this part of the
research is to develop ANNs that
would predict temperatures for
subsequent 12 hours beginning with
1 hour and at hourly interval .



Objectives


Determine the most important input
parameters.


Determine duration of prior data needed


Develop the best ANN architecture once the
prior objectives have been met.


Determine whether input data of 15 minute
resolution or hourly resolution yields more
accurate forecasts.


Determine whether ANNs developed for a
specific location can be used to make
forecasts at other locations.

Methodology


Organizing Data



Data used from Fort Valley



Date from January to April



Δ

values;
Δ

= present value
–

past value



Input Parameters
-

Temperature, Relative Humidity,






Rainfall, Solar Activity and Wind






Speed and their prior and





Δ

values, Time of Day



Output Parameters
–

Subsequent Temperatures



Prior Data
–

1 hour prior to 6 hour Prior



Production Set
–

Data from years 2001 and 2002



Training and Testing Set
–

Prior to 2001.


Error Measure
–

Mean Absolute Error.

Methodology


Determining which inputs are most important



Prior data is kept constant.



Start with Temperature terms only and then add

one set of parameter terms, one at a time.

Determine most important, and then keep this

constant with

temperature and experiment with

remaining terms adding one at a time and the

process continues.


Determining best prior data configuration



Input parameters kept constant.


Determining best of 15 minute resolution or
hourly resolution data


Robustness of the Models.

Results

Table 1: Impact of various input parameters on the accuracy of the predictions


Temperature

Relative

Humidity

Wind

Solar

Radiation

Rainfall

MAE*

9 Nodes

9 Nodes

9 Nodes

9 Nodes

9 Nodes

(In degree C)

X









1.41

X

X







1.19

X



X





1.34

X





X



1.35

X







X

1.44

X

X

X





1.15

X

X



X



1.17

X

X





X

1.21

X

X

X

X



1.12

X

X

X



X

1.16

*
-

The Mean Absolute Error (MAE) was calculated for only temperatures below 5
o
C in the production set

X
–

Denotes that which input parameters were used. For any given parameter there are 4 prior data values (corresponding to 4 hou
rs of prior data corresponding to 4 input nodes),

4


values (corresponding to another 4 input nodes) and current value (corresponding

to one more input nodes)

Results

Table 2: Impact of prior data on the accuracy of the predictions


Prior Data


(Hour)

Prediction Duration

(Hour)

MAE

(C)

2

1

0.59

4

1

0.64

6

1

0.66

2

12

2.49

4

12

2.48

6

12

2.47

Future Work


15 minute vs Hourly ?


Different Locations?

Classification Approach


The overall goal of this part of the
research is to develop ANNs that
would classify whether temperatures
for subsequent ‘n’ hours beginning
with 1 hour and at hourly interval
would be freeze temperatures or not.


Objectives


The objectives of the previous
approach apply to this one too
–

though the outcome may differ
–

the
set of inputs needed to predict the
temperature may not be the best
ones to predict freeze
–

though we
expect this to be unlikely.


Methodology


The process of meeting the objectives
and getting “the best” network
involves evaluating a network
–

we
are currently working on a
formalization of the evaluation of
different networks.


The evaluation of a network is not
straight forward in this approach


Evaluation of the network


The problem is to classify the current
weather data as corresponding to one that
precedes a “freeze event” or not.


The freeze events are very scarce and are
not evenly distributed.


Failing to predict a freeze event is worse
than doing a false prediction of a freeze
event.


Importance of predicting correctly or
incorrectly depends on the current
temperature.


Evaluation of the network


Importance of missing a freeze event
depends on the length of time interval from
now to the freeze event.


Evaluation of output should give different
weights to a missed prediction when the
network is sure of its wrong output and one
when the network is in doubt


Evaluation should give different weights to
a missed prediction when the temperature
just reached 0
º and rose again immediately
and a freeze event which reached lower
temperatures and lasted longer.


Results


0.0
–
0.1 0.1
–

0.2 0.2
–

0.3 0.3
-
0.4 0.4
-
0.5 0.5
-
0.6 0.6
-
0.7 0.7
-
0.8 0.8
-
0.9 0.9
-
1.
0

Future Work


Different Locations


This is the real challenge as different
locations typically have different
distribution of freeze events
–

and it is
important that the network is a general
one and is not sensitive to the number of
freeze events.


Questions??




Thank You