Chapter 11 . Input Selection for an Aigal Bloom Model
r··- ............... _._._ .. _ ............. _........ . .................... _ ....... _ ......... .
4500 ,
I
i I - Acluol I
-
I
- - - -
-
I
~ 1500 j - - . - - - - _. -
g
1
~
I
If 3000 .,
.g
!
~ 2500 t
g
I
o
'
LI 2000 + -: 1500 ~
1 !
~ 1000 j
'<
I
500 "I
20
40
60
Time (wecks)
80
100
229
Figure 11.4.
Forecasts and actual concentrations of Anabaena spp. for the
validation period for model 3 (input determination: PCA + GA-ANN).
Table 11.3. RMSE for the 4-week Forecasts
apriori
PCA
SOM
Data Set
identification
GA
Stepwis GA
Stepwis GA
Stepwis
e
e
e
Model No.
1
2
3
4
5
6
Training
256.3
382.0
371.6
439.8
420.3
398.3
Set
Testing Set 505.4
491.3
561.3
577.9
504.0
502.6
Validation
386.2
517.1
436.4
549.4
436.4
602.1
Set
RMSE are given in cells/ml
dimensionality of the input set. However, the results for the test set suggest that
the SOM ANN models perform slightly better than the PCA ANN models. In this
theoretical study, the performance of each model has been based on the
independent validation set, however, in a real-world application, the decision of
wh ich input determination method to use would be based entirely on the test set
performance. Thus, this contradictory result requires further investigation but may
in part be due to the training, testing and validation sets not being entirely