Geology Reference
In-Depth Information
for the ANN models to
find the best number of hidden layer nodes for modelling.
A feed forward ANN was constructed with nine hidden neurons and trained using
the conjugate gradient algorithm and the training results are presented in Figs.
5.12
and
5.13
. It can be seen that the large number of inputs tends to make a good model
in training but does not guarantee superior results in testing. The best scenario is a
model with three inputs (as indicated by the GT) and 1,010 data points (as indicated
by the entropy theory) for training, regardless of whichever model is used (LLR or
ANN). The presented results in Figs.
5.14
and
5.15
shows that the ANN model
performed better on the testing data set than the LLR model, albeit the general
conclusions on the input variable selection and training data length are the same.
Figures
5.16
and
5.17
show the scatter plots of the computed using ANN model and
the observed daily solar radiation for the best scenario during the training and
validation periods. The LLR modelled solar radiation and the observed solar
radiations for the best scenario (Scenario 2) with the testing data set are shown in
Fig.
5.18
.
Fig. 5.16 Observed versus ANN modelled solar radiation in the training data set for scenario 2
