Geology Reference
In-Depth Information
Fig. 5.32 The modelling performance of different SVMs with different kernels in solar radiation
modelling
ʵ
-SV regression. For daily solar radiation modelling the SVMs were trained using the
above mentioned list of input vectors and GT recommended training data length as
training data. The study has used an interactive Fortran program to reformat the
original data into an appropriate list of vectors, and to normalise/scale the list of
vectors. The scaling of the input lists are important in SVMs as the difference between
extreme values is reduced, which makes it easier and fast to run the SVM algorithm.
The proper identi
cation of kernel function is so important in SVM based modelling
as kernels are the components which simplify the learning process. There are four
major kernels predominantly used in SVMs such as linear, polynomial, radial and
sigmoid. Each kernel was applied to the SVMs like the
-SVR. A total
of 20 different models with different kernel function and regressors were investigated
to
ʵ
-SVR and the
ʽ
find the best one. The analysis results are shown in Fig.
5.32
.
It has found that the
-SVR with linear kernel function is the best model for the
daily solar radiation modelling. The performance of nu-SVR with linear kernel is
comparable to that of
ʵ
ʵ
-SVR, but the least error and higher accuracy observed in the
case of
-SVR. The cost parameter (C) of error assigns a penalty for the number of
vectors falling between the two hyperplanes in the SVM hypothesis. Estimation of
the optimum number of cost is very important as it has an in
ʵ
uence on the quality
of the data used for the modelling. To ascertain the optimum cost value, the support
vector machine made from the best model
-SVR regression algorithm with linear
kernel was run several times, with differing values of C. This was done for available
6 input vectors. The performance of the models was compared by calculating the
root mean squared error of the output daily solar radiation given by the SVM
hypothesis with that of the actual observed daily solar radiation. The effect of cost
factor on modelling result is shown in Fig.
5.33
. The cost value was chosen to be
ʵ
