Geology Reference
In-Depth Information
Fig. 8.14 ANN modeling results during the validation phase
8.3.4 Application of Support Vector Machines
The conventional SVM model has used the
first 566 data points of the available
data points as the training data set, using four input data time series as per the
recommendations of the Gamma Test. The scaling of the input lists are important in
SVM modeling as the difference between extreme values is reduced, which makes it
easier and faster to run the SVM algorithm. So, we have normalized the complete
data sets in a zero to one range. The proper identi
cation of the kernel function out
of the four functions is important in SVM-based modeling as kernels are the
components which simplify the learning process. Trial and error modeling was
adopted to identify the suitable kernel functions and it was found that the v-SVR
with polynomial kernel function is the best model for the discharge modeling in the
Beas basin for the selected input space (results are not shown here). 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 cost is very
important as it has an in
uence on the quality of the data used for the modeling. To
ascertain the optimum cost value, the SVM made from the best model v-SVR
regression algorithm with polynomial kernel was run several times and the mean
least error is lowest when the C parameter is 2. These parameter settings were used
for the v-SVR model for calibration and validation of the data sets, and to compare
with the prediction capabilities of ANNs and SB.
The modeling results from the SVM models in the training and validation phases
are shown in Figs. 8.15 and 8.16 . One can note from the pictorial results that the
model has failed to simulate sudden variations of
flow during the monsoon
season but has better agreement during the non-monsoon months. The performance
of the model was compared using major statistical indices such as coef
ood
cient of
correlation (CC), slope, Nash and Sutcliffe Ef
ciency (NS), and percentage bias
error. The Nash coef
cient of ef
ciency of the conventional SVM model, which
 
Search WWH ::




Custom Search