Geology Reference
In-Depth Information
inherent ability to escape local minima. Its capabilities in eliminating unnecessary
input vectors are evident from the comprehensive modelling even though the success
is application/model-speci
is effectiveness and
rapidness should be acknowledged which are well ahead with other many approaches
to structural optimisation (e.g.: pruning) calibration data length selection (e.g.: data
splitting) and input selection (e.g.: cross correlation).
The study has demonstrated the use of nonlinear modelling methods like the LLR
and ANNs with BFGS neural network training algorithm, Levenberg
c. In general the GT methodology
'
Marquardt
algorithm and conjugate gradient training algorithms in conjunction with GT. Both
radial BFGS neural network training algorithm and Conjugate Gradient training
algorithms networks performed reasonably well in modelling the validation data but
both failed to reach the highest possible values compared to that of Levenberg
-
-
Marquardt algorithm based ANN model. Among them, the Conjugate Gradient
training algorithms is superior because of its better performance than radial BFGS
algorithm. In the meantime, the LLR technique was able to provide more reliable
estimations compared with the ANN models. It would be interesting to explore this
further in other catchments to con
rm if similar results could be repeated. In model
input section regards, the Gamma Test presented in this study would have a huge
potential to help hydrologists in solving many uncertainty issues in hydrological
modelling process.
This case study has also introduced the use different hybrid models like wavelet
neural networks (NW), wavelet support vector machines (W-SVM) and wavelet
neuro fuzzy models (W-ANFIS) for nonlinear function approximation and mod-
elling in conjunction with GT. The study has performed the training method
through Levenberg
Marquardt algorithm implemented in a back propagation
scheme in the case of NW model. Though, the model is a bit complex, overall
training procedure appeared simple for NW model and the results were reliable.
This study has identi
-
-SVR (with cost
value of 10 and slack value of 1) is the best SVM architecture for daily solar
radiation modeling for the Brue catchment. The same structure was used for
W-SVM with decomposed sub series of the above mentioned six data series as
inputs. The modeling performance of these hybrid models were compared with that
of basic structures like ANN-LM, ANFIS and SVMs. The conventional way of
wavelet based modeling uses the summed decomposed subseries with higher cor-
relation coef
ed that a combination of linear kernel with
สต
cient with original and desired data series to be modeled. The study
suggests decomposing the data series into three resolution levels and to use
rst
three details and third approximation series for modeling as mathematically it can
reproduce the original data. The wavelet incorporation for modeling has been
observed as very useful in conjunction with ANN and SVMs but a slight degrade is
observed in the case of W-ANFIS. It has observed that data properties and sudden
change (could be noise or effect of natural phenomenon) in trend of data can be
observed from the wavelet components at different scales than the original signal
and predictions are more accurate compared with those obtained directly from the
