Geology Reference
In-Depth Information
with 3,500 data points, whereas the training curve had a slight decreasing change at
that point.
However, looking at the overall performance of the approaches used in this
section, one could say that none of the compared approaches are as strong as GT in
locating the training data length. The entropy theory failed to locate a proper
training length within the available data range, unlike previous case studies
explained in this topic. The traditional approaches such as data splitting, AIC, and
BIC have exhibited an indication of optimum data points as somewhere near 3,500.
All approaches mentioned in this section have identi
ed that evaporation modeling
requires only three inputs [W, T, Ed] for reasonable evaporation modeling in this
study area. Thus, in this case study, we chose the input data combination [W, T, Ed]
(i.e.,
scenario with daily wind speed (W), daily air temperature (T), and
daily saturation vapor pressure de
'
No RH
'
cit (Ed) as inputs) for modeling. We trusted the
Gamma Test identi
ed data length (i.e., 2,327 data points) and that optimum length
has been used for training of different data-based models in subsequent sections.
The remaining data points out that a total of 4,019 were used for testing the models,
including the hybrid wavelet intelligent systems.
7.6 Data-Based Modeling in Evaporation Modeling
In this section we discuss data based evaporation modeling using different basic
data-based models such as ANNs, SVMs, and hybrid models such as NNARX,
ANFIS, NW, wavelet-ANFIS, and wavelet-SVM, concentrating on the recom-
mendations of the Gamma Test mentioned in the previous section.
7.6.1 Data-Based Evaporation Modeling with LLR, ANNs,
ANFIS, and SVMs
The best input combination (with W, T, Ed) derived from the Gamma Test is used
to develop LLR, ANNs (with different algorithms such as BFGS, Conjugate Gra-
dient, and Levenberg
Marquardt), and SVM models. To assess their relative per-
formance, the analysis results based on different statistical parameters are tabulated
in Table 7.9 . As mentioned earlier, the models were trained using the training data
length of 2,327 and the remainder was for testing. The analysis on ANN models
have shown that the ANN model with Levenberg
-
Marquardt performed better than
the other two algorithm-based models with RMSE of 2.98 mm/day during the
training phase and 3.08 mm/day during the validation phase. Other statistical
parameters can be found in Table 7.9 . It is also observed that the performance of
ANN-Conjugate Gradient models is comparable to that of ANN-LM models in
terms of most of the statistical terms. The scatter plots of the different ANN-LM
-
Search WWH ::




Custom Search