Geology Reference
In-Depth Information
Fig. 5.43 Solar radiation as observed and estimated using the W-ANFIS model for the validation
data set. a scatter plot, b line diagram
values in the form of the line plot are given in Fig.
5.42
b. The corresponding
gures
of W-ANFIS during the validation phase are give in Fig.
5.43
a, b.
5.5 Model Selection in Daily Solar Radiation Estimation
in Terms of Overall Model Utility
In this section we have used many data based models ranging from ANN to the very
complex structured wavelet based support vector machines (W-SVM). One of the
growing concerns in the different realms of modelling is whether a more complex
model gives better performance. But the term complexity is very intrinsic to the
parameters and computational processes involved in the modeling system. It is rather
dif
cult to measure or express as a numerical value. In this chapter we follow the
hypothesis that more complex models simulate the processes in better way but with
high variability in sensitivity and relatively less error. Another hypothesis is the
complexity of the data based models could be connected to the training time required
to achieve a particular desired accuracy. We have used a simple index of utility,
de
es the
models in terms of model complexity (we used training time as the indicator of
complexity) model sensitivity (output response to changes in training input) and
model error (closeness of simulation to measurement). Figure
5.44
shows the vari-
ation of error (RMSE) with the model complexity for the Case Study: Solar Radiation
modelling. The RMSE decreases with increasing complexity as expected, but the
variation is not showing a perfect trend. However, the more complex W-SVM and
W-SVMmodels have relatively less erroneous prediction with very high complexity
values. Both the models took huge amount of time for training to achieve a desirable
accuracy. The complexity of NW is observed relatively less compared to other two
wavelet based hybrid models. The better prediction in terms of root mean square error
(combining validation and training) both training and is exhibited by the NW model,
followed by ANFIS, W-SVM, LLR and ANN-LM models. Even though, the
hypothetical relation between complexity and error is not a perfect straight line, we
observed a decreasing linear relation with R
2
value of 0.1305.
ned in the Chap.
2
which evaluates the performance of models and classi
