Parametric optimization of the algorithm (SVM and RF)
ML models’ performance is influenced by their parameters, making it
essential to optimize the parameters by pre-setting the value of MSE.
There are several methods adopted to address the poor prediction
accuracy and acquire the optimal parameter.
(1) There are two ways to adjust the network parameters of the BPNN: (a)
Determining the minimum MSE value by comparing different numbers of
hidden layers or nodes in the hidden layer to obtain the optimal
parameters. (b) Algorithms of PSO and GWO were combined for parameter
optimization and the optimization ability of two methods was compared inFigure 17 .
(2) The penalty coefficient c and gamma g play an
important role in the SVM model, where c determines the
generalization ability and g affects the prediction accuracy. The
libSVM toolbox (Chang and Lin, 2001) is adopted for parameter
optimization of c and g , the process of which is shown inFigure 9 (a).
(3) In RF model, leaves’ number has an impact on the prediction accuracy
and grown trees determine whether the model will be over-fitted. The
process of optimization are shown in Figure 9 (b).
(4) The optimal number of nodes of the hidden layer is needed to be
determined in the ELM. Therefore, the models’ optimization search is
completed by pre-setting MSE.
(5) The XGB optimizes parameters through regularization,
cross-validation, and so on.
4.2 ML predictive performance
analysis
[Insert Figure 10]
Figure 10 Comparison of
predicted and observed Q0.3
[Insert Figure 11]
Figure 11 Comparison of
predicted and observed Q5
[Insert Figure 12]