3.3.2 Deep Learning Model
The deep learning method is applied. The training set R2 is 0.927 and
the test set R2 is 0.862. It is studied by a simple four-layer fully
connected neural network (Figure 12). Rectified linear unit (ReLu) was
regarded as the activation function . We have fewer training samples. In
neural network training, the trained models are likely to overfitting.
Dropout can be used as a skill in training deep neural networks. In
every training batch, if part of the feature detector is ignored (i.e,
making the value of the partially hidden layer node 0), overfitting can
be reduced. We used the dropout technique in the second and third layers
(the second layer a yerdl is 0.4), the first Three layers of 0.6 to
reduce overfitting. Then mean-square error (MSE) is the loss function.
The Adam optimizer whose perparameters are well explained is appropriate
for unstable objective functions and usually require little or no
adjustment.The learning rate of the Adam optimizer was set to 0.0001
and120 different attempts was performed and we obtained credible
results.