Prediction performance showed that all the approaches
provide a quite satisfactory accuracy, showing a correlation
about 0.98. The optimal setup for predictions of the
test set observations was for a network used from Levenberg–
Marquardt training algorithm with two hidden layers
and the number of ten neurons in each layer. The highest
accuracy of simulation was obtained for this network with
an error of 6.22%. It should be noted that except variable
learning rate training algorithm, Levenberg–Marquardt
training algorithm represented lower performance in
terms of MSE values compared to the others. After Levenberg–
Marquardt, scaled conjugated gradient training algorithm
had the highest network performance concerning
accuracy and SSE values followed by conjugated gradient
training algorithm. The gradient descent was the slowest
training algorithm while the quasi-Newton after Levenberg–
Marquardt was the fast. The performance of the networks
with one hidden layers which used from these six
algorithms represented in Table 3. As seen, the Levenberg–
Marquardt training algorithm had the best performance
like the network with two hidden layers. After it,
the quasi-Newton was the best training algorithm concerning
speed, MSE, accuracy and coefficient of determination.
Unlike the network with two hidden layers, the
Levenberg–Marquardt training algorithm with one hidden
layer had the best performance of MSE.