Page 305 - Contributed Paper Session (CPS) - Volume 7
P. 305
CPS2105 Hermansah et al.
Figure 2: The plot of the actual data and forecasting results of ARIMA
models
b. Modeling with Neural Network
The architecture of the Back-propagation Neural Network (BPNN) is
determined by trial and error on several types of architecture. The activation
function used is the bipolar sigmoid function for the input layer to the hidden
layer and the linear function for the hidden layer to the output layer. The
training model in BPNN was chosen, namely resilient back-propagation. The
parameters chosen are based on the training model selected default with the
specified value. Determining the best model is also done by considering the
smallest MSE and MAPE values.
With 225 training data and 12 testing data, the best architecture of BPNN
by trial and error obtained 4 input layers, 7 hidden layers and 1 output layer
with MSE values of 1,2639 and MAPE of 0,0046 or 0,46%. The best model
obtained is used for forecasting testing data. Obtained forecasting data with
MSE of 2,3735 and MAPE of 0,0075 or 0,75%. Because the MAPE value is below
10%, so it can be concluded that the model has a good performance. Figure
forecasting results with actual data are as follows:
Figure 3: The plot of the actual data and forecasting results of Neural
Network models
292 | I S I W S C 2 0 1 9