Neural Networks Approach - Computational Intelligence in Time Series Forecasting - page 131

training sets, N p being the number of adjustable parameters in the network, and I is
the identity matrix of dimension (

NN u )

.

p p

Table 3.1. Nonlinear combination of two forecasts of a temperature series using an artificial
neural network (ANN: Neural networks combined forecast; BJ: Box-Jenkins forecast, HW:
Holt-Winters exponential smoothing)

Serial
No.

Forecast Data sets from HBXIO

matrix

SSE RMSE

1. BJ 151 to 224 (column-1) 0.4516 0.112

2 HW 151 to 224 (column-2) 0.3174 0.0933

3 ANN (2-6-6-1) 1 to 150 (training)

4 ANN (2-6-6-1) 151 to 224 0.1306 0.0594

5 ANN (2-2-6-1) 151 to 224 0.2425 0.0810

The parameter P is multiplied by some factor P inc whenever an iteration step
increases the network performance index ( i.e. sum squared error) and it is divided
by P dec whenever a step reduces the network performance index. Usually the factor
P inc = P dec and in our case it is selected as 10.

Figure 3.19(a). The combination of forecasts using a 2-2-6-1 artificial neural network