282
H. Wilson . F. Recknagel
r-------~====~Ia
T
Observed
!;J Predicted
o Outlier
o
1990
1991
R,-------------------~------~------------~======~~b
o
1985
1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997
Figure 14.9. The effect of 0 hidden nodes on ANN model predictions. a the
distribution of predictions at each time step in 2 years of the time series. b
Bootstrap aggregate model predictions only for the entire time series. Myponga 30
day model trained to 0 error tolerance
For most ANN structures trained, the presence of hidden nodes did not reduce
test set RMS error of either the individual models or the bootstrap aggregate
models. This is despite the fact that the effect of hidden nodes at reducing RMS
error on training sets was profound (as illustrated by the example for the same day
Lake Kamigaura model in Figure 14.11). The exceptions to this rule were the 30
day ahead model for Myponga Reservoir and both the same day and 30 day ahead
models for Burrinjuck Reservoir (Figures 14.8d, 14.7c,14.7d respectively). In
these cases, the boots trap aggregate models showed a dear benefit from hidden
layer nodes, although the individual (unaggregated) Burrinjuck models illustrated
by the boxplots failed to improve.
H. Wilson . F. Recknagel
r-------~====~Ia
T
Observed
!;J Predicted
o Outlier
o
1990
1991
R,-------------------~------~------------~======~~b
o
1985
1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997
Figure 14.9. The effect of 0 hidden nodes on ANN model predictions. a the
distribution of predictions at each time step in 2 years of the time series. b
Bootstrap aggregate model predictions only for the entire time series. Myponga 30
day model trained to 0 error tolerance
For most ANN structures trained, the presence of hidden nodes did not reduce
test set RMS error of either the individual models or the bootstrap aggregate
models. This is despite the fact that the effect of hidden nodes at reducing RMS
error on training sets was profound (as illustrated by the example for the same day
Lake Kamigaura model in Figure 14.11). The exceptions to this rule were the 30
day ahead model for Myponga Reservoir and both the same day and 30 day ahead
models for Burrinjuck Reservoir (Figures 14.8d, 14.7c,14.7d respectively). In
these cases, the boots trap aggregate models showed a dear benefit from hidden
layer nodes, although the individual (unaggregated) Burrinjuck models illustrated
by the boxplots failed to improve.