Chapter 14 . Time-Series Prediction of Chlorophyll a in Lakes
275
14.4
Results
As discussed in section 14.3.3, the ANN structures had no difficulty in
approximating models that were highly accurate on training data. Figure 14.3
shows a time series plot of the observed and predicted algal abundance for the
Lake Kasumigaura same day model utilising 5 hidden nodes and trained with 0
error tolerance.
The very close correlation between the observed and the
predicted traces illustrates the technique's approximation power given sufficient
hidden neurons and training iterations. This level of predictive performance on
training data was general for all the ANNs trained irrespective of the database.
Kasumigaura - same day model (training)
-
Observed
- _. Predicted
o
1983 1984 1985 1986 1987 1988 1989 1990 1991
1992 1993
Figure 14.3. Predictions of model on training data. Lake Kasumigaura same day
model - 5 hidden nodes, 0 error tolerance.
Figures 14.4 to 14.6 show time series plots of the observed and predicted algal
abundances for the 6 case studies investigated on test data. In each case,
predictions of both the same day and the 30 day ahead model has been illustrated.
The model predictions are the bootstrap aggregate models derived from the ANN
structure comprising of 5 hidden nodes and trained with 0 error tolerance. In all
cases, a visual appraisal of these results shows that the generic ANN model
structure was able to predict the trajectory of algal abundance to a greater or lesser
degree. The best performing ANNs were the 30 day ahead models for the
Myponga Reservoir and the Burrinjuck Reservoir (Figures 14.4d and 14.6b),
where the model was succesfully able to forecast the on set, duration, and
magnitude of nearly all significant bloom events with a high degree of accuracy.
In all the other cases, models predicted the onset, duration and magnitude of
some, but not all, of the significant bloom events. In these cases, while the model
frequently predicted increases or decreases in chlorophyll a, they frequently
under-estimated the peak biomass. In a few models, such as the same day and 30
day ahead models for Lake Soyang (Figures 14.6c and 14.6d), there were
significant false positive predictions (i.e. predictions of blooms when there was no
observed event). However, for the rest of the case studies, the generic structure
avoided false positive predictions reasonably weIl.
275
14.4
Results
As discussed in section 14.3.3, the ANN structures had no difficulty in
approximating models that were highly accurate on training data. Figure 14.3
shows a time series plot of the observed and predicted algal abundance for the
Lake Kasumigaura same day model utilising 5 hidden nodes and trained with 0
error tolerance.
The very close correlation between the observed and the
predicted traces illustrates the technique's approximation power given sufficient
hidden neurons and training iterations. This level of predictive performance on
training data was general for all the ANNs trained irrespective of the database.
Kasumigaura - same day model (training)
-
Observed
- _. Predicted
o
1983 1984 1985 1986 1987 1988 1989 1990 1991
1992 1993
Figure 14.3. Predictions of model on training data. Lake Kasumigaura same day
model - 5 hidden nodes, 0 error tolerance.
Figures 14.4 to 14.6 show time series plots of the observed and predicted algal
abundances for the 6 case studies investigated on test data. In each case,
predictions of both the same day and the 30 day ahead model has been illustrated.
The model predictions are the bootstrap aggregate models derived from the ANN
structure comprising of 5 hidden nodes and trained with 0 error tolerance. In all
cases, a visual appraisal of these results shows that the generic ANN model
structure was able to predict the trajectory of algal abundance to a greater or lesser
degree. The best performing ANNs were the 30 day ahead models for the
Myponga Reservoir and the Burrinjuck Reservoir (Figures 14.4d and 14.6b),
where the model was succesfully able to forecast the on set, duration, and
magnitude of nearly all significant bloom events with a high degree of accuracy.
In all the other cases, models predicted the onset, duration and magnitude of
some, but not all, of the significant bloom events. In these cases, while the model
frequently predicted increases or decreases in chlorophyll a, they frequently
under-estimated the peak biomass. In a few models, such as the same day and 30
day ahead models for Lake Soyang (Figures 14.6c and 14.6d), there were
significant false positive predictions (i.e. predictions of blooms when there was no
observed event). However, for the rest of the case studies, the generic structure
avoided false positive predictions reasonably weIl.