Abstract
A feedback artificial neural network model (FBANNM) was applied to the real-time prediction of dissolved oxygen (DO) in an eutrophic lake. The results obtained from an ordinary FBANNM showed that prediction accuracy and possible prediction lead-time varied season by season. In particular, prediction accuracy during summer was relatively low, and the model could only predict variation of DO with two-hour lead-time. To improve prediction accuracy and extend possible prediction lead-time, processing for noise elimination based on wavelet analysis was introduced to the FBANNM. In the processing, the DO time series was divided into low and high frequency components, while noise was eliminated from the high frequency component. Using low and high frequency components, FBANNM was constructed to consider the time-frequency characteristics. Reconstructed FBANNM showed remarkable improvement in both prediction accuracy and possible prediction lead-time. Furthermore, parameters representing the strength of lake stratification were incorporated into the input variables of the model. As a result, prediction accuracy and possible prediction lead-time were scarcely improved. It is therefore concluded that DO can only be sufficiently predicted by using its past time-series.
