W. Huang et al. / Ocean Engineering 30 (2003) 22752295
2286
the model weights and bias parameters can be saved for future application. The data
set used for model training should be continuous and without gaps. Due to missing
data gaps in the Long Island field observation data, we were able to use only short-
term continuous data sets for about 30 days in model training.
Comparison of model predictions and observations during training and verification
phases are given in Fig. 6 for Station P2, and Fig. 7 for Station P8. Results show
that the RNN--WL neural network model was satisfactorily trained to determine the
weight parameters in the network so that the inputs match well with the target time
series. Moreover, the backpropagation neural network was trained to recognize the
time series pattern. Keeping the same weight parameters determined in the training
phase, the model was able to provide satisfactory predictions for an independent
data set during the verification phase. The correlation coefficients between model
predictions and observations ranged from 0.968 to 0.985 during the verification phase
for all the stations. The root mean square errors (RMSE) were all approximately
0.06. A summary of the statistics of the model performance is given in Table 3. As
shown in Fig. 7b, the neural network model provides good predictions of water levels
that consist of both tidal and non-tidal signals.
Fig. 7. RNN--WL model training and verification at Station P8 of Fire Island Inlet. The distance
between input NOAA station at Montauk and the inlet is about 90 km.