2278

Artificial neural networks (ANN) have proven their usefulness in a number of

through nodes or neurons. It is capable of directly correlating input time series of

forcing functions to the output variables through the interconnected nodes with train-

able weights and bias. In contrast to traditional harmonic analysis (Ippen, 1966),

which is used only in the predictions of periodic tidal component, the neural network

model can be trained to recognize and predict both nonlinear and non-periodic sig-

nals. Wong and Wilson's (1984) study of 30-day data indicates that sub-tidal sea

level variation plays an important role in the column exchange between estuary and

ocean through the inlets in the Long Island South Shore. The traditional harmonic

analysis method is unable to provide accurate predictions of long-term water level

variations along Long Island where non-tidal sea level variation is significant.

In applying a trained and validated ANN, output variables are directly calculated

without iteration from the input variables and the vectors of weights and bias in the

network nodes. This functioning is similar to directly find the output from a linear

regression function. Therefore, applying an ANN model takes much less compu-

tational time than the traditional fluid mechanic models, as long as data is available

to establish the ANN model. For this reason, some researchers have combined fluid

mechanics modeling with neural networks to improve the efficiency of model appli-

cations (Bibike and Abbot, 1999). Greenman and Roth at the NASA Ames Research

Center incorporated neural networks with finite element fluid mechanics models to

optimize airfoil design (Greenman and Roth, 1999). A fluid mechanics model can

be used to provide time series outputs of system responses under a few study scen-

arios for a period of time. The time series outputs from fluid mechanics model simul-

ations and forcing functions can then be used as "data" for neural network model

development. A validated neural network model can serve as a cost-effective tool

in quickly assessing the system response to the input factors.

The application of neural networks in oceanographic study is relatively new due

mainly to highly nonlinear characteristics. Hsieh and Tang (1998) discussed several

typical obstacles and provided some suggestions to incorporate neural networks with

other time series forecasting approaches. There are some successful ANN appli-

cations in coastal engineering. For example, Bibike et al. (1999) used ANN to encap-

sulate numerical hydrodynamic model simulations for cost-effective forecasting of

water levels. Mase et al. (1995) adopted ANN to assess the stability of armor unit

and rubble-mound breakwater and found satisfactory agreement between obser-

vations and model predictions. Huang and Foo (2000) employed neural networks to

directly correlate time series of salinity to the forcing functions of winds, water

levels, and freshwater inputs in a multiple-inlet estuary of Apalachicola Bay, Florida.

toric data at the same station, which did not address the non-periodic sub-tidal sea

levels and the correlation with tidal data at other stations. Tsai et al's (2002) ANN