Keywords: neural networks, wave spectra, spectral shapes, wave data.

Frequency-domain analysis of ocean structures calls for derivation of a design wave spectrum from the specified wave parameters like the significant wave height and the average wave period. Many alternative forms of empirical equations representing wave spectral density against various wave frequencies are available for this purpose. The choice of a suitable empirical wave spectrum out of these alternatives is subjective. Design codes generally recommend the PM spectrum for the fully developed sea state or the JONSWAP spectrum for the partially developed one. However many investigations in the past (e.g. Narasimhan and Deo [1]) have shown that the above recommendation fails to generalize the actual site conditions. This leaves scope to try out alternative approaches.

All traditional theoretical wave spectra, like the PM and the JONSWAP, are based on the technique of statistical curve fitting to observations. In recent past the technique of neural networks has proved to be a better alternative to statistical schemes, e.g. Karunanithi et al. [2] and Deo et al. [3]. This is presumably due to ability of the neural networks to catch the input-output dependency in a `model-free' and `data-oriented' manner with considerable flexibility and adaptability.

A neural network was developed in order to estimate the wave surface spectral density over a wide range of wave frequencies, out of averaged wave parameters of significant wave height, average zero-cross wave period, spectral width and peakedness parameter. It had four input nodes, 12 hidden nodes and 28 output nodes. The network was trained and validated using two sets of actual wave observations. Out of these one, downloaded from the NDBC (National Data Buoy Center)'s website, pertained to a site, which is 150 NM East of Cape HATTERAS off the U.S. Coast. The other one belonged to a location, which is off the western Indian Coast near Goa.

The network was trained using the standard back-propagation as well as the resilient back-propagation learning algorithm [4]. The trained network when validated for unseen inputs showed that the neural network can be a viable option in order to estimate the shape of the wave spectrum from the specified design wave parameters. (Naithani [5]). This was evident from a qualitative comparison made through spectral plots and scatter diagrams as well as a quantitative comparison achieved by means of coefficients of correlation, coefficients of efficiency and mean square errors between the network-predicted and the target spectral density.

The resilient back-propagation training algorithm provided faster and more accurate learning than the common back-propagation scheme.

The network predictions based on the NDBC data were more satisfactory than the same obtained on the basis of the observations off Goa. This could be due to more systematic and high resolution data collection together with a larger sample size pertaining to the former location.

Comparison with the observed values revealed that the network-predicted spectral shapes were more satisfactory than those yielded by the theoretical spectra of PM, JONSWAP and Scotts.

1

S. Narasimhan, M.C. Deo, "Spectral Analysis of Ocean Waves - A study", Proceddings of the Conference on Civil Engineering in the Oceans IV San Fransisco, ASCE, 1979,pp 877-889.

2

N. Karunanithi, W.J. Grenney, D. Whitley and K. Bovee, "Neural Networks for Riverflow Prediction", ASCE Journal of Computing in Civil Engineering, (2), 1994, pp 201-221.

3

M.C. Deo, A. Jha, A.S. Chaphekar and K. Ravikant, "Neural Networks for Wave Forecasting", Ocean Engineering, 28(7), pp 889-898. doi:10.1016/S0029-8018(00)00027-5

4

Stuttgart Nerual Network Software, version 4.01, University of Stuttgart, Germany, 1995.

5

Reena Naithani, "Derivation of Wave Spectrum Using Neural Networks", M.S. thesis, Indian Institute of technology, Bombay, India, 2002.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £82 +P&P)