Determination of Wave Spectrum with Intelligent Computing

Abstract

Considerable works have been reported over last few decades on determination of wave spectral shapes. The resulting spectral models still involve large uncertainties, especially in handling spectra with two or more peaks. Hence studies indicating new options or re-examining some earlier ones in this regard should be welcome. This paper reports results from a fresh investigation carried out to determine wave spectral shapes with neural network (NN)s, support vector regression (SVR) and model tree (MT)s based on measurements made by wave buoys at two locations in Arabian Sea and Bay of Bengal off the Indian coast. It was found that in general all the three modern data driven approaches worked satisfactorily and they were more useful than the commonly adopted semi-empirical spectra. In order to obtain consistent and stable results the model calibration however needs to be made in a careful manner. Separate training over low and high sea states as well as for cases of double and multi-peaked spectra paid rich dividends in this regard. The spectra with two or more peaks were more efficiently handled by neural networks and SVR than MT. The operation of MT involving physical domain splits without non-linear transformations appeared to have over-simplified the underlying complexities. For obtaining more sustained results neural networks can be recommended, although the user's experience and confidence in using the other two methods may make them still successful in a given application.