Predicting Extreme Waves From Wave Spectral Properties Using Machine Learning

Abstract

An important question in the context of rogue waves is whether the statistical properties of individual waves, and in particular the probability of extreme and rogue waves, can be linked to the properties of the underlying wave spectrum of the relevant sea state. It has been suggested that a narrow wave spectrum (in frequency or direction) combined with a large wave steepness may lead to increased occurrence of extreme waves. Parameters based on the ratio of the wave steepness to the spectral band-widths have therefore been suggested as indicators of increased probability of extreme waves. However, for realistic ocean conditions the success of such parameters seems to be questionable. In this paper, we investigate relations between short-time wave statistics and wave spectral properties by using machine learning methods that can take a much wider range of spectral properties, or even the entire directional wave spectrum, into account. Numerical simulations with a nonlinear wave model that provides phase-resolved wave information are combined with wave spectra from a spectral wave model. Machine learning methods are then employed to investigate how well the wave statistics can be predicted from knowledge about the wave spectrum. The results are discussed in the context of existing parameters suggested as indicators of rogue waves, as well as with respect to potential warning against sea states in which extreme waves are expected to occur, based on wave-forecast from spectral wave models.