Simulation of Significant Wave Height by Neural Networks and Its Application to Extreme Wave Analysis

Abstract

Abstract The derivation of the long-term statistical distribution of significant wave heights (Hss) is discussed in this paper. The distribution parameters are estimated using artificial neural networks (ANNs) trained with the help of a simulated annealing algorithm and operated in an autoregressive mode. The ANNs were utilized in estimating the parameters of a conditional probability distribution related to a desired Hs given its preceding Hss, approximated by a proposed distribution called the hepta-parameter spline. The performance function during training was based on the likelihood function of the statistical method of maximum likelihood estimation (MLE). Given the observed dataset, the most probable weights and biases of the neural networks were determined in such a way that the performance function was optimized. The distribution could be used in the simulation and forecasting of Hss. This paper also presents an extreme wave analysis using the simulated Hss. The extreme analysis conducted in this study using the maxima method offers an alternative approach, avoiding the unrealistic hypothesis that annual Hss are identically distributed, as is conventionally assumed when using the Fisher–Tippet theorem.