Artificial Intelligence Applied to Extreme Value Prediction of Non-Gaussian Processes with Bandwidth Effect and Non-monotonicity

摘要

Extreme value prediction of a short-term non-Gaussian random process like ocean waves has been a tough issue for decades. In the 1990’s Winterstein proposed a cubic Hermite transformation using skewness and kurtosis, which has been widely applied in many areas for its accuracy and robustness. However, this approach is valid for monotonic transformation and narrow-banded processes. When the bandwidth of a random process is wide, no reasonable methods are available for acquiring the extreme value. This paper therefore applies the artificial neural network and genetic algorithm to do the extreme value prediction, without seeking rigorous mathematical derivations. Not only skewness and kurtosis are used, the spectral moments up to 4^{th-order reflecting bandwidth effects are also adopted. The results of many random case studies show that the artificial intelligence method is more accurate than the Hermite method in most of situations, especially for non-monotonic transformations. Besides, the artificial intelligence method has a wider application range.}