IJCNN - Multivariate chaotic models vs neural networks in predicting storm surge dynamics

The recently developed methods in nonlinear dynamics and chaos time series analysis are used in this study to analyze, delineate and quantify the underlying coastal water level and surge dynamics in the North Sea along several locations at the Dutch coast. This study analyzes seven water level and surge data sets, five of which characterize coastal locations and two relate to the open sea locations. Both the water level data and the surge data (with the astronomical tide removed) are analyzed. The main objective of this analysis is to delineate and quantify the underlying dynamics of the coastal water levels and to quantify the variability and predictability of the coastal dynamics along the Dutch coast based on time series of observables. Based on the reconstructed multivariate phase space of the water level and surge dynamics, adaptive multivariate local models were built which typically yield more reliable and accurate short-term predictions compared to neural networks.