Basin scale wind-wave prediction using empirical orthogonal function analysis and neural network models

Abstract

A new method is discussed using neural network models in combination with empirical orthogonal function (EOF) analysis for the basin-scale wind-wave forecast. For the Bay of Bengal region EOF analysis has been performed separately on the significant wave height (SWH) data, zonal (U) and meridional (V) components of wind data. For basin scale forecast the dominant principal component (PC) has been subjected to univariate and multivariate neural network models for future predictions. In the univariate approach, only past values of SWH time series are used and in the multivariate approach, U and V time series are used to predict future SWH values. Efficiency in terms of accuracy and speed of four different backpropagation algorithms, namely, Levenberg-Marquardt (LM), Bayesian Regularization (BR), Scaled Conjugate Gradient (SCG) and Fletcher Conjugate Gradient (CGF) have been compared for 1 to 12 multistep ahead time steps and 1 to 13 neurons. After training the models using varied neurons and the PCs, representing the entire basin, the neurons are fixed at which minimum errors are obtained. Further experiments are conducted using the fixed neurons and the PCs for 1 to 12 time steps ahead SWH prediction. Finally independent datasets consisting of normal and cyclonic wind-wave parameters are tested successfully using the above fixed neurons for delays (1 to 12) corresponding to 3 days or 72 h forecast. The novelty of the study lies is the usage of the PCs which represent the entire basin rather than computations at individual locations which are expensive technically and time consuming.