Multi-step-ahead model error prediction using time-delay neural networks combined with chaos theory

This paper presents a time series prediction scheme using time-delay neural networks combined with chaos theory. To achieve reliable multi-step-ahead prediction, the optimal architecture of networks is determined by average mutual information and false nearest neighbors analyses in chaos theory. The networks are applied to predict the model errors at four measurement stations in the Singapore Regional Model domain, with five prediction horizons ranging from 2 h to 96 h. It is found that the combined scheme significantly improves the accuracy of tidal prediction, with more than 70% of the root mean square errors removed for 2 h tidal forecast and more than 50% for 96 h tidal forecast.