An online sequential extreme learning machine for tidal prediction based on improved Gath–Geva fuzzy segmentation

Abstract

A novel sampling pool selection scheme is proposed for the online sequential extreme learning machine (OS-ELM) based on improved Gath–Geva (IGG) fuzzy segmentation algorithm. Tidal change is a time-varying process whose dynamics vary with changes of internal and environmental factors such as celestial bodies movements, coastal topology and environmental disturbances. When OS-ELM is implemented for identifying time-varying system dynamics, it usually sequentially selects samples with fixed number. Under such circumstance, samples representing different system dynamics are mixed together so that the online representation and prediction abilities of OS-ELM may be deteriorated. To consciously select samples with most representing ability and construct appropriate sampling pool for OS-ELM, in this study, a dynamic sampling pool selection scheme is proposed based on IGG fuzzy segmentation approach. Time series of input and output variables are segmented as per their dynamics characteristics. The change points split up the time series into several segments and the change points themselves represent the changes of system dynamics. Samples within the same segment are considered as possessing homogeneous characteristics. To achieve best representing abilities for current system dynamics, the proposed IGG-based sampling scheme is implemented for selecting sampling pool. The OS-ELM selects homogeneous samples from sampling pool thus possesses better representing ability for current dynamics. In the meantime, conventional harmonic analysis is also applied to represent the influences of celestial bodies and coastal topology. The harmonic method and IGG-based OS-ELM are combined together and the resulted modular prediction scheme is applied for online tidal level prediction of ports of King Point, Mokuoloe and Old Port Tampa in the United States. Simulation results demonstrate the feasibility and effectiveness of the proposed sampling scheme and the modular tidal prediction approach.