A Modular Prediction Mechanism Based on Sequential Extreme Learning Machine with Application to Real-Time Tidal Prediction

Abstract

Neural networks have been proved to be efficient for online identification and prediction of complex nonlinear systems. However, for systems with time-varying dynamics which are common in practice, networks achieved by holistic learning scheme cannot reflect the time-varying local dynamics of system. In this study, a modular prediction scheme is proposed by combining the mechanism model with a neural network predictive model which is online acquired by a sequential learning extreme learning machine (ELM) based on a sliding data window (SDW). The SDW-based ELM (SDW-ELM) is online constructed by learning samples in the real-time updated SDW, is suitable for online identification and prediction of time-varying system dynamics. Tidal prediction is essential for marine safety and efficiency, but changes of tidal level is a typical time-varying system which varies not only with the revolutions of celestial bodies but with the environmental influences such as atmospheric pressure, wind, rainfall and ice. The harmonic analysis method is used to represent the influences of celestial bodies, while the SDW-ELM is used to represent the influences of meteorological factors and other unmodeled factors. Therefore, the proposed modular based on SDW-ELM is applied for real-time tidal level prediction based on measurement data in Port Hardy. Simulation results demonstrate the effectiveness and efficiency of the proposed algorithm and results are compared with that by online sequential ELM (OS-ELM) algorithm and SDW-ELM.