Application of a nonlinear model in landfall number forecasting for tropical cyclones in China

Meteorological phenomena evolve according to both external influences and their own internal physical processes. Nevertheless, multivariate analysis ignores the evolution of individual meteorological events overtime, while time series analysis does not make full use of the implicit information on influencing factors. Instead, the threshold autoregressive model considers not only the additive effects of influencing factors, but also the processes controlling the evolution of the meteorological phenomena. Meanwhile, this approach deals with the nonlinear problems of meteorological processes through piecewise linearization, yielding improved fit to observations and better forecasts. The pooled variance, mean square error, and maximum fitted error of TARSO(2, (1, 1), (1, 3)) are all smaller than those obtained using TAR(2, 1, 2). The errors of the landfall number associated with TARSO(2, (1, 1), (1, 3)) are smaller than those associated with TAR(2, 1, 2). At present, however, time series data for meteorological processes are generally short, such that the corresponding information system is incomplete. Therefore, extrapolation should not be too far-ranging. It is strongly suggested that the current information system should be supplemented by the addition of new information each year, in the hope of improving future model accuracy and forecast skill.