一种新的非平稳时间序列直接建模方法

A NEW METHOD OF MODELING NONSTA-TIONARY TIME SERIES DIRECTLY

This paper focuses cn the problem of modeling a kind of nonstationary time series with random walk and ramp. Based on adaptive Kalman filtering, a new method (MRWR) which can be used to get AR and ARMA models with random parameters directly from nonstationary signals is presented. Considering the time-varying characteristics of nonstationary signal statistics, the paper proposes an engineering order making criterion based on an innovation approach. The experiments of modeling show that, in contrast with RELS, MRWR has higher accuracy for modeling nostatioary time scries with random walk and can model the signals with ramp and random walk simultaneously.