A Study Of Non-stationary Time Series With A Combination Of Two Modeling Thinking

Based on the theories and methods from nonlinear sciences, Regional prediction for non-stationary time series is studied.This article is divided into three parts: In the frist part, based on state space reconstruction and embedding theory, a new regional prediction of non-stationary time series has been presented in this paper. The present methods are to embed external forcing factors in the model, and to look for the space information in the entire attractor which made up of time series of all phase points within the whole region, and to put them into the model as well.In the Second part, starting from the "ideal" time series of non-stationary field, the external forcing information and spatial information is embedded into the system of the reconstruction. Using three method of forced through the stationary mode, the the outer forced mode and space mode conduct the experimental analysis of "ideal" non-stationary time series of33-mode Lorenz system.The results show that:1. External forcing mode of "ideal" non-stationary time series have a better forecasting ability. we can understand that the forcing factor is some kind of repair on the reconstruction of the dynamical system.2. Space-external forcing mode of non-stationary time series have better forecast ability. Introducing the spatial information in reconstruction of power system can make up for the length of time series.In the third part, embedded in external forcing factors (NAO, SOI) and spatial information, we use the method mentioned before to predict the NCEP500hPa anomaly height field. The main results are:3. According to the60predicting cases, the average correlation coefficient between prediction and observed is0.2385, similar to the result from GCM model. It is shown that global approximation method of the time series analysis method has a convincing predictability.4. Embedding the NAO and SOI can effect higher predictability5. Embedding the spatial information can make up for the length of time series to improve the predictability.