基于EMD方法的非线性/非平稳时间序列的气候预测研究

目前，大多数统计预测模型均假设时间序列或观测数据是线性和平稳的。然而，自然界的观测资料是非线性和非平稳的，通常很难用这些数学模型预测它们。本文针对这一问题提出了一个新的预测方案，即首先利用经验模态分解方法将非线性/非平稳时间序列平稳化，得到一系列本征模函数(IMF)；其次用均生函数模型预测各IMF分量；最后以所有IMF的预测值为新样本对源序列作最优子集回归模型的拟合及预测。结果表明每个IMF，尤其是特征IMF(即特征层次)比源序列有更高的可预测性。该方案为气候预测开辟了一条新的有效途径。

On the climate prediction of nonlinear and non-stationary time series with the EMD method

At present, most of the statistical prediction models are built on the basis of the hypothesis that the time series or the observation data are linear and stationary. However, the observations are ordinarily nonlinear and non-stationary in nature, which are very difficult to be predicted by those models. Aiming at the nonlinearity/non-stationarity of the observation data, we introduce a new prediction scheme in this paper, in which firstly using the empirical mode decomposition the observations are stationarized and a variety of intrinsic mode functions (IMF) are obtained; secondly the IMFs are predicted by the mean generating function model separately; finally the predictions are used as new samples to fit and predict the original series. Research results show that the individual IMF, especially the eigen-IMF (namely eigen-hierarchy), has more stable predictability than the traditional methods. The scheme may effectively provide a new approach for the climate prediction.