基于多时间尺度的回归集成预测模型

The Regression Ensemble Predication Model Based on Multi-Time Scale

引入能够将非线性、非平稳过程的数据进行线性化和平稳化处理的EMD方法，对广东降水的时间序列进行时间尺度分离，从复杂的非平稳信号中提取相对简单以不同时间尺度振荡的准周期信号，选取能较好描述降水周期特征的IMF分量作为建模备选因子，然后以均生回归、均生相关、韵律拟合误差和拟合误差4种方法构建预测模型，结果得到采用多尺度因子构建的4种单预测模型近10年Ps评分和降水距平符号同号率平均分在68～73分和50％～58％之间，而采用4种模型构建的回归集成模型两种评分方法的平均分分别高达79．8和68．8％，较单一预测模型评分分别提高了近10分和10％以上。将具有降水指示信号的前冬赤道东太平洋海温因子耦合到回归集成预测模型，其Ps评分结果与纯降水集成模型相当，但同号率评分略高3．1％。从而，提取要素序列的多种时间尺度特征，并采用多模型的集成预报，均能有效提高短期气候预测水平。

The empirical mode decomposition （EMD） method has the advantage of dealing with the nonlin- ear and nonstationary data, making them linearized and stationary. So EMD is adopted to analyze the pre- cipitation data based on the multi-time scale viewpoint, and the relatively simple semi-period signal with different oscillations are decomposed from the complex nonstationary and nonlinear signal. Then the char- acteristic intrinsic mode functions （IMFs） are chosen to construct the rggression ensemble prediction model （REPM）, which is based on the mean generation regression （MGR） method, the mean generation correla- tion （MGC） method, the rhythm fitting error （RFE） method and the fitting error （FE） method. The re- suits show that the average score of the Ps and the same symbol ratio （SSR） are 68--73 and 50%--58%, respectively, among the four kinds of single models during rainy period in Guangdong for the recent 10 years. However, in the REPM, the average Ps and SSR scores have reached 79.8 and 68.8%, respective ly, increasing 10 scores and 10% or so compared with one of the four kinds of single models. Meanwhile,if the SST signals in tropical East Pacific in the previous winter are coupled into the REPM, the Ps and SSR scores have improved, but the SSR scores, 3.1% higher than the former. Therefore, both the multi- time scale information extracting from the meteorological elements and the ensemble model construction can improve the accuracy of short-term climate prediction.