基于周期信息的时间序列缺失值填补方法研究

项目名称： 基于周期信息的时间序列缺失值填补方法研究

项目编号： No.30872182

项目类型： 面上项目

立项/批准年度： 2009

项目学科： 金属学与金属工艺

项目作者： 张晋昕

作者单位： 中山大学

项目金额： 30万元

中文摘要： （1）本项目顺利完成了预定的计划内容。在SAS中编程实现了具有周期性的时间序列缺失数据填补的模型理论和流程算法。从模拟时间序列和实例时间序列两个方面评估了"基于周期信息的时间序列缺失值填补方法"在连续型缺失和随机型缺失中的填补效果，及相应的模型拟合和序列预测方面的探索性研究。研究结果显示：周期性填补方法对于随机型缺失和连续型缺失的填补效果优于spline填补法；尤其对于连续性缺失，周期性填补的效果凸显自身的优势。 （2）开展了移动假日效应相关的研究工作，比较了观测值比例因子模型和天数比例因子模型在时间序列中移动假日效应的调整效果，并开展了医院住院人数序列的春节效应调整方面的工作。结果显示：序列随机波动水平高、移动假日效应强度大时，观测值比例因子模型的移动假日效应调整效果优于天数比例因子模型；采用regARIMA模型对春节效应的调整效果较满意。 （3）进行了适用于周期性时间序列异常点诊断的探索性研究，形成了含周期性时间序列的异常点诊断流程。结果提示：周期性诊断的方法对于独立出现的加性异常点体现了较好的识别效果。

中文关键词： 时间序列；缺失数据；移动假日效应；异常点；周期信息

英文摘要： （1）In the project, we have fulfilled all the contents in the research plan. The SAS programs of modeling theory and imputation method for the missing value estimation based on periodicity in time series have been accomplished. The applicability of methods by simulating time series and actual time series to the method based on periodicity was assessed, and two missing patterns were discussed - - randomly missing pattern and continuously missing pattern. Besides, we also conducted the exploration for models and forecast using the imputed time series. The results showed that the imputation method based on periodicity revealed better efficiency than spline method either for randomly missing pattern or continuously missing pattern; especially for continuously missing pattern, the imputation method based on periodicity demonstrated further advantages. （2）A related research work of moving holiday effect in time-series was conducted. The adjustment effects between a model with observation-based proportion factors and the model with day length-based proportion factors were compared; simultanineouly, we explored the adjustment method of the Spring Festival effect in the number of admission in the hospital. When average level of the series was high and intensity of moving holiday effect was strong, the adjustment of moving holiday effect of the model with observation-based proportion factors showed better effectiveness. Using reg-ARIMA model to do the adjustment of the Spring Festival effect was necessary, with satisfactory forecasting effectiveness. （3）A research for the technological process and applicability of the method based on periodicity is explored, and the outliers' detection process of time series with significant periodicity was established. According to the results of outliers' detection, it was showed that periodicity detection method had better efficiency for the independently-occuring additive outliers.

英文关键词： Time-series; Missing values; Moving holiday effect; Outliers; Significant periodicity