作　　者： (聂斌）; (胡雪）; (王曦）;

机构地区： 天津大学管理与经济学部

出　　处： 《统计研究》 2017年第8期61-70,共10页

摘　　要： 空气质量数据具有在时间上连续、空间上相关的特点,这提高了异常点识别的难度。本文提出在时间维度上运用移动平均法,在空间维度上运用反距离加权法对观测值进行预测并求残差的解决思路,从而将时空数据的异常点识别问题转化为二维残差值的异常点检测问题。通过仿真验证表明新方法具有良好的检出力。最后将新方法应用于北京市实际观测数据,取得了满意的识别效果。 Air quality problems have received worldwide attention in recent years. Since the air quality data has continuous and spatial features, the difficulty of outlier detection has been improved. In this paper, we use Moving Average method in the time dimension and use Inverse Distance Weight method to predict and calculate residual error in the space dimension so that outlier detection problem of Spatio-temporal data is transformed into outlier detection problem of two-dimensional residual error value. In the two-dimensional space of the residual error value, abnormality degree of each point to a plurality of neighboring points is calculated by the Nearest Neighbor algorithm. The outlier is determined when the probability of abnormality degree is greater than the threshold value exceeds the predetermined value. The simulation results show that the new method has favorable detection power. Finally, the new method is applied to the real observation data set in Beijing and the satisfactory recognition result is achieved.

关 键 词： 空气质量 时空数据 异常点识别 最近邻