基于时空多视图BP神经网络的城市空气质量数据补全方法研究

Urban air quality data completion method based on spatio-temporal multi-view BP neural network

针对城市空气质量监测数据缺失的问题,提出一种基于时空多视图BP神经网络的数据补全方法。采用指数移动平均、普通克里金和非凸矩阵完备作为时空多视图特征,结合映射非线性关系的BP神经网络,构建数据补全模型。以北京市36个站点2014年5月1日至2015年4月30日监测的PM2.5、PM10、NO2、CO、O3和SO26种空气污染物小时浓度为实验数据。实验结果表明,在15%缺失率下,随机缺失补全的平均相对误差为0.102~0.154,时间连续缺失补全的平均相对误差为0.161~0.271,空间连续缺失的补全平均相对误差为0.108~0.155,优于典型的单视图预测方法和多视图线性预测方法。研究成果可为城市空气质量数据补全工作提供方法支持,研究思路可为时空数据挖掘提供参考。

In order to solve the problem regarding urban air quality monitoring data missing, a data completion method based on spatio-temporal multi-view BP neural network is proposed. By adopting exponential moving average,ordinary kriging and non-convex matrix completion as spatio-temporal multi-view features, combined with BP neural network in mapping nonlinear relationship, a novel data completion model is constructed. The hourly concentration of six air pollutants PM2.5, PM10, NO2, CO, O3 and SO2 in 36 monitoring stations in Beijing from May 1 st, 2014 to April30 th, 2015 are used as the study data. Our experimental results show that under 15% missing rate, the mean relative error of missing random data completion is in the range from 0.102 to 0.154, the mean relative error of missing temporal data completion is in the range from 0.161 to 0.271, and the mean relative error of missing spatial data completion is in the range from 0.108 to 0.155, which are better than typical single-view methods and multi-view linear methods. This study provides an effective method for data completion of urban air quality, and reference for spatiotemporal data mining.