气流后向轨迹和门限重复单元的PM2.5预报  

PM2.5 concentration real-time forecasting method based on GRU model and back trajectories of air mass

作　　者：梁世文 李琦[1,2] 侯俊雄 冯逍[1,2] LIANG Shiwen;LI Qi;HOU Junxiong;FENG Xiao(Institute of Remote Sensing and Geographic Information System,Peking University,Beijing 100871,China;Smart City Research Center,Peking University,Beijing 100871,China)

机构地区：[1]北京大学遥感与地理信息系统研究所,北京100871 [2]北京大学智慧城市研究中心,北京100871

出　　处：《测绘科学》2020年第3期87-94,共8页Science of Surveying and Mapping

摘　　要：针对目前我国实时的空气质量预报不适合重污染天气的问题,该文提出了一种基于气流后向轨迹模型和门限重复单元神经网络的PM2.5浓度预报方法。该方法通过气流后向轨迹模型将区域异地传输效应进行量化,从而为待预报站点提供额外的区域传输预后因子(预后因子即是对未来情况的预估),将待预报站点区域传输预后因子和气象因子加入预报模型,利用GRU模型模拟区域PM2.5浓度的时序连续变化特征,建立1~72 h的PM2.5浓度实时预报模型。实验结果表明,区域传输预后因子的加入,能够很好地量化其他站点对于待预报站点的PM2.5浓度预报影响,提高PM2.5预报模型整体预报精度。Aiming at the problem that the real-time air quality forecasting system was not suitable for the heavily polluted weather in China,a PM2.5 concentration real-time forecasting method based on GRU and back trajectories of air mass was put forward.Back trajectories of air mass could quantify the effects between other monitoring stations and the station to be forecasted.This could provide extra features for the forecasting model.The real-time forecasting model of PM2.5 concentration from 1 to 72 hours could be established by putting the features of back trajectories of air mass and the features of meteorological into the GRU model.Experiment showed that features of back trajectories of air mass could quantify the effects between different monitoring stations well,and improve the accuracy of the PM2.5 forecasting model.

关 键 词：PM2.5实时预报 气流轨迹 门限重复单元 深度学习 

分 类 号：P237[天文地球—摄影测量与遥感]