融机器学习与WRF大气模式的PM2.5预报方法  

PM2.5 forecasting method based on machine learning and WRF hybrid model

作　　者：侯俊雄 李琦[1,2,3] 朱亚杰 冯逍[1,2] 林绍福[3] 

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 [3]北京未来网络科技高精尖创新中心,北京100124

出　　处：《测绘科学》2018年第2期114-120,141,共8页Science of Surveying and Mapping

摘　　要：针对当前我国重污染天气实时的空气质量预报问题,该文提出了一种融合随机森林算法与WRF大气模式的PM2.5浓度实时预报方法。该方法结合了北京市地面空气质量监测数据和WRF气象数据进行分析,将高层大气状态(如逆温层高度等)融入了预报模型中,建立了0~72h的PM2.5浓度实时预报模型。实验证明,该模型能够对0~72h单站点的PM2.5浓度进行较高精度的实时预报,且在24~72h的长时预报结果上较基于地面空气污染物数据与地表气象站数据的预报方法精度有明显提升,即该方法可以更好地模拟大气物理化学状态,从而更为精准地进行长时PM2.5浓度预报。Aiming at the real-time air quality forecasting system suitable for the heavy-polluted weather in China,a PM2.5 concentration real-time forecasting method based on machine learning and WRF hybrid model was put forward.This method combined the ground air quality monitoring data of Beijing and the WRF meteorological data to analyze the high-level atmospheric conditions（such as the inversion layer height）into the forecasting model,and established the real-time forecasting model of PM2.5 concentration from 0 to 72 hours.Experiments showed that this model could predict the PM2.5 concentration in a single station from 0 to 72 hours in real time with high accuracy and was more accurate than the model only based on air pollutants and weather stations data in the long-term forecast of 24~72 hours,the accuracy of forecasting method was obviously improved,this method could better simulate the physicochemical state of atmosphere to carry out more accurate long-term PM2.5 concentration forecasting.

关 键 词：PM2.5实时预报 WRF大气模式 随机森林 空气质量 

分 类 号：P208[天文地球—地图制图学与地理信息工程]