基于集成学习的北京地区空气质量站点订正模型  

Correction Model for Air Quality Stations in Beijing Based on Ensemble Learning

作　　者：时跃 勾润玲 石耀辉 马一帆 王璐 骆思远 薛禄宇 SHI Yue;GOU Run-ling;SHI Yao-hui;MA Yi-fan;WANG Lu;LUO Si-yuan;XUE Lu-yu(Daxing Meteorological Office of Beijing,Beijing 102600,China;China Household Electric Appliance Research Institute,Beijing 100053,China;Beijing Meteorological Observation Centre,Beijing 100176,China)

机构地区：[1]北京市大兴区气象局,北京102600 [2]中国家用电器研究院,北京100053 [3]北京市气象探测中心,北京100176

出　　处：《四川环境》2025年第1期1-8,共8页Sichuan Environment

基　　金：北京市气象局科技项目(BMBKJ202201015)。

摘　　要：为了更好的利用机器学习方法降低北京市PM_(2.5)的预报误差。利用集成学习策略集成多种机器学习算法,对睿图-化学模式在北京市35个站点的PM_(2.5)预报进行订正,以黄村站为例选取适用于北京南部地区的最佳订正模型。结果显示,集成学习模型(WE_L3)和随机森林算法模型(RFE_L2)240小时(h)预报的相关系数达到了0.58,比原模式的0.45提高了28.8%。当预报时效大于48小时(h)时,订正模型误差均低于原模式,在预报时效为216h时订正效果最佳,集成学习模型(WE_L3)MAE比原模式低14.07μg/m^(3),下降了38.6%,RMSE比原模式低16.48μg/m^(3),下降了29.6%,随机森林算法模型(RFG_L2),MAE比原模式低14.64μg/m^(3),下降了40.1%,RMSE比原模式低17.06μg/m^(3),下降了30.7%。空间检验显示,订正后PM_(2.5)在北京西北部站点偏差较低,向东南部偏差逐渐增加。综合检验指标,集成学习订正模型集成了多种机器学习算法的优势,是黄村站的最佳订正模型。In order to make better use of machine learning method to reduce the prediction error of PM_(2.5)in Beijing.Ensemble learning strategies integrating multiple machine learning algorithms are used in the research to correct the PM_(2.5)forecasts of 35 stations from RMAPS-CHEM in Beijing.Taking the Huangcun Station as an example,the best correction model applicable to the southern region of Beijing is selected.The results showed that the correlation coefficient between the ensemble learning model(WE_L3)and the random forest algorithm model(RFE_L2)for 240-hour forecasting reached 0.58,an increase of 28.8% compared to 0.45 of the original model.When the period of validity is greater than 48 hours,the error of the correction model is less than that of the original model.When the period of validity is 216 hours,the correction model achieves the best performance.The Mean Absolute Error(MAE)of the ensemble learning model(WE_L3)is 14.07μg/m^(3) lower than that of the original model,a decrease of 38.6%.The Root Mean Squared Error(RMSE)is 16.48μg/m^(3) lower than that of the original model,a decrease of 29.6%.The MAE of the random forest algorithm model(RFG_L2)is 14.64μg/m^(3) lower than that of the original model,a decrease of 40.1%.The RMSE is 17.06μg/m^(3) lower than that of the original model,a decrease of 30.7%.The spatial test showed that the corrected PM_(2.5)had lower deviations in stations northwest of Beijing,gradually increasing towards the southeast.The comprehensive inspection index showed that the ensemble learning correction model integrated the advantages of multiple machine learning algorithms and was the best one for Huangcun Station.

关 键 词：PM_(2.5) 集成学习 模式订正 睿图-化学模式 北京 

分 类 号：X513[环境科学与工程—环境工程] TP181[自动化与计算机技术—控制理论与控制工程]