基于机器学习方法的河南省城市夏季臭氧预报研究  

Prediction of Summer Ozone Concentration in Henan Province Based on Machine Learning Method

作　　者：杨盛凯 曹乐 卢西畅 张彤 孔海江 王梦珂 朱晓淳 高萌萌 赵天良 YANG Shengkai;CAO Le;LU Xichang;ZHANG Tong;KONG Haijiang;WANG Mengke;ZHU Xiaochun;GAO Mengmeng;ZHAO Tianliang(School of Atmospheric Physics,Nanjing University of Information Science and Technology,China Meteorological Administration Aerosol-Cloud-Precipitation Key Laboratory,Nanjing 210044,China;Hengshui Meteorological Bureau,Hengshui 053000,China;Raoyang National Climatological Observatory,Hengshui 053900,China;Key Laboratory for Meteorological Disaster Prevention and Mitigation of Shandong,Jinan 250031,China;Shandong Institute of Meteorological Sciences,Jinan 250031,China;Henan Meteorological Observatory,Zhengzhou 450003,China;Anyang National Climate Observatory,Anyang 455000,China;Hebi Meteorological Bureau,Hebi 458000,China)

机构地区：[1]南京信息工程大学大气物理学院,中国气象局气溶胶与云降水重点开放实验室,江苏南京210044 [2]衡水市气象局,河北衡水053000 [3]饶阳国家气候观象台,河北衡水053900 [4]山东省气象防灾减灾重点实验室,山东济南250031 [5]山东省气象科学研究所,山东济南250031 [6]河南省气象台,河南郑州450003 [7]安阳国家气候观象台,河南安阳455000 [8]鹤壁市气象局,河南鹤壁458000

出　　处：《环境科学研究》2025年第4期808-820,共13页Research of Environmental Sciences

基　　金：国家重点研发计划项目(No.2022YFC3701204);安阳国家气候观象台开放研究基金(No.AYNCOF202301);河北省气象局指导项目(No.23zc20)。

摘　　要：臭氧(O_(3))是目前影响我国城市地区空气质量的主要污染物。作为全国重要的综合交通枢纽和人流物流信息流中心,河南省非常有必要持续提高环境空气O_(3)的预报准确性。为进一步提升河南省多城市O_(3)污染的预报准确率,优化其预报时长和效果,该研究基于机器学习方法,利用2017−2021年地面实测污染物数据和逐6 h GFS(Global Forecast System)气象场预报数据,对河南省9个城市(郑州、开封、洛阳、平顶山、安阳、焦作、三门峡、信阳、周口)训练O_(3)浓度预报模型,对2023年夏季(6月、7月)O_(3)浓度进行每日迭代7 d的预报,并定量评估模型预报效果。结果表明:①训练的O_(3)预报模型对河南省9个城市未来7 d O_(3)逐小时浓度的预报效果较好,与观测值的相关系数(r)最高可达0.91。郑州市提前1、2、3 d的O_(3)浓度日最大8 h平均值(O_(3)MDA8)预报值与观测值的r分别达到0.85、0.84、0.84,较现有研究具有更优的预报效果。②模型对9个城市O_(3)的预报表现有所差异,对河南省两个边界城市(信阳市和三门峡市)的预报效果较好,其提前6 d的O_(3)MDA8浓度预报值的均方根误差(RMSE)保持在27μg/m3左右,平均相对误差(MRB)保持在18%左右;但对于在夏季O_(3)浓度受省外传输和偏南风影响较大的安阳市,则预报效果较差。③模型提前1~3 d的预报效果较好,提前4~7 d的预报效果变差;且模型对每日14时O_(3)小时浓度(即13:00−14:00平均浓度)的预报效果较O_(3)MDA8差。④O_(3)MDA8处于0~100、100~160、160~215、215~265μg/m3时,模型提前1 d预报的MRB分别为21.65%、13.63%、9.63%、12.81%,具有较高的精确度。研究显示,基于大气主要污染物同步迭代的机器学习预报方法可为河南省各城市提供未来7 d及时且准确的O_(3)预报,实现区域性O_(3)污染事件的及时预报预警。对于预报效果较差的城市,期望未来通过增加其他O_(3)前体物数据、使用更�Ozone(O_(3))is currently the primary pollutant affecting air quality in urban areas of China.Henan Province,serving as a crucial national transportation hub and center for people flow,logistics flow,and information flow,necessitates continuous improvement in O_(3)forecasting accuracy.In order to further improve the forecasting accuracy of ozone pollution in multiple cities across Henan Province and optimize the forecasting duration and effectiveness,we used pollutant data and 6-hour Global Forecast System(GFS)meteorological field data during 2017-2021 to train an O_(3)forecast model for 9 cities in Henan Province based on the machine learning method,and the model was applied to forecast O_(3)concentrations in summer(June and July)of 2023.The forecasting duration of the model can reach 7 d,and the forecasting performance of the model was also evaluated quantitatively in this study.The results indicate:(1)The 7 d iteration model has good forecasting performance,with a correlation coefficient(r)between predicted and observed values reaching up to 0.91.And the values of r for the maximum daily 8-hour average ozone concentration(O_(3)MDA8)forecasted for Zhengzhou City in 1-3 d lead time are 0.85,0.84 and 0.84,respectively,which demonstrates a superior forecasting performance compared to existing studies.(2)The model performance in forecasting ozone varied across different cities.It performed better in the two boundary cities of Henan Province,Xinyang and Sanmenxia,with the root mean square error(RMSE)of the O_(3)MDA8 in 6 d lead time remaining at about 27μg/m3,and the mean relative bias(MRB)staying around 18%.However,the model performance is poorer for Anyang City,where summer O_(3)concentrations are significantly affected by external provincial transport and southerly winds.(3)The model's forecast performance was better for a lead time of 1-3 d,but deteriorated for a lead time of 4-7 d.Moreover,its forecast performance for 1 h average concentration of O_(3)at 14:00(averaged over 13:00-14:00)was worse than that for O

关 键 词：臭氧(O3) 机器学习 XGBRegressor 河南省 小时浓度预报 

分 类 号：X511[环境科学与工程—环境工程]