• 关键词：PM_2.5  数据校正  光散射法  随机森林  去噪

• 基金项目：湖北省重点研发计划项目（No.2020BCB074）；中国科学院精密测量科学与技术创新研究院多学科交叉培育项目（No.S21S7102）

• 宋金文
• 中国科学院精密测量科学与技术创新研究院， 环境与灾害监测评估湖北省重点实验室，武汉 430071;中国科学院大学，北京 100049

• 何报寅
• 中国科学院精密测量科学与技术创新研究院， 环境与灾害监测评估湖北省重点实验室，武汉 430071;中国科学院大学，北京 100049

• 胡柯
• 武汉市生态环境监控中心，武汉 430015

• 万祎
• 武汉市生态环境局，武汉 430022

• 符祖文
• 武汉市生态环境局东西湖区分局，武汉 430040

• 冯奇
• 中国科学院精密测量科学与技术创新研究院， 环境与灾害监测评估湖北省重点实验室，武汉 430071;中国科学院大学，北京 100049

• 杨帆
• 中国科学院精密测量科学与技术创新研究院， 环境与灾害监测评估湖北省重点实验室，武汉 430071;中国科学院大学，北京 100049

• 摘要：光散射法传感器微站以其体积小、反应迅速、成本低等优点，已成为城市PM_2.5规模化移动监测的新选择.由于其标准与传统标准台站不同，必须对这类微站的监测数据进行准确地校正.本研究利用2021年06月—2022年02月武汉市江夏区标准台站及同期传感器微站监测数据，探讨传感器微站监测误差与温度、相对湿度的关系，并通过随机森林回归（Random Forest Regressor，RFR）校正传感器微站PM_2.5监测数据.对比单一RFR模型、按气象因素分类后RFR模型、“小波去噪+RFR”组合模型、“加权滑动平均去噪+RFR”组合模型校正效果，结果表明：RFR模型和分类后RFR模型均出现泛化能力差的问题，不能满足校正需求；“小波去噪+RFR”组合模型、“加权滑动平均去噪+RFR”组合模型平均绝对误差分别为8.77 μg·m^-3和4.78 μg·m^-3，平均相对误差分别为40.80%和18.13%.去噪组合模型能满足校正需求，且“加权滑动平均+RFR”组合模型校正效果明显优于“小波去噪+RFR”组合模型.研究结果可为光散射法传感器微站PM_2.5监测值校正提供有益参考.

• Abstract：Light scattering sensor micro-stations have become a new choice for the large-scale mobile monitoring of PM_2.5 in cities， mainly by virtue of their small size， rapid response， and low cost. Due to the differences between light scattering sensor micro-stations and traditional stations in monitoring standards， it is necessary to accurately correct the monitoring data of such micro-stations. Relying on the monitoring data of the standard station and micro-station in Jiangxia District， Wuhan from June 2021 to February 2022， this study explored the relationships of the monitoring errors of the sensor micro-station with temperature and relative humidity， and corrected its PM_2.5 monitoring data using random forest regression （RFR）. Then， this study compared the correction effects of the single RFR model， the RFR model based on classification of meteorological factors， the combined model of "wavelet denoising + RFR"， and the combined model of "weighted moving average denoising + RFR". The results showed that both the single RFR model and the post-classification RFR model had poor generalization ability， which could not meet the correction requirements. The mean absolute errors of the combined model of "wavelet denoising + RFR" and the combined model of "weighted moving average denoising + RFR" were 8.77 μg?m^-3 and 4.78 μg?m^-3， respectively. In addition， their mean relative errors were 40.80% and 18.13%， respectively. The denoising combination model can meet correction requirements. The correction effect of the "weighted moving average + RFR" combination model was obviously superior to that of the "wavelet denoising + RFR" combination model. The results provide useful references for correcting the PM_2.5 monitoring data of light scattering sensor micro-stations.

• Key words：PM_2.5  data correction  light scattering  Random forest  denoising

• 
  摘要点击次数： 205 全文下载次数： 293