兰州市CMAQ近地面臭氧模拟结果的订正方法——基于机器学习方法  

A revised approach to CMAQ near-surface ozone modelling results in Lanzhou-Based on machine learning methods

作　　者：周恒左 陈恒蕤 廖鹏 孔祥如 潘峰[1] 杨宏[1] ZHOU Heng-zuo;CHEN Heng-rui;LIAO Peng;KONG Xiang-rui;PAN Feng;YANG Hong(College of Atmospheric Sciences,Lanzhou University,Lanzhou 730000,China)

机构地区：[1]兰州大学大气科学学院,甘肃兰州730000

出　　处：《中国环境科学》2022年第12期5472-5483,共12页China Environmental Science

基　　金：兰州大学中央高校基本科研业务费专项资金自由探索项目(lzujbky-2017-65)。

摘　　要：为能更加准确地模拟出兰州市近地面臭氧浓度,在CMAQ(社区多尺度空气质量建模系统)的基础上,利用机器学习方法中的XGBoost(极限梯度提升)模型及LSTM(长短期记忆)神经网络模型建立近地面臭氧模拟结果的订正模型,并以两种方法为基础,利用误差变权倒数组合方法构建LSTM-XGBoost组合模型,以期进一步提高订正效果.本文选取兰州市4个国控站点(兰炼宾馆,铁路设计院,榆中校区,生物制品所)2019年7、8月环境空气质量监测数据及兰州市气象站同期气象数据,对CMAQ模拟的同时段兰州市近地面臭氧浓度进行订正.结果表明,CMAQ能够模拟出兰州市近地面臭氧浓度的空间及时间分布特征,但整体上对浓度有所低估.利用上述方法构建的订正模型中,LSTM-XGBoost组合模型的订正效果最好,臭氧相关性由CMAQ模拟的0.61~0.76提升至0.89~0.95,臭氧8h平均相关性由0.65~0.79提升至0.81~0.88,臭氧RMSE由44.83~70.17μg/m^(3)提升至15.21~26.53μg/m^(3),臭氧8h平均RMSE由40.07~67.57μg/m^(3)提升至14.24~28.54μg/m^(3).该研究表明利用机器学习方法对CMAQ模拟结果订正可行,可以改善环境空气质量模式模拟结果.To better simulate the near-surface ozone concentration in Lanzhou,the XGBoost(eXtreme Gradient Boosting)model and the LSTM(Long and Short-Term Memory)neural network model in the machine learning method were used on the basis of CMAQ(The Community Multiscale Air Quality Modeling System)to establish a revised model of near-surface ozone simulation results,and the combined LSTM-XGBoost model was constructed based on the two methods with the combination of the inverse of error variance weights to further improve the revision effect.In this study,four national monitoring sites in Lanzhou(Lan Lian Hotel,Railway Design Institute,Yuzhong Campus,and Biological Products Institute)were selected,and ambient air quality monitoring data and meteorological data in July and August 2019 were used to revise the near-surface ozone concentrations simulated by CMAQ.Results showed that the CMAQ model could simulate the spatial and temporal distributions of the near-surface ozone concentrations in Lanzhou,but the concentrations was underestimated.Among the revised models mentioned above,the XGBoost combined model revised best.Compared with the simulation results of CMAQ,the correlation of ozone concentration improved from 0.61~0.76 to 0.89~0.95,the correlation of 8h ozone concentration improved from 0.65~0.79 to 0.81~0.88,the ozone RMSE improved from 44.83~70.17μg/m^(3) to 15.21~26.53μg/m^(3),8h ozone RMSE improved from 40.07~67.57μg/m^(3) to 14.24~28.54μg/m^(3).This study indicated that it is feasible to revise the model simulation results using machine learning methods to improve the air quality model.

关 键 词：CMAQ 近地面臭氧 机器学习 LSTM XGBoost 误差变权倒数组合 

分 类 号：X511[环境科学与工程—环境工程]