基于变量筛选的机器学习对兰州市近地面臭氧模拟研究

Study on the Simulation of Near-Surface Ozone in Lanzhou City based on Machine Learning of Variable Screening

为快速准确模拟兰州市臭氧浓度,利用随机森林(RF)和极端梯度提升(XGB)2种机器学习模型,结合ERA5气象数据、MEIC清单数据和兰州市空气质量监测数据,对兰州市2020年近地面8 h滑动平均臭氧浓度ρ(O3_8h)进行模拟。通过SHAP方法评估变量的重要性,筛选出对模型贡献最大的变量,分别构建简化模型RF7和XGB7,并比较其与全变量模型(RF_A和XGB_A)的模拟效果和运行效率。结果表明,RF7和XGB7模型在4个监测站点的模拟效果与全变量模型接近,但计算效率显著提高,模拟时间减少87.97%,96.68%;不同站点对ρ(O3_8h)的影响因素存在差异,说明在简化模型训练数据时需因地制宜。研究表明使用SHAP简化变量的方法在提高模拟效率的同时,能够保持模型的准确性,为兰州市臭氧污染的快速预测提供途径。

In order to quickly and accurately simulate the ozone concentration in Lanzhou City,two machine learning models,Random Forest(RF)and Extreme Gradient Boosting(XGB),were used to simulate the near-surface 8-hour sliding mean ozone concentrationρ(O3_8h)in Lanzhou City in 2020 by combining the ERA5 meteorological data,the MEIC inventory data,and Lanzhou City air quality monitoring data.The importance of variables was assessed by the SHAP method,and the variables with the largest contribution to the model were screened out to construct the simplified models RF7 and XGB7,respectively,and compare their simulation effects and operation efficiency with the all-variable models(RF_A and XGB_A).The results show that the simulation effects of the RF7 and XGB7 models at the four monitoring stations are close to those of the full-variable model,but the computational efficiencies are significantly improved,and the simulation times are reduced by 87.97%and 96.68%,respectively;the differences in the influencing factors ofρ(O3_8h)at different stations indicate that the simplified models need to be customized to fit the needs of different sites when training the data.The study shows that the method of using SHAP to simplify the variables can improve the simulation efficiency while maintaining the accuracy of the model,which provides a way for the rapid prediction of ozone pollution in Lanzhou City.