3种机器学习方法对空气污染预报效果的对比

Comparison of three machine learning methods for air pollution predicti

以中国西北5个省会城市为研究区域,利用2015年1月1日-2020年7月21日空气质量监测资料与气象数据构建基于多元逐步回归(MSR)、随机森林(RF)与支持向量机(SVM)的空气污染预报模型,以平均绝对误差、均方根误差及拟合度指数I_(A)为评判指标,对模型的模拟精度进行对比分析.结果表明,针对5个城市构建空气污染预报模型时,RF重要性评估法比Spearman相关系数法更适用于筛选预报因子.MSR、RF与SVM模型的预报性能由强至弱依次为MSR>RF>SVM.3种模型预报结果均I_(A)>0.8,预报值与实际观测值之间的相关程度较高.

Five provincial capital cities in Northwest China were used as the study area,and the air quality monitoring and meteorological data from 2015-01-01 to 2020-07-21 used to construct a multiple stepwise regression(MSR),random forest(RF)and support vector machine(SVM)three air pollution forecasting models,with the mean absolute error,root mean square error and fit index(I_(A))as the evaluation indicators,so as to compare and analyze the model simulation accuracy.The results showed that,in the construction of air pollution forecasting models for the five cities,the RF importance assessment method was more suitable for screening forecasting factors than the Spearman correlation coefficient method.The order from strong to weak was MSR>RF>SVM,the forecast results of the three models all I_(A)>0.8,and the correlation between the forecast value and the actual observation value was high.