基于机器学习的2016-2020年南京市ρ(O_(3))预报

A prediction of Nanjing Cityρ(O_(3))from 2016 to 2020 based on machine learning

利用南京市2016年1月1日-2020年12月31日城市空气质量监测站点的ρ(NO_(2))、ρ(O_(3))、ρ(PM_(10))、ρ(CO)和ρ(PM2.5)逐小时观测资料,分析南京的O_(3)污染特征.使用随机森林重要度评估法筛选出对ρ(O_(3))影响较大的污染物因子、污染持续性因子和气象因子作为机器学习预报模型的输入变量。采用随机森林算法(RF)和支持向量机算法(SVM)建立南京市不同季节的ρ(O_(3))预报模型,对比分析两种模型的预报效果,结果表明,近5a南京市ρ(O_(3))的日变化整体呈单峰型,午后较高;ρ(O_(3))具有明显的季节变化特征,夏季较高;2016-2020年ρ(O_(3))_(8h)的年际变化总体呈下降趋势,且ρ(O_(3))的超标天数在减少,说明南京市政府实施的O_(3)治理措施有明显效果;RF和SVM算法均能较好地预测ρ(O_(3)),预测值与观测值较为接近;相较于RF模型,SVM四季模型的预测值与观测值的相关系数更接近1,均方根误差与平均绝对误差都较小,预报效果较优。

The hourly observation data onρ(NO_(3)),ρ(O_(3)),ρ(PM_(10)),ρ(CO)andρ(PM_(2.5))in Nanjing City quality monitoring points from 2016-01-01 to 2020-12-31 were used.The characteristics of O_(3),pollution in the city were analyzed.The random forest importance assessment method was used to screen out the pollutant factors,pollution persistence factors and meteorological factors that had a great impact onρ(O_(3))as the input variables of the machine learning prediction model.Random forest algorithm(RF)and sup-port vector machine algorithm(SVM)were used to establish theρ(O_(3))prediction model in different seasons in Nanjing City,and the prediction effects of the two models were compared and analyzed.The results showed that the diurnal variation ofρ(O_(3))in recent 5 years presented a single-peak pattern,and the afternoonρ(O_(3))was high.Theρ(O_(3))had an obvious seasonal variation,and theρ(O_(3))was higher in summer.From 2016 to 2020,the interannual variation of ρ(O_(3))_(8h) generally showed a downward trend,and the number of days of O_(3)exceeding the standard also decreased,indicating that the O_(3)control measures implemented by the Nanjing government had had obvious effects.Both RF and SVM algorithm could predictρ(O_(3))well,and the predicted values were close to the actual observed values.The correlation coefficient between the predicted and the actual observed values of the SvM four season models were closer to 1,the root mean square error and the mean absolute error smaller,and the prediction effect optimized.