海口市臭氧浓度统计预报模型的构建与效果评估

Establishment and Effective Evaluation of Haikou Ozone Concentration Statistical Prediction Model

主要利用2015~2020年海口市臭氧(O_(3))浓度资料和ERA5再分析资料,基于污染物浓度控制方程挑选出海口市O_(3)-8h浓度(日最大8 h滑动平均)的15个关键预报因子,构建了多元线性回归模型(MLR)、支持向量机模型(SVM)和BP神经网络模型(BPNN),并对2021年海口市O_(3)-8h浓度进行预测和效果检验.结果表明,O_(3)-8h浓度与关键预报因子的相关系数绝对值主要分布在0.2~0.507之间,其中1000 hPa的相对湿度(RH1000)和风向(WD1000),875 hPa的经向风(v875)的相关系数绝对值超过了0.4,具有较好的指示作用.3个预报模型基本能预报出海口市O_(3)-8h浓度冬半年偏高,夏半年偏低的变化趋势,其中BPNN模型的标准误差(RMSE)数值最小(22.29μg·m^(−3)).实测值与3个统计模型预报值的相关系数从大到小排列为:0.733(BPNN)>0.724(SVM)>0.591(MLR),均通过了99.9%的信度检验.O_(3)-8h浓度等级预报的结果检验表明,3个预报模型的TS评分均随着O_(3)-8h浓度等级的上升而下降,而漏报率(PO)和空报率(NH)随着O_(3)-8h浓度等级的上升而上升.SVM和BPNN模型在3个等级预报中TS评分均略高于MLR模型,特别是在轻度污染等级,TS评分还维持在70%以上,具有较好的预报性能.

This study selected 15 key predictors of the maximum of 8-hour averaged ozone(O_(3))concentration(O_(3)-8h),using the O_(3) concentration of Haikou and ERA5 reanalysis data from 2015 to 2020,and constructed a multiple linear regression(MLR)model,support vector machine(SVM)model,and BP neural network(BPNN)model,to predict and test the O_(3)-8h concentration of Haikou in 2021.The results showed that the absolute value of correlation coefficients between the O_(3)-8h and related key prediction factors was mainly among 0.2 and 0.507.The 1000 hPa relative humidity(RH1000),wind direction(WD1000),and 875 hPa meridional wind(v875)showed a good indicative effect on the O_(3)-8h,with the absolute correlation value exceeding 0.4.The three prediction models could predict the seasonal variation in the O_(3)-8h in Haikou,which was larger in the winter half year and smaller in the summer half year.The root mean square error(RMSE)was the smallest(22.29μg·m^(−3))in the BPNN model.The correlation coefficients between the predicted values of three statistical models and observations were ranked as 0.733(BPNN)>0.724(SVM)>0.591(MLR),all passing the 99.9%significance test.For the prediction of the O_(3)-8h level,we found that TS scores of these three prediction models decreased with the increase in O_(3)-8h concentration level.Relatively,the point over rate and not hit rate increased with the rise in O_(3)-8h concentration level.TS scores of the SVM and BPNN model were relatively larger than those of MLR,especially in the light pollution level with TS scores remaining above 70%,indicating a better prediction capability.