合肥市冬季PM2.5统计预报方法初试与比较研究

A Preliminary and Comparative Study of PM2.5 Statistical Forecasting in Hefei City During Winter

利用合肥市2015-2018年冬季PM2.5观测资料和FNL再分析资料,文章综合考虑地面及边界层高度范围内各气象要素作用,针对目前空气质量统计预报方法的不足,根据阈值分析筛选预报因子,同时将风向数据转化为对应的八方位上历史污染物浓度均值输入,最后结合BP神经网络对PM2.5浓度进行逐6 h预报。结果表明,所建模型(TA-BP方案)中对PM2.5预测值与观测值的相关系数(R)高达0.85,平均绝对误差(MAE)为21.31μg/m^3,均方根误差(RMSE)为28.20μg/m^3。阈值分析能够有效筛选与污染物浓度呈非线性关系的气象预报因子和高空预报因子。较BP模型,TA-BP模型的R和一致性指数(IA)分别提升14.12%和8.33%,MAE、平均相对误差(MAPE)和RMSE分别降低22.87%、17.86%和23.78%。同时,与其他不同输入变量模型及线性模型对比结果表明:仅考虑气象因子作用的MTA-BP方案限制了预报模型的准确性,以临近6 h的PM2.5浓度代替各气象因子作用的PTA-BP方案能够实现较好的预报效果,但滞后性严重。另外,综合考虑气象因子与污染因子作用的非线性TA-BP模型要优于线性MSR模型。

Considering the limits of the current air quality statistical forecasting methods,based on PM2.5 observation data and FNL reanalysis data of Hefei in winter from 2015 to 2018,by taking into full account of the effects of meteorological elements in the height range of the ground and boundary layers,the author screens the forecasting factors according to the threshold analysis and inputs the mean of the concentrations of previous pollutants which is a result of the wind dispersing in eight directions.Finally,with BP neural network,the PM2.5 concentration is predicted by 6 h.The results show that the correlation coefficient R between the predicted value and the observed value of PM2.5 in TA-BP model is as high as 0.85,the MAE(mean average error)is 21.31μg/m^3,and the RMSE(root mean square error)is 28.20μg/m^3.Threshold analysis can effectively screen meteorological forecasting factors and high-altitude forecasting factors that are nonlinear in relation to concentrations of pollutants.Compared with the BP model,the coefficient R and the consistency index IA in the TA-BP model increased by 14.12%and 8.33%,respectively,and the MAE,MAPE and RMSE decreased by 22.87%,17.86%and 23.78%,respectively.At the same time,compared with other different input variable models and linear models,the MTA-BP model considering only the action of meteorological factors limits the accuracy of the prediction model.The PTA-BP model in which PM2.5 concentration near 6 h replaces the action of various meteorological factors can achieve better prediction results.However,the hysteresis is serious in practical applications.In addition,the nonlinear TA-BP model that takes into account of the effects of meteorological factors and pollution factors is superior to the linear MSR model.