多模式集成方法在安徽地区PM2.5预报中的应用研究

分别采用算术平均、权重平均、多元线性回归和神经网络的集成方法，对3种空气质量模式在安徽地区2017年2月-2018年2月PM_2.5预报结果进行集成释用.结果表明：各模式和订正产品的预报值与实况值之间均能达到显著相关，相较于WRF-Chem，多元线性回归的均方根误差（RMSE）下降了21.7%，归一化平均偏差（NMB）下降了6%，且在16个地市中NMB均下降至-25%~25%之间；从不同时次的预报效果来看，在3个代表性城市（淮北、合肥和芜湖），多元线性回归均能大幅度降低RMSE和NMB，但从时间和空间效果来看，其对于提升预报值与实况值之间的相关性方面，略差于权重平均的集成方法；多元线性回归方法对于重污染天气PM_2.5预报评分（TS）最高，为0.46.该方法能较为有效地提升不同模式的预报效果，可为重污染天气预报预警提供参考.

Application study of multi-mode integration method in PM2.5 forecast in Anhui Province

The integration methods of arithmetic average, weighted average, multiple linear regression and neural network were used to integrate and interpret the PM_2.5 forecasts at 16 cities in Anhui province by three air quality models during the period from February 2017 to February 2018. The results show that the forecasted values of the three models and integrated products were highly correlated with observations. Compared with WRF-Chem, the RMSE of multiple regression decreased by 21.7%, NMB decreased by 6%, and NMB dropped to between -25% and 25% in all cities. Multiple regression significantly reduced RMSE and NMB in the three representative cities, Huaibei, Hefei and Wuhu; however, in terms of the spatiotemporal correlation, it was slightly worse than the weighted average for improving the correlation between the forecast value and the observed value. As for the forecast of PM_2.5 heavy pollution weather, the multiple regression method performed best with the highest Ts, indicating that this integrated method can improve the forecasting effect of different models effectively, and provide a reference for early warning of heavy pollution weather.