京津冀地区作为我国工业发展重点区域，大气污染物的排放量尤其显著，以细颗粒物（PM_2.5）为代表的大气复合污染物成为该地区大气污染防治工作重点。在大气污染形势如此严峻的情况下，需要合理有效的空气质量预报来帮助政府部门做出相应决策，引导公众规避污染峰值。目前空气质量预报主要采用空气质量数值模式预报、统计预报方法和人工经验预报方法。随着科学技术的发展，多模式集成预报不断完善以减少由空气质量模式的不确定性带来的偏差对预报准确率的影响。本文深入对比评估中国环境监测总站原有业务预报系统和最新的全耦合多模式系统成员模式的模拟性能，在此基础之上提出空气质量模式与官方预报经验混合方法，结合模式与基于人工经验的官方预报优势从而提高预报准确率；与此同时，本文采用敏感性试验探究多种集成方法的适用性，基于敏感性试验结果，采用合适的集成方法优化全耦合多模式系统不同成员模式对京津冀13个城市的PM_2.5的预报效果并以此发展多模式集合预报和优化方法，取得主要结论有：

（1）全耦合空气质量模式系统成员模式模拟效果改善效果明显，不同模式成员在不同时效上的预报效果均优于中国环境监测总站原有业务预报系统，预报与实测的相关性提升，预报偏差减小，备受关注的污染预报准确率提高，预报时效24h内，污染预报准确率由原有业务预报系统的72%～77%提升至全耦合模式系统的81%～83%，即污染预报准确率超80%。

（2）全耦合模式系统中CMAQ模式在秦皇岛、北京、廊坊、沧州、承德和衡水预报效果较好，相关系数R稳定在0.6以上，合理性比率FAC2达80%以上，污染预报准确率在80%以上。所选取的模式成员NAQPMS、CMAQ和CAMx模式均在秦皇岛表现最佳，邢台表现最差，说明了排放优化在模式系统模拟性能中的主导作用。

（3）采用单向嵌套的CMAQ模式预报效果与模式系统区域设置和边界到区域内的输送处理有关，CMAQ_BJ09对北京预报整体偏高，CMAQ_BJ01区域预报整体偏低， CMAQ_BJ03预报效果最好，在6个指标表现上显著优于CMAQ_BJ01和CMAQ_BJ09，其中24小时内相关系数达0.78，平均偏差为-0.66μg/m³，合理性比率FAC2达89%，达到本文所定义的“完美”预报水平，即最佳预报效果。

（4）基于最佳模式结果CMAQ_BJ03和官方预报的优势进行经验混合预报，结果表明经验混合预报在采用的评估指标上模拟效果进一步提升，有效提高了预报精度：一致性指标IOA由两者间的最佳0.87进一步提升到了0.92，均方根误差由两者最佳的35.02 μg/m³进一步降低至29.20μg/m³，归一化平均误差NME从90%下降到37%，合理性比率FAC2达到89%。 

（5）多元线性回归方法受噪声数据影响较小，是鲁棒性较强的集成方法；动态权重集成方法受噪声数据有一定的影响，而均值集成方法受噪声数据影响显著；采用小时数据的动态权重集成优于采用日均数据的预报效果，多元线性回归采用日均数据效果更佳；采用日均数据的多元线性回归集成和采用小时数据的动态权重集成均能有效提高空气质量多模式的污染预报准确率。

（6）采用动态权重集成方法对全耦合多模式系统测试版本成员的PM_2.5小时数据进行集成，结果表明京津冀13个城市的相关系数从原有业务预报系统的0.28～0.73提升至0.35～0.75。对京津冀13个城市的小时浓度污染预报准确率提升1.0%~4.4%。该集成方法缩小了各个城市间预报效果的差距，有效地提高了各个城市的预报效果。

（7）结合新一代全耦合多模式系统中的CMAQ和CAMx模式不同水平网格分辨率预报结果进行集成，引入考虑源反演排放源的WRF-Chem和NAQPMS-AMR模式结果，结果表明两者都引入的预报效果达到最佳，其相关系数达0.88，平均偏差仅2.06μg/m³，污染预报准确率提升至93%，有效提升污染命中率，降低假警报率。

As the Beijing-Tianjin-Hebei region is one of the key areas of industrial development in China, the emission of air pollutants is particularly significant. The air compound pollutants represented by fine particulate matter (PM_2.5) have become the focus of air pollution prevention and control. Under such a severe situation of air pollution, reasonable air quality forecast is needed to help government departments make corresponding decisions, guide the public to avoid the peak period of pollution. At present, air quality forecast is mainly based on air quality numerical model, statistical forecasting method and empirical forecasting method. With the development of science and technology, the multi-model ensemble forecast is constantly improved to reduce the influence of the deviation caused by the uncertainty of the model on the accuracy of the forecast. To better improve the air quality forecast accuracy, our study deeply evaluated performance of original multi-model forecast system applied in China's environmental monitoring station and the new generation fully coupled model system. Based on the evaluation results, a hybrid method combining the advantages of model and official forecast is proposed to improve accuracy of forecast; At the same time, sensitivity tests were set up to explore the applicability of various ensemble methods. Based on the sensitivity test results, appropriate ensemble method was used to optimize the PM_2.5 forecast performance of the new generation fully coupled model system test version for 13 cities in Beijing-Tianjin-Hebei region. The main conclusions were as follows:

(1) The new generation fully coupled model system has improved forecast effect significantly. The performance of different model members on different forecast aging is better than that of the original ensemble forecast system. The correlation between prediction and observation has been improved, and the forecast deviation has been reduced. The accuracy of pollution forecasting has been improved. Within 24 hours, the accuracy of pollution forecasting has increased from 72%~77% to 81%~83%, which means that the accuracy of pollution forecasting exceeds 80%.

(2) In the fully coupled model system, CMAQ model performs well in Qinhuangdao, Beijing, Langfang, Cangzhou, Chengde and Hengshui with correlation coefficient above 0.6, FAC2 above 80%, and pollution prediction accuracy above 80%. The three models all perform best in Qinhuangdao and worst in Xingtai, indicating the dominant role of emission in the performance of the model system.

(3) The performance of the one-way nested CMAQ model is related to the model system setting and the transportation processing from the boundary to the area. The BJ09 domain forecast is generally higher than the observation while the BJ01 domain forecast is generally lower than the observation. The CMAQ_BJ03 performs best which is significantly better than CMAQ_BJ01 and CMAQ_BJ09 in six metrics. The correlation coefficient of CMAQ_BJ03 is 0.78, the mean deviation is -0.66μg/m³, and the rationality ratio FAC2 is 89% which reaches the "perfect" forecast level defined in this paper.

(4) Based on the advantages of the best performance model result CMAQ_BJ03 and the official forecast, the empirical hybrid forecast method is carried out. The results show that the empirical hybrid forecast is the best in the evaluation metrics which are adopted, and the forecast accuracy is effectively improved: The index of agreement IOA increases from the best of 0.87 to 0.92, the root mean square error RMSE decreases from the best of 35.02 μg/m³ to 29.20μg/m³, the normalized mean error NME decreases from 90% to 37%, and the consistency ratio FAC2 reaches 89%.

(5) The multiple linear regression method is a robust ensemble method which is less affected by noise data. The dynamic weight ensemble method is affected by noise data to some extent, while the mean ensemble method is significantly affected by noise data. The dynamic weight ensemble with hourly data is better than that with daily data. The multiple linear regression ensemble with daily mean data performs better than of with hourly data.

(6) The dynamic weight ensemble method was used to ensemble the PM_2.5 hourly data of the new generation fully coupled model system. The results showe that the correlation coefficient of 13 cities in Beijing-Tianjin-Hebei increases from 0.28~0.73 to 0.35~0.75. The accuracy of pollution forecast in Beijing-Tianjin-Hebei cities increases by 1.0%~4.4%. This ensemble method narrows the difference of model performance among cities and improves the forecast accuracy of each city effectively.

(7) Combining WRF-Chem and NAQPMS-AMR with CMAQ and CAMX models in the new generation fully coupled multi-model system to ensemble. The results showed that with both two models performs the best, with a correlation coefficient of 0.88 and an average deviation of only 2.06μg/m³, and the pollution forecast accuracy was improved to 93%. Effectively improve the pollution hit rate and reduce the false alarm rate.