427 Application of Ensemble Forecast and Linear Regression Method in Improving PM10 Forecast over Beijing Areas

Monday, 11 January 2016
Si Huang, IBM Research, Beijing, China

Abstract: In this study, ensemble forecast combined with linear regression method is used to reduce the uncertainty in air quality models. Firstly, the PM10 forecast skill of the three models (NAQPMS、CAMx、CMAQ) in EMS-Beijing are evaluated over Beijing areas. In order to improve the forecast skill, the linear regression method (REG) is used to combine the forecast results of the three models and is compared with the ensemble mean method. The results show that. (1) For single model forecast, great difference exists among different models and no model performs much better than the other two models for all statistic indexes. Overall, CMAQ performs better in tendency prediction, while NAQPMS has smaller root mean square errors than the other two models. (2) Ensemble mean method presents poor performance in improving the PM10 forecast of the three models, because neither correlation coefficients nor root mean square errors of the ensemble mean forecast results are entirely better than those of the single model forecast. On the other hand, REG brings significant improvement of the PM10 forecast. When an appropriate training length (36 days) is chosen, the root mean square errors of PM10 forecast over 28 stations of Beijing is reduced by 32%~43% through using REG and the bias decreased considerably to 5.8ug/m3 . This result implies REG can bring much better forecast skill than single model and ensemble mean forecast. Furthermore, the REG also greatly improve the hit ratio of pollution episode forecast.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner