不同集成方法对云南日最低气温的集成预报试验
Experiment on the Daily Minimum Temperature Forecast in Yunnan by Different Ensemble Methods

通过前期对ECMWF-thin地面2 m气温在云南不同地区日最低气温各时效预报产品的检验，结果显示数值模式对云南不同地区日最低气温的预报性能，随预报时效延长并非一致性下降，而是呈现波动性降低。本文在此基础上，采用算术平均集成、相关权重集成、误差订正集成、相关权重与误差订正综合集成、预报准确率评分权重集成、准确率评分与误差订正综合集成等方法分别构建6种集成预报模型，并开展对云南省不同地区站点的日最低气温的集成预报试验。结果表明，6种集成方法对云南省不同地区站点日最低气温的集成预报效果均较未经集成的原始预报有明显提升，但不同集成方法对同一地区日最低气温的改进效果不同，同一种集成方法在不同地区的集成预报效果也有所不同；6种集成模型中简单的算术平均集成预报效果相对较差，而既考虑前期相关性和技巧评分，同时又考虑前期预报误差变化的综合集成效果相对较好；值得注意的是随着预报时效的延长，各种集成方法对原始预报效果的改进更加显著，具有很好的参考意义。
Through the previous verification of the time-dependent prediction products of ECMWF-thin surface 2 m temperature at the daily minimum temperature in different regions of Yunnan province, the results show that the prediction performance of the numerical model for the daily minimum temperature in different regions of Yunnan does not decrease consistently with the prolongation of the prediction time-dependent, but shows a decrease in volatility. On this basis, the arithmetic mean integration, correlation weight integration, error correction integration, correlation weight and er-ror correction integration, forecast accuracy score weight integration, accuracy score and error correction integration are adopted to construct six ensemble forecast models respectively, and the ensemble forecast experiments of daily minimum air temperature at stations in different regions of Yunnan Province are carried out. The results show that the effect of the six ensemble methods on the ensemble forecast of daily minimum temperature at stations in different regions of Yunnan Province is significantly improved as compared with that of the original forecast without integration. However, the improvement effects of different ensemble methods on the daily minimum temperature in the same region are different, and the ensemble forecast effects of the same ensemble method in different regions are also different. The simple arithmetic mean ensemble prediction among the six ensemble models is relatively poor in effect, and the ensemble considering the early correlation and skill score as well as the variation of the early prediction error is relatively good in effect. It should be noted that with the prolongation of prediction time-effect, the improvement of original prediction effect by various ensemble methods is more significant, which has good reference significance.