新疆夏季降水和冬季气温预测方法及效果评估

本文利用4个国内外先进的气候模式（国家气候中心、ECMWF、NCEP和JMA）业务预测数据,采用2种多模式集合方法（等权平均和超级集合）、3种降尺度方法（BP-CCA、EOF迭代、高相关回归集成）和3种统计方法（CCA、最优气候值、高相关回归集成）以及降尺度集成和降尺度-统计方法集成，分析了目前季节模式、多模式集合、降尺度、统计方法、降尺度-统计集合等目前常用气候预测技术对新疆夏季降水和冬季气温的业务预测能力。 研究表明,以上技术方法对新疆夏季降水和冬季气温的预测预测能力有较大差别。目前先进的气候业务模式的预测技巧普遍很低,多模式超级集合和降尺度方法的技巧常高于单个模式,并且最佳的降尺度方法通常技巧高于最佳多模式集合方法。同时,统计方法和降尺度方法的预测技巧通常较为接近,而对二者进行超级集合可以具有相对很高的预测技巧。此外,现有常用气候预测技术方法对新疆夏季降水和冬季气温的趋势有一定的预测能力，但对气候异常的空间分布基本无预测能力。建议新疆气候预测技术围绕统计和降尺度方法集合发展。

The prediction of summer precipitation and winter temperature over Xinjiang

The operational prediction ability of summer precipitation and winter temperature over Xinjiang is explored with the common techniques including seasonal models, multi-model ensemble, statistical downscaling, statistical methods, ensemble of both downscaling and statistical methods. The operational seasonal models from National Climate Center, ECMWF, NCEP, JMA, two muliti-model ensemble methods (the average and super-ensemble), three downscaling methods (BP-CCA, EOF-ITE, HCRE) and three statistical methods (BP-CCA, OCN, HCRE) used in National Climate Center are employed in this paper. Our study shows that the above techniques and methods hold much different prediction ability on the summer precipitation and winter temperature over Xinjiang. The skill scores of the leading operational seasonal models are very low. Meanwhile,the super-ensemble of models and downscaling methods are often better than single model while the best downscaling method shows higher score than the best multi-model ensemble method. Besides, the skills of statistical methods are similar to the downscaling ones. The super-ensemble of both the downscaling and statistical methods often holds quite higher prediction skills. What is more, it is indicated that the common methods used in present operation are of certain ability on the prediction of the trend but of few ability on the spatial distribution of the climate anomalies over Xinjiang. It is advised that the technique on seasonal prediction over Xinjiang will be developed focusing on the ensemble of statistical and downscaling methods.