统计降尺度方法在青海省冬季最低温度预测中的应用  

Application of Statistical Downscaling in Minimum Temperature Prediction over Qinghai Province in Winter

作　　者：曾晓青[1] 王式功[1] 刘还珠[2] 尚可政[1] 

机构地区：[1]兰州大学大气科学学院/甘肃省干旱气候变化与减灾重点实验室,甘肃兰州730000 [2]中国气象局国家气象中心,北京100081

出　　处：《高原气象》2009年第6期1471-1477,共7页Plateau Meteorology

基　　金：国家科技支撑计划项目(2007BAC29B03;2008BAC40B04);国家重点基础研究发展计划项目(2006CB400501)共同资助

摘　　要：利用2003—2007年国家气象中心T213L31全球中期数值预报模式逐日输出产品与青海地区25个气象站的观测数据作为试验资料,利用相关系数和逐步回归进行因子选择,并以单隐层神经网络和多元回归作为降尺度方法进行对比研究,用2003 2006年间的11月1日～次年3月1日的资料作为训练样本,以数值预报产品和前一日观测的最低温度作为因子,建立青海省25个气候站的冬季最低温度的24,48,72 h预报模型,并且以2006年12月和2007年的1、2月作为24,48,72 h逐日最低温度预报试验时段。试验表明,对于青海地区来说,青海北部地区的预报命中率总体好于南部高原地区;在4种对比方案中,以选择数值预报资料结合前一日地面观测的最低温度作为主要因子的方法相对较优,随着预报时效的延长,24 h历史实况的作用逐渐减弱;对于所有台站来说,这4种方案各有优缺点,没有一种方案可以完全代替其他所有方案;在实际业务运行中,对不同的台站应采用不同的预报方案进行实际业务预报。Using the daily numerical weather prediction output products from National Meteorological Center global medium-term model（T213L31） from 2003 to 2007 and observation data of 25 meteorological stations in Qinghai region as test data. The grid output products include 14 clay-to-day weather elements, 13 levels and 5 forecast times for each elements The correlation coefficient and stepwise regression are used as methods of preprocessing for building model. Single hidden layer neural network is used as the statistical downscaling methods. The train of network employs the back-propagation that performs a gradient descent algorithm. Hidden layer function is Sigmoid function and output activation function is the line function. According to the six kinds of options, 24 h, 48 h and 72 h forecast models are established for the minimum temperature at 25 weather stations of Qinghai Province in winter. Using December of 2006 and from January to February of 2007 as 24 h, 48 h, 72 h are as daily minimum temperature forecast test period. The results showed that the forecasting hit ratio in the northern part of Qinghai is better than that in the southern Plateau. Among the four kinds of options, selecting the numerical prediction results of grids around the site as a major factor combined with the history observation data is relatively optimum method. With the extension of forecast time, the effect of 24 h-history observation as predictor grow gradually weaker. However, No one scheme can completely replace all the other schemes in six kinds of schemes for all stations. In actual operations, the different schemes for different stations are taken to forecast the real forecast.

关 键 词：青海省 统计降尺度 神经网络 最低温度 命中率 

分 类 号：P456[天文地球—大气科学及气象学]