基于自组织映射神经网络的吉林省春夏期降水统计模拟研究
投稿时间:2018-05-07  修订日期:2018-06-04  点此下载全文
引用本文:
摘要点击次数: 501
全文下载次数: 0
作者单位E-mail
吴香华 南京信息工程大学 wuxianghua@nuist.edu.cn 
蒙芳秀 南京信息工程大学 mfxnuist@163.com 
于华英 南京信息工程大学  
燕妮 南京信息工程大学  
刘伟奇 南京信息工程大学  
基金项目:国家自然科学基金项目
中文摘要:利用1997—2015年吉林省春夏期(4—7月)气象站地面观测日值资料,以气温、气压、相对湿度、水汽压、风速为协变量,建立各站点逐日降水量的基于自组织映射神经网络(Self-Organizing Maps,SOM)的统计预测模型。分析了吉林省春夏期的主要天气模态,研究了逐日降水和天气模态之间的关系,并基于此关系建立了逐日降水量的蒙特卡罗模拟方法。结果表明:SOM对天气模态的分型质量较好,临近天气模态的累积概率分布较相似,距离较远的天气模态累计概率分布差异较大。各天气模态下无降水的概率与日降水量区间宽度的相关系数为-0.94,显著性水平小于0.01。基于降水量累积概率分布,20种天气模态被划分成四类,并与降水易发程度和逐日降水量完全对应。在此基础上,对吉林省24个站点逐日降水量进行蒙特卡罗模拟,并进行预测性能分析。MAE和RMSE的中位数分别为3.12 mm和6.13 mm,SBrier和Ssig为0.06和0.51,站点的逐日降水量预测性能整体较好。MAE和RMSE分布呈现东南大西北小,去除降水自然变异差异的影响,所有站点的误差都较小;SBrier和Ssig没有明显的空间分布特征。
中文关键词:春夏期降水  自组织映射神经网络  天气模态  蒙特卡罗模拟
 
A statistical simulation study on daily precipitation during the spring-summer season in Jilin province using Self-Organizing Maps
Abstract:Based on the daily ground observations in meteorological stations of Jilin Province during April-July 1997-2015, with the covariates (such as temperature, air pressure, relative humidity, vapour pressure, and wind speed), a statistical simulation model on daily precipitation has been established using Self-Organizing Maps (SOM). This paper showed analysis on major synoptic patterns in Jilin province, the relationship between daily precipitation and the patterns, and Monte Carlo simulation on daily precipitation. Results demonstrate that SOM has high classification quality of synoptic modes, and the cumulative probability distributions of adjacent modes are similar, while those in the distance are quite different. The correlation coefficient between the probability of no precipitation and the corresponding width of the daily precipitation interval in the mode is -0.94 at the significance level less than 0.01. According to the accumulative probability distributions of precipitation, 20 types of synoptic modes have been divided into four categories matching the occurrence rate and the quantity of precipitation. Then, the Monte Carlo simulations on daily precipitation in 24 stations have been carried out, and the performance analysis has been performed. The median values of MAE, RMSE,SBrier and Ssig were 3.12 mm, 6.13 mm, 0.06 and 0.51, respectively, which indicate a good forecast performance of the method in general. The spatial distributions of MAE and RMSE are large in southeast and small in northwest, and models at all sites have small errors after the effects of natural variation of precipitation have been considered. SBrier and Ssig have no obvious characteristics in spatial distributions.
keywords:spring-summer precipitation  Self-Organizing Maps  synoptic patterns  Monte Carlo simulation
  查看/发表评论  下载PDF阅读器