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

利用1997—2015年吉林省春夏期(4—7月)逐日气象站地面观测资料,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;分析吉林省春夏期的主要天气模态,研究逐日降水和天气模态之间的关系,并基于此关系提出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,邻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4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。平均绝对误差(Mean Absolute Error,MAE)和均方根误差(Root Mean Square Error,RM SE)的中位数分别为3. 12 mm和6. 13 mm,SBrier和Ssig分别为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。

A statistical simulation study on spring-summer precipitation in Jilin Province using self-organizing maps

Based on the daily ground observations in meteorological stations of Jilin Province during April-July 1997-2015,taking temperature,air pressure,relative humidity,water vapor pressure and wind speed as covariates,this paper established a statistical prediction model of daily precipitation based on self-organizing maps(SOM).This paper studied major synoptic patterns in Jilin Province and the relationship between daily precipitation and the patterns,and based on this relationship,proposed a Monte Carlo simulation method for daily precipitation.Results demonstrate that SOM has high classification quality of synoptic patterns,and the accumulative probability distributions of adjacent synoptic patterns are similar,while those of synoptic patterns far away are quite different.The correlation coefficient between the probability of no precipitation and the corresponding width of daily precipitation interval in the synoptic patterns is -0.94,and the significance level is less than 0.01.According to the accumulative probability distribution of precipitation,20 types of synoptic patterns are divided into four categories,which match the occurrence rate of precipitation and the daily precipitation.On this basis,this paper carried out Monte Carlo simulation of daily precipitation in 24 stations of Jilin Province,and analyzed the forecast performance.The median values of MAE(mean absolute error),RMSE(root mean square error),S_Brier and S_sig are 3.12 mm,6.13 mm,0.06 and 0.51,respectively,which indicates that the method has a good forecast performance in general.The distribution of MAE and RMSE is large in the southeast and small in the northwest,and all stations have smaller errors after removing the effect of the natural fluctuation of precipitation.S_Brier and S_sig have no obvious spatial distribution characteristics.