基于改进的三层机器学习搜索法极端降雨区域频率估计  

An improved three-layer machine learning searching algorithm for regional frequency analysis of extreme rainfall events

作　　者：杨哲 杨昆 吕娟[2,3] 左惠强[4] 尹建明 YANG Zhe;YANG Kun;LYU Juan;ZUO Huiqiang;YIN Jianming(Post-doc workstation,China Reinsurance(Group)Co.,Ltd.,Beijing 100033;China Institute of Water resources and Hydropower Research,Beijing 100038;Research Center on Flood and Drought Disaster Reduction of the Ministry of Water Resources,Beijing 100038;China Property&Casualty Reinsurance Company LTD,Beijing 100033;China Re CRM,Chongqing 400020)

机构地区：[1]中国再保险(集团)股份有限公司博士后工作站,北京100033 [2]中国水利水电科学研究院,北京100038 [3]水利部防洪抗旱减灾工程技术研究中心,北京100038 [4]中国财产再保险有限责任公司,北京100033 [5]中再巨灾风险管理股份有限公司,重庆400020

出　　处：《中国防汛抗旱》2023年第1期44-51,共8页China Flood & Drought Management

摘　　要：根据极端降雨形成的物理机制,结合机器学习技术如等距特征映射和影响区域,在基于三层搜索法的区域频率分析法基础上提出改进算法,降低了降雨强度-历时-频率(IDF)模型分位点值的不确定性。根据目标站点的气候和地形特征,现有的三层搜索方法运用特征选择法,选取最具代表性的地理或气象要素作为组建同质化群体的相似因子,进而得到具有较低不确定性的降雨强度分位点数值。考虑特征要素之间的非线性相关性,在现有的三层搜索法基础上加入特征抽取法和监督聚类法,减少同质群体聚类受特征要素间非线性相关性的影响,以及其对输入站点数据的依赖性。该方法在加拿大不列颠哥伦比亚省的降雨站点进行了试验,试验结果表明,该方法能进一步降低IDF估计结果的不确定性。According to the physical mechanism of extreme rainfall formation,combined with machine learning techniques such as isometric feature mapping and influence region,an improved algorithm is proposed based on the regional frequency analysis method of three-layer searching method,which reduces the uncertainty of the quantile value of the rainfall intensity-duration-frequency(IDF)model.According to the climate and terrain characteristics of the target site,the existing three-layer searching method uses the feature selection method to select the most representative geographical or meteorological elements as the similarity factors to form a homogeneous population,and then obtain the rainfall intensity quantile value with low uncertainty.Considering the nonlinear correlation between feature elements,feature extraction method and supervised clustering method are added on the basis of the existing three-layer searching method to reduce the influence of the nonlinear correlation between feature elements on homogeneous group clustering and its dependence on the input site data.The method was tested at rainfall stations in British Columbia,Canada,and the test results show that the method can further reduce the uncertainty of IDF estimation results.

关 键 词：区域频率估计 极端降雨事件 数据降维 同质群体 空间差异性 

分 类 号：P426[天文地球—大气科学及气象学]