中国农业气象 ›› 2026, Vol. 47 ›› Issue (3): 443-455.doi: 10.3969/j.issn.1000-6362.2026.03.011

• 农业气象保险栏目 • 上一篇    下一篇

基于机器学习确定湖南省降水保险理赔阈值

周威,廖春花,王瑶,官建文,李好   

  1. 1. 湖南省气象服务中心/气象防灾减灾湖南省重点实验室,长沙 410118;2. 中国气象局秦岭和黄土高原生态环境重点开放实验室,西安 710000;3. 中国人民财产保险有限公司湖南省分公司,长沙 410119
  • 收稿日期:2025-06-19 出版日期:2026-03-20 发布日期:2026-03-17
  • 作者简介:周威,E-mail:z-will@163.com
  • 基金资助:
    湖南省自然科学基金项目(2025JJ80285);中国气象局秦岭和黄土高原生态环境气象重点开放实验室2024年基金课题(2024G−37)

Claim Threshold of Precipitation Insurance in Hunan Province Based on Machine Learning

ZHOU Wei, LIAO Chun-hua, WANG Yao, GUAN Jian-wen, LI Hao   

  1. 1. Hunan Provincial Meteorological Service Center/Hunan Key Laboratory of Meteorological Disaster Prevention and Reduction, Changsha 410118, China; 2. China Meteorological Administration Eco-Environment and Meteorology for the Qinling Mountains and Loess Plateau Key Laboratory, Xi’an 710000; 3. People´s Insurance Corporation of China Hunan Branch, Changsha 410119
  • Received:2025-06-19 Online:2026-03-20 Published:2026-03-17

摘要:

为厘清降水事件引发保险理赔的关系,提升保险气象服务能力,基于湖南省20212024703次保险理赔数据和逐小时降水量资料,利用随机森林等3个模型构建湖南省县(市、区)级降水保险理赔阈值,并进行显著性检验。结果表明:(1)随机森林模型的精确率、临界成功指数、准确率等均高于XGBoostLightGBM模型,保险理赔阈值模拟效果更优,可较好评估不同位置、不同时间窗口的降水保险理赔阈值;(2)保险理赔前120h96h72h累计降水量和120h最大降水量的特征重要性评分较高,是影响理赔的关键因素;(3)降水保险理赔阈值的空间异质性显著,“降水越大越易理赔”的结论在空间上不成立,低阈值区域(120h累计降水50mm)占理赔区域比重的40%,对降水事件敏感,易致灾,高阈值区域(96h累计降水>65mm)占比16%表明强降水事件导致保险理赔;(4)卡方检验表明基于随机森林构建的6个地级市、19个县(市、区)级的降水保险理赔阈值具有显著的统计关系,能有效区分不同降水量级的理赔触发条件。

关键词: 机器学习, 降水量, 保险, 阈值, 湖南

Abstract:

In order to clarify the relationship between precipitation events and insurance claims and enhance meteorological insurance services, this study utilized 703 insurance claim records and hourly precipitation data from Hunan province (20212024). A random forest model was used to construct thresholds for precipitation insurance claims in countylevel urban areas in Hunan province, followed by a significance test. The results showed that: 1the random forest model outperformed XGBoost and LightGBM in precision, Critical success index (CSI) and accuracy, demonstrating superior simulation of insurance claim thresholds. It effectively evaluated the thresholds across different locations and time windows. 2Cumulative precipitation 120h, 96h and 72h prior to claims, along with the maximum hourly precipitation whthin 120h, showed high feature importance scores, identifying them as key claiminfluencing factors. 3Significant spatial heterogeneity existed in precipitation claim thresholds. The conclusion that 'higher precipitation intensity let to a higher likelihood of claims' did not hold spatially. Lowthreshold regions (cumulative 120h precipitation <50mm) accounted for 40% of claimed areas, indicating high sensitivity to precipitation events and vulnerability to disasters. Highthreshold regions (cumulative 96h precipitation >65mm) constituted 16%, suggesting claims occur only under intense precipitation. 4The Chisquare tests confirmed statistically significant relationships between the precipitation insurance claim thresholds of 6 prefecturelevel urban areas and 19 countylevel cities (districts) level urban areas constructed based on random forests, effective discrimination of claim triggers across precipitation intensities. The research results reveal an inherent correlation between the spatiotemporal distribution of precipitation and insurance claims, which can provide a scientific basis for insurance meteorological warnings and disaster prevention and reduction in Hunan province. 

Key words: Machine learning, Precipitation, Insurance, Threshold, Hunan province