基于大语言模型的洪涝灾害统计调查

Statistical survey on flood disasters based on large language model

摘要

【目的】洪涝灾害统计调查是变化环境下区域防灾减灾和洪涝风险管理的重要依据。大语言模型在水利工程领域展现出强大应用潜力，为充分利用其语义理解和信息抽取能力，提出基于大语言模型的洪涝灾害统计调查方法，全过程实现大语言模型自主的互联网数据统计调查。【方法】面向北京“7·21”“23·7”和珠江流域“22·6”、南方地区“24·4”暴雨洪涝灾害，采用大语言模型统计调查危险程度和损失情况。【结果】结果表明，统计调查准确率随大语言模型温度参数升高而降低，当温度参数为0时准确率最高。北京“7·21”暴雨洪涝灾害的房屋倒塌数量、受灾人口和直接经济损失准确率超过90%,农作物受灾面积和洪峰流量的空检率超过40%;“23·7”暴雨洪涝灾害准确率更高，房屋倒塌数量、死亡失踪人口、受灾人口、农作物受灾面积和洪峰流量的准确率均超过90%。珠江流域“22·6”暴雨洪涝灾害的石角站洪峰流量和飞来峡最大入库流量准确率超过90%,北江流域最大小时降雨量和平均降雨量的准确率分别为83%和61%;南方地区“24·4”暴雨洪涝灾害石角站洪峰流量的准确率为89%,飞来峡最大入库流量、北江流域最大小时降雨量和平均降雨量的空检率超过70%。【结论】大语言模型适用于洪涝灾害数据统计调查，可以为水旱灾害防御工作提供数据支撑。

Abstract

[Objective] The statistical survey on flood disasters is an important basis for regional disaster prevention, reduction and flood risk management under changing environmental conditions. The large language model(LLM) has shown potential in the field of hydraulic engineering. In order to utilize the capabilities of LLMs on semantic understanding and information extraction, the LLM-based method of statistical survey on flood disasters is proposed. All procedures of the statistical survey based on Internet data are completed by the LLM. [Methods] For the flood disasters on 21 July 2012 in Beijing(“7·21”), July 2023 in Beijing(“23·7”), on June 2022 in Pearl River Basin(“22·6”) and April 2024 in Southern China(“24·4”), the proposed LLM-based method is used to investigate the risk and loss. [Results] The accuracy of the statistical survey is the highest when the temperature of the LLM is 0 and decreases as the temperature increases. For the “7·21” flood disaster, the accuracy of the number of collapsed dwellings, the affected population and the direct economic losses is more than 90%. The proportion of no retrieved result of affected cropland area and peak discharge is more than 40%. For the “23·7” flood disaster, the accuracy is generally higher. The accuracy of the number of collapsed dwellings, the number of deaths and missing persons, the affected population, the affected cropland area and the peak discharge is more than 90%. For the “24·6” flood disaster in the Beijiang river basin, the accuracy of the maximum hourly precipitation, the average precipitation and the peak discharge is 83%, 61% and more than 90%. For the “22·4” flood disaster, the accuracy of the peak discharge at Shijiao station is 89%. the proportion of no retrieved result of the peak discharge at Feilaixia reservoir, the maximum hourly precipitation and the average precipitation exceeds 70%. [Conclusion] The LLM is suitable for the statistical survey of flood disasters and can provide data support for the management of flood and drought disasters.

关键词

洪涝灾害 / 统计调查 / 大语言模型 / 互联网数据 / 信息抽取 / 气候变化 / 风险评估 / 降雨

Key words

flood disaster / statistical survey / large language model / Internet data / information extraction / climate change / risk assessment / rainfall

基金资助

国家重点研发计划项目(2023YFF0804900)

国家自然科学基金项目(52379033)