| 李强,赵铜铁钢,徐博,李昱,田雨.面向暴雨事件的气象大模型适用性评估[J].水利学报,2025,56(10):1280-1291 |
| 面向暴雨事件的气象大模型适用性评估 |
| Applicability evaluation of large weather models for heavy rainfall events |
| 投稿时间：2025-03-29 修订日期：2025-08-29 |
| DOI：10.13243/j.cnki.slxb.20250114 |
| 中文关键词: 极端暴雨 气象大模型 降水预报 人工智能 适用性评估 |
| 英文关键词: extreme precipitation global climate model precipitation forecast artificial intelligence forecast verification |
| 基金项目:国家自然科学基金项目（52379033） |
|
| 摘要点击次数: 54 |
| 全文下载次数: 39 |
| 中文摘要: |
| 暴雨洪涝灾害是我国最严重的自然灾害之一。气象大模型在预报准确性、时效性和计算速度上呈现巨大潜力，本文评估不同气象大模型对暴雨预报的适用性，以期为洪水预报提供更高精度、更长预见期和更易获取的气象预报数据。面向珠江流域北江2022年6月（“2022.6”）和2024年4月（“2024.4”）两场极端暴雨洪水，评估了GraphCast、FuXi和AIFS三种气象大模型在不同预见期的暴雨预报性能，并与欧洲中期天气预报中心的高分辨率气象预报产品（HRES）进行对比。结果表明：GraphCast、FuXi、AIFS和HRES的降水预报性能与降水强度呈负相关关系。相比HRES，GraphCast、FuXi和AIFS对降水落区形态和位置的预报更准确，但低估了强降水落区范围和降水量。对于“2022.6”暴雨，起报时间为1 d前时，GraphCast最大6 h降水量和最大累积降水量预报分别为20.1和271.7 mm，FuXi的分别为21.8和222.5 mm，AIFS的分别为24.8和303.0 mm，均低估了实况降水110.9和416.3 mm。对于“2024.4”暴雨，起报时间为3 d前时，GraphCast、FuXi、AIFS和HRES预报最大6 h降水发生时间的误差分别为24、-6、72和96 h。整体上，气象大模型在暴雨预报中展现出巨大潜力，可为洪水预报提供具有业务参考价值的气象预报数据，服务于防灾减灾和应急管理的实际需求。 |
| 英文摘要: |
| Heavy rainfall and floods are among the most severe natural disasters in China. As large weather models （LWMs）show great potential in terms of accuracy，timeliness and efficiency in weather forecasting，this paper evaluates the applicability of different LWMs for heavy rainfall forecasting to provide more accurate，longer lead-time and more accessible weather forecasting data for flood forecasting. Focusing on two extreme floods in the Beijiang River Basin of the Pearl River Basin in June 2022（"2022.6"）and April 2024（"2024.4"），the predictive performance of three LWMs，i.e.，the GraphCast，FuXi，and AIFS，in forecasting heavy rainfall under different lead times has been assessed and compared with the high-resolution forecasts（HRES） from the European Centre for Medium-range Weather Forecast（ECMWF）. The results show that the predictive performance of GraphCast，FuXi，AIFS and HRES is negatively correlated with the intensity of precipitation. Compared with HRES，the performance of GraphCast，FuXi，and AIFS is more accurate in forecasting the morphology and location of precipitation，but they underestimate the areas and amounts of the two heavy rainfall events. For the "2022.6" event，when the forecast initialization time is 1 day ahead，the forecasted maximum 6-hour and cumulative precipitation are respectively 20.1 mm and 271.7 mm for the GraphCast，21.8 mm and 222.5 mm for the FuXi and 24.8 mm and 303.0 mm for the AIFS，all of which underestimate the observed precipitation of 110.9 mm and 416.3 mm. As the lead time increases，the performance of GraphCast，FuXi，and AIFS deteriorates，but they remain more accurate than HRES in forecasting the spatial location and occurrence time of heavy rainfall events. For the "2024.4" event，when the forecast initialization time is 3 days ahead，the errors in forecasting the occurrence time of the maximum 6-hour precipitation are respectively 24，-6，72，and 96 hours for GraphCast，FuXi，AIFS，and HRES，respectively. These LWMs demonstrate great potential in forecasting heavy rainfall and can provide valuable weather forecasts for flood forecasting，serving the practical needs of disaster prevention，mitigation，and emergency management. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |