摘要
利用2018—2019年期间10 min定量降水估计(Quantitative Precipitation Estimation,QPE)实况观测,构建基于U-Net的分钟级临近降水预报模型,实现了京津冀地区未来0~2 h逐10 min降水量滚动预报。以TS、BIAS、POD、SR、FAR作为评价指标,通过检验2020和2021年6—9月长序列以及分析2020年8月12日和2021年7月1日两次强降水个例,表明U-Net模型预报接近实况,局部伴随着一定程度的空报,相较光流法、持续性预报及CMA-MESO模式预报效果有明显提升。具体表现为:当分钟级降水预报不超过10 mm/(10 min)时,U-Net模型明显优于光流法和持续性预报;当小时预报不超过25 mm·h^(-1),U-Net模型优于CMA-MESO模式和光流法。然而,当降水强度超过10 mm/(10 min)或25 mm·h^(-1)时,U-Net模型存在预报偏弱的情况,可能与强降水样本较少有关。
In order to strengthen the application of deep learning in precipitation nowcasting over the Beijing-Tianjin-Hebei region,the 10-minute QPE observation during 2018-2019 was used to construct a minute-level precipitation nowcasting model based on U-Net,which realizes 10-minute rolling precipitation forecast in the future 0-2 hours.By verifying the long series from June to September in 2020 and 2021 and analyzing two cases of heavy precipitation on both August 12,2020 and July 1,2021 based on evaluation indicators such as TS,BIAS,POD,SR and FAR,results show that the prediction of U-Net model is close to observation accompanied by false alarms to some extent and its forecasting effect is significantly improved compared with the optical flow method,persistent forecast and CMA-MESO model.Specific performance as follows:when the minute-level precipitation forecast does not exceed 10 mm/(10 min),the U-Net model is significantly better than the optical flow method and persistence forecast;when the hourly forecast does not exceed 25 mm·h^(-1),the U-Net model is significantly better than the CMA-MESO model and the optical flow method.However,when the precipitation intensity exceeds 10 mm/(10 min)or 25 mm·h^(-1),U-Net has a weak forecast,which may be related to fewer samples of heavy precipitation.
作者
徐成鹏
曹勇
张恒德
刘海知
梅双丽
XU Chengpeng;CAO Yong;ZHANG Hengde;LIU Haizhi;MEI Shuangli(National Meteorological Center,Beijing 100081,China)
出处
《气象科学》
北大核心
2022年第6期781-792,共12页
Journal of the Meteorological Sciences
基金
国家重点研发计划项目(2021YFC3000903)
风云卫星应用先行计划(2022)FY-APP-2022.0113。
关键词
深度学习
定量降水
临近预报
deep learning
quantitative precipitation
nowcasting