基于多要素3D特征提取的短期定量降水预报技术研究  

Short⁃term quantitative precipitation forecasting based on multi⁃factor 3D feature extraction

作　　者：熊文睿 张恒德 陆振宇[1] 郭云谦 XIONG Wenrui;ZHANG Hengde;LU Zhenyu;GUO Yunqian(School of Artificial Intelligence,Nanjing University of Information Science&Technology,Nanjing 210044,China;National Meteorological Center,Beijing 100081,China)

机构地区：[1]南京信息工程大学人工智能学院,南京210044 [2]国家气象中心,北京100081

出　　处：《南京信息工程大学学报》2025年第1期125-137,共13页Journal of Nanjing University of Information Science & Technology

基　　金：新疆维吾尔自治区重点研发任务专项(2022B03027);国家重点研发计划(2021YFC3000903);国家自然科学基金重点项目(U20B2061);中国气象局创新发展项目(CXFZ2023J001)。

摘　　要：由于空间分辨率有限、物理参数化方案不够完善、泛化性较弱等原因,使得传统业务数值天气模式(NWP)在定量降水预报中存在固有偏差,而深度学习神经网络具有强大的非线性拟合能力、能够自主性学习到任务相关的关键特征、泛化性较高等优势,有望改善现状.为此,本文提出一种基于多要素3D特征提取的短期定量降水预报技术.基于欧洲中期天气预报中心(ECMWF)提供的高分辨率ECMWF-HRES(EC-Hres)模式预报数据,构建3D-QPF(3D-Quantitative Precipitation Forecast)语义分割模型,通过先分类后回归的耦合框架,捕捉多种降水相关要素数据的3D空间特征,得到与降水实况数据间的非线性关系,并增加准确率和召回率损失函数,进一步提升模型对偏态数据的预报效果.实验结果表明,3D-QPF的逐日累积降水预报不仅在晴雨量级(0.1 mm/(24 h))准确率评分稳定增长,在暴雨量级(50 mm/(24 h))的准确率评分也有明显提升,暴雨量级较EC-Hres的TS评分最高提升了15.8%,RMSE优化达到18.71%.经过长期检验,3D-QPF模型与EC-Hres、中国气象局全球模式(CMA-GFS)预报以及2D-Unet和3D-Unet等经典网络模型相比做出了有效的预报订正效果.此外,随着预报时效延长至72 h,模型的优化效果仍能够保持相对稳定.Traditional Numerical Weather Prediction(NWP)models suffer from inherent biases in Quantitative Precipitation Forecasting(QPF)tasks due to limited spatial resolution,incomplete physical parameterization schemes,and poor generalization capabilities.Deep learning neural networks,with their robust nonlinear fitting capabilities,autonomous learning of task⁃specific features,and high generalization,hold the potential to address these issues and improve the current state of NWP.Here,we propose a novel short⁃term QPF approach based on multi⁃factor 3D feature extraction.Leveraging high⁃resolution ECMWF HRES(EC⁃Hres)model forecasting data provided by the European Centre for Medium Range Weather Forecasts(ECMWF),we construct a 3D⁃QPF semantic segmentation model.This model employs a coupled framework of classification and regression to capture the 3D spatial features of various precipitation⁃related element data,establishing a nonlinear relationship with actual precipitation data.Furthermore,we incorporate the PR(Precision⁃Recall)loss function to further enhance the model's predictive performance,particularly for skewed data.Experimental results show that the 3D⁃QPF model achieves a steady increase in accuracy score for daily cumulative precipitation forecast,not only at the light rain threshold(0.1 mm/(24 h))but also significantly at the rainstorm threshold(50 mm/(24 h)),with a maximum improvement of 15.8%in TS(Threat Score)compared to that of EC⁃Hres and a reduction in RMSE(Root Mean Square Error)by 18.71%.Upon extended validation,the 3D⁃QPF model outperforms EC⁃Hres,China Meteorological Administration Global Model(CMA⁃GFS)forecasting,and classic network models such as 2D⁃Unet and 3D⁃Unet,demonstrating effective prediction correction.Notably,the model's optimization performance remains relatively stable even when the forecast lead time is extended to 72 hours.

关 键 词：定量降水预报 语义分割 偏态数据 耦合方式 

分 类 号：P457[天文地球—大气科学及气象学]