基于U-Net的淮河流域降水预报统计后处理

Statistical postprocessing of precipitation forecasts in the Huai River basin using U-Net

摘要

为校正数值天气预报降水产品中的系统性偏差，采用一种融合U-Net深度学习架构与删失移位伽马分布参数化的统计后处理方法，提升淮河流域中短期降水预报的准确性、分辨力与可靠性。基于GEFSv12再预报数据与0.25°×0.25°高精度观测数据，构建融合多源辅助因子（位势高度、可降水量、DEM）并考虑降水场空间特征的U-Net模型，其输出层直接预测移位伽马分布的形状、尺度与位置参数。模型训练采用k折交叉验证策略，以连续排名概率得分（continuous ranked probability score，CRPS）为损失函数进行优化，在2000−2019年淮河流域4−10月降水数据上进行系统评估，并与传统贝叶斯联合概率模型（Bayesian joint probability，BJP）后处理方法进行对比。结果表明：U-Net后处理方法优于原始预报与BJP方法，E_RMS中值平均降低4.88%（BJP仅为0.30%）；相对作用特征（relative operating characteristic，ROC）曲线分析显示，U-Net模型对极端降水事件的分辨能力显著增强，ROC曲线下面积（area under curve，AUC）平均提升0.08～0.12，而BJP仅在弱降水下略有改进。研究通过移位伽马分布生成降水集合预报，为水文集合预报提供降水预报的不确定性信息，对提升流域防洪减灾决策支持能力具有重要应用价值。

Abstract

Numerical weather prediction models frequently exhibit systematic biases in precipitation forecasts, which can significantly reduce the accuracy and reliability of medium-range precipitation predictions. Addressing these limitations is critical for improving flood prevention and water resource management, especially in vulnerable areas such as China's Huai River basin. This study was carried out to create an advanced statistical postprocessing technique aimed at improving the accuracy, resolution, and reliability of short- to medium-term precipitation forecasts in this basin. The study concentrated on using advanced machine learning techniques in combination with probabilistic distribution modeling to reduce systematic errors present in raw numerical outputs. Given the basin's vulnerability to extreme weather events and its significance for agricultural and hydrological planning, improving precipitation forecasts was deemed critical for facilitating evidence-based decision-making in disaster risk reduction. To correct systematic biases in precipitation forecasts, a novel statistical postprocessing framework was developed. It combines a U-Net deep learning architecture with a censored shifted gamma distribution parameterization. Primary inputs included high-resolution reforecast data from the Global Ensemble Forecast System version 12 and observational data with a spatial resolution of 0.25°. The U-Net model was specifically designed to include a variety of auxiliary factors, such as geopotential height, precipitable water, and the digital elevation model, while explicitly accounting for the spatial characteristics of precipitation fields. Probabilistic forecasting was made possible by the model's output layer, which directly predicted the shifted gamma distribution's shape, scale, and location parameters. To minimize forecast errors, model training was optimized withk-fold cross-validation, with the continuous ranked probability score serving as the loss function. Precipitation data from 2000 to 2019 were used in a systematic evaluation, with an emphasis on rainy season data gathered from April to October throughout the Huai River basin. To evaluate relative performance improvements, a thorough comparative analysis was carried out against the conventional Bayesian joint probability method. The U-Net postprocessing method demonstrated superior performance compared to both the raw numerical forecasts and the Bayesian joint probability approach. Specifically, the median root mean square error was decreased by an average of 4.88% for the U-Net method, whereas the Bayesian joint probability method achieved only a minimal reduction of 0.30%. Relative operating characteristic curve analysis revealed a substantial enhancement in the ability to distinguish extreme precipitation events, with the area under the curve increasing by 0.08 to 0.12 on average for the U-Net model. In contrast, the Bayesian joint probability method improved only marginally for light precipitation scenarios and made no significant gains for extreme events. Furthermore, the shifted gamma distribution parameterization facilitated the generation of ensemble precipitation forecasts, which effectively quantified prediction uncertainty. This capability allowed for more reliable evaluations of potential flood risks by providing hydrological ensemble forecasting systems with crucial probabilistic information. The results consistently demonstrated the U-Net model's ability to capture complex spatial patterns and improve forecasting skill across a range of precipitation intensities. Findings revealed that combining U-Net deep learning with shifted gamma distribution parameterization significantly improves the accuracy of precipitation forecasts in the Huai River basin. In addition to lowering systematic errors, the suggested approach improved prediction resolution and dependability, especially for high-impact extreme events. By generating ensemble forecasts that convey essential uncertainty information, the approach offers practical value for operational hydrological modeling and flood forecasting systems. These advancements directly support improved decision-making in flood prevention and disaster reduction strategies, contributing to more resilient water resource management in the basin. The findings highlight the ability of hybrid machine learning and statistical techniques to transform numerical weather prediction outputs into actionable insights, which has broader implications for improving weather-related risk mitigation in similar regions around the world. Future work should explore the scalability of this framework to other basins and its integration with real-time forecasting operations.

关键词

深度学习 / U-Net / 伽马分布 / 统计后处理 / 降水预报 / 淮河流域 / 集合预报

Key words

deep learning / U-Net / Gamma distribution / statistical postprocessing / precipitation forecasting / Huai River basin / ensemble forecasting

参考文献

[1] Li Zehua, Wu Yanfeng, Zhang Guangxin, et al. China's black soil granary is approaching the climax phase of agricultural water security risk[J]. Agricultural Water Management, 2025, 319: 109780. DOI: 10.1016/j.agwat.2025.109780.

[2] 罗兴树, 章数语, 郭园, 等. 极端降水对陕西省农业生产的影响[J]. 南水北调与水利科技（中英文）, 2024, 22(2): 339-347. DOI: 10.13476/j.cnki.nsbdqk.2024.0035.

[3] Yin Wen, Yang Xiaolin, Liu Wenfeng. Sustainable management and regulation of agricultural water resources in the context of global climate change[J]. Sustainability, 2025, 17(6): 2760. DOI: 10.3390/su17062760.

[4] 韩忠青, 刘招, 肖瑜, 等. 未来气候情景下引汉济渭工程水源区水文情势演变规律研究[J]. 水利水电技术（中英文）, 2025, 56(1): 26-38. DOI: 10.13928/j.cnki.wrahe.2025.01.003.

[5] 陶伯钧, 桑国庆, 张永平, 等. 基于SWMM的平原区小城镇排涝体系评估与优化[J]. 南水北调与水利科技（中英文）, 2022, 20(6): 1197-1207. DOI: 10.13476/j.cnki.nsbdqk.2022.0118.

[6] 武慧敏, 吕爱锋, 张文翔. 巴音河流域水文干旱对气象干旱的响应[J]. 南水北调与水利科技（中英文）, 2022, 20(3): 459-467. DOI: 10.13476/j.cnki.nsbdqk.2022.0046.

[7] Salameh E, Abdallat G. The impacts of climate change on the availability of surface water resources in Jordan[J]. Journal of Geoscience and Environment Protection, 2020, 8(10): 52-72. DOI: 10.4236/gep.2020.810004.

[8] Li Wentao, Pan Baoxiang, Xia Jiangjiang, et al. Convolutional neural network-based statistical post-processing of ensemble precipitation forecasts[J]. Journal of Hydrology, 2022, 605: 127301. DOI: 10.1016/j.jhydrol.2021.127301.

[9] Eze J, Salihu B, Isong A, et al. Climate change impact on agriculture and water resources: A review[J]. Badeggi Journal of Agricultural Research and Environment, 2022, 4: 72-85. DOI: 10.35849/BJARE202202012.

[10] 李梦杰, 刘琨, 牟海磊, 等. 结合多重假设检验的随机森林长期降水预测方法及应用[J]. 南水北调与水利科技（中英文）, 2024, 22(5): 920-926. DOI: 10.13476/j.cnki.nsbdqk.2024.0091.

[11] 黄泽青, 赵铜铁钢, 田雨, 等. 全球气象模型GFDL-SPEAR月降水预报订正与检验[J]. 南水北调与水利科技（中英文）, 2024, 22(1): 56-66. DOI: 10.13476/j.cnki.nsbdqk.2024.0008.

[12] 崔洪铭, 刘梅冰, 甘贝贝, 等. 闽江流域极端降水时空演变特征及气候影响因子[J]. 水利水电技术（中英文）, 2025, 56(11): 59-70. DOI: 10.13928/j.cnki.wrahe.2025.11.005.

[13] 谢漪云, 王建捷. GRAPES千米尺度模式在西南复杂地形区降水预报偏差与成因初探[J]. 气象学报, 2021, 79(5): 732-749. DOI: 10.11676/qxxb2021.053.

[14] Xu Hongxiong, Zhao Yang, Zhao Dajun, et al. Improvement of disastrous extreme precipitation forecasting in north China by Pangu-Weather AI-driven regional WRF model[J]. Environmental Research Letters, 2024, 19(5): 054051. DOI: 10.1088/1748-9326/ad41f0.

[15] 张永轩, 黄兴友, 王雪婧, 等. 结合Causal-LSTM单元的CrevNet深度学习模型在对流降水临近预报中的试验研究[J]. 水利水电技术（中英文）, 2024, 55(11): 1-14. DOI: 10.13928/j.cnki.wrahe.2024.11.001.

[16] Wang Qinjun, Robertson D E, Chiew F H S. A Bayesian joint probability modeling approach for seasonal forecasting of streamflows at multiple sites[J]. Water Resources Research, 2009, 45(5): W05407. DOI: 10.1029/2008WR007355.

[17] Shrestha D L, Robertson D E, Bennett J C, et al. Improving precipitation forecasts by generating ensembles through postprocessing[J]. Monthly Weather Review, 2015, 143(9): 3642-3663. DOI: 10.1175/MWR-D-14-00329.1.

[18] 吴国丽, 鲍艳松, 林青, 等. 基于BMA-QM的福建区域降水预报订正及暴雨检验研究[J]. 应用数学进展, 2022(9): 6755-6767. DOI: 10.12677/AAM.2022.119716.

[19] 周正, 叶爱中, 马凤, 等. 基于贝叶斯理论的水文多模型预报[J]. 南水北调与水利科技, 2017, 15(1): 43-48. DOI: 10.13476/j.cnki.nsbdqk.2017.01.008.

[20] 吉璐莹. 基于多模式集成和动力降尺度的东亚夏季降水数值预报研究[D]. 南京: 南京信息工程大学, 2023.

[21] 星寅聪, 林依雪, 李艳忠, 等. 基于EMOS方法的多源遥感降水数据融合与性能评估: 以黄河中游为例[J]. 南水北调与水利科技（中英文）, 2021, 19(2): 226-236. DOI: 10.13476/j.cnki.nsbdqk.2021.0024.

[22] 马思敏, 舒志亮, 常倬林, 等. 宁夏六盘山区地面雨滴谱特征统计分析[J]. 干旱区研究, 2023, 40(8): 1203-1214. DOI: 10.13866/j.azr.2023.08.01.

[23] 顾家辉, 郭家力, 李颖, 等. 降水产品偏差校正方法在长江子流域应用效果[J]. 南水北调与水利科技（中英文）, 2024, 22(2): 319-330. DOI: 10.13476/j.cnki.nsbdqk.2024.0034.

[24] Scheuerer M, Hamill T M. Statistical postprocessing of ensemble precipitation forecasts by fitting censored, shifted gamma distributions[J]. Monthly Weather Review, 2015, 143(11): 4578-4596. DOI: 10.1175/MWR-D-15-0061.1.

[25] 张普宁. 不同时空尺度下降雨量空间插值方法研究[D]. 焦作: 河南理工大学, 2023.

[26] Tong Xuan, Zhou Wen, Xia Jiangjiang. Improving Boreal summer precipitation predictions from the global NMME through Res34-Unet[J]. Geophysical Research Letters, 2024, 51(2): e2023GL106391. DOI: 10.1029/2023GL106391.

[27] Ronneberger O, Fischer P, Brox T. U-Net: Convolutional networks for biomedical image segmentation[A]// Navab N, Hornegger J, Wells W M, et al. Medical Image Computing and Computer-Assisted Intervention[C]// MICCAI 2015. Lecture Notes in Computer Science, vol 9351. Cham: Springer, 2015: 234-241. DOI: 10.1007/978-3-662-54345-0_3.

[28] Pic R, Dombry C, Naveau P, et al. Distributional regression U-Nets for the postprocessing of precipitation ensemble forecasts[J]. Artificial Intelligence for the Earth Systems, 2025, 4(4): 240067. DOI: 10.1175/AIES-D-24-0067.1.

[29] 黄赛男, 李文韬, 段青云. GEFSv12降水再预报数据在淮河流域的适用性评估[J]. 南水北调与水利科技（中英文）, 2022, 20(5): 925-934. DOI: 10.13476/j.cnki.nsbdqk.2022.0092.

[30] 于群, 吴炜, 周发琇, 等. 中国东部降水的气候模态及雨季划分[J]. 大气科学学报, 2014, 37(3): 378-384. DOI: 10.13878/j.cnki.dqkxxb.2014.03.012.

[31] 智协飞, 张珂珺, 田烨, 等. 基于神经网络和地理信息的华东及华南地区降水概率预报[J]. 大气科学学报, 2021, 44(3): 381-393. DOI: 10.13878/j.cnki.dqkxxb.20210117001.

[32] Adachi M, Fujisawa M, Osborne M A. Fixing the pitfalls of probabilistic time-series forecasting evaluation by kernel quadrature[EB/OL]. (2025-09-16). https://arxiv.org/abs/2503.06079. DOI: 10.48550/arXiv.2503.06079.

[33] Gompel J V, Claessens B, Develder C. Probabilistic forecasting of power system imbalance using neural network-based ensembles[J]. Applied Energy, 2025, 401(partB): 126714. DOI: 10.1016/j.apenergy.2025.126714.

[34] Hu Weiming, Ghazvinian M, Chapman W E, et al. Deep learning forecast uncertainty for precipitation over the western United States[J]. Monthly Weather Review, 2023, 151(6): 1367-1385. DOI: 10.1175/MWR-D-22-0268.1.

[35] Koochali A, Tahaei E, Dengel A, et al. VAEneu: A new avenue for VAE application on probabilistic forecasting[J]. Applied Intelligence, 2025, 55(7): 483. DOI: 10.1007/s10489-024-06203-5.

[36] Olafsdottir H K, Rootzén H, Bolin D. Locally tail-scale invariant scoring rules for evaluation of extreme value forecasts[J]. International Journal of Forecasting, 2024, 40(4): 1701-1720. DOI: 10.1016/j.ijforecast.2024.02.007.

[37] 杜晓婉, 陈曦, 郑宏伟, 等. 一种降水降尺度模型: 地形约束下的条件生成对抗网络[J]. 地球信息科学学报, 2023, 25(8): 1586-1600. DOI: 10.12082/dqxxkx.2023.230033.

[38] 段青云, 叶爱中. 改善水文气象预报的统计后处理[J]. 水资源研究, 2012, 1(4): 161-168. DOI: 10.12677/JWRR.2012.14023.

[39] 王晓燕, 刘刚, 刘依嘉, 等. 基于机器学习模型的塔里木河流域卫星降水融合校正[J]. 水土保持学报, 2025, 39(6): 1-12. DOI: 10.13870/j.cnki.stbcxb.2025.06.018.

[40] 赵忱. 我国地面降水的多模式多成员集成预报及检验研究[D]. 南京: 南京信息工程大学, 2020.

[41] 彭飞, 李晓莉, 陈静, 等. CMA全球集合预报系统误差增长及预报性能的尺度依赖特征诊断分析[J]. 气象学报, 2023, 81(4): 605-618. DOI: 10.11676/qxxb2023.20220139.

[42] 牛嫣静, 徐向军, 郭欢, 等. 三种全球预报产品中国区近地面气温短期预报效果检验[J]. 气象与环境学报, 2020, 36(4): 18-27. DOI: 10.3969/j.issn.1673-503X.2020.04.003.

[43] 胡海川, 董林. 一种基于集合数值预报产品的台风强度订正方法[J]. 气象学报, 2023, 81(2): 316-327. DOI: 10.11676/qxxb2023.20220026.

[44] Ahmadian R, Ghatee M, Wahlström J. Superior scoring rules for probabilistic evaluation of single-label multi-class classification tasks[J]. International Journal of Approximate Reasoning, 2025, 182: 109421. DOI: 10.1016/j.ijar.2025.109421.

[45] 姚梦颖, 娄小芬, 刘雪晴, 等. 2020年超长梅汛期降水概率预报应用与检验[J]. 气象科技, 2024, 52(3): 367-379. DOI: 10.19517/j.1671-6345.20230172.

[46] 罗聪, 时洋, 吴乃庚, 等. 基于邻域最优概率的GRAPES快速更新循环同化预报系统逐时降水订正方法[J]. 热带气象学报, 2021, 37(4): 569-578. DOI: 10.16032/j.issn.1004-4965.2021.054.

[47] 朱科锋, 张晨悦, 薛明, 等. 对流可分辨尺度集合预报对河南“21·7”极端降水事件可预报性研究[J]. 中国科学: 地球科学, 2022, 52(10): 1905-1928. DOI: 10.1360/N072022-0046.

[48] 张颖. 降水集合预报的检验分析及在洪水预报中的应用研究[D]. 武汉: 华中科技大学, 2019.

[49] 杨寅, 林建, 包红军. 基于集合降水预报的长江流域面雨量预报与应用[J]. 暴雨灾害, 2022, 41(5): 571-579. DOI: 10.12406/byzh.2022-029.

[50] 张振东, 罗斌, 覃晖, 等. 风光水互补系统时间序列变量概率预报框架[J]. 水利学报, 2022, 53(8): 949-963. DOI: 10.13243/j.cnki.slxb.20220155.

[51] Wang Qinjun, Robertson D E. Multisite probabilistic forecasting of seasonal flows for streams with zero value occurrences[J]. Water Resources Research, 2011, 47(2): 155-170. DOI: 10.1029/2010WR009333.

[52] Renard B, Kavetski D, Kuczera G, et al. Understanding predictive uncertainty in hydrologic modeling: the challenge of identifying input and structural errors[J]. Water Resources Research, 2010, 46(5): W05521. DOI: 10.1029/2009WR008328.

基金资助

国家自然科学基金项目(W2431029)

河海大学校内基本科研业务费资助项目(B240201114)

河海大学校内基本科研业务费资助项目(B240203007)