Research on Quantitative Precipitation Correction Forecasting Based on the Improved U-Net Model in Hunan
-
摘要: 利用2017—2022年汛期(4—9月)湖南省1 912个地面观测站降水实况和欧洲中心中期天气预报一体化预报系统(European Centre for Medium-Range Weather Forecasts-Integrated Forecasting System，ECMWFIFS)最优因子集，在一种U型语义分割网络(U-Net模型)基础上结合残差网络和注意力机制网络，构建了逐时降水订正预报模型(SARU)，并将模型2023年汛期预报结果与最优TS评分订正法(OTS)以及中国气象局上海数值预报模式系统(CMA-SH9)进行对比。(1) SARU模型整体的晴雨准确率、相关系数(COR)、平均绝对误差(MAE)、偏差(BIAS)分别为0.87、0.17、0.35、0.73，皆优于OTS模型和CMA-SH9模式，尤其在湘中地区，SARU模型对于预报趋势和量级有明显优势，其空报率和漏报率比例基本相当，OTS模型漏报率远超空报率，CMA-SH9则正好相反。(2) SARU模型对于分级降水频次的预报较OTS模型和CMA-SH9模式更接近实况，尤其是20 mm以上降水频次，预报偏少27.29%，远优于OTS模型预报的偏少85.54% 和CMA-SH9模式的偏多95.50%。(3) 对于小时雨量[5，10)、[10，20)和≥ 20 mm这三个级别的降水，SARU模型TS、命中率(POD)、空报率(FAR)、漏报率(MAR)皆最优，尤其短时强降水，SARU模型较CMA-SH9模式有明显优势，而OTS模型的预报能力则明显不足。(4) 湖南存在明显的夜雨特征，夜间时段(北京时02—08时)短时强降水频次明显高于其他时段。SARU模型很好地把握了夜间短时强降水特征，TS在夜间明显升高，尤其是在北京时05时达到峰值(0.07左右)，明显优于CMA-SH9模式和OTS模型。Abstract: This study presents a quantitative precipitation forecasting correction experiment in Hunan Province based on an improved U-Net model. Utilizing precipitation data from 1912 ground observation stations in Hunan during the rainy season from April to September between 2017 and 2022, along with the optimal factor set of the European Centre for Medium-Range Weather Forecast-Integrated Forecasting System (ECMWF-IFS) model, we developed an hourly precipitation forecasting correction model (SARU) on the basis of the U-Net model, integrating residual networks and attention mechanism networks. The model' s forecast results for the 2023 rainy season were compared with those corrected by using the optimal threat score (OTS) method and those from the CMA-SH9 model. The results show that: (1) The SARU model' s overall accuracy in clear/rainy forecasts, correlation coefficient, mean absolute error, and bias were 0.87, 0.17, 0.35, and 0.73, respectively, all of which outperformed the OTS model and the CMASH9 model, especially in central Hunan, where SARU showed a clear advantage in forecasting trends and magnitudes; its false alarm and missed forecast ratios were nearly equal, contrasting with the OTS model' s significantly higher missed forecast ratio and the CMA-SH9 model' s opposite trend. (2) The SARU model' s forecasts of categorized precipitation frequency were closer to actual observations than the OTS model and the CMA-SH9 model, particularly for precipitation exceeding 20 mm, where it was under-predicted by 27.29%, significantly better than the OTS model' s under-prediction of 85.54% and the CMA-SH9 model' s over-prediction of 95.50%. (3) For hourly precipitation at levels of [5, 10), [10, 20), and ≥20 mm, the SARU model achieved the best threat score (TS), probability of detection, false alarm ratio, and missed alarm ratio, particularly for short-term heavy precipitation, where it significantly outperformed the CMASH9 model, while the OTS model' s forecasting capability was notably insufficient. (4) Nocturnal rainfall was pronounced in Hunan, with a noticeable increase in the frequency of short-term heavy rainfall during the night (0200-0800 BJT). The SARU model effectively captured the nocturnal characteristics of shortterm heavy rainfall, with TS notably increasing at night, peaking around 0500 BJT at approximately 0.07, markedly outperforming the CMA-SH9 model and OTS model.
-
Key words:
- SARU /
- Correction /
- U-Net /
- OTS /
- hourly precipitation forecasting /
- nocturnal rainfall characteristics
-
表 1 最优预报因子集
要素 物理意义 地面(13个) 2T 地面两米气温 MN2T3 3 h最低气温 MX2T3 3 h最高气温 SKT 体感温度 LCC 自由对流高度 TCC 总云量 TCW 整层气柱水含量 TCWV 总体积水汽含量 MSL 平均海平面气压 VIS 能见度 TP 总降水量 CP 云压 LSP 自由对流高度 高空(46个) U(150/250/400/700/800/850 hPa) 150/250/400/700/800/850 hPa U风 V(150/250/400/700/800/850 hPa) 150/250/400/700/800/850 hPa V风 GH (800/850/900/925/950/1 000 hPa) 800/850/900/925/950/1 000 hPa高度 T(800/850/900/925/950/1 000 hPa) 800/850/900/925/950/1 000 hPa温度 R(200/250/300/400/500/600/700/800/900/925/950/1 000 hPa) 200/250/300/400/500/600/700/800/900/925/950/1 000 hPa相对湿度 Q(300/400/500/600/700/800/850/900/925/1 000 hPa) 300/400/500/600/700/800/850/900/925/1 000 hPa比湿 
下载: 导出CSV
表 3 逐小时降水预报整体评估
指标 SARU OTS CMA-SH9 晴雨准确率 0.868 0.789 0.847 COR 0.166 0.146 0.088 MAE 0.349 0.377 0.446 BIAS 0.731 0.145 1.954
下载: 导出CSV
表 4 2023年汛期逐小时分级降水划分和占比情况
分级 降水量划分/mm 占比 降水说明 级别1 < 0.1 89.491% 无降雨 级别2 [0.1，5) 9.505% - 级别3 [5，10) 0.583% - 级别4 [10，20) 0.298% - 级别5 ≥20 0.124% 短时强降雨
下载: 导出CSV
表 5 2023年汛期逐小时分级降水预报频次与实况对比
量级划分 < 0.1 [0.1，5) [5，10) [10，20) ≥20 观测 频次 12 513 431 1 329 075 81 485 41 627 17 343 SARU 频次 12 110 086 1 720 572 102 749 36 943 12 611 相对误差 -3.223% 29.456% 26.096% -11.252% -27.285% OTS 频次 10 859 910 2 968 872 111 731 39 940 2 508 相对误差 -13.214% 123.379% 37.118% -4.053% -85.539% CMA-SH9 频次 12 032 641 1 723 544 123 608 69 263 33 905 相对误差 -3.842% 29.680% 51.694% 66.390% 95.497%
下载: 导出CSV
表 6 2023年汛期逐小时分级降水预报评估
分级/mm 指标 SARU OTS CMA-SH9 [0.1，5) TS 0.257 0.191 0.207 POD POD 0.526 0.399 FAR 0.640 0.769 0.699 MAR 0.527 0.474 0.601 [5，10) TS 0.040 0.032 0.024 POD 0.087 0.074 0.060 FAR 0.931 0.946 0.962 MAR 0.913 0.926 0.940 [10，20) TS 0.020 0.017 0.014 POD 0.037 0.032 0.036 FAR 0.958 0.966 0.978 MAR 0.963 0.968 0.964 ≥20 TS 0.015 0.001 0.007 POD 0.026 0.001 0.021 FAR 0.964 0.992 0.989 MAR 0.974 0.999 0.977
下载: 导出CSV
-
[1] 黄晓远, 李旭, 杜梦莹, 等. 融合相似预报方法在陇东南短期强降水预报中的应用[J/OL]. 高原气象, 2025, 44(1): 214-223. [2] 张亚萍, 张焱, 翟丹华, 等. 深度学习方法在强降水临近预报技术中的应用及其改进思考[J]. 暴雨灾害, 2022, 41(5): 506-514. [3] 李超, 裘薇, 娄小芬, 等. 浙江省温度和相对湿度释用技术及其效果检验分析[J]. 气象科学, 2016, 36(4): 556-561. [4] 吴乃庚, 曾沁, 刘段灵, 等. 日极端气温的主客观预报能力评估及多模式集成网格释用[J]. 气象, 2017, 43(5): 581-590. [5] 张成军, 纪晓玲, 马金仁, 等. 多种数值预报及其释用产品在宁夏天气预报业务中的检验评估[J]. 干旱气象, 2017, 35(1): 148-156. [6] 曾晓青, 薛峰, 赵瑞霞, 等. 几种格点化温度滚动订正预报方案对比研究[J]. 气象, 2019, 45(7): 1 009-1 018. [7] 盛春岩, 范苏丹, 荣艳敏, 等. 几种气温客观预报方法对比及最优集成预报研究[J]. 气象, 2020, 46(10): 1 351-1 361. [8] 吕林宜, 王新敏, 栗晗. 华东区域模式对河南"7·19" 特大暴雨的数值检验与分析[J]. 气象与环境科学, 2019, 42(1): 101-109. [9] 郑婧, 夏侯杰, 陈娟, 等. 基于ECMWF模式的定量降水客观订正方法[J]. 高原气象, 2020, 39(4): 830-839. [10] 肖瑶, 杜明哲, 徐鹏. 两种订正方法在河南高速公路沿线气温精细化预报中的应用[J]. 干旱气象, 2020, 38(6): 1 023-1 030. [11] 赵宁坤, 张秀年, 孙俊奎, 等. 高分辨率区域模式降水预报在云南的检验[J]. 暴雨灾害, 2021, 40(1): 78-86. [12] 陈豫英, 陈楠, 马金仁, 等. 宁夏汛期分级降水客观预报方法应用检验[J]. 暴雨灾害, 2016, 35(5): 546-553. [13] 杨璐, 南刚强, 陈明轩, 等. 基于三种机器学习方法的降水相态高分辨率格点预报模型的构建及对比分析[J]. 气象学报, 2021, 79(6): 1 022-1 034. [14] KLEIN W H, LEWIS B M, ENGER I. Objective prediction of five-day mean temperatures during winter[J]. J Meteor, 1959, 16(6): 672-682. [15] GLAHN H R, LOWRY D A. The use of model output statistics (MOS) in objective weather forecasting[J]. J Appl Meteor, 1972, 11(8): 1 203-1 211. [16] HOMLEID M. Diurnal corrections of short-term surface temperature forecasts using the Kalman filter[J]. Wea Forecasting, 1995, 10(4): 689-707. [17] DOSWELL C A. Flash flood forecasting: An ingredient-based methodology[J]. Wea Forecasting, 1996, 11(4): 560-581. [18] 张小玲, 陶诗言, 孙建华. 基于"配料" 的暴雨预报[J]. 大气科学, 2010, 34(4): 754-764. [19] ELBERT E E. Ability of a poor man's ensemble to predict the probability and distribution of precipitation[J]. Mon Wea Rev, 2001, 129(10): 2 461-2 479. [20] ZHU Y J, LUO Y. Precipitation calibration based on the frequency matching method[J]. Wea Forecasting, 2015, 30(5): 1 109-1 124. [21] 李俊, 杜钧, 陈超君. 降水偏差订正的频率(或面积)匹配方法介绍和分析[J]. 气象, 2014, 40(5): 580-588. [22] 孙靖, 程光光, 张小玲. 一种改进的数值预报降水偏差订正方法及应用[J]. 应用气象学报, 2015, 26(2): 173-184. [23] 吴启树, 韩美, 刘铭, 等. 基于评分最优化的模式降水预报订正算法对比[J]. 应用气象学报, 2017, 28(3): 306-317. [24] 危国飞, 刘会军, 吴启树, 等. 多模式降水分级最优化权重集成预报技术[J]. 应用气象学报, 2020, 31(6): 668-680. [25] 彭昱忠, 王谦, 元昌安, 等. 数据挖掘技术在气象预报研究中的应用[J]. 干旱气象, 2015, 33 (1): 19-27. [26] 沈皓俊, 罗勇, 赵宗慈, 等. 基于LSTM网络的中国夏季降水预测研究[J]. 气候变化研究进展, 2020, 16(3): 263-275. [27] 刘娜, 黄武斌, 杨建才, 等. 基于SCTP-RF算法的甘肃省短期定量降水客观预报方法研究[J]. 干旱气象, 2022, 40(1): 146-155. [28] RODRIGUES E R, OLIVEIRA I, CUNHA R L F, et al. Deep Downscale: A deep learning strategy for high-resolution weather forecast[C]// 2018 IEEE 14th International Conference on e-Science (e-Science). Amsterdam, Netherlands IEEE, 2018: 415-422. [29] SABOUR S, FROSST N, HINTON G E. Dynamic routing between capsules[J]. Advances in Neural Information Processing Systems, 2017: 3 856-3 866. [30] 李扬, 刘玉宝, 许小峰. 基于深度学习改进数值天气预报模式和预报的研究及挑战[J]. 气象科技进展, 2021, 11(3): 103-112. [31] HOCHREITER S, SCHMIDHUBER J. Long short-term memory[J]. Neural Comput, 1997, 9(8): 1 735-1 780. [32] CHO K, MERRIËNBOER B, GULCEHRE C, et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation[C]//Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). Doha, Qatar: Association for Computational Linguistics, 2014: 1 724-1 734. [33] GOODFELLOW I, BENGIO Y, COURVILLE A. Deep learning[M]. New York: MIT Press, 2016. [34] SHI X J, CHEN Z, WANG H, et al. Convolutional LSTM network: A machine learning approach for precipitation nowcasting[J]. Advances in Neural Information Processing Systems, 2015: 802-810. [35] SHI X J, GAO Z, LAUSEN L, et al. Deep learning for precipitation nowcasting: A benchmark and a new model[J]. Advances in Neural Information Processing Systems, 2017: 5 617-5 627. [36] WANG Y B, LONG M S, WANG J M, et al. PredRNN: Recurrent neural networks for predictive learning using spatiotemporal LSTMs[C]// Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach, USA: Curran Associates Inc, 2017: 879-888. [37] WANG Y B, GAO Z F, LONG M S, et al. PredRNN++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning[C]//Proceedings of the 35th International Conference on Machine Learning, PMLR 80, 2018: 5 123-5 132. [38] WANG Y B, ZHANG J J, ZHU H Y, et al. Memory in Memory: A predictive neural network for learning higher-order non-stationarity from spatiotemporal dynamics[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, CA, USA: IEEE, 2019: 9146-9154. [39] CHEN L, CAO Y, MA L M, et al. A deep learning-based methodology for precipitation nowcasting with radar[J]. Earth Space Sci, 2020, 7 (2): e2019EA000812 [40] HE K, ZHAN X, REN S, et al. Deep Residual Learning for Image Recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 2016: 770-778. doi: 10.1109/CVPR.2016.90 .[41] HE K, ZHANG X, REN S, et al. Identity mappings in deep residual networks[J]. 2016, arXiv: 1603.05027v3. [42] RASP S, THUEREY N. Data-driven medium-range weather prediction with a Resnet pretrained on climate simulations: A new model for weatherbench[J]. Journal of Advances in Modeling Earth Systems. 2021, 13(2), p. e2020MS002405. [43] DUPUY F, MESTRE O, SERRURIER M, et al. ARPEGE cloud cover forecast postprocessing with convolutional neural network[J]. Wea Forecasting, 2021, 36(2): 567-586. [44] KUDO A. Statistical post-processing for gridded temperature prediction using encoder-decoder-based deep convolutional neural networks [CP/OL]. 2021, arXiv: 2103.01479. [45] HAN L, CHEN M X, CHEN K K, et al. A deep learning method for bias correction of ECMWF 24-240 h forecasts[J]. Adv Atmos Sci, 2021, 38(9): 1 444-1 459. [46] 张延彪, 陈明轩, 韩雷, 等. 数值天气预报多要素深度学习融合订正方法[J]. 气象学报, 2022, 80(1): 153-167. [47] HUANG G, LIU Z, VAN DER MAATEN L, et al. Densely connected convolutional networks[C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 2 261-2 269. [48] 杨绚, 代刊, 朱跃建. 深度学习技术在智能网格天气预报中的应用进展与挑战[J]. 气象学报, 2022, 80(5): 649-667. [49] 蔡康龙, 胡志群, 谭浩波, 等. 利用卷积神经网络开展偏振雷达定量降水估测研究[J]. 热带气象学报, 2024, 40(1): 64-74. [50] 杨璐, 南刚强, 陈明轩, 等. 基于三种机器学习方法的降水相态高分辨率格点预报模型的构建及对比分析[J]. 气象学报, 2021, 79(6): 1 022-1 034. [51] YOU X X, LIANG Z M, WANG Y Q, et al. A study on loss function against data imbalance in deep learning correction of precipitation forecasts[J]. Atmospheric Research, 2023, 281: 106500. -
点击查看大图
图(9) / 表(6)
计量
- 文章访问数: 11
- HTML全文浏览量: 5
- PDF下载量: 2
- 被引次数: 0
粤公网安备 4401069904700003号