基于机器学习的次季节尺度降水集合预报误差订正方法

Error correction for subseasonal ensemble precipitation forecasting based on machine learning

摘要

为提高降水集合预测精度和可靠性，需对数值模式预测降水进行误差订正。以汉江上游流域为研究区域，构建机器学习降水预测误差订正方法即分位数回归森林（quantile regression forest，QRF），对欧洲中期天气预报中心（European Centre for Medium-Range Weather Forecasts, ECMWF）模式次季节尺度逐日降水预测结果进行误差订正，并与传统降水预测误差订正方法中的分位数映射（quantile mapping，QM）、贝叶斯联合概率（Bayesian joint probability，BJP）进行对比。从集合预报角度，分析不同误差订正方法精度和可靠性差异。结果表明：与 ECMWF 模式预测结果相比，QM、BJP、QRF 方法均可以有效提高原始降水集合预测精度，BJP 方法订正后预测精度略高于 QRF 方法，QM 方法最低；不同误差订正方法预测降水精度均随着预见期的延长而降低，当预见期达到 20 d 以上时，BJP 和 QRF 方法预测精度和气候预测相当；QM 方法和 ECMWF 模式预测结果低于气候预测精度；ECMWF 模式预测和 QM 方法订正后预测降水的可靠性均较低。BJP 和 QRF 方法均能够提高降水集合预测的可靠性，且 QRF 方法的可靠性高于 BJP 方法。研究成果可为汉江上游流域水旱灾害防御、水资源优化配置提供重要支撑。

Abstract

Water resource managers and users received crucial information from skillful sub-seasonal precipitation forecasts (between two weeks and three months). Sub-seasonal forecasts remained difficult because predictability from initial atmospheric conditions faded after two weeks and variability in the lower boundary was too short to be effective at this time scale. With a better understanding of potential predictability sources and the development of dynamical climate models, sub-seasonal forecasts have advanced dramatically. In recent years, many operational centers have developed Global Climate Models (GCMs) to provide routine subseasonal precipitation forecasts. Nevertheless, the GCM precipitation forecasts were always biased, as discrepancies existed between the numerical approximations of the Earth's system and the real world. Furthermore, the ensemble spread of GCMs was frequently insufficient to provide accurate forecasts. It was necessary to correct GCM precipitation forecasts to improve their accuracy and reliability before they could be used for hydrological applications such as flood and drought disaster prevention and water resource management. Over the last two decades, several traditional statistical error correction methods have been proposed to improve forecast accuracy and reliability, including Bayesian joint probability, censored and shifted gamma distribution, quantile mapping (QM), and others. However, statistical error correction methods were based on probabilistic distribution assumptions, which proved problematic for skewed precipitation data. Machine learning methods could capture the nonlinear relationships among variables and were gradually applied in the field of error correction for precipitation forecasts in recent years, including random forest and eXtreme gradient boosting. The machine learning methods did not rely on the statistical assumptions and were of great potential application. However, because uncertainty was not taken into account, the majority of machine learning techniques were unable to produce probabilistic forecasts. The quantile regression forests (QRF), a variant of random forests (RF), could not only predict the conditional mean, but also generate the complete conditional probability distribution. Studies proved that the QRF method could improve wind and temperature forecast skill. However, the application of QRF on correcting precipitation forecasts was less. Meanwhile, the impact of various error correction methods on sub-seasonal precipitation ensemble forecasts remained to be investigated. QRF method was applied to correct the daily precipitation forecast results of the ECMWF (European Centre for Medium-Range Weather Forecasts) model at the sub-seasonal scale over the Upper Hanjiang River basin. The forecast accuracy and reliability were assessed through a leave-one-year-out cross-validation using continuous ranked probability skill score andα-index. We further compared the performance of the QRF method with that of two traditional error correction methods for precipitation forecasts, including QM and BJP. The QM, BJP, and QRF methods all significantly improved the accuracy of the original ensemble precipitation forecasts when compared to those derived directly from the ECMWF model. However, these methods differed in some respects. The BJP method yielded slightly higher forecast accuracy after correction than the QRF method, and the QM method performed the poorest. The precipitation forecast accuracy of all error correction methods decreased with the extension of lead time. When the lead time exceeded 20 days, the forecast accuracy of the BJP and QRF methods was comparable to that of climatic forecasts, while the forecast accuracy of the QM method and the original ECMWF model was lower than that of climatic forecasts. Both the precipitation forecasts of the ECMWF model and those corrected by the QM method showed low reliability. In contrast, the BJP and QRF methods could both improve the reliability of ensemble precipitation forecasts, with the QRF method outperforming the BJP method in terms of reliability. These findings provided important support for flood and drought disaster prevention and optimal allocation of water resources in the Upper Hanjiang River basin.

关键词

降水预测 / 误差订正 / 机器学习 / 分位数回归森林 / 汉江上游流域

Key words

precipitation forecast / error correction / machine learning / quantile regression forest / upper Hanjiang River basin

参考文献

[1] 赵铜铁钢, 张弛, 田雨, 等. 全球气象预报驱动流域水文预报研究进展与展望[J]. 水科学进展, 2024, 35(1): 156-166. DOI: 10.14042/j.cnki.32.1309.2024.01.014.

[2] 杨子昕, 王佳, 刘家宏, 等. 城市场次降雨情景库构建及预报降雨匹配技术[J]. 水利水电技术（中英文）, 2024, 55(10): 13-24. DOI: 10.13928/j.cnki.wrahe.2024.10.002.

[3] Vitart F, Robertson A W, Brookshaw A, et al. The WWRP/WCRP S2S project and its achievements[J]. Bulletin of the American Meteorological Society, 2025, 106(5): 791-808. DOI: 10.1175/BAMS-D-24-0047.1.

[4] Yao Junchen, Liu Xiangwen, Wu Tongwen, et al. Progress of MJO prediction at CMA from phase Ⅰ to phase Ⅱ of the sub-seasonal to seasonal prediction project[J]. Advances in Atmospheric Sciences, 2023, 40(10): 1799-1815. DOI: 10.1007/s00376-023-2351-z.

[5] 高斌. 长江上游极端降水非平稳拟合及致灾效应分析[J]. 水利水电技术（中英文）, 2025, 56(2): 1-14. DOI: 10.13928/j.cnki.wrahe.2025.02.001.

[6] De Andrade F M, Coelho C A S, Cavalcanti I F A. Global precipitation hindcast quality assessment of the subseasonal to seasonal (S2S) prediction project models[J]. Climate Dynamics, 2019, 52(9): 5451-5475. DOI: 10.1007/s00382-018-4457-z.

[7] 谢洁宏, 林巧美, 胡娅敏, 等. 次季节-季节预报模式对华南前后汛期降水的预报技巧评估[J]. 热带气象学报, 2025, 41(2): 224-235. DOI: 10.16032/j.issn.1004-4965.2025.021.

[8] 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.

[9] 韦经豪, 黄迎春, 姚成. 降水预报产品在不同水文气象分区中小流域的适应性评估[J]. 南水北调与水利科技（中英文）, 2022, 20(6): 1208-1219. DOI: 10.13476/j.cnki.nsbdqk.2022.0119.

[10] 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.

[11] 苏翔, 袁慧玲. 集合预报统计学后处理技术研究进展[J]. 气象科技进展, 2020, 10(2): 30-41. DOI: 10.3969/j.issn.2095-1973.2020.02.005.

[12] Vannitsem S, Bremnes J B, Demaever J, et al. Statistical post-processing for weather forecasts: Review, challenges, and avenues in a big data world[J]. Bull. American Meteorological Society, 2021(102): 681-699. DOI: 10.1175/BAMS-D-19-0308.1.

[13] Maraun D. Bias correction, quantile mapping, and downscaling: Revisiting the inflation issue[J]. Journal of Climate, 2013, 26(6): 2137-2143. DOI: 10.1175/JCLI-D-12-00821.1.

[14] Raftery A E, Gneiting T, Balabdaoui F, et al. Using Bayesian model averaging to calibrate forecast ensembles[J]. Monthly Weather Review, 2005, 133(5): 1155-1174. DOI: 10.1175/MWR2906.1.

[15] Scheuerer M, Hamill T M. Statistical postprocessing of ensemble precipitation forecasts by fitting censored, shifted gamma distributions[J]. Monthly Weather Review, 2015, 141(11): 4578-4596. DOI: 10.1175/MWR-D-15-0061.1.

[16] Schepen A, Zhao Tongtiegang, Wang Quanjun, et al. A Bayesian modelling method for post-processing daily sub-seasonal to seasonal rainfall forecasts from global climate models and evaluation for 12 Australian catchments[J]. Hydrology and Earth System Sciences, 2018, 22(2): 1615-1628. DOI: 10.5194/hess-22-1615-2018.

[17] 曹萍萍, 肖递祥, 龙柯吉, 等. 基于分位数映射法的四川省 ECMWF 模式降水预报误差订正分析[J]. 干旱气象, 2023, 41(4): 666-675. DOI: 10.11755/j.issn.1006-7639（2023）-04-0666.

[18] Li Yuan, Wu Zhiyong, He Hai, et al. Post-processing sub-seasonal precipitation forecasts at various spatiotemporal scales across China during boreal summer monsoon[J]. Journal of Hydrology, 2021, 598： 125742. DOI: 10.1016/j.jhydrol.2020.125742.

[19] 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.

[20] 李伶杰, 王银堂, 云兆得, 等. 顾及分类与定量误差订正的数值预报降水统计后处理方法[J]. 水科学进展, 2024, 35(1): 48-61. DOI: 10.14042/j.cnki.32.1309.2024.01.005.

[21] 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.

[22] 钟学敏, 文军, 王卓元, 等. 黄河源区径流预报的机器学习算法研究[J]. 南水北调与水利科技（中英文）, 2026, 24(1): 111-122. DOI: 10.13476/j.cnki.nsbdqk.2026.0011.

[23] 尹明波, 窦明, 刘玉倩, 等. 基于 XGBoost-MIKE21 耦合模型的缺资料水库富营养化模拟[J]. 南水北调与水利科技（中英文）, 2026, 24(1): 224-234. DOI: 10.13476/j.cnki.nsbdqk.2026.0022.

[24] 姜煦, 胡义明, 曹子恒, 等. 基于多机器学习模型综合集成的丹江口流域中长期降雨预报[J]. 南水北调与水利科技（中英文）, 2026, 24(1): 187-195. DOI: 10.13476/j.cnki.nsbdqk.2026.0018.

[25] 曹子恒, 李永坤, 胡义明, 等. 基于机器学习模型的数值降雨预报校正[J]. 南水北调与水利科技（中英文）, 2023, 21(5): 843-861. DOI: 10.13476/j.cnki.nsbdqk.2023.0083.

[26] 钱磊, 孙明馨, 刘晓蓓, 等. 基于随机森林算法的模式晴雨预报技术及应用效果检验评估[J]. 暴雨灾害, 2026, 45(3): 405-416.

[27] 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.

[28] Zhang Yuhang, Ye Aizhong, Nguyen P, et al. QRF4P-NRT: Probabilistic postprocessing of near-real-time satellite precipitation estimates using quantile regression forests[J]. Water Resources Research, 2022, 58(5): e2022WR032117. DOI: 10.1029/2022WR032117.

[29] Taillardat M, Mestre O, Zamo M, et al. Calibrated ensemble forecasts using quantile regression forests and ensemble model output statistics[J]. Monthly Weather Review, 2016, 144(6): 2375-2393. DOI: 10.1175/MWR-D-15-0260.1.

[30] Zhang Yuhang, Ye Aizhong, Analui B, et al. Comparing quantile regression forest and mixture density long short-term memory models for probabilistic post-processing of satellite precipitation-driven streamflow simulations[J]. Hydrology and Earth System Sciences, 2023, 27(24): 4529-4550. DOI: 10.5194/hess-27-4529-2023.

[31] 胡洪烨, 杨晨曦, 姜煦, 等. 气候变化背景下丹江口水库入库年径流预估[J]. 南水北调与水利科技（中英文）, 2026, 24(1): 206-213. DOI: 10.13476/j.cnki.nsbdqk.2026.0020.

[32] 万育生, 王栋, 黄朝君. 丹江口水库来水情势分析与径流预测[J]. 南水北调与水利科技（中英文）, 2021, 19(3): 417-426. DOI: 10.13476/j.cnki.nsbdqk.2021.0045.

[33] Vitart F, Ardilouze C, Bonet A, et al. The subseasonal to seasonal (S2S) prediction project database[J]. Bulletin of the American Meteorological Society, 2017, 98(1): 163-173. DOI: 10.1175/BAMS-D-16-0017.1.

[34] 陈亚玲, 文军, 杨传国, 等. 不同降水产品及 WRF-Hydro 模式在黄河源区的适用性分析[J]. 大气科学, 2024, 48(3): 923-937. DOI: 10.3878/j.issn.1006-9895.2205.22057.

[35] 王文雨, 刘刚, 王晓燕, 等. 多源降水产品在天山典型流域的评估及径流模拟[J]. 水土保持学, 2025, 39(1): 284-293. DOI: 10.13870/j.cnki.stbcxb.2025.01.005.

[36] 童尧, 高学杰, 韩振宇, 等. 基于 RegCM4 模式的中国区域日尺度降水模拟误差订正[J]. 大气科学, 2017, 41(6): 1156-1166. DOI: 10.3878/j.issn.1006-9895.1704.16275.

[37] Wang Qijun, Shrestha D L, Robertson D E, et al. A log-sinh transformation for data normalization and variance stabilization[J]. Water Resources Research, 2012, 48(5): W05514. DOI: 10.1029/2011WR010973.

[38] Wang Q J, Shao Yawen, Song Yong, et al. An evaluation of ECMWF SEAS5 seasonal climate forecasts for Australia using a new forecast calibration algorithm[J]. Environmental Modelling & Software, 2019, 122： 104550. DOI: 10.1016/j.envsoft.2019.104550.

[39] Li Yuan, Xu Gaoliang, Xu Jian, et al. Comparing the impact of precipitation pre-processing and streamflow post-processing for daily sub-seasonal streamflow forecasts over the Gan River basin[J]. Journal of Hydrology, 2026, 666： 134773. DOI: 10.1016/j.jhydrol.2025.134773.

[40] Maxwell A E, Warner T A, Fang F. Implementation of machine-learning classification in remote sensing: An applied review[J]. International Journal of Remote Sensing, 2018, 39(9): 2784-2817. DOI: 10.1080/01431161.2018.1433343.

[41] Jia Xiaolong, Chen Lijuan, Ren Fumin, et al. Impacts of the MJO on winter rainfall and circulation in China[J]. Advances in Atmospheric Sciences, 2011, 28(3): 521-533. DOI: 10.1007/s00376-010-9118-z.

[42] Ren Pengfei, Ren Hongli, Fu Xiuhua, et al. Impact of boreal summer intraseasonal oscillation on rainfall extremes in southeastern China and its predictability in CFSv2[J]. Journal of Geophysical Research: Atmospheres, 2018, 123(9): 4423-4442. DOI: 10.1029/2017JD028043.

[43] 叶宇辰, 陈海山, 朱司光, 等. 基于机器学习的中国夏季降水延伸期预报及土壤湿度的可能贡献[J]. 高原气象, 2024, 43(1): 184-198. DOI: 10.7522/j.issn.1000-0534.2023.00025.

[44] Zhu Zhiwei, Wu Junting, Huang Hongjie. The influence of 10-30-day boreal summer intraseasonal oscillation on the forecast skill of extreme rainfall over southern China[J]. Climate Dynamics, 2024, 62(1): 69-86. DOI: 10.1007/s00382-023-06900-w.

[45] Wu Jie, Ren Hongli, Zhang Peiqun, et al. The dynamical-statistical subseasonal prediction of precipitation over China based on the BCC new-generation coupled model[J]. Climate Dynamics, 2022, 59(3-4): 1213-1232. DOI: 10.1007/s00382-022-06187-3.

基金资助

国家自然科学基金面上项目(52579004)