| 74 | 0 | 89 |
| 下载次数 | 被引频次 | 阅读次数 |
【目的】针对小流域山洪起涨快、处置窗口短,以及现有站网观测难以转化为可用短临预警信息的问题,评估仅依赖历史多站点降雨和当前水文状态时的山洪预报能力。【方法】基于罗玉沟1991—2024年12站降雨和1990—2024年沟口流量资料,构建时空图Transformer(Topo-STGT),联合预测未来30~180 min流量、试验流量阈值超阈概率、洪峰和峰现时间,并在2021—2024年年代外独立测试集上开展7类模型对比、空间结构消融、事件级回算及缺站试验。【结果】30 min和60 min流量NSE分别为0.742和0.603;Topo-STGT相对普通Transformer的MAE降低0.190~0.396 m³·s⁻¹,表明多站点时空编码具有稳定增益。5 m³·s⁻¹试验阈值下,命中过程的中位有效提前量为32.6 min。59个可评价观测块的洪峰MAE为1.878 m³·s⁻¹、峰现时间MAE为68.9 min;随机缺失3站和5站后,MAE分别增至完整站网的1.541倍和2.181倍。STGT-Distance在60~180 min均取得更低MAE,说明固定DEM拓扑尚未形成额外预测增益。【结论】Topo-STGT能够在现有站网条件下为罗玉沟提供30~60 min定量流量预报和高流量风险筛查;本文建立的预见期分级、事件回算和缺站降级方法可为小流域山洪预警研判与站网运行提供量化依据。
Abstract:[Objective] To address the rapid rise of flash floods, short response windows, and the difficulty of converting existing gauge observations into usable nowcasting information in small watersheds, the flash-flood nowcasting capability based only on historical multi-gauge rainfall and current hydrological state is evaluated. [Methods] A spatiotemporal graph Transformer (Topo-STGT) was developed using rainfall records from 12 gauges in the Luoyugou Basin during 1991–2024 and outlet-discharge observations during 1990–2024. The model jointly predicted discharge at lead times of 30–180 min, exceedance probabilities for experimental discharge thresholds, peak discharge, and time to peak. Seven models were compared, and spatial-structure ablations, event-level retrospective analyses, and missing-gauge experiments were conducted on the temporally independent test set during 2021–2024. [Results] The NSE values were 0.742 and 0.603 at 30 and 60 min, respectively. Compared with a standard Transformer, Topo-STGT reduced MAE by 0.190–0.396 m³·s⁻¹, showing a stable benefit from multi-gauge spatiotemporal encoding. At the experimental threshold of 5 m³·s⁻¹, the median effective lead time among detected episodes was 32.6 min. Across 59 evaluable observation blocks, peak-discharge MAE was 1.878 m³·s⁻¹ and time-to-peak MAE was 68.9 min. Randomly masking three and five gauges increased MAE to 1.541 and 2.181 times the complete-network value, respectively. STGT-Distance achieved lower MAE at all lead times from 60 min to 180 min, indicating that the fixed DEM-derived topology provided no additional predictive gain. [Conclusion] Topo-STGT can support 30–60 min quantitative discharge forecasting and high-flow risk screening for the Luoyugou Basin under the existing gauge network. The lead-time grading, event-level retrospective analysis, and missing-gauge degradation methods developed in this study provide quantitative support for flash-flood warning assessment and gauge-network operation in small watersheds.
[1] 刘洪岫, 张晓蕾, 路江鑫, 等. 2024年全国山洪灾害防御工作[J]. 中国防汛抗旱, 2024, 34(12): 5-8.
[2] 吴娟, 林荷娟, 姜桂花, 等. 太湖流域超标特大洪水风险预警系统建设及应用[J]. 河海大学学报(自然科学版), 2023, 51(2): 164-170.
[3] 胡琳琳, 邱志章, 陆小勇. 浙江省山洪灾害气象风险预警模型研究及应用[J]. 中国防汛抗旱, 2025, 35(10): 46-50.
[4] 翟晓燕,李淼,刘昌军,等. 基于机器学习和水文模拟的不同山洪类型行为特征指标预测[J]. 南水北调与水利科技(中英文),2025,23(6):1468-1479.
[5] 庄潇然, 郑玉, 王亚强, 等. 基于深度学习的融合降水临近预报方法及其在中国东部地区的应用研究[J]. 气象学报, 2023, 81(2): 286-303.
[6] ZHANG Y C, LONG M S, CHEN K Y, et al. Skilful nowcasting of extreme precipitation with NowcastNet[J]. Nature, 2023, 619(7970): 526-532.
[7] RAVURI S, LENC K, WILLSON M, et al. Skilful precipitation nowcasting using deep generative models of radar[J]. Nature, 2021, 597(7878): 672-677.
[8] 王磊,牛庚,桑学锋,等. 基于径流组分约束的机器学习融水径流模型[J]. 南水北调与水利科技(中英文),2025,23(5):1173-1184.
[9] 田烨, 谭伟丽, 王国庆, 等. LSTM变体模型在径流预测中的性能及其可解释性[J]. 水资源保护, 2023, 39(3): 188-194.
[10] 李文忠, 刘成帅, 邬强, 等. GRU-Transformer洪水预报模型构建与应用[J]. 中国农村水利水电, 2023(11): 35-44.
[11] 刘成帅, 解添宁, 李文忠, 等. 考虑径流过程矢量化的机器学习洪水预报模型[J]. 水科学进展, 2024, 35(3): 420-429.
[12] 许月萍, 周欣磊, 王若桐, 等. 基于集成学习与深度学习的洪水径流预报研究[J]. 人民长江, 2024, 55(9): 18-25.
[13] 胡乐怡, 蒋晓蕾, 周嘉慧, 等. 考虑记忆时间的LSTM模型在赣江流域径流预报中的应用[J]. 湖泊科学, 2024, 36(4): 1241-1251.
[14] 崔震, 郭生练, 向鑫, 等. 基于AI深度学习的向家坝—三峡区间流域洪水预报[J]. 人民长江, 2025, 56(4): 49-55.
[15] LUO Y X, ZHOU Y L, CHEN H, et al. Exploring a spatiotemporal hetero graph-based long short-term memory model for multi-step-ahead flood forecasting[J]. Journal of Hydrology, 2024, 633: 130937.
[16] LI W Z, LIU C S, XU Y Y, et al. An interpretable hybrid deep learning model for flood forecasting based on Transformer and LSTM[J]. Journal of Hydrology: Regional Studies, 2024, 54: 101873.
[17] 张玮, 郑雅莲, 刘志武, 等. 物理机制引导的水库调度深度学习模型研究[J]. 水力发电学报, 2023, 42(3): 13-25.
[18] 崔震, 郭生练, 王俊, 等. 基于混合深度学习模型的洪水过程概率预报研究[J]. 水利学报, 2023, 54(8): 889-897.
[19] 崔震, 郭生练, 汪芸, 等. 三峡入库洪水概率预报的深度学习模型[J]. 水力发电学报, 2023, 42(4): 1-10.
[20] 解添宁, 胡彩虹, 刘成帅, 等. 基于洪水过程矢量方向的深度学习洪水概率预报方法[J]. 应用基础与工程科学学报, 2026, 34(2): 487-499.
[21] NEVO S, MORIN E, GERZI ROSENTHAL A, et al. Flood forecasting with machine learning models in an operational framework[J]. Hydrology and Earth System Sciences, 2022, 26(15): 4013-4032.
[22] NEARING G, COHEN D, DUBE V, et al. Global prediction of extreme floods in ungauged watersheds[J]. Nature, 2024, 627(8004): 559-563.
[23] ZHOU X Y, HUANG X R, JIANG X, et al. Real-time error correction of multiple-hour-ahead flash flood forecasting based on the sliding runoff-rain data and deep learning models[J]. Journal of Hydrology, 2025, 655: 132918.
[24] ROUDBARI N S, PUNEKAR S R, PATTERSON Z, et al. From data to action in flood forecasting leveraging graph neural networks and digital twin visualization[J]. Scientific Reports, 2024, 14: 18571.
[25] SAITO T, REHMSMEIER M. The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets[J]. PLoS One, 2015, 10(3): e0118432.
[26] GNEITING T, BALABDAOUI F, RAFTERY A E. Probabilistic forecasts, calibration and sharpness[J]. Journal of the Royal Statistical Society Series B: Statistical Methodology, 2007, 69(2): 243-268.
[27] KRATZERT F, KLOTZ D, BRENNER C, et al. Rainfall-runoff modelling using Long Short-Term Memory (LSTM) networks[J]. Hydrology and Earth System Sciences, 2018, 22(11): 6005-6022.
[28] PÖLZ A, BLASCHKE A P, KOMMA J, et al. Transformer versus LSTM: A comparison of deep learning models for karst spring discharge forecasting[J]. Water Resources Research, 2024, 60(4): e2022WR032602.
[29] ROGELIS M C, WERNER M. Streamflow forecasts from WRF precipitation for flood early warning in mountain tropical areas[J]. Hydrology and Earth System Sciences, 2018, 22(1): 853-870.
[30] POLETTI M L, SILVESTRO F, DAVOLIO S, et al. Using nowcasting technique and data assimilation in a meteorological model to improve very short range hydrological forecasts[J]. Hydrology and Earth System Sciences, 2019, 23(9): 3823-3841.
[31] 刘正疆, 张浩, 马丽娜, 等. 城市地表雨洪特征的数值模拟研究 [J] . 水利水电技术 (中英文) , 2025, 56 (10) : 46-57.
[32] LI J Z, LI L J, ZHANG T, et al. Flood forecasting based on radar precipitation nowcasting using U-Net and its improved models[J]. Journal of Hydrology, 2024, 632: 130871.
[33] HOSS F, FISCHBECK P S. Performance and robustness of probabilistic river forecasts computed with quantile regression based on multiple independent variables[J]. Hydrology and Earth System Sciences, 2015, 19(9): 3969-3990.
[34] 吴泽斌, 田济扬, 刘荣华, 等. 山洪灾害临近预报预警技术及实践应用[J]. 中国防汛抗旱, 2024, 34(12): 41-45.
[35] 刘荣华, 窦延虹, 高兰兰, 等. 山洪灾害预警模型改进与效果评估[J]. 中国防汛抗旱, 2025, 35(12): 45-50.
[36] FRAME J M, KRATZERT F, KLOTZ D, et al. Deep learning rainfall–runoff predictions of extreme events[J]. Hydrology and Earth System Sciences, 2022, 26(13): 3377-3392.
[37] SUN A Y, JIANG P S, MUDUNURU M K, et al. Explore spatio-temporal learning of large sample hydrology using graph neural networks[J]. Water Resources Research, 2021, 57(12): e2021WR030394.
[38] SUN A Y, JIANG P S, YANG Z L, et al. A graph neural network (GNN) approach to basin-scale river network learning: The role of physics-based connectivity and data fusion[J]. Hydrology and Earth System Sciences, 2022, 26(19): 5163-5184.
[39] LIU Y Q, HOU G B, HUANG F, et al. Directed graph deep neural network for multi-step daily streamflow forecasting[J]. Journal of Hydrology, 2022, 607: 127515.
[40] 王明阳, 王恩志, 罗火钱, 等. 中小流域径流预报的图神经网络模型: 以福建沙溪流域为例[J]. 水力发电学报, 2025, 44(6): 50-61.
[41] WANG H J, CHEN J Y, ZHENG Y Q, et al. Accelerating flood warnings by 10 hours: The power of river network topology in AI-enhanced flood forecasting[J]. npj Natural Hazards, 2025, 2: 45.
[42] XIA T Y, ZHOU Y L, XU C Y, et al. A spatiotemporal pathformer-based deep learning framework for watershed flood forecasting[J]. Water Resources Research, 2025, 61(12): e2025WR040193.
基本信息:
中图分类号:P338
引用信息:
[1]胡祥祥,石亚亚,安乐平,等.基于时空图Transformer的小流域山洪短临预报:预见期边界、缺站降级与预警效用评估[J].水利水电技术(中英文)().
基金信息:
国家自然科学基金项目(42361020); 甘肃省教育厅高校教师创新基金项目(2026A-158); 天水市科技局项目
2026-09-24
2026-09-24
2026-09-24