典型集总式水文模型误差优化方法对比评估

Comparative assessment of error optimization methods for typical lumped hydrological models

摘要

选取多个流域对比评估 3 种集总式水文模型（HMETS、GR4J、新安江）的预报精度，采用误差优化与耦合模拟方法，探讨各类方法对模型精度的提升效果及影响因素。结果表明：3 种模型在不同流域间的模拟精度存在显著差异，其中 HMETS 模型整体表现最优，但 GR4J 与新安江模型在部分流域仍具优势，表明模型选取应结合实际情况判定；LSTM 误差优化方法可有效提升模型在率定期的精度，但在验证期表现受限于数据长度与人类干扰程度，在观测数据较长的自然流域效果更为稳定；多模型耦合方法中，贝叶斯模式平均法在高精度单一模型基础上表现更优，而 LSTM 耦合方法适用于样本丰富的区域。本研究通过对比典型集总式水文模型的模拟性能，验证多模型协同优化的可行性与优势，为相关模型的精度提升与集成应用提供方法参考。

Abstract

Accurate runoff prediction is an important nonstructural measure for ensuring watershed water security and informing water management decisions, such as flood control and agricultural irrigation. However, hydrological model performance varies significantly across basins due to differences in climatic and anthropogenic conditions. Thus, this study aims to conduct a systematic inter-model comparison and assess the effectiveness of advanced post-processing techniques in improving forecasting precision. A number of representative watersheds with various hydrological features are chosen. A comprehensive assessment of three lumped conceptual hydrological models is conducted, including the hydrological model ecole de technologie supérieure (HMETS), the génie rural à 4 paramètres journalier (GR4J), and the Xinanjiang (XAJ) model. Model performance is evaluated using a multi-metric framework encompassing the Kling-Gupta efficiency (E_KG), Nash-Sutcliffe efficiency (E_NS), percent bias (P_BIAS), and normalized root mean square error (E_NRMS). Furthermore, three distinct optimization strategies are implemented and evaluated: error correction using a long short-term memory network; multi-model coupling using the Bayesian model averaging (BMA) method; and multi-model coupling using a long short-term memory (LSTM) network. Significant differences in simulation accuracy were observed between the three models across different watersheds. The XAJ model demonstrated a unique strength in simulating flood periods, while the HMETS model generally performed best. By contrast, the GR4J model exhibited limitations in capturing complex hydrological processes due to its simple structure. This suggested that the choice of a hydrological model should be context-dependent, influenced by the watershed's characteristics and data availability. It was also noted that lumped models struggle to accurately simulate regions undergoing Karst development or experiencing substantial anthropogenic impacts. Nevertheless, they can achieve satisfactory results in watersheds characterized by highly fluctuating flow regimes. An LSTM-based error correction method was applied to improve hydrological simulations. Repeated experiments were performed to compare simulation performance before and after error correction, assessing its effectiveness and stability. The LSTM-based error-correction method effectively improved the accuracy of all models during the calibration period. However, its validation performance proved highly dependent on both data quantity and watershed characteristics. As a result, it generated stable optimization in natural watersheds with longer records, but risked degrading performance in data-scarce or heavily impacted basins. Among the multi-model coupling methods, the BMA method was most effective when applied to individual models with higher accuracy (i.e., those withE_KG > 0.87), enhancing the coupled forecast accuracy. However, because static weights were used in the BMA method, it was unable to accurately capture dynamic variations across seasons or hydrological conditions. In contrast, the LSTM-based coupling approach, which can extract dynamic error features from multi-source model outputs, is mainly suitable for watersheds with sufficient data. Longer training data sequences facilitate the LSTM model's ability to learn the error patterns present in the streamflow predictions generated by the lumped hydrological models. Furthermore, the effectiveness of LSTM coupling was closely linked to the simulation performance during flood periods. The performance of the LSTM coupling optimization is also connected to the accuracy of lumped models during the validation period. Therefore, the LSTM coupling process must be carefully managed to prevent error propagation between the models. Effective multi-model coupling forecasting depends on strategic trade-offs and scenario-specific optimization. This study empirically validated the viability and potential of multi-model optimization frameworks for streamflow forecasting. The findings underscore that both model selection and post-processing techniques are highly context-specific, shaped by hydrological complexity, data constraints, and anthropogenic influences. Future research should prioritize several directions: (1) deepening the understanding of uncertainties in data-driven components like the long short-term memory network; (2) expanding the application of distributed and semi-distributed hydrological models to better represent spatial heterogeneity; and (3) developing physically coupled approaches that integrate deep learning with hydrological modeling principles. Furthermore, the use of multi-criteria decision-making frameworks for model selection and optimization is strongly advised to avoid the biases introduced by single-metric methods.

关键词

径流预测 / 集总式水文模型 / 误差优化 / 多模型耦合 / 模型比对

Key words

runoff prediction / lumped hydrological model / error correction / multi-model ensemble / model comparison

参考文献

[1] 赵人俊. 降雨径流流域模型中的水源划分问题[J]. 水文, 1981(3): 27-30. DOI: 10.19797/j.cnki.1000-0852.1981.03.005.

[2] Sezen C, Šraj M. Improving the simulations of the hydrological model in the Karst catchment by integrating the conceptual model with machine learning models[J]. Science of The Total Environment, 2024, 926： 171684. DOI: 10.1016/j.scitotenv.2024.171684.

[3] Ayele G T, Yu Bofu. Comparative analysis of lumped and semi-distributed hydrological models for an upland watershed in Ethiopia[J]. Journal of Hydrology: Regional Studies, 2025, 60： 102486. DOI: 10.1016/j.ejrh.2025.102486.

[4] Collenteur R A, Moeck C, Schirmer M, et al. Analysis of nationwide groundwater monitoring networks using lumped-parameter models[J]. Journal of Hydrology, 2023, 626： 130120. DOI: 10.1016/j.jhydrol.2023.130120.

[5] 张润庆, 何蒙, 姚成, 等. 新安江-海河模型在烟台市典型流域的径流模拟[J]. 南水北调与水利科技（中英文）, 2020, 18(6): 78-84. DOI: 10.13476/j.cnki.nsbdqk.2020.0117.

[6] 周妍, 白国营, 赵洪岩, 等. 分布式地表水-地下水耦合数值模型研究进展[J]. 南水北调与水利科技（中英文）, 2023, 21(3): 435-446. DOI: 10.13476/j.cnki.nsbdqk.2023.0044.

[7] 田辽西, 覃华清. 基于多变量 LSTM 神经网络的地下水位预测方法研究[J]. 干旱区资源与环境, 2024, 38(9): 138-146. DOI: 10.13448/j.cnki.jalre.2024.192.

[8] 廖如婷, 徐宗学, 叶陈雷, 等. 典型 LID 设施多尺度径流调控效应研究[J]. 水利水电技术（中英文）, 2025, 56(3): 61-75. DOI: 10.13928/j.cnki.wrahe.2025.03.005.

[9] 石彬, 李巧玲, 王越, 等. 考虑暴雨中心的子流域划分在洪水预报中的应用[J]. 南水北调与水利科技（中英文）, 2022, 20(5): 944-952. DOI: 10.13476/j.cnki.nsbdqk.2022.0094.

[10] 张挺, 马睿佳, 程泳铭, 等. 基于 VMD-LSTM 模型的流域径流预报技术[J]. 南水北调与水利科技（中英文）, 2025, 23(5): 1196-1203. DOI: 10.13476/j.cnki.nsbdqk.2025.0119.

[11] Thalli Mani S, Kolluru V, Amai M, et al. Enhanced streamflow simulations using nudging based optimization coupled with data-driven and hydrological models[J]. Journal of Hydrology: Regional Studies, 2022, 43： 101190. DOI: 10.1016/j.ejrh.2022.101190.

[12] Cao Chen, Ying Miaomiao. Comparative study of daily streamflow prediction based on coupling SWAT+ with interpretable machine learning algorithms[J]. Ecological Informatics, 2025, 91： 103406. DOI: 10.1016/j.ecoinf.2025.103406.

[13] Guo Jun, Liu Yi, Zou Qiang, et al. Study on optimization and combination strategy of multiple daily runoff prediction models coupled with physical mechanism and LSTM[J]. Journal of Hydrology, 2023, 624： 129969. DOI: 10.1016/j.jhydrol.2023.129969.

[14] Ishida K, Ercan A, Nagasato T, et al. Use of one-dimensional CNN for input data size reduction in LSTM for improved computational efficiency and accuracy in hourly rainfall-runoff modeling[J]. Journal of Environmental Management, 2024, 359： 120931. DOI: 10.1016/j.jenvman.2024.120931.

[15] Kapoor A, Pathiraja S, Marshall L, et al. Deep GR4J: A deep learning hybridization approach for conceptual rainfall-runoff modelling[J]. Environmental Modelling & Software, 2023, 169： 105831. DOI: 10.1016/j.envsoft.2023.105831.

[16] Koutsoyiannis D, Montanari A. Bluecat: A local uncertainty estimator for deterministic simulations and predictions[J]. Water Resources Research, 2022, 58(1）: e2021WR031215. DOI: 10.1029/2021WR031215.

[17] Okkan U, Ersoy Z B, Ali Kumanlioglu A, et al. Embedding machine learning techniques into a conceptual model to improve monthly runoff simulation: A nested hybrid rainfall-runoff modeling[J]. Journal of Hydrology, 2021, 598： 126433. DOI: 10.1016/j.jhydrol.2021.126433.

[18] Xu Wenxin, Chen Jie, Corzo G, et al. Coupling deep learning and physically based hydrological models for monthly streamflow predictions[J]. Water Resources Research, 2024, 60(2）: e2023WR035618. DOI: 10.1029/2023WR035618.

[19] Valdez E S, Anctil F, Ramos M H. Choosing between post-processing precipitation forecasts or chaining several uncertainty quantification tools in hydrological forecasting systems[J]. Hydrology and Earth System Sciences, 2022, 26(1): 197-220. DOI: 10.5194/hess-26-197-2022.

[20] Zubelzu S, Ghalkha A, Ben issaid C, et al. Coupling machine learning and physical modelling for predicting runoff at catchment scale[J]. Journal of Environmental Management, 2024, 354： 120404. DOI: 10.1016/j.jenvman.2024.120404.

[21] Schaumann P, Hess R, Rempel M, et al. A calibrated and consistent combination of probabilistic forecasts for the exceedance of several precipitation thresholds using neural networks[J]. Weather and Forecasting, 2021, 36(3): 1079-1096. DOI: 10.1175/WAF-D-20-0188.1.

[22] 陈晓宏, 颜依寒, 李诚, 等. 溶蚀丘陵型岩溶流域概念性水文模型及其应用[J]. 水科学进展, 2020, 31(1): 1-9. DOI: 10.14042/j.cnki.32.1309.2020.01.001.

[23] 刘伟, 王世杰, 罗维均. 贵州荔波岩溶峰丛区表层岩溶泉对大气降雨的响应及其指示意义[J]. 地球化学, 2011, 40(5): 487-496. DOI: 10.19700/j.0379-1726.2011.05.009.

[24] 韩瑞光, 冯平. 流域下垫面变化对洪水径流影响的研究[J]. 干旱区资源与环境, 2010, 24(8): 27-30. DOI: 10.13448/j.cnki.jalre.2010.08.020.

[25] 郝偌楠, 刘登嵩, 孙英军, 等. 基于概率分布模型的流量预报及参数动态识别[J]. 水力发电学报, 2020, 39(6): 39-48. DOI: 10.11660/slfdxb.20200604.

[26] 汤佳, 胡希军, 韦宝婧, 等. 基于 FLUS 模型的流域土地利用变化预测及水文响应评估[J]. 长江科学院院报, 2022, 39(4): 63-69. DOI: 10.11988/ckyyb.20210182.

[27] 傅春, 毛安琪. 抚河流域径流对土地利用变化时空响应[J]. 长江流域资源与环境, 2021, 30(2): 342-350. DOI: 10.11870/cjlyzyyhj202102009.

[28] 王真, 姜三元, 张奇, 等. 赣江流域水文情势时空差异性及变化分析[J]. 人民长江, 2024, 55(3): 105-116. DOI: 10.16232/j.cnki.1001-4179.2024.03.015.

[29] 张丹, 要威, 肖华, 等. 基于因果分析的金沙江流域气象-水文干旱传播特征及其未来情景预估[J]. 水资源保护, 2025, 41(2): 158-164. DOI: 10.3880/j.issn.1004-6933.2025.02.018.

[30] Gu Lei, Yin Jiabo, Wang Sheng, et al. How well do the multi-satellite and atmospheric reanalysis products perform in hydrological modelling[J]. Journal of Hydrology, 2023, 617： 128920. DOI: 10.1016/j.jhydrol.2022.128920.

[31] Yang Jing, Chen Fulong, Long Aihua, et al. Runoff simulation of the Kaidu River basin based on the GR4J-6 and GR4J-6-LSTM models[J]. Journal of Hydrology: Regional Studies, 2024, 56： 102034. DOI: 10.1016/j.ejrh.2024.102034.

[32] Yi Bin, Liu Huiyuan, Chen Lu, et al. Stepwise calibration of a lumped hydrological model for Xun River basin, China, considering the heterogeneity of soil moisture[J]. Journal of Hydrology: Regional Studies, 2025, 58： 102293. DOI: 10.1016/j.ejrh.2025.102293.

[33] Fowler K, Peel M, Western A, et al. Improved rainfall-runoff calibration for drying climate: Choice of objective function[J]. Water Resources Research, 2018, 54(5): 3392-3408. DOI: 10.1029/2017WR022466.

[34] Lamontagne J R, Barber C A, Vogel R M. Improved estimators of model performance efficiency for skewed hydrologic data[J]. Water Resources Research, 2020, 56(9）: e2020WR02701. DOI: 10.1029/2020WR027101.

[35] 向鑫, 敖天其, 肖钦太. 典型参数优化算法在新安江模型中的应用对比[J]. 水文, 2023, 43(3): 16-22. DOI: 10.19797/j.cnki.1000-0852.20210445.

[36] Williams G P. Friends don't let friends use Nash-Sutcliffe efficiency (NSE) or KGE for hydrologic model accuracy evaluation: A rant with data and suggestions for better practice[J]. Environmental Modelling & Software, 2025: 106665. DOI: 10.1016/j.envsoft.2025.106665.

[37] Hu Yuqian, Li Heng, Zhang Chunxiao, et al. Investigate the rainfall-runoff relationship and hydrological concepts inside LSTM[J]. Environmental Modelling & Software, 2025, 192： 106527. DOI: 10.1016/j.envsoft.2025.106527.

[38] He Shaokun, Chen Kebing, Liu Zhangjun, et al. Exploring the impacts of climate change and human activities on future runoff variations at the seasonal scale[J]. Journal of Hydrology, 2023, 619： 129382. DOI: 10.1016/j.jhydrol.2023.129382.

[39] Lan Tian, Du Xinyue, Xie Xue, et al. Exploring controls on precipitation-runoff dependencies: Implications for non-stationary and spatially heterogeneous analyses[J]. Journal of Hydrology, 2025, 659： 133333. DOI: 10.1016/j.jhydrol.2025.133333.

[40] 杨兴豪, 刘晓阳, 赵丽平, 等. 集合预报降雨信息在寸滩−三峡区间的精度评价和可利用性评估[J]. 南水北调与水利科技（中英文）, 2024, 22(增刊 1): 17-22.

[41] Xu Yue, Tang Guoqiang, Zhu Siyu, et al. Optimization of feature inputs in machine learning-based multi-source precipitation merging[J]. Journal of Hydrology, 2025, 663： 134185. DOI: 10.1016/j.jhydrol.2025.134185.

[42] Song J H, Kim Hakkwan. Impact of precipitation station density and homogeneous regions on hydrological modeling performance[J]. Journal of Hydrology, 2025, 661： 133618. DOI: 10.1016/j.jhydrol.2025.133618.

[43] Jin Chen, Xu Qin, Chen Xing, et al. Enhancing runoff prediction accuracy of deep learning model using baseflow separation method and timestamp information[J]. Journal of Hydrology, 2025, 662： 134044. DOI: 10.1016/j.jhydrol.2025.134044.

[44] Hunt K M, Matthews G R, Pappenberger F, et al. Using a long short-term memory (LSTM) neural network to boost river streamflow forecasts over the western United States[J]. Hydrology and Earth System Sciences, 2022, 26(21): 5449-5472. DOI: 10.5194/hess-26-5449-2022.

基金资助

湖南省水利科技重大项目(XSKJ2024064-2)

湖南省水利科技项目(XSKJ2024064-19)