In-Flow Forecasting for the Gezhen Reservoir Guided by Physical Mechanisms

Zhang Yonghui; Wang Jian; Liu Taiwen; Luo Qiyao; Huanghanliang; Zhang Jingwen; Wu Chengjun; Lai Zhiqiang

doi:null

English Version Journal Login

您当前的位置：

首页 >

文章列表页 >

In-Flow Forecasting for the Gezhen Reservoir Guided by Physical Mechanisms

更新时间：2026-09-28

- In-Flow Forecasting for the Gezhen Reservoir Guided by Physical Mechanisms
- PEARL RIVER Pages: 1-13(2026)
- 作者机构：
  
  1.华能海南发电股份有限公司东方电厂，海南东方 572600
  2.中山大学土木工程学院，广东广州 510275
  3.黄河水利委员会黄河水利科学研究院，河南郑州 450003
- 作者简介：
- 基金信息：
- DOI：
  CLC： TV87
- Received：15 July 2026，
  
  Revised：2026-09-09，
  
  Accepted：20 September 2026，
  
  Online First：28 September 2026，
- 稿件说明：
移动端阅览
张永辉，王健，刘泰文等.物理机制引导机器学习的戈枕入库洪水预报［J］.人民珠江，DOI：10.3969/j.issn.1001-9235.XXXX.XX.001.

Zhang Yonghui,Wang Jian,Liu Taiwen,et al.In-Flow Forecasting for the Gezhen Reservoir Guided by Physical Mechanisms[J].PEARL RIVER,
张永辉，王健，刘泰文等.物理机制引导机器学习的戈枕入库洪水预报［J］.人民珠江，DOI：10.3969/j.issn.1001-9235.XXXX.XX.001. DOI：

Zhang Yonghui,Wang Jian,Liu Taiwen,et al.In-Flow Forecasting for the Gezhen Reservoir Guided by Physical Mechanisms[J].PEARL RIVER, DOI：10.3969/j.issn.1001-9235...001.

摘要

针对热带季风区复杂洪水过程中传统新安江模型易产生系统偏差的问题，旨在提出一种融合物理机制与机器学习残差修正的戈枕水库入库洪水预报方法，以提高预报精度并增强模型的可解释性。研究选取60场洪水，采用PSO算法率定三水源新安江模型，以2024年前44场洪水训练、2024年及以后16场洪水独立验证；利用XGBoost学习历史残差、降雨和上游出库流量与模型误差的非线性关系，并开展模型对比、消融试验和SHAP分析。修正后验证期平均NSE达0.958，单场NSE为0.82～0.98，洪峰与洪量相对误差多控制在±10%以内；残差MAE和RMSE分别由130.49 m³/s和193.25 m³/s降至84.51 m³/s和147.92 m³/s。历史残差为最关键输入，降雨提供补充信息。新安江模型与XGBoost残差修正框架可有效降低系统偏差，提高复杂洪水预报精度与稳定性，可为水库实时预报调度提供技术支撑。

Abstract

To address the systematic biases of the conventional Xin'anjiang model in forecasting complex flood events in tropical monsoon regions

this study proposes a physics-guided machine-learning framework for inflow flood forecasting at Gezhen Reservoir

aiming to improve forecasting accuracy and model interpretability. A total of 60 flood events were selected. The three-component Xin'anjiang model was calibrated using particle swarm optimization (PSO). Forty-four flood events before 2024 were used for model calibration and training

while 16 events occurring in and after 2024 were used for independent validation. XGBoost was employed to predict the residual errors of the Xin'anjiang model using model outputs

historical residuals

rainfall

and upstream reservoir outflow as input features. The predicted residuals were added to the Xin'anjiang model simulations to obtain the corrected inflow forecasts. Model comparison

input-feature ablation experiments

residual analysis

and SHAP analysis were conducted to evaluate model performance and interpret the contribution of different information sources. After residual correction

the mean Nash-Sutcliffe efficiency coefficient (NSE) during the validation period reached 0.958

with event-specific NSE values ranging from 0.82 to 0.98. The relative errors of flood peak discharge and flood volume were mostly within ±10%. The mean absolute error (MAE) and root mean square error (RMSE) of the residuals decreased from 130.49 m³/s and 193.25 m³/s to 84.51 m³/s and 147.92 m³/s

respectively. Ablation experiments showed that historical residuals were the most important inputs for error correction

whereas rainfall and upstream reservoir outflow alone provided limited improvement but contributed additional information when combined with residual features. SHAP analysis further indicated that recent residuals contributed most strongly to model predictions

reflecting the short-term persistence of hydrological-model errors.The Xin'anjiang model coupled with XGBoost residual correction effectively combines physical-process representation with nonlinear data-driven error learning. The proposed framework reduces systematic bias and improves the accuracy and stability of complex flood forecasting

while retaining reasonable interpretability. It can provide technical support for real-time reservoir inflow forecasting and flood-control operation in tropical monsoon regions.

关键词

Keywords

references

李峰平，章光新，董李勤 . 气候变化对水循环与水资源的影响研究综述［J］. 地理科学， 2013 ， 33 （ 4 ）： 457 - 464 .

Li Fengping ， Zhang Guangxin ， Dong Liqin . Review of the impacts of climate change on the water cycle and water resources ［J］. Scientia Geographica Sinica ， 2013 ， 33 （ 4 ）： 457 - 464 . （in Chinese）

秦拓，孙子茱，宋馨媛 . 基于流溪河模型的巴家咀水库入库洪水预报［J］. 陕西水利， 2025 （ 9 ）： 29 - 32 .

Qin Tuo ， Sun Zizhu ， Song Xinyuan . Inflow flood forecasting for Bajiazui Reservoir based on the Liuxihe model ［J］. Shaanxi Water Resources ， 2025 （ 9 ）： 29 - 32 . （in Chinese）

苑瑞芳，廖卫红，康爱卿，等 . 基于新安江模型的台风期间福州山区水库入库洪水预报［J］. 水资源与水工程学报， 2020 ， 31 （ 3 ）： 24 - 29 .

Yuan Ruifang ， Liao Weihong ， Kang Aiqing ， et al . Inflow flood forecasting for mountainous reservoirs in Fuzhou during typhoon periods based on the Xin’anjiang model ［J］. Journal of Water Resources and Water Engineering ， 2020 ， 31 （ 3 ）： 24 - 29 . （in Chinese）

陈奕帆，张靖文，李泽君，等 . 基于半分布式GUH模型的台风期水库入库径流模拟［J/OL］. 人民长江， 2026 ［ 2026-07-19 ］.

Chen Yifan ， Zhang Jingwen ， Li Zejun ， et al . Reservoir inflow runoff simulation during typhoon periods based on a semi-distributed GUH model ［J/OL］. Yangtze River ， 2026 ［ 2026-07-19 ］. （in Chinese）

官学文 . 长江水文预报现代化建设的实践与思考［J］. 人民长江， 2025 ， 56 （ 10 ）： 151 - 156 .

Guan Xuewen . Practice and reflections on the modernization of hydrological forecasting in the Yangtze River ［J］. Yangtze River ， 2025 ， 56 （ 10 ）： 151 - 156 . （in Chinese）

赵丽平，任明磊，刘昌军，等 . 小型水库洪水预报模型研究及应用［J］. 中国防汛抗旱， 2024 ， 34 （ 4 ）： 77 - 82 .

Zhao Liping ， Ren Minglei ， Liu Changjun ， et al . Research and application of flood forecasting models for small reservoirs ［J］. China Flood & Drought Management ， 2024 ， 34 （ 4 ）： 77 - 82 . （in Chinese）

黄泽禧，孙伟，陈新林，等 . 耦合物理模型与深度学习的流量预测方法［J］. 人民珠江， 2025 ， 46 （ 5 ）： 51 - 62 .

Huang Zexi ， Sun Wei ， Chen Xinlin ， et al . A flow forecasting method coupling a physical model with deep learning ［J］. Pearl River ， 2025 ， 46 （ 5 ）： 51 - 62 . （in Chinese）

何禹阳，冯星昱，徐源浩，等 . 耦合可解释机器学习的水文模型在洪水模拟误差校正中的应用［J］. 人民珠江， 2026 ， 47 （ 5 ）： 21 - 33 .

He Yuyang ， Feng Xingyu ， Xu Yuanhao ， et al . Application of a hydrological model coupled with interpretable machine learning to error correction in flood simulation ［J］. Pearl River ， 2026 ， 47 （ 5 ）： 21 - 33 . （in Chinese）

梁丽莎，蒋晓蕾，胡乐怡，等 . 基于XGBoost与SWAT径流模拟对比及校正研究［J］. 人民珠江， 2025 ， 46 （ 1 ）： 50 - 58 .

Liang Lisha ， Jiang Xiaolei ， Hu Leyi ， et al . Comparison and correction of runoff simulations based on XGBoost and SWAT ［J］. Pearl River ， 2025 ， 46 （ 1 ）： 50 - 58 . （in Chinese）

朱畅畅，黄华平 . 基于VIC-BJP模型的黄河源区水文模拟及误差校正研究［J］. 人民珠江， 2026 ， 47 （ 1 ）： 112 - 119 .

Zhu Changchang ， Huang Huaping . Hydrological simulation and error correction in the source region of the Yellow River based on the VIC-BJP model ［J］. Pearl River ， 2026 ， 47 （ 1 ）： 112 - 119 . （in Chinese）

赵人俊，王佩兰 . 新安江模型参数的分析［J/OL］. 水文， 1988 （ 6 ）： 2 - 9 .

Zhao Renjun ， Wang Peilan . Analysis of the parameters of the Xin’anjiang model ［J/OL］. Journal of China Hydrology ， 1988 （ 6 ）： 2 - 9 . （in Chinese）

代堂刚，宋昭义，谢开荣 . 基于马斯京根及新安江模型的应用研究［J］. 环境科学与技术， 2016 ， 39 （ S2 ）： 470 - 473 .

Dai Tanggang ， Song Zhaoyi ， Xie Kairong . Application study based on the Muskingum and Xin’anjiang models ［J］. Environmental Science & Technology ， 2016 ， 39 （ S2 ）： 470 - 473 . （in Chinese）

王岁花，冯乃勤，李爱国 . 一类新颖的粒子群优化算法［J］. 计算机工程与应用， 2003 （ 13 ）： 109 - 110， 134 .

Wang Suihua ， Feng Naiqin ， Li Aiguo . A novel particle swarm optimization algorithm ［J］. Computer Engineering and Applications ， 2003 （ 13 ）： 109 - 110， 134 . （in Chinese）

Ammann L ， Fenicia F ， Reichert P . A likelihood framework for deterministic hydrological models and the importance of non-stationary autocorrelation ［J］. Hydrology and Earth System Sciences ， 2019 ， 23 （ 4 ）： 2147 - 2172 .

Li Ming ， Wang Qianjin ， Bennett J C ， et al . A strategy to overcome adverse effects of autoregressive updating of streamflow forecasts ［J］. Hydrology and Earth System Sciences ， 2015 ， 19 （ 1 ）： 1 - 15 .

Liu Jun ， Koch J ， Stisen S ， et al . A national-scale hybrid model for enhanced streamflow estimation–consolidating a physically based hydrological model with long short-term memory （LSTM） networks ［J］. Hydrology and Earth System Sciences ， 2024 ， 28 （ 13 ）： 2871 - 2893 .

Li Dayang ， Marshall L ， Liang Zhongmin ， et al . Characterizing distributed hydrological model residual errors using a probabilistic long short-term memory network ［J］. Journal of Hydrology ， 2021 ， 603 ： 126888 .

Views

下载量

CSCD

Alert me when the article has been cited

提交

Tools

Publicity Resources

Application of Fusion Neural Network Based on XGBoost Correction in Precipitation Prediction

Comparison and Correction of Runoff Simulation Based on XGBoost and SWAT

Study on Flood Simulation and Its Real-time Correction of Qinhuai River Basin Based on Grid-Xin'anjiang Model

通什水幸福河湖建设与流域水生态环境修复研究——以五指山市南圣河为例

库区水面光伏物理-LSTM协同预警算法模型

Related Author

ZHANG Daiqing

ZHU Yuxiao

FU Xiaolei

ZHANG Liping

SHEN Jing

HU Leyi

JIANG Xiaolei

LIANG Lisha

Related Institution

Faculty of Electric Power Engineering, Kunming University of Science and Technology

College of Hydraulic Science and Engineering, Yangzhou University

College of Hydrology and Water Resources, Hohai University

Hydrology Monitor and Forecast Center,Ministry of Water Resources of China

AI问答

⁰