Page 37 - 2025年第56卷第6期

［ 17］ 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）：
.
889-897，909.
［ 18］ 孙明博，闫宝伟，常建波，等 . 基于物理机制和深度学习的混合模型及应用研究［J］ 中国农村水利水电，
.
2024（8）：67-72，80.
［ 19］ 廖 耀 星 ， 高 玮 志 ， 张 轩 ， 等 . 基 于 深 度 学 习 和 水 动 力 模 型 的 洪 水 演 进 快 速 模 拟 方 法［J］ 中 国 防 汛 抗 旱 ，
.
2024，34（2）：16-22.
［ 20］ FUKUSHIMA K. Neocognitron： A self-organizing neural network model for a mechanism of pattern recognition unaf⁃
.
fected by shift in position［J］ Biological cybernetics，1980，36（4）：193-202.
.
［ 21］ HOCHREITER S，SCHMIDHUBER J. Long short-term memory［J］ Neural Computation，1997，9（8）：1735-1780.
［ 22］ CHO K， Van MERRIËNBOER B， GULCEHRE C， et al. Learning phrase representations using RNN encoder-
decoder for statistical machine translation［C］/Proceeding of the 2014 Conference on Empirical Methods in Natural
/
Language Processing （EMNLP） Doha，Qatar： Association for Computational Linguistics，2014.
.
［ 23］ KARIM F，ARMIN M，AHMEDT A D，et al. A review of hydrodynamic and machine learning approaches for flood
.
inundation modeling［J］ Water，2023，15（3）：566.
［ 24］ YU M，LIU X，WOOD P，et al. Simulation of Poyang Lake water levels and outflow under historical extreme hydro⁃
logical scenarios［J］ Journal of Flood Risk Management，2022，15（3）：e12806.
.
［ 25］ CUI C，DONG Z，LUO Y，et al. Lake impoundment in advance of post-flood period based on large-scale numeri⁃
.
cal simulation［J］ Journal of Water Resources Planning and Management，2023，149（8）：05023004.

Flood simulation and forecasting in the Yangtze River-Dongting Lake
Basin based on deep learning

2
1
3
1
1

LI Chenglong ，GUO Shenglian ，LIANG Zhiming ，CUI Zhen ，XIANG Xin ，ZHANG Jun 3

（1. State Key Laboratory of Water Resources Engineering and Management，Wuhan University，Wuhan 430072，China；

2. China Yangtze Power Co.，Ltd.，Wuhan 430010，China；3. The Hydrological Bureau of the Yangtze

River Water Resources Commission，Wuhan 430010，China）

Abstract：Accurate and fast flood forecasting in the Yangtze River-Dongting Lake basin is vitally important for flood

control operation of the Three Gorges Reservoir. Three deep learning models，namely the Convolutional Neural Net⁃
work （CNN），the Long Short-Term Memory neural network （LSTM），and the Gated Recurrent Unit （GRU），were

constructed. The recorded flow discharges and rainfalls of eight hydrological stations and 301 rain gauge stations

during flood season （from May to September）from 2010 to 2023 were used to calibrate and verify these models. The
results show that the flood simulation and forecasting accuracy of these deep learning models are all sufficiently high.
Among them，the GRU model performs the best，with the Nash-Sutcliffe Efficiency （NSE）in the 24-hour forecast

horizon reaching 0.993 during the training period and 0.988 during the testing period，and the total runoff relative

3
errors being 0.25% and -0.26%，respectively. For the 6 flood events with peak flows exceeding 20，000 m /s during

the testing period，the forecast results of the three models vary significantly. The GRU model is superior to LSTM and

CNN models，with the NSE values greater than 0.85，and the absolute errors of peak flow less than 2%. The GRU and

LSTM models have high simulation accuracy and strong generalization ability，which can provide a new approach for

complex flood forecasting in the Yangtze River-Dongting Lake Basin.

Keywords：hydrological forecasting；deep learning；convolutional neural network；Long Short-Term Memory neural

network；Gated Recurrent Unit；Yangtze River-Dongting Lake Basin

（责任编辑：耿庆斋）
— 725 —