中国农村水利水电Issue(6)：47-53,7.DOI:10.12396/znsd.2501133
基于LSTM的洪水预报残差修正方法研究
A Study on an LSTM-Based Residual Correction Method for Flood Forecasting
摘要
Abstract
Flood forecasting for small mountainous catchments presents a significant challenge within hydrological science.Conceptual hydrological models,typified by the Xin'anjiang(XAJ)model,often struggle to accurately capture the intricate runoff generation and routing processes in mountainous terrain due to their simplified structures,frequently resulting in suboptimal forecasting accuracy.While data-driven models,such as Long Short-Term Memory(LSTM)networks,exhibit powerful fitting capabilities,their inherent"black-box"nature provides limited support from physical mechanisms.To overcome these limitations,this study develops a hybrid physics-data model that leverages an LSTM network to correct the forecast residuals of the XAJ model and incorporates SHAP for an interpretable attribution analysis.The model was validated using 15 flood events from 2015 to 2018 in the Qiaodong Village catchment,a typical small mountainous watershed in Zhejiang Province.Results indicate the following:① All XAJ-LSTM hybrid models demonstrated a significant improvement over the standalone XAJ model,which served as a baseline.The configuration that integrated the XAJ-forecasted discharge and soil moisture state variables yielded the optimal performance,with the Nash-Sutcliffe efficiency coefficient markedly increasing from 0.55 to 0.77.② The introduction of more information does not necessarily enhance model performance;supplementing the optimal configuration with additional runoff component data resulted in performance degradation due to informational redundancy.③ From a mechanistic perspective,the SHAP analysis confirmed that antecedent observed discharge and XAJ-forecasted discharge are the key drivers influencing the residual correction model's decisions.This reveals that the integration of physical information effectively guides the data-driven model's learning process,shifting its focus from mere statistical fitting to reliance on more robust physical constraints.The proposed physics-data hybrid model,which combines high accuracy with strong interpretability,offers a new paradigm for the development of reliable and trustworthy intelligent hydrological forecasting models.关键词
径流模拟/新安江模型/长短期记忆神经网络/残差修正/SHAPKey words
streamflow simulation/Xin'anjiang(XAJ)model/LSTM/residual correction/SHAP分类
建筑与水利引用本文复制引用
崔晨璐,张轩,吴永祥,王高旭..基于LSTM的洪水预报残差修正方法研究[J].中国农村水利水电,2026,(6):47-53,7.基金项目
国家重点研发计划项目(2023YFC3006501) （2023YFC3006501）
南京水利科学研究院研究生学位论文创新基金项目(Yy525002) （Yy525002）
中央级公益性科研院所基本科研业务费专项资金重点基金项目(Y524008 （）
Y525015). （）
