中国水利

针对华北半干旱半湿润流域洪水预报中传统概念模型洪峰低估、数据驱动模型物理约束不足的问题，以海河流域大清河水系之阜平流域为研究对象，构建了新安江模型（XAJ）与双向长短期记忆网络（BiLSTM）耦合的洪水预报残差修正模型，并结合SHAP方法揭示模型补偿机理。以XAJ输出的预报流量、三层土壤含水量及径流分量等物理变量为输入，采用递进式特征组合构建BiLSTM残差修正模型，系统评估不同物理信息对预报效果的影响。结果表明：XAJ在阜平流域存在明显洪峰低估和时序偏差；将XAJ输出的基准预报流量作为先验特征输入BiLSTM后，混合模型整体拟合效果显著提升，且物理边界约束可有效抑制纯数据驱动模型在极端条件下的数值失真；进一步引入三层土壤含水量后，模型洪峰捕捉能力最优，平均洪峰相对误差控制在10%以内。SHAP归因分析表明，土壤水分状态变量是驱动残差正向补偿的关键因子，径流分量在本研究样本条件下未表现出稳定增益，其附加信息可能因误差传播和信息冗余削弱模型泛化能力。研究表明，基于关键物理状态约束的XAJ-BiLSTM混合框架能够有效提高北方复杂流域洪峰预报精度，并为物理-数据融合水文模型的可解释构建提供参考。

To address the issues of flood forecasting for flood-peak underestimation in traditional conceptual models and insufficient physical constraints in data-driven models for semi-arid and semi-humid catchments in northern China, a residual correction model for flood forecasting was developed by coupling the Xin'anjiang model(XAJ) with a bidirectional long short-term memory network(BiLSTM) for the Fuping Catchment in the Daqing River subsystem of the Haihe River basin. The SHAP method was introduced to interpret the model's compensation mechanism. Physical variables derived from XAJ, including forecasted discharge, three-layer soil moisture, and runoff components, were used as inputs. A progressive feature combination approach was adopted to construct the BiLSTM residual correction model and to evaluate the effects of different physical information on forecasting performance. The results indicate that XAJ exhibits noticeable flood-peak underestimation and temporal deviation in the Fuping catchment. When the baseline forecasted discharge from XAJ is used as a prior input to BiLSTM, the hybrid model achieves significantly improved overall fitting performance, and the physically constrained framework effectively suppresses numerical distortions in purely data-driven models under extreme conditions. After further incorporating the three-layer soil moisture, the model captures flood peaks most accurately, with the mean relative error of flood peaks reduced to within 10%. SHAP attribution analysis reveals that soil moisture state variables are key drivers of positive residual compensation. Runoff component variables do not provide stable additional gains under the current sample conditions, and their inclusion may weaken the model's generalization ability due to error propagation and information redundancy. The research reveals that the proposed XAJ-BiLSTM hybrid framework, constrained by key physical state variables, effectively improves flood-peak forecasting accuracy in complex catchments in northern China and provides a reference for the interpretable construction of physics-data fusion hydrological models.