在气候变化与快速城镇化背景下，城市暴雨内涝风险持续加剧。现有机理模型虽具有较好的物理基础，但在高分辨率、多情景推演条件下计算效率受限，纯数据驱动模型在深水区预测中亦存在稳定性不足问题。针对上述问题，以南京市主城区为研究对象，基于水文水动力耦合机理模型生成72组降雨潮位情景样本，提出融合混合损失函数和物理非负约束的Flood LSTM模型。结果表明，在相同硬件条件下，Flood LSTM模型平均耗时为9.31 s，仅为本研究中机理模型计算耗时的0.50%；测试集的RMSE、NSE和Top5_MAE分别为0.010 8、0.991 8、0.019 7 m，深水区（水深≥1.00 m）NSE达到0.946 8，多数场景CSI>0.85，模型准确性较高。SHAP模型分析表明，最大雨强变化率和平均落潮率是关键驱动因子，并揭示了南京内涝“雨-潮-雨”风险逐步增大的三阶段分层驱动特征。研究结果丰富了城市洪涝致灾机理理论，可为城市洪涝快速预警与精细化调度提供技术支撑。

[Objective] The integration of mechanism-based models and deep learning emerges as a pivotal direction in urban flood prediction. While mechanism-based models possess a robust physical foundation，their computational efficiency is often constrained under high-resolution，multi-scenario inference conditions. Conversely，purely data-driven models frequently suffer from instability in predicting deep-water inundation. Focusing on the compound scenario of urban rainfall and high Yangtze River water levels，this study aims to synergistically validate the prediction accuracy，physical rationality，and driving mechanisms of a deep learning model. It seeks to deepen the understanding of the contributions and interactions of rainfall and tidal factors across varying risk levels. [Methods] Taking the main urban area of Nanjing as the study site，this research proceeds along two main lines： “accelerated computation” and “mechanism interpretation.” First，a coupled hydrological-hydrodynamic mechanism model was constructed，calibrated，and validated to generate 72 scenarios of rainfall-tide combinations. Second，a Flood Long Short-Term Memory （Flood LSTM） network，incorporating a hybrid loss function and physical non-negative constraints，was proposed and systematically compared against various machine learning and deep learning baselines. Finally，the SHapley Additive exPlanations （SHAP） framework was introduced to identify the contributions of rainfall and tidal drivers and their evolutionary characteristics across different risk levels. [Results] The Flood LSTM model effectively reproduces spatial inundation patterns under typical rain-tide scenarios with significantly improved computational efficiency. Under identical hardware conditions，the average inference time was 9.31 s，representing only 0.50% of the computational cost of the mechanism model. The model demonstrated high accuracy，with test set RMSE，NSE，and Top-5 MAE values of 0.0108 m，0.9918 and 0.0197 m，respectively. The NSE for deep-water zones （≥1.00 m） reached 0.946 8，and the Critical Success Index （CSI） exceeded 0.85 in most scenarios. SHAP analysis revealed a distinct compound rain-tide driving characteristic in the formation of urban waterlogging. The maximum rainfall intensity change rate （RChg） and the average ebb tide rate （TFAv） were identified as key drivers. The evolution of waterlogging in Nanjing generally follows a three-stage stratified driving pattern of “Rain-Tide-Rain，” characterized by progressively increasing risk. [Conclusion] The proposed Flood LSTM model exhibits high precision in the rapid prediction of maximum urban flood inundation depths. These findings enrich the theoretical understanding of urban flood disaster mechanisms and provide a reference for rapid early warning，engineering dispatch，and refined risk management. While the model shows promise，limitations remain regarding its reliance on simulated training data and its focus on static maximum depth prediction. Future work will incorporate more field observations and dynamic process data to further enhance model applicability.