Scholarlink

<p>To address rapid forecasting of urban pluvial flooding in dike-ringed plain cities, this study develops for central Changshu an integrated framework that couples a database-backed complex river-network representation, 1D/2D hydrodynamics, and deep-learning surrogates. A topology library, boundary library, and operations library standardize pipes, channels, and control structures. Runoff generation and overland routing are described by Horton infiltration and a nonlinear reservoir scheme; hydraulics are solved via a 1D Saint-Venant–2D shallow-water coupling, and a node-based formulation ensures mass-conserving 1D–2D exchange. Pump–gate operations are replayed from rule sets based on water-level thresholds, interior–exterior head differences, multi-condition triggers, pump–gate coupling, and discharge constraints. Multiple historical and design storm events are simulated to produce high-fidelity reference datasets, and particle swarm optimization (PSO) is used to identify key parameters. Two surrogates are trained: a long short-term memory (LSTM) network for water-depth time series at critical points and a convolutional neural network (CNN) for city-scale maps of maximum water depth, with physical metadata channels—normalized coordinates, pump/gate proximity, and inlet/pipe density—concatenated to the input tensors. On independent tests, performance reaches Nash–Sutcliffe efficiency &gt; 0.90, intersection-over-union ≈ 0.87, and mean depth error ≈ 0.04&#xa0;m. Inference is three to four orders of magnitude faster than mechanistic simulation (24-h event: LSTM ≈ 0.05&#xa0;s, CNN &lt; 1&#xa0;s). We also report Monte Carlo dropout uncertainty bands and ablation results demonstrating that operation rules and physical metadata improve boundary delineation and model generalization.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Rapid Urban Flood Simulation and Prediction Using Integrated Hydrodynamic Modeling and Deep Learning Approaches

  • 期刊论文

  • 发表日期：

    2026年1月10日

查看全文
  • Houlin Chen,
  • Jiahu Wang

摘要

To address rapid forecasting of urban pluvial flooding in dike-ringed plain cities, this study develops for central Changshu an integrated framework that couples a database-backed complex river-network representation, 1D/2D hydrodynamics, and deep-learning surrogates. A topology library, boundary library, and operations library standardize pipes, channels, and control structures. Runoff generation and overland routing are described by Horton infiltration and a nonlinear reservoir scheme; hydraulics are solved via a 1D Saint-Venant–2D shallow-water coupling, and a node-based formulation ensures mass-conserving 1D–2D exchange. Pump–gate operations are replayed from rule sets based on water-level thresholds, interior–exterior head differences, multi-condition triggers, pump–gate coupling, and discharge constraints. Multiple historical and design storm events are simulated to produce high-fidelity reference datasets, and particle swarm optimization (PSO) is used to identify key parameters. Two surrogates are trained: a long short-term memory (LSTM) network for water-depth time series at critical points and a convolutional neural network (CNN) for city-scale maps of maximum water depth, with physical metadata channels—normalized coordinates, pump/gate proximity, and inlet/pipe density—concatenated to the input tensors. On independent tests, performance reaches Nash–Sutcliffe efficiency > 0.90, intersection-over-union ≈ 0.87, and mean depth error ≈ 0.04 m. Inference is three to four orders of magnitude faster than mechanistic simulation (24-h event: LSTM ≈ 0.05 s, CNN < 1 s). We also report Monte Carlo dropout uncertainty bands and ablation results demonstrating that operation rules and physical metadata improve boundary delineation and model generalization.