Chen et al. (2026) Rapid Urban Flood Simulation and Prediction Using Integrated Hydrodynamic Modeling and Deep Learning Approaches

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-01-01
• Authors: Houlin Chen, Jiahu Wang
• DOI: 10.1007/s11269-025-04400-6

Research Groups

College of Hydrology and Water Resources, Hohai University, Nanjing, China

Short Summary

This study develops an integrated framework for rapid urban flood simulation and prediction in dike-ringed plain cities, coupling 1D/2D hydrodynamic modeling with deep learning surrogates (LSTM and CNN) to achieve high-fidelity, operationally relevant forecasts with significantly reduced computational time. The framework, applied to central Changshu, demonstrates accurate prediction of water depths and inundation extents, achieving speed-ups of 10^3–10^4 times compared to mechanistic models.

Objective

• To develop an integrated framework that couples database-backed complex river-network representation, 1D/2D hydrodynamics, and deep-learning surrogates for rapid forecasting of urban pluvial flooding in dike-ringed plain cities.

Study Configuration

• Spatial Scale: Central urban district of Changshu, China (60.7 km²), characterized by a dense interior river and sewer network (~150 interconnected channels) and a levee-encircled plain polder.
• Temporal Scale: Simulations of multiple historical and design storm events, typically 24-hour events, with data collected at hourly or 15-minute intervals. Surrogate models provide predictions in milliseconds to seconds.

Methodology and Data

• Models used:
  • Mechanistic Model: InfoWorks ICM, coupling 2D overland flow, 1D storm-sewer network, 1D open-channel river network, and rainfall–runoff processes.
    • Runoff generation: Horton infiltration model for pervious/semi-pervious surfaces, constant runoff coefficient for impervious surfaces.
    • Overland routing: Nonlinear reservoir method.
    • 1D hydraulics: 1D Saint-Venant (shallow-water) system, discretized using the Preissmann four-point implicit scheme, solved with the tridiagonal Thomas algorithm.
    • 1D–2D exchange: Mass-conserving exchange at node/inlet interfaces.
  • Optimization: Particle Swarm Optimization (PSO) for parameter calibration (e.g., Horton infiltration coefficients, Manning roughness).
  • Deep Learning Surrogates:
    • Long Short-Term Memory (LSTM) network: For point-wise water-depth time series prediction at critical locations.
    • Convolutional Neural Network (CNN): For domain-wide maximum water-depth maps, using an encoder–decoder architecture.
• Data sources:
  • Construction Data: Digital Elevation Model (DEM), land use, channel centerlines and cross-sections, attributes of pumping stations, gates, culverts, operation schedules, outfalls.
  • Calibration and Validation Data: Hourly or 15-minute water levels and concurrent records of pump and gate operations at external boundaries (B1–B8) and interior checkpoints (H1–H12).
  • Design Scenarios: Twelve synthetic storm events covering various intensities and time distributions, used to generate high-fidelity reference datasets from the calibrated mechanistic model.
  • Physical Metadata Channels: Normalized coordinates, proximity to pumps/gates, inlet/pipe density, and distance to main channels, integrated into surrogate model inputs.

Main Results

• The integrated mechanistic model robustly reproduced water-level hydrographs, achieving station-wise Nash–Sutcliffe Efficiency (NSE) ranging from 0.898 to 0.990, Pearson correlation (r) from 0.95 to 0.995, and percent bias (PBIAS) from -2.08% to 0.37% during calibration.
• The LSTM surrogate model achieved mean absolute errors (MAE) of approximately 3.05 cm for point-wise water-depth time series on independent test sets.
• The CNN surrogate model achieved an Intersection-over-Union (IoU) of approximately 0.87 and a mean water-depth error of about 4.3 cm for domain-wide peak water-depth maps on independent test sets.
• Surrogate inference demonstrated a computational speed-up of 10^3–10^4 times compared to the mechanistic model (e.g., a 24-hour event: LSTM ≈ 0.05 s, CNN < 1 s, ICM > 30 min).
• Physical metadata channels and explicit representation of pump–gate operation rules significantly improved boundary delineation and model generalization.

Contributions

• Development of a tri-library (topology, boundary, operations) + 1D/2D coupling + compiled operations modeling system specifically for dike-ringed, operation-dominated, low-slope, dense-network cities, enhancing standardization and reusability of boundary and operations representations.
• Implementation of a node-driven 1D solver and a mass-conserving 1D–2D exchange strategy that maintains stability and mass conservation under low-gradient conditions.
• Creation of a time–space dual surrogate (LSTM + CNN) enhanced by physical metadata channels, improving the representation of spatial heterogeneity and interpretability.
• Comprehensive multi-event calibration/validation with quantitative metrics (NSE, IoU, depth error) and demonstrated speed-ups relative to mechanistic simulation, enabling minute-scale forecasting and rapid multi-scenario evaluation.

Funding

This research was supported by no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Citation

@article{Chen2026Rapid,
  author = {Chen, Houlin and Wang, Jiahu},
  title = {Rapid Urban Flood Simulation and Prediction Using Integrated Hydrodynamic Modeling and Deep Learning Approaches},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-025-04400-6},
  url = {https://doi.org/10.1007/s11269-025-04400-6}
}