Liu et al. (2025) PINN framework for urban flood depth prediction: integrating data-driven insights with physical constraints

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-12-05
• Authors: Wenli Liu, Qian Wu, Zihan Liu, Zheng Xu, Tianxiang Liu, Han Gao
• DOI: 10.1016/j.jhydrol.2025.134693

Research Groups

• School of Civil and Hydraulic Engineering, Huazhong University of Science and Technology, Wuhan, China
• National Center of Technology Innovation for Digital Construction, Wuhan, China
• International Joint Research Laboratory of Smart Construction, Wuhan, China
• School of Civil and Environmental Engineering, University of New South Wales, Sydney, Australia

Short Summary

This study developed SWEPINN, a Physics-Informed Neural Network integrating shallow water equations and data-driven insights, to provide efficient, accurate, and interpretable urban flood depth predictions, outperforming traditional data-driven models.

Objective

• To develop and validate a Shallow Water Equation-based Physics-Informed Neural Network (SWEPINN) that integrates physical constraints and data-driven insights for efficient, accurate, and interpretable urban flood depth prediction.

Study Configuration

• Spatial Scale: Wuhan Hongshan District, China (case study)
• Temporal Scale: 50 distinct rainstorm scenarios

Methodology and Data

• Models used: SWEPINN (developed), SWMM/TELEMAC (for data generation), Deep Neural Network (DNN), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM) (for comparison).
• Data sources: A synthetic dataset generated by a coupled SWMM/TELEMAC hydrodynamic model, covering 50 rainstorm scenarios. Nine spatial and fourteen rainfall features were determined using geographic detectors and backward selection.

Main Results

• SWEPINN achieved superior performance in urban flood depth prediction compared to DNN, CNN, and LSTM models.
• SWEPINN metrics: Coefficient of Determination (R²) of 0.952, Mean Absolute Error (MAE) of 0.068 meters, and Mean Squared Error (MSE) of 0.013 square meters.
• SHapley Additive exPlanations (SHAP) analysis identified Digital Elevation Model (DEM) and MANHOLE as key features influencing flood depth predictions.

Contributions

• Development of a novel Physics-Informed Neural Network (SWEPINN) that effectively integrates physical laws (Shallow Water Equations) with data-driven insights for urban flood depth prediction.
• Addresses the limitations of traditional hydrodynamic models (computational cost, data intensity) and purely data-driven models (lack of physical consistency).
• Demonstrates superior accuracy and efficiency compared to state-of-the-art data-driven models (DNN, CNN, LSTM).
• Enhances the interpretability of urban flood prediction models through Explainable AI (SHAP analysis), revealing key influencing factors.
• Provides a rapid, reliable, and interpretable framework for urban flood forecasting, supporting risk assessment and management.

Funding

• Not specified in the provided text.

Citation

@article{Liu2025PINN,
  author = {Liu, Wenli and Wu, Qian and Liu, Zihan and Xu, Zheng and Liu, Tianxiang and Gao, Han},
  title = {PINN framework for urban flood depth prediction: integrating data-driven insights with physical constraints},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134693},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134693}
}