Pang et al. (2025) A skillful self-evolving deep-learning framework for pluvial flood process forecasting in urban areas

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-11-11
• Authors: Y. Pang, Jian He, Andrea Canlas, Luyu Ju, Kai Fei, Limin Zhang
• DOI: 10.1016/j.jhydrol.2025.134593

Research Groups

• Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology, HKSAR, China
• School of Engineering and Technology, China University of Geosciences (Beijing), Beijing, China
• State key Laboratory of Climate Resilience for Coastal Cities, The Hong Kong University of Science and Technology, HKSAR, China

Short Summary

This paper presents a self-evolving deep-learning framework, based on two autoregressive convolutional neural networks, for real-time pluvial flood process forecasting. The framework accurately predicts flood depth and velocity fields with significantly reduced computational time compared to conventional hydrodynamic models and improved accuracy over existing end-to-end surrogate models.

Objective

• To develop a real-time, accurate, and computationally efficient deep-learning framework for forecasting pluvial flood depth and velocity fields in urban areas, overcoming the limitations of conventional hydrodynamic models and existing end-to-end surrogate models.

Study Configuration

• Spatial Scale: 79 km² urban area (Hong Kong) at 30 m resolution.
• Temporal Scale: Historical storm events (1984–2018) for training and validation; simulates a 96-hour flood event.

Methodology and Data

• Models used: Self-evolving framework based on two autoregressive convolutional neural networks (AR-CNNs). Performance was compared against conventional hydrodynamic solvers and a comparable end-to-end deep learning model.
• Data sources: Historical storm events (1984–2018) in Hong Kong, likely including rainfall data and corresponding flood simulations from hydrodynamic models for training and validation.

Main Results

• The framework simulates a 96-hour flood event in 1 s.
• It achieves computational speedups of 850 times (GPU-based) and 3,000 times (CPU-based) over conventional hydrodynamic solvers.
• It maintains mean absolute errors of 0.0007 m for flow depth and 0.0003 m/s for flow velocity.
• These errors are 16% (flow depth) and 50% (flow velocity) of the errors produced by a comparable end-to-end model.

Contributions

• Introduces a novel self-evolving deep-learning framework utilizing autoregressive convolutional neural networks for iterative and accurate updates of flood depth and velocity fields.
• Achieves unprecedented computational speedups (up to 3,000 times) for pluvial flood forecasting, enabling real-time prediction.
• Demonstrates significantly higher accuracy in predicting flood depth and velocity compared to existing end-to-end deep learning surrogate models.
• Provides a robust solution for real-time flood early warning and forecasting in urban flood-prone areas.

Funding

• Not explicitly mentioned in the provided text.

Citation

@article{Pang2025skillful,
  author = {Pang, Y. and He, Jian and Canlas, Andrea and Ju, Luyu and Fei, Kai and Zhang, Limin},
  title = {A skillful self-evolving deep-learning framework for pluvial flood process forecasting in urban areas},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134593},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134593}
}