Han et al. (2025) Enhancing multi-step-ahead prediction of wave propagation with the CAE-LSTM model: a novel deep learning-based approach to flood dynamics

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Geomatics Natural Hazards and Risk
• Year: 2025
• Date: 2025-11-25
• Authors: Zheng Han, George E.M. Long, Changli Li, Yange Li, Bin Su, Linrong Xu, Weidong Wang, Guangqi Chen
• DOI: 10.1080/19475705.2025.2588708

Research Groups

Not explicitly mentioned in the paper.

Short Summary

This paper introduces a novel Convolutional Autoencoder (CAE)-integrated Long Short-Term Memory (LSTM) model to enhance the learning ability and generalization of Physics-Informed Neural Networks (PINNs) for long-term wave propagation in flood dynamics, demonstrating superior accuracy and computational efficiency compared to traditional finite volume methods.

Objective

• To develop a novel CAE-LSTM model that integrates spatial and temporal dimensions to enhance the capture and prediction ability for long-term wave propagation processes, addressing limitations of traditional fully connected neural network-based PINNs in learning ability and generalization.

Study Configuration

• Spatial Scale: Localized, focusing on dam-break scenarios and wave propagation dynamics.
• Temporal Scale: Long-term, validated over 3,000 prediction steps for flood wave evolution.

Methodology and Data

• Models used: Convolutional Autoencoder (CAE)-integrated Long Short-Term Memory (LSTM) model, Physics-Informed Neural Networks (PINNs) (general framework), Finite Difference Method (inspiration), Finite Volume Method (FVM) (for comparison), Shallow Water Equations (underlying physics).
• Data sources: Simulated data from four dam-break scenarios.

Main Results

• The CAE-LSTM model generally achieves a Root Mean Square Error (RMSE) less than 0.5 after 3,000 steps of rolling prediction.
• The model demonstrates computational efficiency approximately 200 times higher than traditional finite volume method (FVM) simulations.
• It significantly enhances the capture and prediction ability for long-term wave propagation processes.

Contributions

• Introduction of a novel CAE-LSTM model specifically designed to improve the learning ability and generalization of PINNs for long-term wave propagation in flood dynamics.
• Integration of spatial feature extraction (CAE) and temporal dependency capturing (LSTM) to create compact latent representations and precise predictions.
• Demonstration of superior accuracy (RMSE < 0.5) and significantly higher computational efficiency (200 times faster) compared to traditional finite volume methods.
• Addresses key limitations of traditional fully connected neural network-based PINNs in handling long-term wave propagation and generalization to untrained scenarios.

Funding

Not explicitly mentioned in the paper.

Citation

@article{Han2025Enhancing,
  author = {Han, Zheng and Long, George E.M. and Li, Changli and Li, Yange and Su, Bin and Xu, Linrong and Wang, Weidong and Chen, Guangqi},
  title = {Enhancing multi-step-ahead prediction of wave propagation with the CAE-LSTM model: a novel deep learning-based approach to flood dynamics},
  journal = {Geomatics Natural Hazards and Risk},
  year = {2025},
  doi = {10.1080/19475705.2025.2588708},
  url = {https://doi.org/10.1080/19475705.2025.2588708}
}