Lei et al. (2025) PANet: a physics and action informed network for water level prediction in canal systems

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-10-30
• Authors: Xiaohui Lei, Jiahao Wu, Yan Long, Lingqiang Chen, Meng Wang, Weikang Zhao
• DOI: 10.1016/j.jhydrol.2025.134485

Research Groups

• School of Information and Electrical Engineering, Hebei University of Engineering, Handan 056038, China
• College of Water Conservancy and Hydropower, Hebei University of Engineering, Handan 056038, China
• Hebei Key Laboratory of Smart Water Conservancy, Handan, Hebei 056009, China

Short Summary

This study introduces PANet, a multi-step water level prediction model for canal systems that integrates Integral Delay theory and action-informed structures to simultaneously model hydrological delays, Markovian properties, and the impact of operational interventions. PANet achieves superior accuracy and stability, significantly outperforming mainstream deep learning models in multi-step forecasting tasks.

Objective

• To develop a multi-step water level prediction model (PANet) that overcomes limitations of existing physically based and data-driven methods by simultaneously modeling hydrological delays, Markovian properties, and the causal relationship between human-regulated control operations and hydraulic responses in canal systems.

Study Configuration

• Spatial Scale: Inter-basin water transfer projects, specifically canal systems like the South-to-North Water Diversion Project.
• Temporal Scale: Multi-step water level prediction, capturing both short-term hydraulic dynamics and long-term water level evolution.

Methodology and Data

• Models used: PANet (Physics- and Action-informed water prediction Network) comprising an Action Attention Encoder (A2E) and a Dual-branch Decoder (DBD). Compared against Informer, Transformer, LSTM, and traditional hydrodynamic models (e.g., Saint-Venant equations).
• Data sources: Operational signals (e.g., gate openings) and historical water level data.

Main Results

• PANet achieved a mean absolute error (MAE) of 0.0139 m in multi-step water level prediction.
• It reduced prediction errors by 32.11 % compared to Informer.
• It reduced prediction errors by 76.79 % compared to Transformer.
• It reduced prediction errors by 65.11 % compared to LSTM.
• PANet demonstrated superior accuracy and stability in multi-step forecasting tasks.

Contributions

• Proposes PANet, a novel multi-step water level prediction model that integrates Integral Delay (ID) theory as a physical mechanism with action-informed structures.
• Enables the simultaneous modeling of hydrological delays, Markovian properties, and the causal relationship between operational interventions and hydraulic responses, addressing limitations of existing models.
• Introduces a unique architecture with an Action Attention Encoder (A2E) to enhance sensitivity to anthropogenic interventions and a Dual-branch Decoder (DBD) to capture multi-temporal scale dynamics.
• Significantly outperforms mainstream deep learning models (Informer, Transformer, LSTM) in accuracy and stability for multi-step water level forecasting.

Funding

• Not specified in the provided text.

Citation

@article{Lei2025PANet,
  author = {Lei, Xiaohui and Wu, Jiahao and Long, Yan and Chen, Lingqiang and Wang, Meng and Zhao, Weikang},
  title = {PANet: a physics and action informed network for water level prediction in canal systems},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134485},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134485}
}