Lu et al. (2025) A lake water level prediction method based on data augmentation and Physics-Informed Neural Networks with imbalanced data

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-11-23
• Authors: Lingjiang Lu, Tao Yan, Yongcan Chen, Haoran Wang, Tong Yang, Zhaowei Liu
• DOI: 10.1016/j.jhydrol.2025.134660

Research Groups

• State Key Laboratory of Hydroscience and Engineering, Tsinghua University
• Tianfu Yongxing Laboratory
• Sichuan Energy Internet Research Institute, Tsinghua University
• Chongqing University

Short Summary

This study proposes a novel Physics-Informed Neural Network (PINN) framework that integrates data augmentation and physically guided hyper-parameter selection to accurately predict lake water levels, specifically addressing challenges posed by imbalanced extreme event data and high computational costs. The framework demonstrates superior accuracy and efficiency compared to traditional models, achieving an RMSE of 0.021 m and requiring significantly less computational time.

Objective

• To develop an accurate and efficient lake water level prediction method that addresses two major limitations of existing deep learning surrogates: the lack of physically informed guidance for hyper-parameter selection (which increases computational costs) and the scarcity of extreme water level samples (which leads to imbalanced datasets and reduced accuracy).

Study Configuration

• Spatial Scale: Lower Lake of Nansi Lake, China.
• Temporal Scale: Not explicitly defined for the study's data or prediction horizon, but addresses long-term fluctuations over decades.

Methodology and Data

• Models used: Physics-Informed Neural Network (PINN) framework, incorporating mass-conservation constraints and a clustering-based data augmentation method. Performance was compared against a classical Long Short-Term Memory (LSTM) model and traditional hydrodynamic models (e.g., MIKE21, Delft3D, HEC-RAS for computational time comparison).
• Data sources: Boundary water level time series (implied observational data).

Main Results

• The proposed PINN framework, by incorporating physical constraints, robustly improves predictive accuracy, surpassing the performance of a classical LSTM model.
• Physically guided hyper-parameter selection significantly enhances both training efficiency and accuracy.
• The proposed clustering-based data augmentation method reduces Root Mean Squared Error (RMSE) by 69.1 % under extreme conditions.
• Compared to an existing augmentation method, the proposed method shortens training time by 63.35 % while achieving better prediction performance.
• The final surrogate model achieves an RMSE of 0.021 m and a Nash–Sutcliffe Efficiency (NSE) greater than 0.94 against observations.
• The framework requires only 2.42 % of the computational time of a traditional hydrodynamic model.

Contributions

• Introduction of a novel Physics-Informed Neural Network (PINN) framework that integrates data augmentation with physically guided hyper-parameter selection for lake water level prediction.
• Development of a clustering-based data augmentation method specifically designed to enrich extreme event samples in imbalanced datasets, leading to improved prediction accuracy under extreme conditions.
• Demonstration of significant improvements in predictive accuracy and computational efficiency by incorporating mass-conservation constraints and physically guided hyper-parameter selection.
• Validation of the framework's effectiveness in a real-world application (Lower Lake of Nansi Lake), showing superior performance over classical deep learning models and substantial computational savings compared to traditional hydrodynamic models.

Funding

• Not explicitly mentioned in the provided text.

Citation

@article{Lu2025lake,
  author = {Lu, Lingjiang and Yan, Tao and Chen, Yongcan and Wang, Haoran and Yang, Tong and Liu, Zhaowei},
  title = {A lake water level prediction method based on data augmentation and Physics-Informed Neural Networks with imbalanced data},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134660},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134660}
}