Lv et al. (2026) An adaptive decomposition-denoising and temporal context fusion framework for multi-station water-level forecasting

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-03-21
• Authors: Haifeng Lv, Chaoyu Shi, Xiaoyu Ji, Shan Shi, Yong Ding
• DOI: 10.1016/j.jhydrol.2026.135339

Research Groups

• Guangxi Key Laboratory of Cryptography and Information Security, School of Computer Science and Information Security, Guilin University of Electronic Technology, Guilin, Guangxi, China
• Guangxi Key Laboratory of Machine Vision and Intelligent Control, Wuzhou University, Wuzhou, Guangxi, China
• Wuzhou Hydrological Center, Wuzhou, Guangxi, China

Short Summary

This study constructs a high-resolution hydrological dataset for the Mengjiang River Basin and proposes a hybrid deep learning framework, CEEMDANVF-WD-TCF-LSTM, for multi-station water-level forecasting. The framework demonstrates superior accuracy and stability, particularly in mitigating multi-step error accumulation for both short-term and medium-term predictions.

Objective

• To enhance short-term (3-hour) and medium-term (6-hour) multi-station water-level forecasting by developing a robust hybrid deep learning framework that addresses the nonlinearity, non-stationarity, and noise inherent in hydrological time series.

Study Configuration

• Spatial Scale: Mengjiang River Basin, focusing on multiple hydrological stations.
• Temporal Scale: Forecasting horizons of 3 hours (short-term) and 6 hours (medium-term).

Methodology and Data

• Models used: CEEMDANVF-WD-TCF-LSTM (CE-WD-TCF-LSTM), which integrates:
  • Complete Ensemble Empirical Mode Decomposition with Adaptive Noise and Variance Filtering (CEEMDANVF)
  • Targeted Wavelet Denoising (WD)
  • Temporal Context Fusion Long Short-Term Memory network (TCF-LSTM)
• Data sources: A high-resolution hydrological benchmark dataset (MJWD) constructed for the Mengjiang River Basin.

Main Results

• The proposed CE-WD-TCF-LSTM model consistently outperformed baseline models (CNN, LSTM) and other hybrid variants.
• For 3-hour forecasts, the coefficient of determination (R²) values ranged from 0.87 to 0.96 across all stations.
• For 6-hour forecasts, the R² values ranged from 0.81 to 0.94 across all stations.
• Kling–Gupta Efficiency (KGE) values remained above 0.86 for all forecasts.
• The framework showed more significant improvement in 6-hour forecasting, effectively mitigating multi-step error accumulation through its decomposition–denoising and multi-scale temporal feature fusion capabilities.

Contributions

• Proposed a novel hybrid deep learning framework (CE-WD-TCF-LSTM) specifically designed for multi-station water-level forecasting.
• Developed an adaptive variance-filtering strategy within CEEMDAN to efficiently separate signal from noise in hydrological time series.
• Designed a Temporal Context Fusion Long Short-Term Memory (TCF-LSTM) module to effectively fuse multi-scale temporal contexts for robust prediction.
• Constructed a high-resolution benchmark dataset (MJWD) for the Mengjiang River Basin to address data fragmentation challenges in hydrological forecasting.
• Demonstrated superior accuracy and stability, particularly in mitigating multi-step error accumulation, providing a generalizable solution for water resource management and flood risk mitigation.

Funding

• National Science Foundation

Citation

@article{Lv2026adaptive,
  author = {Lv, Haifeng and Shi, Chaoyu and Ji, Xiaoyu and Shi, Shan and Ding, Yong},
  title = {An adaptive decomposition-denoising and temporal context fusion framework for multi-station water-level forecasting},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135339},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135339}
}