Li et al. (2025) Improving medium-long-term streamflow forecasts by exploiting multi-scale Temporal patterns with deep learning

Identification

• Journal: Natural Hazards
• Year: 2025
• Date: 2025-12-22
• Authors: T Li, Jiali Guo, Jihua Chen, Yan Huang, Jincheng Han, Biao Xiong
• DOI: 10.1007/s11069-025-07734-x

Research Groups

• Hubei Engineering Technology Research Center for Farmland Environment Monitoring, China Three Gorges University
• College of Computer and Information Technology, China Three Gorges University
• College of Hydraulic and Environmental Engineering, China Three Gorges University
• Engineering Research Center of Eco-Environment in Three Gorges Reservoir Region, Ministry of Education, China Three Gorges University

Short Summary

This study proposes a novel EMD-TCN-GRU deep learning framework to improve medium-long-term streamflow forecasting by exploiting multi-scale temporal patterns. The model achieves high accuracy and robustness for forecast horizons up to 15 days in the Yangtze River Basin, significantly outperforming existing deep learning and operational models.

Objective

• To address challenges in medium-long-term streamflow forecasting, specifically non-stationarity, limitations in capturing long-range dependencies, and errors from simply increasing network depth or attention mechanisms, by proposing an EMD-TCN-GRU framework.

Study Configuration

• Spatial Scale: Wuhan hydrometric station, middle Yangtze River Basin, China.
• Temporal Scale: Daily discharge data from 1 January 2013 to 31 December 2022 (10 years). Forecast horizons of 3 days, 7 days, and 15 days.

Methodology and Data

• Models used:
  • Proposed: EMD-TCN-GRU (Empirical Mode Decomposition, Temporal Convolutional Networks, Gated Recurrent Units).
  • Benchmark models: LSTM, GRU, TCN-GRU, EMD-GRU, EMD-TCN-LSTM.
• Data sources: Continuous daily-discharge dataset from the Wuhan hydrometric station, obtained from the Hydrological Bureau of the Changjiang Water Resources Commission, Ministry of Water Resources, P. R. China.

Main Results

• The EMD-TCN-GRU model demonstrated superior performance across all forecast horizons:
  • 3-day horizon: R² = 0.9951, Mean Absolute Percentage Error (MAPE) = 2.87%. Mean Absolute Error (MAE) reduced by 63% compared to a standard GRU.
  • 7-day horizon: R² = 0.9925, MAPE = 3.22%. MAE reduced by 51% compared to a standard GRU.
  • 15-day horizon: R² = 0.9922, MAPE = 3.25%. MAE reduced by 58% compared to a standard GRU.
• Performance decay over time was negligible, indicating robust long-term predictive capacity.
• Ablation studies confirmed that the decomposition-convolution-gating pipeline is the primary factor for the observed accuracy increase.
• The model accurately preserves extreme flow peaks (> 99th percentile), with MAPE remaining below 3%.
• Achieved a 7-day Nash-Sutcliffe Efficiency (NSE) of 0.992, surpassing the 99.5th percentile (0.97) of LSTM models in global comparisons.
• The 15-day MAPE of 3.25% represents an 80% reduction compared to the average error of current operational mid-term forecasting systems.

Contributions

• Introduces a novel EMD-TCN-GRU framework that synergistically integrates empirical mode decomposition for non-stationarity reduction, temporal convolutional networks for multi-scale feature extraction, and gated recurrent units for multi-step prediction.
• Effectively addresses long-standing challenges in medium-long-term streamflow forecasting, including non-stationarity, intricate long-range dependencies, and extreme-event sparsity, without requiring meteorological forcings.
• Achieves state-of-the-art predictive accuracy and robustness for streamflow forecasts up to 15 days, with negligible performance degradation over extended horizons.
• Demonstrates superior performance against various deep learning benchmarks and significantly improves upon current operational forecasting systems, particularly in preserving accuracy during extreme flow conditions.
• Offers a robust, transferable, and purely data-driven solution that reduces data dependency and computational costs, making it highly suitable for operational deployment in complex river networks and data-scarce regions.
• Provides significant practical implications for flood-risk management and water resource allocation, enabling proactive strategies and an estimated reduction of direct economic losses by CNY 1.7 billion in the Wuhan region.

Funding

• National Natural Science Foundation of China (No: 22136003)
• Hubei Provincial Natural Science Foundation of China (No: 2024AFB217)

Citation

@article{Li2025Improving,
  author = {Li, T and Guo, Jiali and Chen, Jihua and Huang, Yan and Han, Jincheng and Xiong, Biao},
  title = {Improving medium-long-term streamflow forecasts by exploiting multi-scale Temporal patterns with deep learning},
  journal = {Natural Hazards},
  year = {2025},
  doi = {10.1007/s11069-025-07734-x},
  url = {https://doi.org/10.1007/s11069-025-07734-x}
}