Chen et al. (2025) Improving estuarine discharge forecasting with a KAN-augmented LSTM model: A case study of the Yangtze River Estuary

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2025
• Date: 2025-11-21
• Authors: Zhigao Chen, Yan Zong, Sheng-Ping Wang, Dajun Li
• DOI: 10.1016/j.ejrh.2025.102961

Research Groups

• Key Laboratory of Mine Environmental Monitoring and Improving around Poyang Lake of Ministry of Natural Resources, East China University of Technology, Nanchang 330013, China
• Jiangxi Key Laboratory of Watershed Ecological Process and Information, East China University of Technology, Platform No. 2023SSY01051, Nanchang 330013, China

Short Summary

This paper proposes a novel KAN-augmented LSTM (LSTM-KAN) hybrid deep learning model to enhance estuarine discharge forecasting in the Yangtze River Estuary, demonstrating significantly improved accuracy across short-, medium-, and long-term horizons compared to traditional and state-of-the-art methods.

Objective

• To develop and evaluate a novel KAN-augmented LSTM (LSTM-KAN) hybrid deep learning model for accurate estuarine discharge forecasting in tidal rivers, specifically assessing its performance against traditional and state-of-the-art models in the Xuliujing section of the Yangtze River Estuary.

Study Configuration

• Spatial Scale: Xuliujing section of the Yangtze River Estuary, China, approximately 110 km from the river mouth and spanning nearly 6 km in width.
• Temporal Scale:
  • Data period: January 1, 2022, to December 31, 2023 (2 years), with a sampling interval of 0.5 hours.
  • Forecast horizons: Short-term (6 hours), Medium-term (24 hours), and Long-term (48 hours).
  • Input look-back windows: 12 hours for short-term, 48 hours for medium-term, and 96 hours for long-term forecasts.

Methodology and Data

• Models used:
  • Proposed: LSTM-KAN (Long Short-Term Memory - Kolmogorov-Arnold Network)
  • Comparative: Harmonic Analysis (HA), XGBoost (Extreme Gradient Boosting), DLinear, Informer, LSTM (standard).
• Data sources:
  • Discharge data: Real-time monitoring system deployed by the Yangtze River Estuary Hydrological Bureau at the Xuliujing section, using 300 kHz ADCPs (Acoustic Doppler Current Profilers) sampled every 0.5 hours across multiple vertical lines.
  • Meteorological data: Daily total precipitation (mm) and daily average temperature (°C) for Qingpu District of Shanghai from Weather24.com (collected but excluded from model inputs due to weak correlation).
  • Input features for models: Flow velocity, water level, and cross-sectional area.
  • Output: Discharge (m³/s).

Main Results

• The LSTM-KAN model consistently outperformed all comparative methods (HA, LSTM, XGBoost, DLinear, Informer) in estuarine discharge forecasting across short-term (6 hours), medium-term (24 hours), and long-term (48 hours) horizons.
• Relative accuracy improvements for LSTM-KAN ranged from 12.1 % to 35.2 % over Harmonic Analysis and 7 % to 52.8 % over the traditional LSTM model.
• In medium-term forecasting, LSTM-KAN achieved the highest accuracy, with Relative Standard Deviation (RSD) values 15.1 %–31.1 % lower than HA, 5.4 %–15.5 % lower than LSTM, 1.6 %–3.7 % lower than XGBoost, 6 %–23 % lower than DLinear, and 1.1 %–2.5 % lower than Informer.
• For long-term forecasting, LSTM-KAN maintained the best performance, with RSD values 15.9 %–34.4 % lower than HA, 7 %–52.8 % lower than LSTM, 0.9 %–2.8 % lower than XGBoost, 6.7 %–27.3 % lower than DLinear, and 0.2 %–5.6 % lower than Informer.
• Statistical significance testing (Wilcoxon signed-rank test) confirmed LSTM-KAN's highly significant performance advantage (p < 0.001) over T_tide, LSTM, and DLinear across all forecast horizons. It also showed significant superiority over XGBoost in short-term (p < 0.05) and Informer in short-term and long-term (p < 0.05) forecasting.
• SHAP analysis revealed that the model dynamically adjusts feature importance based on tidal conditions: flow velocity was dominant during neap and moderate tides, while water level and cross-sectional area became more pronounced during spring tides, reflecting the underlying hydrodynamics.
• The model's highest forecast accuracy (RSD of 5.4 %) was achieved with a 48-hour forecast lead time, demonstrating its ability to leverage longer historical data to capture stable patterns while filtering out short-term noise.

Contributions

• Model Innovation: Introduction of a novel hybrid deep learning architecture, LSTM-KAN, which synergizes the sequential encoding power of LSTM with the superior, adaptive function approximation capabilities of KAN, specifically tailored for complex, periodic time series data in estuarine environments.
• Methodological Advancement: First application and systematic validation of the LSTM-KAN architecture in hydrological forecasting, using the challenging real-world case study of the Xuliujing section of the Yangtze River Estuary.
• Enhanced Interpretability: The KAN architecture offers a pathway to mechanistic insights by allowing visualization of learned functions for each feature connection, potentially revealing how the model adapts to complex physical relationships like tidal hysteresis.
• Improved Robustness and Efficiency: Demonstrates superior parameter efficiency and robustness, consistently outperforming standard LSTM and other models, particularly in demanding long-term forecasts and highly volatile tidal conditions, suggesting better generalization capabilities.

Funding

• National Natural Science Foundation of China (42266006)
• Jiangxi Provincial Natural Science Foundation (20232BAB204089)
• Key Laboratory of Marine Environmental Survey Technology and Application, Ministry of Natural Resources, Open Fund Project (MESTA-2023-A003)

Citation

@article{Chen2025Improving,
  author = {Chen, Zhigao and Zong, Yan and Wang, Sheng-Ping and Li, Dajun},
  title = {Improving estuarine discharge forecasting with a KAN-augmented LSTM model: A case study of the Yangtze River Estuary},
  journal = {Journal of Hydrology Regional Studies},
  year = {2025},
  doi = {10.1016/j.ejrh.2025.102961},
  url = {https://doi.org/10.1016/j.ejrh.2025.102961}
}