Xu et al. (2025) RMC: advancing daily runoff forecasting with a unified cross-scale deep learning approach

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-12-05
• Authors: Dong-mei Xu, Qingle Zeng, Wenchuan Wang, Xu-tong Zhang, Hong-fei Zang
• DOI: 10.1016/j.jhydrol.2025.134722

Research Groups

College of Water Resources, North China University of Water Resources and Electric Power, Zhengzhou, China

Short Summary

This paper introduces Res-Mamba-Causal (RMC), a novel deep learning architecture designed to improve daily runoff forecasting by unifying multi-scale hydrologic feature modeling. RMC consistently outperforms existing models like LSTM and Transformer across various performance metrics on four U.S. watersheds, demonstrating enhanced accuracy and the ability to capture complex hydrologic dynamics.

Objective

• To develop a novel hybrid deep learning architecture (RMC) that effectively captures nonlinear dynamics, non-stationary behaviors, and complex interactions among multi-source forcings for accurate daily hydrologic time-series forecasting.

Study Configuration

• Spatial Scale: Four U.S. watersheds: Fish River, Redstone, Johnson Creek, and McKenzie River.
• Temporal Scale: Daily runoff forecasting.

Methodology and Data

• Models used: Res-Mamba-Causal (RMC) model, which integrates a residual convolutional layer, a Mamba state-space kernel, and a causal mask within a sliding window. Compared against Long Short-Term Memory (LSTM), Transformer, Mamba, and Convolutional Neural Network (CNN).
• Data sources: Hydrologic time-series data from the four U.S. watersheds. (Specific data types like precipitation, temperature, etc., are implied as "multi-source forcings" but not explicitly detailed as data sources in the abstract).

Main Results

• RMC consistently outperformed LSTM, Transformer, and other baselines across Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), Nash-Sutcliffe Efficiency (NSE), and Kling-Gupta Efficiency (KGE).
• At Fish River, RMC achieved a MAPE of 3.75, which is 71.1 % lower than Mamba (12.97), and a KGE of 0.992, 11.2 % higher than Mamba.
• At Redstone, RMC reduced Mean Absolute Error (MAE) by 52.9 % compared to Transformer.
• At McKenzie River, RMC attained an NSE of 0.980, 12.5 % higher than CNN.
• The RMC model effectively integrates local–short–long-range features, offering a scalable and interpretable paradigm for complex hydrologic forecasting.

Contributions

• Introduction of RMC, a novel hybrid deep learning architecture that unifies multi-scale hydrologic feature modeling through a three-stage progressive framework.
• Integration of residual convolutional layers for fine-grained local patterns, a Mamba state-space kernel for long-range dependencies with linear complexity, and a causal mask for enhanced local expression and physical causal response.
• Demonstrated superior performance of RMC over state-of-the-art deep learning models (LSTM, Transformer, Mamba, CNN) in daily runoff forecasting across diverse U.S. watersheds.
• Provides a scalable and interpretable paradigm for addressing the challenges of nonlinear dynamics, non-stationary behaviors, and complex interactions in hydrologic time-series forecasting.

Funding

[No funding information was provided in the abstract or introduction section of the paper text.]

Citation

@article{Xu2025RMC,
  author = {Xu, Dong-mei and Zeng, Qingle and Wang, Wenchuan and Zhang, Xu-tong and Zang, Hong-fei},
  title = {RMC: advancing daily runoff forecasting with a unified cross-scale deep learning approach},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134722},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134722}
}