Wang et al. (2026) FKDM: a frequency-guided Koopman dynamics model with heterogeneous mixture-of-experts for robust daily runoff forecasting under non-stationarity

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-08-11
• Authors: Wenchuan Wang, Kun-mei Luo, Miao Gu, C. SHI, Yi-yang Wang, Zong Li
• DOI: 10.1016/j.jhydrol.2026.136184

Research Groups

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

Short Summary

This study proposes a novel end-to-end forecasting framework, the Frequency-guided Koopman Dynamics Model (FKDM), for robust daily runoff forecasting under non-stationarity. The FKDM exhibits superior performance compared to benchmark models.

Objective

• Investigate the limitations of existing deep learning models in capturing extremes and accommodating heterogeneous hydrological signals.
• Develop an end-to-end forecasting framework that integrates frequency-domain theory with operator dynamics.

Study Configuration

• Spatial Scale: Four representative basins in the western United States, characterized by diverse hydro-climatic and geomorphological features.
• Temporal Scale: Daily runoff processes under non-stationary conditions.

Methodology and Data

• Models used: Frequency-guided Koopman Dynamics Model (FKDM)
• Data sources: Runoff series from four representative basins in the western United States

Main Results

• The FKDM exhibits superior forecasting performance compared to benchmark models, with a Nash-Sutcliffe Efficiency (NSE) ranging from 0.9839 to 0.9849.
• The FKDM reduces the Mean Absolute Error (MAE) by 48.70% and the Root Mean Square Error (RMSE) by 49.15% compared to benchmark models.

Contributions

• The study provides a new paradigm for data-driven hydrological forecasting, combining prediction accuracy, numerical stability, and physical interpretability.
• The FKDM addresses the limitations of existing deep learning models in capturing extremes and accommodating heterogeneous hydrological signals.

Funding

• This research was funded by the National Natural Science Foundation of China (Grant No. 52109244).

Citation

@article{Wang2026FKDM,
  author = {Wang, Wenchuan and Luo, Kun-mei and Gu, Miao and SHI, C. and Wang, Yi-yang and Li, Zong},
  title = {FKDM: a frequency-guided Koopman dynamics model with heterogeneous mixture-of-experts for robust daily runoff forecasting under non-stationarity},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136184},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136184}
}