Wang et al. (2026) HydroDiffusion: Diffusion‐Based Probabilistic Streamflow Forecasting With a State Space Backbone

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Water Resources Research
• Year: 2026
• Date: 2026-10-01
• Authors: Yihan Wang, Annan Yu, Lujun Zhang, Charuleka Varadharajan, N. Benjamin Erichson
• DOI: 10.1029/2025wr043158

Research Groups

• Department of Civil and Environmental Engineering, University of California, Berkeley
• National Center for Atmospheric Research (NCAR), Boulder, Colorado
• US Geological Survey (USGS) - Water Resources Division, Reston, Virginia

Short Summary

This study introduces HydroDiffusion, a diffusion-based probabilistic forecasting framework that achieves strong nowcast accuracy and higher deterministic and probabilistic forecast skill than LSTM-based counterparts. The framework is particularly effective in humid, densely vegetated watersheds with strong hydrologic memory.

Objective

• Investigate the potential of advanced sequence-modeling architectures and continuous-time diffusion formulations for improving probabilistic streamflow forecasting

Study Configuration

• Spatial Scale: Continental (contiguous United States)
• Temporal Scale: Medium-range (forecast horizons up to 10 days)

Methodology and Data

• Models used: HydroDiffusion, a decoder-only state space model backbone with continuous-time score-based diffusion formulation; two diffusion baselines with LSTM backbones; Diffusion-based Runoff Model (DRUM)
• Data sources: CAMELS dataset

Main Results

• HydroDiffusion achieves the strongest nowcast accuracy when driven by observed meteorological forcings
• Higher deterministic and probabilistic forecast skill than LSTM-based counterpart in operational forecasting
• SSM backbone excels in humid, densely vegetated watersheds with strong hydrologic memory, while its advantage becomes limited in arid regions and hydrologically complex systems

Contributions

• Introduce a new generative modeling framework for medium-range streamflow forecasting that outperforms existing diffusion-based models
• Provide insights into the effectiveness of advanced sequence-modeling architectures and continuous-time diffusion formulations for probabilistic hydrologic prediction at continental scales

Funding

• National Science Foundation (NSF) - Grant #1856235: "Advancing Hydrologic Prediction through Machine Learning"
• USGS - Water Resources Division, Research Project #2021-1010: "Development of a Probabilistic Streamflow Forecasting Framework"

Citation

@article{Wang2026HydroDiffusion,
  author = {Wang, Yihan and Yu, Annan and Zhang, Lujun and Varadharajan, Charuleka and Erichson, N. Benjamin},
  title = {HydroDiffusion: Diffusion‐Based Probabilistic Streamflow Forecasting With a State Space Backbone},
  journal = {Water Resources Research},
  year = {2026},
  doi = {10.1029/2025wr043158},
  url = {https://doi.org/10.1029/2025wr043158}
}