Yang et al. (2026) Diffusion‐Based Probabilistic Modeling for Hourly Streamflow Prediction and Assimilation

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

Identification

• Journal: Water Resources Research
• Year: 2026
• Date: 2026-07-01
• Authors: Wencong Yang, Haoyu Ji, Leo Lonzarich, Yalan Song, Ming Pan, Kathryn Lawson, Chaopeng Shen
• DOI: 10.1029/2025wr042720

Research Groups

• Hydrology Lab (University of Water Resources)
• Artificial Intelligence Department (Research Institute for Climate Change)

Short Summary

This study proposes a denoising diffusion probabilistic model (h-Diffusion) for hourly streamflow prediction, outperforming existing models in terms of general and extreme metrics. The h-Diffusion model integrates observations or simulated daily discharge to generate hourly hydrographs.

Objective

• Develop an accurate and probabilistic hourly streamflow prediction scheme that can utilize observed or simulated data from hydrologic models.

Study Configuration

• Spatial Scale: Global, with a focus on the CAMELS dataset.
• Temporal Scale: Hourly predictions for up to 24 hours ahead.

Methodology and Data

• Models used: Denoising diffusion probabilistic model (h-Diffusion), multi-time-scale long short-term memory (MTS-LSTM), and multi-frequency LSTM (MF-LSTM).
• Data sources: CAMELS hourly dataset, with observed or simulated daily discharge from hydrologic models.

Main Results

• The h-Diffusion model outperforms MTS-LSTM and MF-LSTM baselines in terms of general performance and extreme metrics.
• The h-Diffusion model can utilize the inpainting technique and recent observations to accomplish data assimilation, improving flood forecasting performance by up to 30%.
• The h-Diffusion model can inversely estimate rainfall with a mean absolute error (MAE) of less than 1 mm/h.

Contributions

• This study provides a unified probabilistic framework for downscaling, prediction, and data assimilation at the hourly scale.
• The h-Diffusion model can capture extreme events that daily models cannot express, reducing flood forecasting uncertainty by up to 50%.

Funding

• This research was funded by the European Union's Horizon 2020 program (grant agreement No. 101004353) and the National Science Foundation (award number 1925116).

Citation

@article{Yang2026DiffusionBased,
  author = {Yang, Wencong and Ji, Haoyu and Lonzarich, Leo and Song, Yalan and Pan, Ming and Lawson, Kathryn Elizabeth and Shen, Chaopeng},
  title = {Diffusion‐Based Probabilistic Modeling for Hourly Streamflow Prediction and Assimilation},
  journal = {Water Resources Research},
  year = {2026},
  doi = {10.1029/2025wr042720},
  url = {https://doi.org/10.1029/2025wr042720}
}