Xu et al. (2025) Elevating predictive reliability: time-varying parameter bayesian deep learning techniques for flood probability forecasting

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-11-11
• Authors: Hanbing Xu, Yanlai Zhou, Tianyu Xia, Hua Chen, Fi‐John Chang, Chong‐Yu Xu
• DOI: 10.1016/j.jhydrol.2025.134597

Research Groups

• State Key Laboratory of Water Resources Engineering and Management, Wuhan University, Wuhan 430072, China
• Department of Bioenvironmental Systems Engineering, National Taiwan University, Taipei 10617, Taiwan
• Department of Geosciences, University of Oslo, P.O. Box 1047, Blindern, N-0316 Oslo, Norway

Short Summary

This study introduces a Fourier-based Time-Varying Parameter Bayesian Long Short-Term Memory (F-TV-BLSTM) model, integrating FourCastNet precipitation forecasts, to enhance multi-step-ahead probabilistic flood forecasting reliability in non-stationary environments. Applied to the Yalong River Basin, the model demonstrates superior performance in accuracy and dependability, particularly for extreme flood events, by dynamically adjusting parameters.

Objective

• To develop a robust flood probability forecasting model that can effectively capture the non-stationary and highly dynamic nature of rainfall-runoff relationships in changing environments.
• To improve the accuracy and dependability of multi-step-ahead probabilistic flood forecasts by dynamically adjusting model parameters using Fourier basis functions and incorporating forward-looking hydrometeorological information from FourCastNet.

Study Configuration

• Spatial Scale: Yalong River Basin, China; data at 0.25° × 0.25° spatial resolution.
• Temporal Scale: 2013 to 2023; 6-hourly data.

Methodology and Data

• Models used:
  • Fourier-based Time-Varying Parameter Bayesian Long Short-Term Memory (F-TV-BLSTM)
  • Time-invariant Bayesian Long Short-Term Memory (BLSTM) (benchmark)
  • FourCastNet-based BLSTM (F-BLSTM) (benchmark)
  • FourCastNet (for high-resolution precipitation forecasts)
• Data sources:
  • Observed streamflow
  • European Centre for Medium-Range Weather Forecasts Reanalysis V5 (ERA5) precipitation
  • FourCastNet forecast data (precipitation)

Main Results

• Incorporating precipitation forecasts from FourCastNet significantly improved predictive performance.
• The F-TV-BLSTM achieved a maximum of 6.3 % improvement in Nash-Sutcliffe Efficiency Coefficient (NSE) compared to benchmark models.
• The F-TV-BLSTM showed maximum reductions of 34.3 % in Root Mean Square Error (RMSE) and 32.5 % in Mean Absolute Error (MAE).
• The model provided superior probabilistic forecasts with a 12.7 % maximum increase in Containment Ratio.
• The F-TV-BLSTM achieved a 16.4 % maximum reduction in Average Bandwidth and a 13.5 % maximum reduction in Continuous Ranked Probability Score (CRPS).
• In extreme flood scenarios, the model provided more accurate peak flow predictions and narrower, more reliable confidence intervals.
• The F-TV-BLSTM effectively captures non-stationary streamflow behavior, offering a robust solution for real-time flood forecasting.

Contributions

• Introduction of a novel Fourier-based Time-Varying Parameter Bayesian Long Short-Term Memory (F-TV-BLSTM) model, specifically designed to address the challenges of non-stationary hydrological systems by dynamically adjusting model parameters.
• Integration of high-resolution precipitation forecasts from FourCastNet into a Bayesian deep learning framework, significantly enhancing the accuracy and dependability of multi-step-ahead probabilistic flood forecasts.
• Demonstrated substantial improvements in both deterministic and probabilistic flood forecasting metrics, providing more accurate peak flow predictions and reliable confidence intervals, particularly crucial for extreme flood events and real-time disaster mitigation.

Funding

• Not explicitly mentioned in the provided text.

Citation

@article{Xu2025Elevating,
  author = {Xu, Hanbing and Zhou, Yanlai and Xia, Tianyu and Chen, Hua and Chang, Fi‐John and Xu, Chong‐Yu},
  title = {Elevating predictive reliability: time-varying parameter bayesian deep learning techniques for flood probability forecasting},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134597},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134597}
}