Li et al. (2025) A hybrid framework for sub-seasonal to seasonal streamflow prediction: integrating numerical and statistical models

Identification

• Journal: npj Climate and Atmospheric Science
• Year: 2025
• Date: 2025-11-28
• Authors: Lingfeng Li, Huan Wu, Lulu Jiang, Yiwen Mei, John S. Kimball, Lorenzo Alfieri, Zhijun Huang, Ying Hu, Sirong Chen, Shaorou Dong, Yueqiang Hu, Wei Wu
• DOI: 10.1038/s41612-025-01273-9

Research Groups

• Southern Marine Science and Engineering Laboratory (Zhuhai), Sun Yat-sen University, Zhuhai, China
• School of Atmospheric Sciences, Sun Yat-sen University, Zhuhai, China
• Guangdong Province Key Laboratory for Climate Change and Natural Disaster Studies, Sun Yat-sen University, Zhuhai, China
• Earth System Science Interdisciplinary Center, University of Maryland, College Park, Maryland, USA
• Carbon-Water Observation and Research Station in Karst Regions of Northern Guangdong, School of Geography and Planning, Sun Yat-sen University, Guangzhou, China
• Numerical Terradynamic Simulation Group, W.A. Franke College of Forestry & Conservation, University of Montana, Missoula, MT, USA
• CIMA Research Foundation, Savona, Italy
• Climate Center of Guangxi, Nanning, China
• Climate Center of Guangdong, Guangzhou, China
• National Disaster Reduction Center of China, Ministry of Emergency Management, Beijing, China

Short Summary

This study develops a hybrid framework integrating a distributed hydrological model (DRIVE) with a probabilistic statistical model (BJP) to enhance sub-seasonal to seasonal (S2S) streamflow prediction, demonstrating improved forecast skill for flood events in the complex Pearl River Basin.

Objective

• To develop and validate a hybrid framework that integrates a distributed hydrological model (DRIVE) with a probabilistic statistical model (BJP) to improve sub-seasonal to seasonal (S2S) streamflow prediction, particularly for flood forecasting, by assimilating statistical hydroclimate relationships.

Study Configuration

• Spatial Scale: Pearl River Basin (PRB), China, covering an area of 453,690 square kilometers. The study was validated at 24 hydrological gauging stations across the basin. Hydrological modeling used a grid cell resolution of 0.125° × 0.125°.
• Temporal Scale: Sub-seasonal to seasonal (S2S) streamflow prediction, with forecast lead times ranging from 1 to 6 weeks (up to 44 days). The study period for data and cross-validation was from June 2019 to December 2022 (approximately 3.5 years).

Methodology and Data

• Models used:
  • Distributed Hydrological Model: Dominant River Tracing-Routing Integrated with Variable Infiltration Capacity Environment (DRIVE), which integrates the Variable Infiltration Capacity Macroscale Hydrologic Model (VIC) for surface and subsurface runoff calculations and the Dominant River Tracing based Routing Model (DRTR) for river routing.
  • Probabilistic Statistical Model: Bayesian Joint Probability (BJP) model, extended for sub-seasonal to seasonal streamflow prediction.
  • Hybrid Schemes: E1 (BJP using DRIVE-forecasted streamflow as sole predictor) and E2 (BJP using DRIVE-forecasted streamflow, observed initial streamflow, and S2S precipitation forecasts as predictors).
• Data sources:
  • Streamflow data: Daily streamflow observations from 24 gauging stations in the Pearl River Basin, obtained from the National Hydrological and Rainfall Information Network, Ministry of Water Resources of China.
  • Satellite-derived precipitation: Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS v2.0), with a spatial resolution of 0.05° and daily temporal resolution.
  • Forecast precipitation: ECMWF public dataset S2S prediction, comprising 13 models (specifically ECMWF, KMA, UKMO, and NCEP were used) providing forecasts up to 60 days at a spatial resolution of 1.5 degrees.
  • Other atmospheric variables: ERA-5 reanalysis data (air temperature, downward solar and longwave radiation, relative humidity, surface pressure, and wind speed), with hourly temporal resolution and 0.25-degree spatial resolution.

Main Results

• The hybrid ensemble schemes (E1 and E2) significantly improve streamflow prediction performance compared to standalone physical (DRIVE) or statistical (BJP) models, especially for longer lead times (2-6 weeks).
• For the 2-6 week forecast period, the mean Nash-Sutcliffe Efficiency (NSE) scores were 0.23 for E2 and 0.22 for E1, outperforming DRIVE (0.08) and BJP (0.12).
• The E2 scheme achieved mean NSE scores of 0.45 (week 1), 0.36 (week 2), 0.28 (week 3), 0.20 (week 4), 0.16 (week 5), and 0.16 (week 6), demonstrating valuable forecasting skill up to six weeks. Individual methods showed little to no skill for weeks 5-6.
• The E2 scheme showed a 15% improvement in Continuous Ranked Probability Score Skill (CRPSS) compared to the standalone DRIVE model and over 8% compared to the standalone BJP model across all lead times.
• E2 consistently exhibited a superior balance between flood event detection (high Probability of Detection, POD) and false alarms (low False Alarm Ratio, FAR) across most lead time intervals, making it particularly suitable for S2S flood forecasting.
• The ensemble approach effectively mitigates performance degradation observed in upstream catchments by BJP and in downstream catchments by DRIVE over extended lead times.

Contributions

• Proposes and validates a novel hybrid framework that integrates a distributed hydrological model (DRIVE) with a probabilistic statistical model (BJP) for sub-seasonal to seasonal (S2S) streamflow prediction.
• Demonstrates significant improvements in S2S streamflow forecast skill, extending valuable predictability up to six weeks in a complex, rain-dominant, and flood-prone basin (Pearl River Basin), surpassing the performance of individual physical or statistical models.
• Highlights the critical value of combining process understanding from physical models with data-driven correction from statistical models to enhance both deterministic accuracy and probabilistic reliability, particularly at longer lead times where atmospheric predictability is limited.
• Offers a practical and robust pathway to enhance flood forecast skill and lead times in challenging hydrological regions, addressing a significant gap in current S2S flood forecasting capabilities.

Funding

• National Key R&D Program of China (Grants: 2024YFC3013302)
• National Natural Science Foundation of China (Grants: 42275019, 42088101)
• CMA’s Open Fund Project for Heavy Rain of China (Grants: BYKJ2024Z10)
• Hainan R&D Program (CXFZ2022J074, SCSF202203)
• Guangdong Province Key Laboratory for Climate Change and Natural Disaster Studies (Grants: 2020B1212060025)

Citation

@article{Li2025hybrid,
  author = {Li, Lingfeng and Wu, Huan and Jiang, Lulu and Mei, Yiwen and Kimball, John S. and Alfieri, Lorenzo and Huang, Zhijun and Hu, Ying and Chen, Sirong and Dong, Shaorou and Hu, Yueqiang and Wu, Wei},
  title = {A hybrid framework for sub-seasonal to seasonal streamflow prediction: integrating numerical and statistical models},
  journal = {npj Climate and Atmospheric Science},
  year = {2025},
  doi = {10.1038/s41612-025-01273-9},
  url = {https://doi.org/10.1038/s41612-025-01273-9}
}