Chu et al. (2026) Streamflow Forecasting Using a Hybrid Modelling Coupled with Different Components

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-01-01
• Authors: Haibo Chu, Yulin Jiang, Wei Zhang, Jiahua Wei
• DOI: 10.1007/s11269-025-04395-0

Research Groups

• College of Architecture and Civil Engineering, Beijing University of Technology, Beijing, China
• State Key Laboratory of Hydroscience & Engineering, Tsinghua University, Beijing, China

Short Summary

This paper introduces a novel hybrid modeling approach that integrates Long Short-Term Memory (LSTM), the Snowmelt Runoff Model (SRM), and a degree-day model to enhance streamflow forecasting accuracy by explicitly incorporating rainfall, snowmelt, and glacier melt components. The hybrid model significantly outperforms individual models, particularly in snow and glacier melt-dominated catchments and during medium to high flow conditions, by better capturing complex hydrological dynamics.

Objective

• To develop and rigorously evaluate a novel, physically-guided multi-component hybrid model for streamflow forecasting that explicitly incorporates rainfall, snowmelt, and glacier melt contributions.
• To assess the performance of this hybrid framework, which combines data-driven and process-based components, across catchments with different dominant streamflow regimes (rainfall- and snowmelt-driven flows).

Study Configuration

• Spatial Scale: Two watersheds:
  • Yellow River source area, China: approximately 12,200 km² (32.15°–36.24°N, 95.78°–103.43°E).
  • North Platte River basin, USA: approximately 265,000 km² (40.3125°−41.9375°N, 105.9375°−107.0625°W).
• Temporal Scale:
  • Yellow River source area: Meteorological and streamflow data from 1960–2014 (LSTM trained 1960–2004, validated 2005–2014); snow depth data 1979–2023 (focus 2005–2014); MODIS snow cover data 2005–2014.
  • North Platte River: Daily streamflow data from 1975–2020; meteorological observations from 1975–2010 (training), remainder for testing; MODIS snow cover data 2011–2020.

Methodology and Data

• Models used:
  • Hybrid Model: Combines QLSTM (rainfall-runoff), QSnowmelt (snowmelt-driven streamflow), and Q_DDF (glacier melt contributions).
  • Long Short-Term Memory (LSTM) network: For rainfall-runoff prediction.
  • Snowmelt Runoff Model (SRM): For snowmelt-driven streamflow simulation, using degree-day logic and MODIS snow-covered area.
  • Degree-Day Model: For glacier melt estimation based on accumulated temperature data.
• Data sources:
  • Yellow River source area: Historical meteorological observations (rainfall, temperature) from Xinghai and Maqu stations; streamflow records from Tangnaihai station; snow depth data from the National Tibetan Plateau Data Center; MODIS 8-day composite snow cover data (MOD10A2) from NASA; Digital Elevation Model (DEM) from SRTM3 (90 m resolution) from the Geospatial Data Cloud.
  • North Platte River: Daily streamflow data from an undisturbed USGS gauge; meteorological observations from NOAA Laramie 2 NW station; MODIS snow cover data (MOD10A2); DEM data (90 m resolution) from SRTM from the Geospatial Data Cloud.

Main Results

• The hybrid model consistently outperformed the standalone LSTM model in both the Yellow River source area and the North Platte River basin.
• Yellow River source area: The hybrid model achieved R² = 0.92, RMSE = 150.38 m³/s, NSE = 0.92, and KGE = 0.95, showing improvements of 2.22% in R² and NSE, a 12.62% reduction in RMSE, and a 10.47% increase in KGE compared to the LSTM model.
• North Platte River: The hybrid model achieved R² = 0.88, RMSE = 176.52 m³/s, NSE = 0.88, and KGE = 0.90, demonstrating improvements of 4.76% in R² and NSE, a 10.15% reduction in RMSE, and a 9.88% increase in KGE compared to the LSTM model.
• The hybrid model showed enhanced predictive accuracy during snowmelt periods (e.g., KGE improved by 6% and RMSE reduced by 4.6% in the Yellow River source area during snowmelt season).
• The hybrid model particularly excelled under medium and high flow conditions, where snowmelt and glacier melt contributions are significant, while the LSTM model performed relatively better under low flow conditions.
• The model's performance varied seasonally, achieving highest accuracy in winter for rainfall-dominated basins and consistent performance across all seasons in snowmelt-dominated regions.

Contributions

• Introduces a novel, physically-guided multi-component hybrid model that effectively integrates data-driven (LSTM) and empirical/physical (SRM, degree-day) models for comprehensive streamflow forecasting.
• Explicitly accounts for rainfall, snowmelt, and glacier melt contributions, addressing a critical gap in hydrological modeling, especially in cold regions.
• Demonstrates significant improvements in streamflow forecasting accuracy and stability compared to standalone LSTM models, particularly in snowmelt-dominated catchments and during medium to high flow events.
• Provides a valuable framework for applications requiring a clear distinction among water sources, such as hydropower inflow forecasting, water allocation management, and climate impact assessment.

Funding

• Major Science and Technology Projects of Qinghai Province (2021-SF-A6)
• National Natural Science Foundation (42207070)
• National Key Research and Development Program (2023YFC3206700)
• National Natural Science Foundation of China (U2243232)

Citation

@article{Chu2026Streamflow,
  author = {Chu, Haibo and Jiang, Yulin and Zhang, Wei and Wei, Jiahua},
  title = {Streamflow Forecasting Using a Hybrid Modelling Coupled with Different Components},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-025-04395-0},
  url = {https://doi.org/10.1007/s11269-025-04395-0}
}