Zuo et al. (2026) Exploring viable approaches for long-term seasonal streamflow forecasting under different forcing mechanisms

Identification

• Journal: Stochastic Environmental Research and Risk Assessment
• Year: 2026
• Date: 2026-01-01
• Authors: Ganggang Zuo, Tingting Wang, Yani Lian, Ni Wang, Jiancang Xie
• DOI: 10.1007/s00477-025-03141-7

Research Groups

State Key Laboratory of Water Engineering Ecology and Environment in Arid Area, Xi’an University of Technology, Xi’an, China.

Short Summary

This study systematically evaluates autoregressive, data-driven, process-driven, and hybrid modeling approaches for long-term seasonal streamflow forecasting (up to 12 months lead time) in snow-dominated catchments. It finds that simple autoregressive models and climate forcings derived from hydrological similarity years significantly outperform GCM-driven projections, offering more practical and reliable predictions for water resource management.

Objective

• To develop and evaluate a comprehensive framework of autoregressive, data-driven, process-driven, and hybrid modeling approaches for long-term seasonal streamflow forecasting (up to 12 months lead time).
• To identify practical and accurate approaches for predicting the monthly streamflow process in large, snow-dominated catchments.
• To evaluate forecasting paradigms using both standard Coupled Model Intercomparison Project (CMIP) projections and novel climate forcings derived from historical data using a newly developed comprehensive similarity evaluation index.

Study Configuration

• Spatial Scale: Three snow-dominated catchments in the upper reaches of the Yellow River, China (Tangnaihai, Guide, and Xunhua stations). The study area spans 145,459 square kilometers with elevations above 3000 meters.
• Temporal Scale: Long-term seasonal streamflow forecasting with lead times up to 12 months. The experimental dataset covers monthly observed streamflow and daily climate data from January 1969 to December 2019. Calibration period: January 1969 to December 2009. Validation period: January 2010 to December 2014. Forecasting period: January 2015 to December 2019.

Methodology and Data

• Models used:
  • Autoregressive: SARIMA (Seasonal Autoregressive Integrated Moving Average)
  • Process-driven: SWAT+ (Soil and Water Assessment Tool Plus)
  • Data-driven: XGBoost (eXtreme Gradient Boosting), LSTM (Long Short-Term Memory)
  • Hybrid: SWAT+ XGBoost, SWAT+ LSTM
  • Similarity Index (SI) combining Euclidean Distance (ED), Dynamic Time Warping (DTW), Fréchet Distance (FD), Hausdorff Distance (HD), Longest Common Subsequence Similarity (LCSS), and Cosine Similarity (CS).
  • Variance Inflation Factor (VIF) for multicollinearity analysis.
  • Bayesian Optimization (BO) for hyperparameter tuning of XGBoost and LSTM.
  • Muskingum method for flow routing and Penman–Monteith method for potential evapotranspiration (within SWAT+).
• Data sources:
  • Observed monthly streamflow data from Tangnaihai, Guide, and Xunhua stations (Yellow River Network, Hydrological Yearbook).
  • Historical daily climate forcings (precipitation, maximum and minimum temperatures, relative humidity, average wind speed, solar radiation) from 13 meteorological stations (China Meteorological Data Service Center).
  • CMIP6 climate projections from ACCESS-ESM1.5 (daily precipitation, maximum and minimum near-surface air temperature, near-surface relative humidity, surface downwelling shortwave radiation, near-surface wind speed) for SSP126, SSP245, and SSP585 scenarios (2015-2050).
  • Digital Elevation Model (DEM) data (30 meter resolution) from Shuttle Radar Topography Mission.
  • Soil map (30 arc-second resolution) from Harmonized World Soil Database (HWSD v1.2).
  • Land use data (1 kilometer resolution) from the Resource and Environmental Science Data Platform of China.

Main Results

• Simple autoregressive models (SARIMA) provided a robust baseline, consistently achieving a Normalized Nash–Sutcliffe Efficiency (NNSE) exceeding 0.75 across both simulation and forecasting periods, and demonstrating the best overall forecasting capability.
• Climate data from hydrological similarity years offered a superior alternative to GCM-driven forecasting, outperforming CMIP projections with NNSE improvements of up to 26.28% and NRMSE reductions of up to 21.82%.
• Data-driven and hybrid approaches consistently outperformed the standalone process-driven model (SWAT+), achieving superior predictive skill with an average NNSE of 0.63.
• XGBoost models (standalone and hybrid) generally exhibited better simulation and forecasting performance compared to LSTM models.
• All forecasting schemes struggled to accurately capture streamflow dynamics during flood events (July and September peaks) compared to non-flood periods, indicated by significantly higher Peak Percentage Threshold Statistic (PPTS) values than Low Percentage Threshold Statistic (LPTS) values.
• SSP-forced schemes, despite bias correction, tended to significantly overestimate actual streamflow, while raw SSP precipitation often underestimated observed rainfall.
• Hydrological similarity-based climate data yielded more accurate forecasts compared to meteorological similarity-based data, highlighting the importance of a global perspective on flow memory effects in large watersheds.

Contributions

• Development and application of a comprehensive similarity evaluation index to generate robust, observation-based climate forcings.
• Benchmarking of a wide array of modeling paradigms (autoregressive, process-driven, data-driven, and physics-informed hybrid models) within a unified framework.
• Rigorous re-evaluation of the utility of standard CMIP projections against the similarity-based forcing alternative.
• Identification of operationally viable forecasting schemes capable of simultaneously predicting the entire 12-month streamflow process.

Funding

• National Natural Science Foundation of China (Grant Nos. 52309034)
• China Postdoctoral Science Foundation (2022M722561)
• Water Conservancy Science and Technology Project of Shaanxi Province (Program No. 2025slkj-10)

Citation

@article{Zuo2026Exploring,
  author = {Zuo, Ganggang and Wang, Tingting and Lian, Yani and Wang, Ni and Xie, Jiancang},
  title = {Exploring viable approaches for long-term seasonal streamflow forecasting under different forcing mechanisms},
  journal = {Stochastic Environmental Research and Risk Assessment},
  year = {2026},
  doi = {10.1007/s00477-025-03141-7},
  url = {https://doi.org/10.1007/s00477-025-03141-7}
}