Wang et al. (2026) A Hybrid Multi-Strategy Monthly Runoff Forecasting Model Based on Parallel CNN-GRU Architecture, SSA Optimization, and Error Correction Mechanisms

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-01-01
• Authors: Liyuan Wang, Xuebin Wang, Jianxia Chang, Xuejiao Meng, Yimin Wang, Chengqing Ren, Junhao Zhang
• DOI: 10.1007/s11269-025-04457-3

Research Groups

State Key Laboratory of Eco-Hydraulics in Northwest Arid Region of China, Xi’an University of Technology, Xi’an, 710048, China

Short Summary

This study proposes SVPsEC, a novel hybrid multi-strategy model integrating Variational Mode Decomposition (VMD), parallel CNN-GRU architecture, Sparrow Search Algorithm (SSA) optimization, and an error correction mechanism, to enhance monthly runoff forecasting accuracy and stability in non-stationary hydrological systems. Evaluations at three hydrological stations demonstrate that SVPsEC consistently produces highly accurate forecasts, significantly improving the prediction of peak flow events compared to benchmark methods.

Objective

• To address the significant challenges posed by the high nonlinearity and non-stationarity of hydrological processes in reliable medium- to long-term runoff forecasting.
• To develop an integrated forecasting framework (SVPsEC) that embeds multi-module collaboration and cross-subsequence spatiotemporal coupling to systematically enhance the accuracy and robustness of monthly runoff predictions.

Study Configuration

• Spatial Scale: Wei River Basin, China, specifically at three hydrological stations: Beidao (24,871 km² catchment area), Zhangjiashan (43,216 km² drainage area), and Huaxian (106,498 km² management area).
• Temporal Scale: Monthly runoff forecasting, covering medium- to long-term periods.

Methodology and Data

• Models used:
  • SVPsEC (Proposed Model): Integrates Variational Mode Decomposition (VMD), Parallel Convolutional Neural Network (CNN) – Gated Recurrent Unit (GRU) architecture, Sparrow Search Algorithm (SSA) for optimization, and an Error Correction strategy.
  • Benchmark Models: CNN, GRU, PCG (Parallel CNN-GRU), PSO-PCG, SSA-PCG, EMD-PCG, CEEMD-PCG, SVMD-PCG (SVP), SVMD-SPCG (SVPS), SVMD-Transform (SVTF), SVMD-NBEATS (SVNBEATS), SVMD-Inform (SVIF).
• Data sources: Observed monthly runoff series from Beidao, Zhangjiashan, and Huaxian hydrological stations in the Wei River Basin.

Main Results

• The proposed SVPsEC model consistently achieved high accuracy and stability in monthly runoff forecasting across all three hydrological stations.
• SVPsEC demonstrated Nash–Sutcliffe Efficiency Coefficient (NSC) values exceeding 0.98 (0.9830 for Zhangjiashan, 0.9971 for Huaxian, 0.9975 for Beidao) and Coefficient of Determination (R) values above 0.99 (0.9930 for Zhangjiashan, 0.9977 for Huaxian, 0.9984 for Beidao).
• The model significantly reduced Root Mean Square Error (RMSE) compared to benchmark methods.
• Peak flow forecasting performance was notably improved; average errors for the five largest extreme values were reduced by 62.27% at Zhangjiashan, 91.18% at Huaxian, and 88.59% at Beidao, compared to the SVP model.
• The parallel CNN-GRU (PCG) architecture showed significant advantages over standalone CNN and GRU models in capturing spatiotemporal features (e.g., at Zhangjiashan, PCG improved NSC by 59.91% over CNN and 109.31% over GRU).
• SSA optimization exhibited faster convergence and superior global search capability compared to Particle Swarm Optimization (PSO), enhancing model generalization and predictive accuracy.
• Variational Mode Decomposition (VMD) effectively reduced the complexity of runoff series, significantly improving prediction accuracy (e.g., at Zhangjiashan, SVP reduced RMSE by 55.05% and MAE by 58.70% compared to baseline PCG).
• The error correction mechanism effectively suppressed the cumulative propagation of prediction errors, further enhancing forecasting stability.

Contributions

• Proposes a novel multi-strategy hybrid model (SVPsEC) for regional monthly runoff forecasting, integrating VMD, parallel CNN-GRU, SSA optimization, and an error correction mechanism into a unified "decompose–predict–correct" framework.
• Introduces a systematic modeling approach for handling non-stationary hydrological series by leveraging multi-module collaboration and iterative error correction, enhancing model adaptability and robustness.
• Demonstrates the effectiveness of a parallel CNN-GRU architecture in capturing both local spatial features and long-term temporal dependencies from decomposed runoff components.
• Validates the superior performance of the Sparrow Search Algorithm (SSA) for hyperparameter optimization in complex hydrological models.
• Provides a reliable and scalable analytical tool for accurate water resource allocation, hydrological risk management (e.g., flood and drought management), and decision-making in river basins, particularly in data-scarce contexts.
• Significantly improves the capability to predict extreme hydrological events, such as peak runoff.

Funding

• National Key R&D Program of China (No. 2021YFC3000204)
• Natural Science Foundation of China (No. 52109032)

Citation

@article{Wang2026Hybrid,
  author = {Wang, Liyuan and Wang, Xuebin and Chang, Jianxia and Meng, Xuejiao and Wang, Yimin and Ren, Chengqing and Zhang, Junhao},
  title = {A Hybrid Multi-Strategy Monthly Runoff Forecasting Model Based on Parallel CNN-GRU Architecture, SSA Optimization, and Error Correction Mechanisms},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-025-04457-3},
  url = {https://doi.org/10.1007/s11269-025-04457-3}
}