Zhang et al. (2026) Asymmetric Feature Reconstruction and Improved Transformer for Multi-step River Streamflow Prediction

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-01-01
• Authors: Yinghao Zhang, Zhanying Li, Yu Gao, Wenhao Fu, Mingyu Wang
• DOI: 10.1007/s11269-025-04470-6

Research Groups

School of Information Science and Engineering, Dalian Polytechnic University, Dalian, China

Short Summary

This study proposes a novel method integrating asymmetric feature reconstruction with an enhanced Transformer architecture for multi-step streamflow prediction of the Yangtze River. The approach significantly improves predictive accuracy by over 67% in Mean Absolute Error compared to the strongest baseline, demonstrating superior robustness across various prediction horizons.

Objective

• To develop an improved deep learning model for accurate multi-step river streamflow prediction by enhancing input feature quality and strengthening the Transformer's logical inference capabilities.

Study Configuration

• Spatial Scale: Jiujiang hydrological station in the Yangtze River Basin, China. Precipitation data from 50 locations within 90°–117° E and 26°–34° N.
• Temporal Scale: 9 years (2015–2023) of data with a 1-hour resolution. Multi-step predictions for horizons of 6, 9, 12, and 15 hours.

Methodology and Data

• Models used:
  • Proposed Model: Enhanced Transformer architecture incorporating Long Short-term Memory (LSTM) and Dilated Causal Convolution (DCC) modules.
  • Feature Reconstruction: Kernel Principal Components Analysis (KPCA) for high-dimensional precipitation features and Variational Mode Decomposition (VMD) for one-dimensional streamflow features.
  • Baselines: CNN-Transformer, Transformer, CNN-Attention-LSTM, CNN-LSTM, LSTM.
• Data sources:
  • Streamflow data: Jiujiang station, provided by the Water Resources Bureau of Hubei Province.
  • Precipitation data: NASA publicly available datasets for the Yangtze River Basin.

Main Results

• The proposed model achieved substantial accuracy gains over baseline models.
• Compared to the strongest baseline (CNN-Transformer), the average Mean Absolute Error (MAE) decreased by:
  • 69.12% for a 6-hour prediction length.
  • 67.79% for a 9-hour prediction length.
  • 68.30% for a 12-hour prediction length.
  • 67.57% for a 15-hour prediction length.
• The average Mean Absolute Percentage Error (MAPE) decreased by:
  • 70.24% for a 6-hour prediction length.
  • 68.51% for a 9-hour prediction length.
  • 68.47% for a 12-hour prediction length.
  • 67.77% for a 15-hour prediction length.
• The model consistently outperformed other models across all evaluation periods and demonstrated superior robustness, exhibiting the smallest MAPE variation across different months.
• Nash-Sutcliffe Efficiency (NSE) values remained above 0.97, even during periods of minimal streamflow variability.

Contributions

• First to reconstruct two classes of asymmetrically dimensioned features (precipitation via KPCA, streamflow via VMD) as inputs for streamflow forecasting models.
• Achieved multivariate feature fusion using an attention mechanism, which mitigates feature anisotropy and reduces overfitting.
• Redesigned the Transformer architecture to enhance its capacity for logical reasoning and to better capture relationships among multiple features.
• Integrated LSTM and Dilated Causal Convolution (DCC) modules into the Transformer to strengthen its ability to capture temporal variations and infer hydrological dynamics.
• Provided a systematic investigation and analysis of streamflow forecasting performance across various prediction horizons.

Funding

Special Fund for Basic Scientific Research Operations of Liaoning Provincial Universities (No.LJ212510152029).

Citation

@article{Zhang2026Asymmetric,
  author = {Zhang, Yinghao and Li, Zhanying and Gao, Yu and Fu, Wenhao and Wang, Mingyu},
  title = {Asymmetric Feature Reconstruction and Improved Transformer for Multi-step River Streamflow Prediction},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-025-04470-6},
  url = {https://doi.org/10.1007/s11269-025-04470-6}
}