Scholarlink

<p>Accurate streamflow prediction remains both essential and challenging in hydrology. This study proposes a method that integrates asymmetric feature reconstruction with an enhanced Transformer architecture for multi-step forecasting of the Yangtze River’s streamflow. We combine two datasets and construct a multivariate feature space to better represent the mechanisms underlying hydrological variability, enabling the deep learning model to extract and learn these dynamics more effectively. Our approach improves predictive accuracy through two main strategies. First, we enhance the quality of the input features by using Kernel Principal Components Analysis (KPCA) and Variational Mode Decomposition (VMD) to reconstruct two classes of features with inherently asymmetric dimensions. Second, we redesign the Transformer to strengthen its capacity for logical inference. The modified architecture improves the modeling of relationships across multiple features, and we incorporate Long Short-term Memory (LSTM) and Dilated Causal Convolution (DCC) modules to better capture temporal variation. Experimental results show that the proposed model yields substantial accuracy gains over baseline models. Compared with the strongest baseline, the Convolutional Neural Network (CNN)-Transformer, the average Mean Absolute Error(MAE) decreases by 69.12%, 67.79%, 68.30%, and 67.57% for prediction lengths of 6, 9, 12, and 15 steps, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Asymmetric Feature Reconstruction and Improved Transformer for Multi-step River Streamflow Prediction

  • 期刊论文

  • 发表日期：

    2026年1月10日

查看全文
  • Yinghao Zhang,
  • Zhanying Li,
  • Yu Gao,
  • Wenhao Fu,
  • Mingyu Wang

摘要

Accurate streamflow prediction remains both essential and challenging in hydrology. This study proposes a method that integrates asymmetric feature reconstruction with an enhanced Transformer architecture for multi-step forecasting of the Yangtze River’s streamflow. We combine two datasets and construct a multivariate feature space to better represent the mechanisms underlying hydrological variability, enabling the deep learning model to extract and learn these dynamics more effectively. Our approach improves predictive accuracy through two main strategies. First, we enhance the quality of the input features by using Kernel Principal Components Analysis (KPCA) and Variational Mode Decomposition (VMD) to reconstruct two classes of features with inherently asymmetric dimensions. Second, we redesign the Transformer to strengthen its capacity for logical inference. The modified architecture improves the modeling of relationships across multiple features, and we incorporate Long Short-term Memory (LSTM) and Dilated Causal Convolution (DCC) modules to better capture temporal variation. Experimental results show that the proposed model yields substantial accuracy gains over baseline models. Compared with the strongest baseline, the Convolutional Neural Network (CNN)-Transformer, the average Mean Absolute Error(MAE) decreases by 69.12%, 67.79%, 68.30%, and 67.57% for prediction lengths of 6, 9, 12, and 15 steps, respectively.