Scholarlink

<p>Accurate daily streamflow forecasting is important for water-resources allocation, flood-risk assessment, and watershed management. This study proposes a variational mode decomposition-based dual-stream temporal convolutional network with gated attention (VMD-DSTCN-GA) to separately represent internal runoff dynamics and external meteorological forcing. Historical runoff is decomposed into frequency-specific components and encoded through a runoff-state stream comprising causal dilated convolutions and residual blocks, while precipitation, maximum and minimum temperature, solar radiation, relative humidity, and wind speed are processed through an independent meteorological-forcing stream. The two feature streams are adaptively integrated using gated fusion and temporal attention, and a peak-sensitive loss is introduced to improve the representation of high-flow events. The model was evaluated using daily hydro-meteorological observations from 1990 to 2019 at the Jingcun hydrological station, with 2010–2019 reserved for independent evaluation. VMD-DSTCN-GA achieved an R² of 0.9872, an RMSE of 7.3734&#xa0;m³/s, a PBIAS of 13.0314%, and an NSE of 0.9631. It obtained the highest R² among all evaluated models, indicating the strongest ability to reproduce observed runoff variability. CNN–LSTM–Attention achieved the lowest RMSE of 6.6113&#xa0;m³/s and the highest NSE of 0.9703, whereas SWAT produced the smallest absolute PBIAS. Monthly aggregation further improved VMD-DSTCN-GA performance, yielding an R² of 0.9929, an NSE of 0.9660, and an RMSE of 6.1879&#xa0;m³/s. These results demonstrate that the proposed framework effectively reproduces runoff variability, although systematic volume bias persists and its cross-basin transferability remains to be evaluated.</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

Daily Streamflow Forecasting using a VMD-Based Dual-Stream Temporal Convolutional Network with Gated Attention

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • Hongye Cao

摘要

Accurate daily streamflow forecasting is important for water-resources allocation, flood-risk assessment, and watershed management. This study proposes a variational mode decomposition-based dual-stream temporal convolutional network with gated attention (VMD-DSTCN-GA) to separately represent internal runoff dynamics and external meteorological forcing. Historical runoff is decomposed into frequency-specific components and encoded through a runoff-state stream comprising causal dilated convolutions and residual blocks, while precipitation, maximum and minimum temperature, solar radiation, relative humidity, and wind speed are processed through an independent meteorological-forcing stream. The two feature streams are adaptively integrated using gated fusion and temporal attention, and a peak-sensitive loss is introduced to improve the representation of high-flow events. The model was evaluated using daily hydro-meteorological observations from 1990 to 2019 at the Jingcun hydrological station, with 2010–2019 reserved for independent evaluation. VMD-DSTCN-GA achieved an R² of 0.9872, an RMSE of 7.3734 m³/s, a PBIAS of 13.0314%, and an NSE of 0.9631. It obtained the highest R² among all evaluated models, indicating the strongest ability to reproduce observed runoff variability. CNN–LSTM–Attention achieved the lowest RMSE of 6.6113 m³/s and the highest NSE of 0.9703, whereas SWAT produced the smallest absolute PBIAS. Monthly aggregation further improved VMD-DSTCN-GA performance, yielding an R² of 0.9929, an NSE of 0.9660, and an RMSE of 6.1879 m³/s. These results demonstrate that the proposed framework effectively reproduces runoff variability, although systematic volume bias persists and its cross-basin transferability remains to be evaluated.