Scholarlink

<p>Hydrological systems exhibit increasingly nonlinear and abrupt responses to climate variability and land use change, yet traditional models struggle to reproduce such dynamics due to parameter rigidity, data scarcity, and substantial uncertainty. We introduce VMD-CT-LSTM, integrating Variational Mode Decomposition (VMD), a correlation test (CT), and Long Short-Term Memory (LSTM) networks, to reduce input redundancy and improve computational efficiency. The model was evaluated on monthly runoff at three stations in the Weihe River basin and evaluated against standard LSTM and VMD-LSTM models. <i>NSE</i> values at all three stations exceeded 0.94, surpassing VMD-LSTM and substantially outperforming LSTM, particularly in capturing abrupt runoff fluctuations, a scenario where VMD-LSTM showed pronounced underestimation. Compared with VMD-LSTM, VMD-CT-LSTM reduced input variables by 64.3–85.7%, decreased the average runtime per mode by 63.7%, and cut total runtime from 246.08&#xa0;s to 89.07&#xa0;s. Residual bias was primarily attributed to <i>QIMF</i><sub><i>1</i></sub>, the low-frequency mode representing the long-term runoff trend, whose simulation was hampered by gradually evolving forcings including reservoir regulation, land-use change, and long-term precipitation decline. The mid-frequency modes (<i>QIMF</i><sub><i>2</i></sub>–<i>QIMF</i><sub><i>5</i></sub>) and the high-frequency residual (<i>QIMF</i><sub><i>r</i></sub>), associated with seasonal-to-interannual climate variability and short-term stochastic disturbances respectively, achieved comparatively higher simulation accuracy. VMD-CT-LSTM offers a computationally efficient and framework for operational runoff forecasting in basins undergoing rapid environmental change.</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

Deep Learning-Based Monthly Runoff Simulation in Changing Environments: Enhancing Accuracy by Reducing Data Redundancy

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Shenghong Liu,
  • Ji Liu,
  • Shaokang Yang,
  • Zhenxin Bao,
  • Shan-e-hyder Soomro,
  • Wenjuan Chang,
  • Qingxia Lin,
  • Qirui Zhang,
  • Gantao Yang,
  • Ruilin Geng,
  • Haiyu Wang,
  • Longteng Dai

摘要

Hydrological systems exhibit increasingly nonlinear and abrupt responses to climate variability and land use change, yet traditional models struggle to reproduce such dynamics due to parameter rigidity, data scarcity, and substantial uncertainty. We introduce VMD-CT-LSTM, integrating Variational Mode Decomposition (VMD), a correlation test (CT), and Long Short-Term Memory (LSTM) networks, to reduce input redundancy and improve computational efficiency. The model was evaluated on monthly runoff at three stations in the Weihe River basin and evaluated against standard LSTM and VMD-LSTM models. NSE values at all three stations exceeded 0.94, surpassing VMD-LSTM and substantially outperforming LSTM, particularly in capturing abrupt runoff fluctuations, a scenario where VMD-LSTM showed pronounced underestimation. Compared with VMD-LSTM, VMD-CT-LSTM reduced input variables by 64.3–85.7%, decreased the average runtime per mode by 63.7%, and cut total runtime from 246.08 s to 89.07 s. Residual bias was primarily attributed to QIMF1, the low-frequency mode representing the long-term runoff trend, whose simulation was hampered by gradually evolving forcings including reservoir regulation, land-use change, and long-term precipitation decline. The mid-frequency modes (QIMF2–QIMF5) and the high-frequency residual (QIMFr), associated with seasonal-to-interannual climate variability and short-term stochastic disturbances respectively, achieved comparatively higher simulation accuracy. VMD-CT-LSTM offers a computationally efficient and framework for operational runoff forecasting in basins undergoing rapid environmental change.