Scholarlink

<p>The complex, nonlinear, and bidirectional flow dynamics of urban river networks in estuarine regions pose significant challenges for traditional control strategies, which are further compounded by scarce monitoring data and high computational demands. To address these limitations, this study proposes a novel hybrid feedforward-feedback adaptive control framework (FMPC-AF) that integrates long short-term memory (LSTM) networks with a condition-aware switching mechanism between proportional-integral (PI) feedback and model predictive control (MPC). A high-fidelity SWMM model is first employed to generate a comprehensive training dataset, enabling the LSTM surrogate model to achieve high predictive accuracy even in data-scarce urban settings. The framework then decouples the control problem by using offline analysis to activate computationally intensive MPC only during critical tidal windows, while relying on efficient PI control during quiescent periods. The proposed methodology is applied to the Baima River network, a representative looped urban system. Results demonstrate that the LSTM model provides highly reliable 24-hour forecasts (NSE &gt; 0.985, MSE &lt; 0.008). Furthermore, the FMPC-AF framework rapidly obtains feasible solutions under limited computational resources and, importantly, exhibits strong robustness to variations in the control step length (a key practical advantage over conventional MPC). By addressing the critical gaps of data scarcity, computational burden, and sensitivity to control parameters, this framework offers an effective and practical solution for real-time water level regulation and ecological management in complex urban river networks.</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

A Novel Framework for Urban Water System Prediction and Regulation Based on LSTM Neural Networks

  • 期刊论文

  • 发表日期：

    2026年4月23日

查看全文
  • Minghan Yang,
  • Shang Gao,
  • Chao Wang,
  • Min Cai,
  • Lina Zhang,
  • Hao Wang,
  • Yongwei Shao

摘要

The complex, nonlinear, and bidirectional flow dynamics of urban river networks in estuarine regions pose significant challenges for traditional control strategies, which are further compounded by scarce monitoring data and high computational demands. To address these limitations, this study proposes a novel hybrid feedforward-feedback adaptive control framework (FMPC-AF) that integrates long short-term memory (LSTM) networks with a condition-aware switching mechanism between proportional-integral (PI) feedback and model predictive control (MPC). A high-fidelity SWMM model is first employed to generate a comprehensive training dataset, enabling the LSTM surrogate model to achieve high predictive accuracy even in data-scarce urban settings. The framework then decouples the control problem by using offline analysis to activate computationally intensive MPC only during critical tidal windows, while relying on efficient PI control during quiescent periods. The proposed methodology is applied to the Baima River network, a representative looped urban system. Results demonstrate that the LSTM model provides highly reliable 24-hour forecasts (NSE > 0.985, MSE < 0.008). Furthermore, the FMPC-AF framework rapidly obtains feasible solutions under limited computational resources and, importantly, exhibits strong robustness to variations in the control step length (a key practical advantage over conventional MPC). By addressing the critical gaps of data scarcity, computational burden, and sensitivity to control parameters, this framework offers an effective and practical solution for real-time water level regulation and ecological management in complex urban river networks.