Scholarlink

<p>Accurate water quality forecasting is essential for reducing operational costs, ensuring sustainable water resource management, and maintaining ecosystem health. Most existing models fall short in capturing the interdependencies among multiple water quality parameters. To address this challenge, a forecasting framework is introduced that integrates Adaptive Gaussian Weighted Smoothing (AGWS), Graph Convolutional Networks (GCN), Convolutional Multi-head Self-Attention (CMSA), and the Mamba architecture, referred to as AGWS-GAMamba. AGWS is first applied to dynamically denoise raw water quality data, enhancing the robustness of the input sequences. GCN is then employed to model the interrelationships among water quality variables, enabling the effective utilization of feature correlations. Subsequently, local temporal dependencies are extracted via convolution-enhanced self-attention, while long-term dependencies are captured using a selective state-space mechanism implemented by the Mamba module, resulting in efficient and scalable sequence modeling. The proposed model is evaluated using water quality datasets collected from multiple monitoring stations in the Chesapeake Bay, United States. Experimental results demonstrate that the proposed model achieves superior performance in both predictive accuracy and robustness compared to long short-term memory (LSTM), Autoformer, and other hybrid methods, underscoring its effectiveness in modeling multivariate dependencies and long-range temporal dynamics.</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 Hybrid Mamba Architecture with Graph Convolution and Convolutional Self-Attention for Multivariate Water Quality Forecasting

  • 期刊论文

  • 发表日期：

    2025年7月28日

查看全文
  • Yulong Bai,
  • Xianbao Tan,
  • Xiaoxin Yue

摘要

Accurate water quality forecasting is essential for reducing operational costs, ensuring sustainable water resource management, and maintaining ecosystem health. Most existing models fall short in capturing the interdependencies among multiple water quality parameters. To address this challenge, a forecasting framework is introduced that integrates Adaptive Gaussian Weighted Smoothing (AGWS), Graph Convolutional Networks (GCN), Convolutional Multi-head Self-Attention (CMSA), and the Mamba architecture, referred to as AGWS-GAMamba. AGWS is first applied to dynamically denoise raw water quality data, enhancing the robustness of the input sequences. GCN is then employed to model the interrelationships among water quality variables, enabling the effective utilization of feature correlations. Subsequently, local temporal dependencies are extracted via convolution-enhanced self-attention, while long-term dependencies are captured using a selective state-space mechanism implemented by the Mamba module, resulting in efficient and scalable sequence modeling. The proposed model is evaluated using water quality datasets collected from multiple monitoring stations in the Chesapeake Bay, United States. Experimental results demonstrate that the proposed model achieves superior performance in both predictive accuracy and robustness compared to long short-term memory (LSTM), Autoformer, and other hybrid methods, underscoring its effectiveness in modeling multivariate dependencies and long-range temporal dynamics.