Scholarlink

<p>Deep analysis of aquaculture water parameters can improve both the accuracy and efficiency of dissolved oxygen (DO) prediction and help identify key environmental factors essential for precision aquaculture management. To enhance short-term forecasting performance, this study proposes an adaptive graph convolutional–improved long short-term memory (AGCN-ILSTM) model optimized using an improved Superb Fairy-wren Optimization Algorithm (ISFOA). The ISFOA employs a hierarchical learning mechanism to strengthen global exploration and convergence stability. The AGCN-ILSTM model integrates adaptive graph convolutional networks with an improved recurrent learning structure, enabling it to capture complex spatial–temporal dependencies between DO and environmental parameters. The proposed model provides 3-h-ahead predictions and is validated using measured seawater DO data collected from an aquaculture zone in Huguang Town, Mazhang District, Zhanjiang City, Guangdong Province, China. Results show that the model achieves NSE, LMI, and WI values of 0.958, 0.978, and 0.989, respectively, with RMSE and MAE as low as 0.084&#xa0;mg/L and 0.069&#xa0;mg/L—significantly outperforming conventional models such as SMI-TCN-BiLSTM, WD-MIC-PSO-SVR, and NGO-CNN-GRU. These findings demonstrate that the ISFOA-AGCN-ILSTM model can accurately perform short-term seawater DO prediction and provide effective support for intelligent and sustainable aquaculture management.</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 ISFOA-AGCN-ILSTM model for short-term prediction of dissolved oxygen in aquaculture systems

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Kun Song,
  • Yuqiang Yang,
  • Huanzhi Luo

摘要

Deep analysis of aquaculture water parameters can improve both the accuracy and efficiency of dissolved oxygen (DO) prediction and help identify key environmental factors essential for precision aquaculture management. To enhance short-term forecasting performance, this study proposes an adaptive graph convolutional–improved long short-term memory (AGCN-ILSTM) model optimized using an improved Superb Fairy-wren Optimization Algorithm (ISFOA). The ISFOA employs a hierarchical learning mechanism to strengthen global exploration and convergence stability. The AGCN-ILSTM model integrates adaptive graph convolutional networks with an improved recurrent learning structure, enabling it to capture complex spatial–temporal dependencies between DO and environmental parameters. The proposed model provides 3-h-ahead predictions and is validated using measured seawater DO data collected from an aquaculture zone in Huguang Town, Mazhang District, Zhanjiang City, Guangdong Province, China. Results show that the model achieves NSE, LMI, and WI values of 0.958, 0.978, and 0.989, respectively, with RMSE and MAE as low as 0.084 mg/L and 0.069 mg/L—significantly outperforming conventional models such as SMI-TCN-BiLSTM, WD-MIC-PSO-SVR, and NGO-CNN-GRU. These findings demonstrate that the ISFOA-AGCN-ILSTM model can accurately perform short-term seawater DO prediction and provide effective support for intelligent and sustainable aquaculture management.