Scholarlink

Blue-green algae (BGA) blooms are a common and harmful occurrence in many water bodies. In this study, a spatial–temporal deep learning prediction model based on Graph Convolutional Network and Gated Recurrent Unit (GCN-GRU) is utilized to predict BGA concentration based on water quality data gathered by an unmanned surface vehicle (USV). To solve the uneven distribution of the water quality sampling positions of USV, the Kriging algorithm is applied to structure the water quality data into a graphical distribution, which can be further input into our proposed GCN-GRU deep learning network to predict short-term BGA concentration. We compare the prediction accuracy of our algorithm with the other four prediction benchmarks (HA, SVR, GRU, LTSM) in four metrics (RMSE, MAE, MAPE, R-Squared). Experimental results indicate that our GCN-GRU model has the best performance in predicting the spatial–temporal distribution of BGA in each metric.

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

Algal Bloom Prediction Based on Graph Convolutional Network and Gated Recurrent Unit Deep Neural Network and Massive Spatial–temporal Water Quality Data

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yixuan Hou,
  • Zixian Zhang,
  • Yichen Wei,
  • Ruoxuan Cao,
  • Yihang Xu,
  • Yong Yue,
  • Ruyu Yan,
  • Xiaohui Zhu

摘要

Blue-green algae (BGA) blooms are a common and harmful occurrence in many water bodies. In this study, a spatial–temporal deep learning prediction model based on Graph Convolutional Network and Gated Recurrent Unit (GCN-GRU) is utilized to predict BGA concentration based on water quality data gathered by an unmanned surface vehicle (USV). To solve the uneven distribution of the water quality sampling positions of USV, the Kriging algorithm is applied to structure the water quality data into a graphical distribution, which can be further input into our proposed GCN-GRU deep learning network to predict short-term BGA concentration. We compare the prediction accuracy of our algorithm with the other four prediction benchmarks (HA, SVR, GRU, LTSM) in four metrics (RMSE, MAE, MAPE, R-Squared). Experimental results indicate that our GCN-GRU model has the best performance in predicting the spatial–temporal distribution of BGA in each metric.