An improved graph convolutional network with feature and temporal attention for multivariate water quality...

Area of Science:

• Environmental Science
• Data Science
• Artificial Intelligence

Background:

• Accurate water quality analysis is crucial for effective water management and pollution control.
• Existing prediction methods often focus on single indicators and struggle with multivariate data, failing to capture inter-indicator correlations.

Purpose of the Study:

• To develop an advanced deep learning model for accurate prediction of multivariate water quality data.
• To address the limitations of conventional models in capturing correlations and temporal dependencies in water quality indicators.

Main Methods:

• Proposed a novel deep learning model: Graph Convolutional Network with Feature and Temporal Attention (FTGCN).
• Incorporated a feature attention mechanism (multi-head self-attention) to identify indicator correlations.
• Developed a temporal prediction module (temporal convolution, bidirectional GRU, temporal attention) for time series dependencies.
• Integrated adaptive graph learning and an auto-regression module to capture hidden associations and non-linearities.
• Optimized model parameters using an evolutionary algorithm.

Main Results:

• The FTGCN model demonstrated superior performance in multivariate water quality prediction compared to existing models.
• Experimental results on four real-world datasets validated the model's effectiveness in forecasting water quality indicators.
• The model successfully captured complex correlations and temporal dynamics within the water quality data.

Conclusions:

• The FTGCN model offers a significant advancement in multivariate water quality prediction.
• This approach enhances water quality management and pollution control strategies through more accurate forecasting.
• The model's ability to handle complex inter-indicator relationships and temporal dependencies makes it a valuable tool for environmental monitoring.