A spatiotemporal dissolved oxygen prediction model based on graph attention networks suitable for missing data

Area of Science:

• Environmental Science
• Water Quality Monitoring
• Data Science

Background:

• Accurate dissolved oxygen (DO) prediction is vital for water pollution control.
• Existing models struggle with missing data and capturing complex spatiotemporal dynamics.
• Effective water resource management requires robust forecasting tools.

Purpose of the Study:

• To develop a novel spatiotemporal prediction model for dissolved oxygen (DO) concentration.
• To address the challenge of missing data in water quality time series.
• To improve the accuracy and robustness of DO prediction models.

Main Methods:

• Utilized neural controlled differential equations (NCDEs) for effective handling of missing data.
• Employed graph attention networks (GATs) to capture complex spatiotemporal relationships in DO data.
• Enhanced model performance through iterative optimization, feature selection (SHAP), and a fusion graph attention mechanism.

Main Results:

• The proposed model demonstrated superior long-term prediction accuracy (step=18) compared to other models.
• Achieved a Mean Absolute Error (MAE) of 0.194, Nash-Sutcliffe Efficiency (NSE) of 0.914, Relative Absolute Error (RAE) of 0.219, and Index of Agreement (IA) of 0.977.
• Validated using water quality data from Hunan Province, China (Jan 2021 - Jun 2022).

Conclusions:

• Constructing appropriate spatial dependencies significantly enhances DO prediction accuracy.
• The NCDE module provides robustness to missing data, a common issue in environmental monitoring.
• The developed model offers a reliable tool for water quality management and pollution control.