Multi-scale dynamic graph neural network for PM2.5 concentration prediction in regional station cluster

Area of Science:

• Environmental Science
• Data Science
• Artificial Intelligence

Background:

• Accurate PM2.5 forecasting is vital for public health and environmental management.
• Existing methods struggle with complex spatiotemporal dependencies, especially with sparse monitoring data.
• Addressing these limitations is crucial for effective air quality management.

Purpose of the Study:

• To develop an advanced model for precise PM2.5 concentration prediction.
• To enhance the capture of multi-scale spatiotemporal dependencies in air quality data.
• To improve forecasting accuracy in areas with limited monitoring infrastructure.

Main Methods:

• Proposed a Multi-Scale Dynamic Graph Neural Network (MSDGNN) for PM2.5 forecasting.
• Incorporated multi-scale temporal modeling (hourly, daily, weekly) and dynamic station grouping.
• Utilized multi-head attention, spatiotemporal graph attention, adaptive adjacency matrices, and Chebyshev graph convolutions.

Main Results:

• MSDGNN demonstrated superior performance in PM2.5 forecasting compared to baseline models.
• Achieved a 6.77% reduction in Mean Absolute Error (MAE).
• Achieved an 8.67% reduction in Root Mean Square Error (RMSE).

Conclusions:

• The MSDGNN model effectively learns complex spatiotemporal dependencies for accurate PM2.5 prediction.
• The model shows significant improvements, especially in diverse spatiotemporal conditions and sparse data environments.
• This approach offers a robust solution for air quality forecasting and environmental management.