A PM_2.5 spatiotemporal prediction model based on mixed graph convolutional GRU and self-attention network

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Rising atmospheric pollution necessitates accurate predictive models.
• Existing models struggle with incorporating data from local/neighboring stations and long-term historical patterns.
• Particulate Matter 2.5 (PM2.5) poses significant health risks, demanding precise forecasting.

Purpose of the Study:

• To develop an advanced spatiotemporal prediction model for PM2.5 concentrations.
• To effectively integrate short-term spatiotemporal dependencies and long-term temporal patterns.
• To improve multi-station, multi-time step PM2.5 forecasting accuracy.

Main Methods:

• Proposed MGCGRU-SAN model combining Mixed Graph Convolutional GRU (MGCGRU) and Self-Attention Network (SAN) modules.
• MGCGRU captures short-term spatiotemporal dependencies across stations.
• SAN analyzes segmented long-term historical PM2.5 data to identify temporal patterns.

Main Results:

• The MGCGRU-SAN model demonstrated significant improvements in multi-step predictions.
• Achieved 9.62% (RSE), 6.33% (MAE), and 4.98% (RMSE) improvements over baseline models.
• Outperformed best baseline by an average of 8.34% (RSE), 6.12% (MAE), 4.06% (RMSE), and 2.60% (Correlation).

Conclusions:

• The MGCGRU-SAN model effectively captures both short-term and long-term patterns for superior PM2.5 prediction.
• The integration of spatiotemporal and long-term temporal analysis enhances forecasting accuracy.
• The model shows strong performance for multi-station, long-term PM2.5 concentration predictions.