Domain knowledge-enhanced multi-spatial multi-temporal PM_2.5 forecasting with integrated monitoring and reanalysis...

Area of Science:

• Environmental Science and Engineering
• Atmospheric Science
• Data Science and Artificial Intelligence

Background:

• Accurate air quality forecasting is vital for public health, environmental monitoring, and urban planning.
• Existing methods struggle to integrate multi-scale spatial and temporal information, and overlook temporal periodicity.
• There's a gap in connecting individual monitoring stations with city-wide air quality scales.

Purpose of the Study:

• To develop a novel air quality forecasting method that effectively utilizes multi-scale spatial and temporal data.
• To bridge the gap in integrating information from individual monitoring stations to city-wide scales and across different temporal scales.
• To improve the accuracy and comprehensiveness of air quality predictions.

Main Methods:

• Introduced a Multi-spatial Multi-temporal air quality forecasting method (M2G2) using Graph Convolutional Networks (GCN) and Gated Recurrent Units (GRU).
• Developed a Multi-scale Spatial GCN (MS-GCN) module for spatial information fusion, incorporating bidirectional and residual structures.
• Developed a Multi-scale Temporal GRU (MT-GRU) module for temporal information integration, adaptively combining information from different temporal scales.

Main Results:

• M2G2 demonstrated superior accuracy compared to nine advanced methods across all tested air quality indicators.
• Achieved significant improvements in Root Mean Square Error (RMSE) for 72-h predictions: PM2.5 (6%-10%), PM10 (5%-7%), NO2 (5%-16%), and O3 (6%-9%).
• Ablation studies confirmed the effectiveness of individual M2G2 modules, and sensitivity analysis revealed that O3 negatively impacts PM2.5 prediction accuracy.

Conclusions:

• The M2G2 method effectively integrates spatio-temporal multi-scale domain knowledge for enhanced air quality forecasting.
• The proposed MS-GCN and MT-GRU modules successfully address limitations in spatial and temporal information fusion.
• M2G2 offers a significant advancement in air quality prediction accuracy, benefiting public health and environmental management.