Spatiotemporal adaptive attention graph convolution network for city-level air quality prediction.

Area of Science:

• Environmental Science
• Data Science
• Computer Science

Background:

• Air pollution poses significant risks to human health globally.
• Accurate air quality forecasting is essential for public health interventions.
• Existing models struggle to effectively capture complex spatiotemporal dependencies in air quality data.

Purpose of the Study:

• To develop a novel deep learning model for enhanced city-level air quality prediction.
• To improve the extraction of spatiotemporal features from complex air pollution data.
• To accurately predict short-term series of PM2.5 concentrations.

Main Methods:

• Proposed a spatiotemporal adaptive attention graph convolution model.
• Encoded multiple spatiotemporal dependencies using station-level attention.
• Employed a Bi-level sharing strategy for efficient extraction of shared inter-station relationships.
• Utilized multiple decoders and a gating mechanism for multi-step predictions.

Main Results:

• The proposed model demonstrated superior performance in city-level air quality prediction.
• Achieved state-of-the-art results on several real-world air quality datasets.
• Effectively captured complex spatiotemporal dependencies, outperforming existing methods.

Conclusions:

• The novel deep learning model significantly advances air quality prediction capabilities.
• The approach offers a more systematic analysis of spatial dependencies for improved forecasting.
• This method provides a robust tool for public health and environmental monitoring.