An Adaptive Spatiotemporal Graph Transformer for Multi-Site PM_2.5 Multi-Step Forecasting with Non-stationary and...

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Science

Background:

• Fine particulate matter (PM2.5) concentration is a key air quality indicator impacting health and ecosystems.
• Accurate multi-site PM2.5 forecasting is hindered by non-stationary distribution drift, uneven monitoring site distribution, and complex spatiotemporal dependencies.

Purpose of the Study:

• To propose an Adaptive Spatiotemporal Graph Transformer (AST-GT) framework for enhanced multi-site, multi-step PM2.5 forecasting.
• To address challenges in PM2.5 prediction, including distribution drift, uneven spatial distribution of monitoring sites, and intricate temporal-spatial interactions.

Main Methods:

• Developed an Adaptive Spatiotemporal Graph Transformer (AST-GT) framework.
• Incorporated adaptive non-stationary normalization, Transformer-based temporal learning, multi-source context encoding, and graph-based dual-scale spatial attention with Temporal-Spatial Cooperative Attention.
• Jointly modeled spatial correlations and spatiotemporal interactions, especially for unevenly distributed sites.

Main Results:

• The AST-GT framework demonstrated effectiveness, accuracy, and robustness on multi-site datasets from Beijing and India.
• The model showed strong generalization capabilities, particularly under irregular spatial distributions of monitoring sites.
• Validated the proposed model's superior performance in multi-site, multi-step air quality prediction.

Conclusions:

• The AST-GT framework provides a reliable solution for multi-site, multi-step air quality prediction.
• The study highlights the model's ability to handle complex spatiotemporal dynamics and data irregularities in PM2.5 forecasting.
• The findings contribute to refined environmental management through improved air quality prediction.