Spatiotemporal Graph Neural Networks for Air Quality Forecasting: A Comprehensive Review and Future Direction

Abstract

Prediction​‍​‌‍​‍‌ of the air that we breathe is the first step in the prevention of health risks and to avoid the negative effects of pollutants, i.e., PM2.5, O3, and NO2. These pollutants exhibit complex spatial and temporal patterns, and their behaviour is influenced by factors such as weather, land use, and emissions. Incorporation of Spatiotemporal Graph Neural Networks (STGNNs) has become a revolutionary solution by utilizing graph structures for representing spatial relationships and temporal mechanisms for capturing the changes. Here, this paper constitutes a major part of the research in this domain and extensively covers methodologies, experiments, and applications. Some of the outstanding features are extreme handling in E-STGCN, physics integration in DGM, and dynamic edges in DST_GNN. The performance terms illustrate MAE reductions from 13% to 57% relative to baselines. The issues, such as scalability and consistency, are acknowledged here, while subsequent ideas for hybrid systems and real-time forecasting are proposed.