Spatiotemporal graph neural networks for predicting mid-to-long-term PM2.5 concentrations

Predicting the concentration of PM2.5 particles is of critical importance in public health management because their small size enables them to penetrate deep into the lungs and even enter the bloodstream. However, achieving accurate predictions is challenging due to the complexity of the transport and dispersion processes involved, which are influenced by multiple factors, including atmospheric pollutants, meteorological conditions, and geographic features. To address this challenge, we developed a novel framework for PM2.5 prediction that utilizes multiple edges for feature extraction and employs a multi-gated graph neural network for feature calculations. The model utilizes multiple edges, created by combining an atmospheric diffusion coefficient with a PM2.5 similarity metric, to elucidate the complex characteristics of PM2.5 interactions between different monitoring stations. Node features are computed by aggregating connected nodes' features, weighted by these multifaceted edges. The experimental results demonstrate that the proposed method outperforms conventional time series prediction models in the mid-to-long-term prediction of PM2.5 concentrations (96 h in advance), with improvements in Root Mean Square Error (RMSE) by 2.613%, Critical Success Index (CSI) by 3.143%, and R squared (R2) by 5.263%. Using the proposed model, the learning of global features can be achieved and issues related to local dependency can be mitigated.