A graph neural network and Transformer-based model for PM2.5 prediction through spatiotemporal correlation

Abstract

It is important for both urban residents and government agencies to accurately predict the concentration of fine particulate matter (PM2.5) in the atmosphere. In existing research, various traditional and hybrid network models have been applied and developed, all of which have played a positive role in the prediction of PM2.5 concentration. Despite Transformer-based networks demonstrating unique advantages in time series prediction tasks, the Transformer architecture faces challenges related to inadequate extraction of spatiotemporal features and susceptibility to interference from irrelevant data. To address these challenges, a graph neural network (GNN) and Transformer-based model for PM2.5 concentration prediction, named GNN-Transformer, is proposed. Firstly, an instantaneous phase synchronization-based estimator is designed to mitigate the negative influence of irrelevant data on prediction performance. Subsequently, a spatial impact modeling layer based on GNN is introduced to extract spatial impacts between the target city and its surrounding cities. Finally, a spatiotemporal prediction module based on Transformer is devised to further extract the spatiotemporal features between the target city and its surrounding cities, and generate more accurate predictions of PM2.5 concentration. Experiments conducted on real-world datasets demonstrate that the proposed GNN-Transformer outperforms other models in both short and long term prediction task. Specifically, for 3-h prediction task, the proposed model achieves the lowest Mean Absolute Error (MAE) of 6.35 and the highest R² of 0.97. Additionally, the proposed model exhibits superior performance in multiscale prediction tasks across different time spans, achieving the best results for 24-h prediction task (MAE = 18.66, R² = 0.76). Furthermore, the proposed method exhibits the capability to accurately predict high PM2.5 concentration, achieving the highest Critical Success Index (CSI) and Probability of Detection (POD), along with the lowest False Alarm Ratio (FAR). This performance may enable early warnings for potential air pollution events.