Integrating multi-dimensional graph attention networks and transformer architecture for predicting air pollution in subway stations

Abstract

Accurate prediction of PM_2.5 concentrations in subway stations is crucial for developing effective air pollution control strategies. However, existing methods struggle to accurately predict PM_2.5 concentrations due to the challenges of multi-step ahead forecasting, modeling long time series, capturing complex spatiotemporal correlations, and handling data quality issues such as missing values. This study proposes MAGICFormer, a novel hybrid end-to-end model for predicting PM_2.5 concentrations. The model comprises key components such as data preprocessing, a multi-dimensional graph attention network (md-GAT) module, as well as an Informer encoder and a Cross Decoder based on the Transformer architecture. The data preprocessing method improves data quality by addressing missing values and correcting anomalies. MAGICFormer integrates spatiotemporal correlations to predict PM_2.5 concentrations. The md-GAT module adaptively captures complex spatial relationships among subway stations across different dimensions, with its output serving as input to the Spatio Decoder. The Informer encoder processes long sequences and extracts temporal features, which are then passed to the Spatio Decoder and Temporal Decoder within the Cross Decoder for information fusion. The Cross Decoder aggregates the outputs of the Spatio and Temporal Decoders using a cross-attention mechanism, leveraging the interdependencies between graph-structured and time-series data to enhance prediction accuracy and improve model performance by effectively fusing spatial and temporal information. Experiments on Seoul subway stations show that MAGICFormer improves prediction accuracy by over 20 % compared to existing methods, demonstrating its effectiveness in long-term PM_2.5 forecasting. The proposed model offers a practical decision support tool for enhancing air quality management strategies in subway systems, particularly for long-term monitoring and control.