A dual-graph spatiotemporal convolutional network with reasoning alignment for PM2.5 concentration multi-horizon forecasting

PM 2.5 , a primary component of haze, is an inhalable form of particulate matter that is hazardous to both the environment and human health. Accurate PM 2.5 concentration forecasting is of great significance for governments in formulating haze prevention and control measures. Previous studies have insufficiently explored the non-Euclidean relationships of monitoring stations and the cross-sample features within PM 2.5 concentration sequences. To fully exploit the spatiotemporal characteristics of PM 2.5 concentrations from both spatial and temporal perspectives, this study proposes a dual-graph spatiotemporal convolutional network with reasoning alignment (DGST-RA). Specifically, through considering non-Euclidean relationships representations for each station, a dual-graph convolutional network is constructed to model spatial dependencies between stations from both static and dynamic dimensions. Subsequently, we reframe temporal feature extraction as a sequence reasoning process. To improve the consistency of the learned temporal representations, the intra-reasoning alignment and inter-reasoning alignment modules are introduced for temporal modeling. Finally, the features extracted by the spatial and temporal modeling modules are fused to obtain the predicted PM 2.5 concentration values. Extensive experiments on both short-term and long-term forecasting demonstrate that, compared to other baseline methods, DGST-RA achieves the best forecasting performance for both short-term and long-term PM 2.5 concentration forecasting. On average, compared to the best baseline model, our DGST-RA model improves RMSE, MAE, and SMAPE by 8.04%, 8.10%, and 11.12%, respectively, for 1‑hour prediction, and CRMSE by 41.44% for 24‑hour prediction. These results confirm that the DGST-RA is an effective model for multi-horizon PM 2.5 concentration forecasting.