A3TGCN: attention temporal graph convolutional network for enhanced PM _2.5 spatiotemporal modeling

Accurate spatiotemporal modeling of PM2.5 is crucial for protecting public health, but remains challenging due to the non-Euclidean distribution of AQM stations and the spatiotemporal dependencies. Existing approaches often ignore the simultaneous extraction of spatiotemporal dependencies and tend to model only local, Euclidean spatial dependencies between AQM stations, overlooking global and non-Euclidean dependencies. To address these limitations, an attention-based temporal graph convolutional network (A3TGCN) is developed to simultaneously extract the spatiotemporal dependencies of PM2.5 across urban AQM stations. Temporal dependencies are extracted using gated recurrent units (GRUs), spatial dependencies are modeled using a graph convolutional network (GCN) built on a non-Euclidean graph. An attention mechanism adaptively weights different time steps to extract global temporal dependencies. Experiments on Tehran AQM data show that A3TGCN reduces RMSE by 11% and MAE by 13% at T + 1 compared to GRU, maintaining 3–5% RMSE improvements up to 24 hours and outperforming state-of-the-art baselines.