Air pollution is a critical issue with significant public health and policy implications. Accurate forecasting of \(PM_{2.5}\) concentrations is essential for effective environmental management. Addressing the challenges of non-Euclidean data distributions, this paper introduces AirGraphNet, a novel spatio-temporal graph attention recurrent neural network for \(PM_{2.5}\) forecasting. We propose a novel graph structure construction method that integrates the geographic distribution of monitoring stations with the data’s inherent structure. Additionally, we design a forecasting architecture combining graph attention networks (GATs) to capture spatial dependencies and gated recurrent units (GRUs) for temporal modeling. Evaluated on the Tehran Air Quality dataset, AirGraphNet outperforms state-of-the-art models, demonstrating superior prediction accuracy. These results highlight the potential of advanced graph-based architectures in addressing environmental challenges and supporting public health initiatives.
