Regional PM 2.5 prediction with hybrid directed graph neural networks and Spatio-temporal fusion of meteorological factors

Traditional statistical prediction methods on PM often focus on a single temporal or spatial dimension, with limited consideration for regional transport interactions among adjacent cities. To address this limitation, we propose a hybrid directed graph neural network method based on deep learning, which utilizes domain features to quantify the influence of neighboring cities and construct a directed graph. The model comprises a historical feature extraction module and a future transmission prediction module, and each module integrates a Graph Neural Network (GNN) and a Long Short-Term Memory Network (LSTM) for spatiotemporal encoding. Compared to other neural network models, our model improves the prediction accuracy of PM concentration and demonstrates superior performance for 48-h prediction in the North China Plain. For 3- to 48-h prediction tasks, the proposed model achieves mean absolute error (MAE) at 7.64 - 14.04 μg/m . In addition, by expanding the modeling scope from different directions and integrating domain information, the model significantly enhances its ability to predict PM trends, seasonal variations, and PM exceedances in heavily polluted urban areas. The proposed model represents a promising advancement in optimizing air quality forecasting and management.