A Prediction Model for Air Pollution Based on GATTL Spatiotemporal Model

For preventing and controlling air pollution, it is essential to accurately predict air pollutant concentrations. This paper proposes a novel prediction model (GATTL), to forecast PM2.5 concentrations. The model consists of three components: (1) the Graph Attention Network (GAT) layer, (2) the Transformer layer, (3) the Long Short-Term Memory (LSTM) layer. This design logically employs Graph Attention Networks (GAT) to capture spatial dependencies between monitoring stations, employs the Transformer model to model long-term temporal dynamics, and utilizes Long Short-Term Memory (LSTM) operation for short-term sequence memory. This hybrid approach effectively combines the strengths of both graph-based and sequence-based models, representing a significant contribution to the field. We evaluated the model using a dataset from five monitoring stations in Beijing, covering the period from January 1, 2022, to December 31, 2024. Experimental results demonstrate that compared to four other models, GATTL reduces the root mean square error (RMSE) and mean absolute error (MAE) by 1.5 to 2 respectively, indicating superior prediction accuracy.