Improved PM_2.5 prediction with spatio-temporal feature extraction and chemical components: The RCG-attention model

Deep learning models are widely used for PM_2.5 prediction. However, neglecting temporal and spatial characteristics leads to low prediction accuracy. In this work, a new deep learning model (RCG - Attention model) was developed, which combines the residual neural network (ResNet) and the convolution gated recurrent network (ConvGRU) and is applied to extract the spatio - temporal features for predicting PM_2.5 concentration over the subsequent 24 h. The ResNet extracts the spatial distribution features of pollutants, and the ConvGRU extracts temporal features. The spatial and temporal features are fused by the multi - head attention mechanism to obtain multi - dimensional features. These features are finally fed into a series of fully connected layers to predict the future results. Incorporating these chemical components enhances the scientific validity of the dataset and strengthens the inherent logical connections among variables. The Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and R - squared (R²) results indicate that the prediction performance of the RCG - Attention model surpasses that of other baseline models. The model demonstrates superior prediction performance across multiple monitoring stations, suggesting robust generalization capabilities and adaptability for various regions in one city. The SHAP results show that PM₁₀, NO₂, RH, NO₃^-, OC and NH⁴⁺ are significant influencing features. The RCG - Attention model provides a comprehensive solution for PM_2.5 concentration prediction by integrating spatial and temporal feature extraction with chemical components.