A multiscale spatial–temporal-variable feature fusion network for predicting multiple air pollutants

Accurate prediction of air quality at urban monitoring stations is vital for improving environmental management and safeguarding. However, most existing studies focus on single pollutants, overlooking the complex interactions among multiple pollutants, thereby limiting the accuracy of predictions. To address this issue, a Multiscale Spatial-Temporal-Variable Feature Fusion Network (MSTVFFN) for predicting multiple air pollutants at air quality monitoring stations is proposed. The encoder employs three dedicated modules to extract temporal, spatial, and variable (i.e., pollutants) features. The temporal feature extraction module models temporal dependencies at individual, local, and global scales. The spatial feature extraction module captures both global and local spatial relationships. The variable feature extraction module explicitly learns inter-correlations among pollutants. The decoder uses a feature fusion module to capture the multi-dimensional interactions across temporal, spatial, and inter-pollutant features. Experiments on the Beijing, London, and Wuhan datasets for 12-hour and 24-hour joint prediction of four air pollutants demonstrated that MSTVFFN achieves 11%-33% reductions in MAE and 3%-16% improvements in R² over state-of-the-art models, showing significant and consistent performance gains across different pollutants and cities. The proposed multi-pollutant prediction framework not only advances methodological development but also provides a valuable tool for more accurate air quality forecasting, thereby supporting informed decision-making and promoting public health protection. The source codes of MSTVFFN are publicly available at https://github.com/HPSCIL/MSTVFFN-multiple-air-pollutants-prediction.