A Novel Multistep Ahead PM$_{2.5}$ Forecasting Approach Using Spatial–Temporal Attention Network

The establishment of a long-term and accurate forecasting approach of urban air pollution is conducive to the implementation of pollution prevention and control policies. However, existing research has not fully taken into account the spatial-temporal pattern characteristics of long-distance air pollution transport. Multistep ahead forecasting faces the challenge of aliasing long-term spatial-temporal correlation and accumulating errors. In this study, a long-term spatial–temporal pattern forecasting model (ASTemCN) of PM $_{2.5}$ air pollutants in Chinese cities was established based on the monitoring sites of the Internet of Things. The model skillfully designs a novel spatial–temporal fusion mechanism to integrate the temporal and spatial characteristics under the spatial–temporal pattern of PM $_{2.5}$ . Compared with other learning paradigms, ASTemCN is more suitable for learning long-term spatial–temporal patterns, has the highest forecasting accuracy and stronger generalization ability, and provides a research direction for the spatial–temporal pattern analysis of air pollution.