Air pollution poses a significant environmental challenge in today’s world. Air pollution monitoring is crucial for improving ecological environments, controlling pollution, and fostering a harmonious relationship between humans and their surroundings. In the process of pollution monitoring, data anomalies often occur due to instrument failures, adverse weather conditions, abnormal pollution sources, and other reasons, which seriously affects the pollution traceability and air quality prediction of researchers. The existing pollution detection methods analyze historical pollution through machine learning and establish models to identify abnormal situations in the environment. It is necessary to make assumptions about the spatial relationship between monitoring stations in advance, but this assumption usually has a large error. It is necessary to model spatiotemporal dynamics and conduct multi-scale analysis of time. Based on this, we raise a self-supervised model PASTGAT that leverages the strengths of both Transformer and GAT. Specifically, We employ the MITT module to capture time characteristics at varying scales, and the SAGA module to extract dynamic dependencies in spatiotemporal data, fully modeling global dependencies. We validated the advantages of our proposed method on air pollution data in the Yangtze River Delta region.