Air pollution is a product of the development of civilization. With the development of society, the diseases and injuries caused by PM2.5 (fine particulate matter) to the human body have become an important issue that countries all over the world want to solve. At present, most air quality index AQI monitoring uses machine learning and deep learning such as LSTM to predict the trend of air pollution, but this method causes the prediction model to always be one hour away from the actual value, which makes it impossible to judge some abnormal turning changes in time. Therefore, this article uses the following methods: (1) Use the deep learning method CNN-GRU classification model for training, use the data of the surrounding micro-sites in the previous hour as input, and the change trend of PM2.5 as output. (2) Using the characteristics of the micro-station changes, and try to select the range of the micro-stations according to the wind direction. We expect to find the characteristics and relationships between micro-stations and the central EPA observation station. We obtain the following change trend of PM2.5 within one hour, and then use the change trend to modify the predicted value of the original prediction model. Although the desired result has not yet been achieved, in the future, the use of wind speed to select the range of the micro-station may be considered, or other methods may be used as the direction of follow-up research.