Ground-level fine particulate matter with a diameter less than 2.5 μm (PM_2.5) has negative impacts on human health and social economy. Most methods obtain PM_2.5 data from indirect products of aerosol optical depth retrieved from satellites or daytime top-of-atmosphere reflectivity. This paper aims to directly utilize infrared data from the advanced geosynchronous radiation imager (AGRI) on the Fengyun-4B satellite and artificial intelligence models to perform near-real-time PM_2.5 retrievals over the Yangtze-Huaihe region throughout the entire day (including day and night), with a spatial resolution of 4 km and a temporal resolution of 15 minutes. Firstly, the seasonal signal response of AGRI brightness temperature to different PM_2.5 levels is explored. Secondly, a study on AGRI brightness temperature retrieval of PM_2.5 is conducted based on the random forest method in different seasons. The experimental results show that the correlation coefficients for PM_2.5 retrievals in all four seasons exceed 0.87. Finally, the model is interpretable using SHapley additive prediction (given the significant contribution of geographic information to PM_2.5), and the application of the proposed product is further explored.