Fine-grained pollutant concentration forecast based on deep learning is a new and promising method, and how to make full use of meteorological, spatial and temporal information is the key point. In order to cooperatively fuse the three pieces of information, we propose a pollutant concentration forecasting model based on a multiscale spatiotemporal graph neural network. This model uses the air quality model to dynamically construct the multiscale spatiotemporal graph neural network to learn the dynamic spatiotemporal relationship between pollutants. Specifically, the graph neural network is used to learn the multiscale spatial relationship between pollutants, the air quality model hybrid single-particle lagrangian integrated trajectory(HYSPLIT) is used to construct the node and edge attributes of the graph, and the attention mechanism-based gate recurret unit(GRU) is used to learn the temporal relationship between pollutant concentrations. The model not only fully considers the three influencing factors of meteorology, space and time, but also integrates the three factors into a framework for collaborative learning. Compared with the traditional mechanism model methods, the proposed method has the characteristics of flexible deployment and easy implementation. Experiments on real project datasets and public databases show that the mean absolute error of pollutant concentration is reduced by about 0.6 and the symmetric mean absolute percentage error is reduced by about 4% compared with the existing advanced method based on graph neural networks.