Learning Spatiotemporal Dependencies Using Dynamic Graph Learning Algorithms for Air Quality Prediction in Lanzhou City

To address urban air pollution, Lanzhou City have deployed 456 air quality micro monitoring stations to collect increasing amounts of data for accurately predicting future air quality conditions. Spatiotemporal graph convolutional networks exhibits excellent performance in air quality prediction methods, therefore this paper regard Lanzhou air pollution monitoring stations as nodes of the graph structure, and the spatial dependence caused by the diffusion of pollutants between stations as edges of the graph structure. However, the diffusion of pollutants is influenced by many external factors such as location, terrain, and meteorological information, which dynamically change over time and are difficult to accurately represent as a graph structure. This paper designs a Dynamic Graph Learning (DGL) module to capture spatial dependencies caused by these external factors from both global and local perspectives, learning dynamic graph adjacency matrices whose edges can represent spatial dependencies caused by pollutant diffusion. Unlike methods that rely on meteorological knowledge to construct graph structures, our approach uses a data-driven method for graph structure learning, which is advantageous when air quality micro monitoring station, constrained by cost, cannot obtain meteorological data for specific locations. Additionally, this paper presents the STGnet model for urban air quality prediction, capable of making multi-step air quality predictions using real pollutant datasets from Lanzhou. The results of combining the two modules for prediction indicate that our method achieves the best performance.