MTLPM: a long-term fine-grained PM2.5 prediction method based on spatio-temporal graph neural network

The concentration of PM2.5 is one of the air quality indicators that the public pays the most attention to. Existing methods for PM2.5 prediction primarily analyze and forecast data from individual monitoring stations, without considering the mutual influence among multiple stations caused by natural environmental factors, e.g., air circulation. Moreover, the existing methods are mostly short-term predictions and perform poorly in long-term forecasting. In this paper, we propose MTLPM, i.e., a spatio-temporal graph neural network model based on an encoder-decoder architecture, which fully exploits the spatial dynamic patterns and long-term dependencies. Firstly, we adopt a message passing mechanism combined with spatial features and complex environmental factors (e.g., temperature, humidity, and wind direction) to update station data, capturing real-time spatial dynamic information. Secondly, we adopt the Multi-head ProbSparse Self-attention to extract temporal features, learning the long-term dependency relationships among the data. Finally, we adopt a generative one-step decoder structure to simultaneously forecast the data for multiple stations over a long period. We conducted experiments on both the project dataset and the publicly available dataset. Compared to existing state-of-the-art methods, MTLPM achieved an average reduction of approximately 1.6 in mean absolute error (MAE) and approximately 0.02 in symmetric mean absolute percentage error (SMAPE) in predicting results. The relevant source code is publicly available on GitHub1.