Research on PM_2.5 concentration prediction based on transformer

Abstract The pollution caused by PM 2.5 can have serious consequences for human health and the environment, making timely prediction of its concentration important. This study proposes a Spatial-Temporal Air quality Transformer (STAFormer) for forecasting PM 2.5 concentration. In the dimension of time, this study designs a novel multi-head attention mechanism that can discover multi-periodicity in time series, thereby extracting more complex temporal features. In terms of spatial analysis, this study designs a graph convolutional neural network that incorporates distance and wind direction information to account for the impact of surrounding cities on PM 2.5 concentration prediction. Additionally, it dynamically captures spatial correlations using an attention mechanism. The prediction results for Zhengzhou show that the model suggested in this study performs better. The RMSE for forecasting PM 2.5 levels for the next hour is 3.51 μg/m 3 , and MAE is 2.38 μg/m 3 .