STFformer: An Effective PM2.5 Prediction Model

PM2.5 is an important indicator of air pollution index, and its accurate prediction is of great significance for urban air pollution prevention and control. In this paper, an effective prediction model named STFformer is proposed for PM2.5. Firstly, the original PM2.5 data is decomposed into seasonal component and trend component, which are predicted separately not as a whole to enhance the prediction accuracy. Then, for the seasonal component with dominant frequency, the Transformer equipped with Fourier attention is proposed to achieve effective feature extraction, and the multi-scale information contraction and expansion blocks using series of convolution are designed to achieve seasonal component prediction. Meanwhile, for the trend component, the MLP network with RevIN block is proposed to process the non-stationary information and the trend component is accurately predicted; Finally, experiment results on the PM2.5 data set and the public seasonal data set WTH show that the proposed STFformer achieves the best prediction accuracy in most cases compared with some typical time series prediction models and reduce prediction error by 2.5% to 3.6% at different time scales.