Probabilistic concentration prediction of PM2.5 in subway stations based on multi-resolution elastic-gated attention mechanism and Gaussian mixture model

The subway has increasingly taken over as the primary method of short-distance travel in the development of modern urban transportation. Due to the poor air mobility in subway stations, it is crucial to monitor and provide alerts on air quality. This work provides a probabilistic prediction framework of PM2.5 concentration to solve the air quality early warning problem in subway stations. Firstly, outliers are discovered and corrected utilizing a probabilistic-based auto-encoder (PAE). Secondly, the multi-resolution elastic-gated attention mechanism is used to address error accumulation and historical information lost during the prediction process. Moreover, the decoder structure of sequence to sequence (Seq2Seq) is improved through multiple output strategy and flexible gate attention mechanism to reduce error accumulation. Finally, the Seq2Seq is equipped with Gaussian mixture models (GMM), allowing it to adapt to more complicated changes and produce a probability distribution of PM2.5 concentrations. According to tests on data from subway stations, the Pinball loss and Winkler score of the proposed model are smaller compared to other models.