Parallel Prediction Model with Multilevel Attention Mechanism for Air Quality Index Forecasting

To mitigate the negative impact of fine particulate matter (PM2.5) on human health, ecology, and global climate change, accurate prediction of the air quality index (AQI) is essential for the environmental protection authorities to understand the future urban air quality and to formulate preventive countermeasures in advance from the historical observations obtained from the monitoring system. To further improve the prediction accuracy of AQI, this study proposes a parallel prediction model based on a multilevel attention mechanism(TF-BiG-CA). That is consists of three modules: the Transformer, the BiGRU optimized by Global-Attention, and the cross-attention. The Transformer takes the data from the six pollutants and AQI data to extract temporal features, while the Global-Attention-optimized BiGRU captures global spatial feature signatures. Cross-attention is introduced to learn the correlation between temporal and spatial features for the fusion of spatio-temporal features in parallel networks. The results show that the performance metrics of the model proposed in this paper, The RMSE was 0.303, and the MAE was 0.197. The model remains robust in <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7}$</tex> days of missing data for every <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 4}$</tex> days of missing values and outperforms the other comparative models, improving the air quality prediction accuracy.