A novel ST-iTransformer model for spatio-temporal ambient air pollution forecasting

Abstract

Abstract Accurate prediction of air pollutant dynamics is crucial for public health, sustainable environmental management, and ecosystem analysis. Air pollutants exhibit intricate spatial and temporal patterns, which pose considerable challenges for accurate prediction. This study proposes a novel Spatio-temporal inverted Transformer (ST-iTransformer) model, which uses a new spatio-temporal embedding and inverted dimensions for attention mechanisms to improve air pollution forecast accuracy. The input is three-dimensional data comprising multi-station, multi-step air pollutants, and meteorological variables. In the first phase, the input data is transformed by the proposed spatio-temporal embedding. Next, attention and feed-forward network strategies are applied to extract spatio-temporal features from the embedded data. Then, the multi-station, multi-step air pollutants forecast results of all the monitoring stations are obtained based on a linear layer. Systematic experiments conducted in Beijing, China, show that ST-iTransformer significantly outperforms benchmark models across the metrics, air pollutants, and stations. Additionally, ablation experiments confirm that the spatio-temporal embedding, the inclusion of spatial data, and the addition of meteorological data all improve the prediction accuracy of the model. These results demonstrate that the proposed model could sufficiently capture the spatio-temporal correlation characteristics of air quality data and be successfully applied for air quality forecasting.

Keywords

• Air pollution
• Spatio-temporal prediction
• Transformer
• Embedding
• Deep learning