Air quality is of vital importance to human health. Accurately predicting air quality, especially its sudden changes, is highly valuable for citizens and governments to make personal and local decisions, design intelligent policies and control pollution at minimal cost. However, none of the existing methods achieves sufficient prediction accuracy for time intervals of sudden pollution change due to inability of existing models to take into account pollution propagation between different areas caused by air mass movement.

For the first time, we consider pollution transfer in the context of short-term air quality prediction and propose to use air flow trajectory data, widely used in environmental sciences, to represent pollution transfer patterns between different locations. By learning trajectory representations, measurement location embedding vectors, and interrelationships between local weather at relevant locations, we propose a new attention based seq2seq model to track pollution propagation for accurate air quality prediction.

We evaluate our model on datasets from Beijing area and compare the results to several state-of-the-art baselines. Experiments show that the proposed approach can successfully capture pollution transfer patterns between different sites in the area. Our model outperforms all the baselines and decreases prediction errors by 9.6% to 22.4%. The method allows interpreting prediction results visually and analytically.

Y. Cheng, O. Saukh, L. Thiele, TIP-Air: Tracking Pollution Transfer for Accurate Air Quality Prediction, in: Proceedings of the 2021 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2021 ACM International Symposium on Wearable Computers