Air pollution prediction via physics-informed neural networks

Abstract

Short-term forecasting of fine particulate matter (PM₂.₅) still remains a challenge in sparsely monitored regions like Finland, which is influenced by episodic pollution sources like re-suspended road dust and burning wood. The state-of-the-art techniques usually focus on the forecasting of single-pollutant or urban monitoring network with high-density, and focus is rarely sparsely monitored region. In this work, we adapt the AQ-Net, a deep spatiotemporal neural network originally proposed for (PM₂.₅) forecasting in northern China, to multi-horizon forecasting in Finnish air quality monitoring networks. Our model extends the original AQ-Net to include multiple pollutants, meteorological factors, utilizes a residual prediction formula, and incorporates traffic volume from Fintraffic to predict air quality. As a demonstration of the efficiency of our model, we adopt approximately ten years of open access data of FMI air quality and meteorological conditions of twenty Finnish stations from 2016 to 2026. Through comprehensive experiments with different configurations of sequence length, normalization method, loss function, and model architecture, we show that the extended AQNet consistently outperforms PatchTST, iTransformer, and a persistence baseline across all four forecasting horizons (6, 12, 24, and 48 hours), with traffic information reducing 24-hour RMSE by 17.4% during the Nordic spring road-dust season.