The theory-guided air quality model solves the mathematical equations of chemical and physical processes in pollution transportation numerically. While the data-driven model, as another scientific research paradigm with powerful extraction of complex high-level abstractions, has shown unique advantages in the PM2.5 prediction applications. In this paper, to combine the two advantages of strong interpretability and feature extraction capability, we integrated the partial differential equation of PM2.5 dispersion with deep learning methods based on the newly proposed DPGN model. We extended its ability to perform long-term multi-step prediction and used advection and diffusion effects as additional constraints for graph neural network training. We used hourly PM2.5 monitoring data to verify the validity of the proposed model, and the experimental results showed that our model achieved higher prediction accuracy than the baseline models. Besides, our model significantly improved the correct prediction rate of pollution exceedance days. Finally, we used the GNNExplainer model to explore the subgraph structure that is most relevant to the prediction to interpret the results. We found that the hybrid model is more biased in selecting stations with Granger causality when predicting.
Keywords: Graph neural network; LSTM; PM(2.5)concentration prediction; Partial differential equation.
Copyright © 2021 Elsevier Ltd. All rights reserved.