A theory-guided graph networks based PM2.5 forecasting method | JoVE Visualize

Area of Science:

Environmental Science
Computer Science
Atmospheric Chemistry

Background:

Traditional theory-guided air quality models numerically solve complex chemical and physical pollution transport equations.
Data-driven models excel at extracting high-level abstractions but often lack interpretability.
Existing methods struggle with long-term, multi-step PM2.5 forecasting and accurately predicting pollution exceedance events.

Purpose of the Study:

To combine the interpretability of theory-guided models with the feature extraction capabilities of data-driven models for PM2.5 prediction.
To develop a novel hybrid model integrating partial differential equations with deep learning for enhanced air quality forecasting.
To improve the accuracy and reliability of long-term, multi-step PM2.5 predictions and pollution exceedance day forecasting.

Main Methods:

Developed the Deep Partial Differential Equation Graph Neural Network (DPGN) model, integrating PM2.5 dispersion PDEs with deep learning.
Extended DPGN for long-term, multi-step prediction, incorporating advection and diffusion effects as graph neural network training constraints.
Utilized hourly PM2.5 monitoring data for model validation and employed GNNExplainer for interpreting prediction drivers.

Main Results:

The proposed hybrid model demonstrated higher prediction accuracy compared to baseline models in forecasting hourly PM2.5 concentrations.
Significantly improved the correct prediction rate for days with pollution exceedances.
GNNExplainer analysis revealed that the hybrid model prioritizes stations exhibiting Granger causality for accurate predictions.

Conclusions:

The integration of physical laws (PDEs) with deep learning (GNNs) offers a powerful approach for accurate and interpretable air quality modeling.
The DPGN model provides a robust framework for long-term, multi-step PM2.5 forecasting and predicting critical pollution events.
Understanding the influence of causal relationships between monitoring stations enhances the interpretability and reliability of AI-driven environmental predictions.