Regional Prediction of Ozone and Fine Particulate Matter Using Diffusion Convolutional Recurrent Neural Network

Area of Science:

• Environmental Science
• Atmospheric Science
• Data Science

Background:

• Current air quality forecasting models often lack regional spatiotemporal analysis.
• Site-specific forecasts limit comprehensive pollution control strategies.

Purpose of the Study:

• To develop and evaluate a novel Diffusion Convolutional Recurrent Neural Network (DCRNN) model for regional air quality forecasting.
• To assess the impact of geographic distance and wind direction on air quality prediction accuracy.

Main Methods:

• Utilized hourly fine particulate matter (PM2.5) and ozone data from 123 monitoring stations in the Yangtze River Delta.
• Implemented a DCRNN model incorporating directed and undirected graphs to capture spatiotemporal dependencies.
• Compared DCRNN performance against baseline models and evaluated graph variations (directed vs. undirected).

Main Results:

• The DCRNN model demonstrated superior prediction accuracy for PM2.5 and ozone compared to baseline models.
• A directed graph model, accounting for wind direction, outperformed an undirected graph model in 24-hour predictions.
• Forecast accuracy was higher in densely monitored regions than in sparsely monitored areas.

Conclusions:

• The DCRNN model offers an effective approach for regional air quality forecasting, improving upon existing methods.
• Incorporating wind direction significantly enhances prediction accuracy, particularly for short-term forecasts.
• The model's performance is influenced by monitoring station density, suggesting benefits for well-instrumented areas.