A New Framework for Medium- to Long-Term PM_2.5 Predictions Using AI-Based High-Resolution Meteorological Forecasts.

Area of Science:

• Environmental science and atmospheric chemistry
• Artificial intelligence and machine learning
• Computational modeling

Background:

• Accurate medium- to long-term PM$_{2.5}$ forecasting is crucial for air quality management and public health.
• Traditional methods struggle with limitations of short-term observations and retrospective meteorological data.

Purpose of the Study:

• To develop a deep learning framework for enhanced PM$_{2.5}$ prediction using AI-based gridded meteorological forecasts.
• To capture spatiotemporal pollutant transport patterns for extended-range forecasting.

Main Methods:

• Coupling AI-based gridded meteorological forecasts with a dynamic graph-based architecture.
• Utilizing spatiotemporal patterns of pollutant transport for prediction.
• Interpretability analysis to understand model behavior and inform graph construction.

Main Results:

• Reduced 3-day RMSE and MAE by 12% compared to models without gridded meteorological inputs.
• Decreased errors by approximately 20% for seasonal 10-day forecasts versus the WRF-CMAQ system.
• Demonstrated high performance across different AI-based meteorological data sources (IFS, PanGu-Weather) without retraining.

Conclusions:

• The proposed deep learning framework offers a computationally efficient and operationally feasible alternative to conventional chemical transport models.
• The model's flexibility allows integration with various AI meteorological models for adaptable air quality forecasting.
• Findings support the development of early warning systems and targeted emission control strategies.