Combining Machine Learning and Numerical Simulation for High-Resolution PM_2.5 Concentration Forecast

Area of Science:

• Environmental Science
• Atmospheric Science
• Data Science

Background:

• Accurate forecasting of ambient fine particulate matter (PM2.5) is essential for public health and pollution episode management, particularly in areas with sparse ground monitoring.
• Existing methods, such as chemical transport models (CTMs) and statistical algorithms, have limitations in spatial coverage or accuracy.

Purpose of the Study:

• To develop a novel PM2.5 forecast framework that provides spatiotemporally continuous predictions.
• To enhance the accuracy of PM2.5 forecasts by integrating a machine learning algorithm with a global CTM product.
• To enable reliable, near-real-time PM2.5 forecasting in resource-limited regions.

Main Methods:

• Developed a PM2.5 forecast framework by combining the Random Forest algorithm with NASA's Goddard Earth Observing System "Composition Forecasting" (GEOS-CF) data.
• Conducted a 5-day forecast experiment over Central China, focusing on the Fenwei Plain.
• Validated the model using spatial cross-validation and assessed metrics such as R-squared and normalized mean bias.

Main Results:

• Achieved validation R-squared values of 0.76 and 0.64 for the first two forecast days, and approximately 0.5 for subsequent days.
• Demonstrated substantial reduction in biases present in the GEOS-CF product, with a normalized mean bias close to 0.
• The framework provides spatiotemporally continuous PM2.5 forecasts at a 1 km resolution.

Conclusions:

• The proposed framework effectively integrates CTM data with machine learning for improved PM2.5 forecasting.
• The model offers a significant improvement over existing methods, providing accurate and spatially continuous air quality predictions.
• This approach is computationally efficient, making it suitable for near-real-time PM2.5 forecasting in environments with limited resources.