A nested machine learning approach to short-term PM_2.5 prediction in metropolitan areas using PM_2.5 data from...

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Science

Background:

• Current PM2.5 prediction models often rely on single, sparsely distributed sensor networks.
• Integrating data from multiple sensor networks for short-term PM2.5 prediction is an underexplored area.
• Accurate short-term PM2.5 forecasting is crucial for public health and environmental management.

Purpose of the Study:

• To develop and evaluate a machine learning approach for predicting ambient PM2.5 concentrations at unmonitored locations.
• To leverage data from multiple sensor networks and social-environmental factors for enhanced prediction accuracy.
• To assess the performance improvement gained by integrating diverse data sources compared to traditional methods.

Main Methods:

• A hybrid Graph Neural Network and Long Short-Term Memory (GNN-LSTM) model was employed for daily PM2.5 prediction using regulatory network data.
• A second GNN-LSTM network processed hourly data from a low-cost sensor network, incorporating daily predictions to generate spatiotemporal features.
• A Fully Connected (FC) network merged these spatiotemporal features with social-environmental data for final hourly PM2.5 concentration prediction.

Main Results:

• The proposed approach successfully predicted hourly PM2.5 concentrations at unmonitored locations.
• Utilizing data from two sensor networks demonstrably improved prediction performance.
• The case study in Denver, CO, validated the enhanced accuracy compared to baseline models.

Conclusions:

• Integrating data from multiple sensor networks significantly enhances the accuracy of short-term PM2.5 prediction.
• The developed machine learning framework offers a promising solution for predicting PM2.5 in areas lacking dense monitoring infrastructure.
• This approach provides a more comprehensive understanding of PM2.5 dynamics by combining diverse data streams.