Hourly forecasting on PM_2.5 concentrations using a deep neural network with meteorology inputs.

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Particulate matter (PM2.5) pollution is a significant global environmental concern.
• Deep learning offers advanced, non-linear regression methods for accurate PM2.5 concentration prediction.

Purpose of the Study:

• To develop a deep learning model for predicting hourly PM2.5 concentrations in Beijing.
• To leverage meteorological data and spatiotemporal correlations for enhanced prediction accuracy.

Main Methods:

• Utilized a deep learning framework with meteorological data as input.
• Incorporated time-series concatenation to capture spatiotemporal correlations.
• Extracted intrinsic features between meteorological variables and PM2.5 concentrations.

Main Results:

• Achieved a maximum R-squared score of 0.98 and an average of 0.9295.
• Demonstrated low average bias: 9 μg/m³ on validation and 1 μg/m³ on training sets.
• Prediction accuracy increases with the prediction period.

Conclusions:

• The deep learning model provides a fast and accurate method for PM2.5 prediction.
• Prediction errors can inform emission change estimations.
• Results offer valuable scientific advice for policymakers and environmental supervisors.