PM_2.5 Prediction with a Novel Multi-Step-Ahead Forecasting Model Based on Dynamic Wind Field Distance

Area of Science:

• Environmental Science
• Atmospheric Science
• Data Science

Background:

• Particulate matter (PM2.5) poses a significant global health risk.
• Existing PM2.5 forecasting models often neglect crucial spatiotemporal data and wind impacts.
• Accurate PM2.5 predictions are vital for public health early warnings.

Purpose of the Study:

• To develop an advanced model for predicting PM2.5 concentration 24 hours in advance.
• To improve the selection of related monitoring sites by considering wind field dynamics.
• To enhance the accuracy of PM2.5 forecasting by integrating spatiotemporal correlations and weather data.

Main Methods:

• Proposed a Long Short-Term Memory-Convolutional Neural Network based on Dynamic Wind Field Distance (LSTM-CNN-DWFD).
• Developed a KNN method using dynamic wind field distance to select relevant sites, accounting for wind's influence.
• Employed a local stateful LSTM for temporal feature extraction and CNN for spatiotemporal feature extraction.
• Integrated weather forecasts to further boost prediction accuracy.

Main Results:

• The LSTM-CNN-DWFD model demonstrated high accuracy in predicting PM2.5 concentrations, with mean R² values ranging from 0.85 (1-hour) to 0.59 (24-hour).
• The model achieved lower mean RMSE (43.90) and MAE (29.17) for 1-6 hour predictions compared to other methods.
• Site selection based on dynamic wind field distance proved more effective than geographical distance for improving prediction accuracy.

Conclusions:

• The proposed LSTM-CNN-DWFD model offers a significant advancement in PM2.5 forecasting accuracy.
• Incorporating dynamic wind field distance for site selection enhances the model's ability to capture relevant spatiotemporal correlations.
• This approach provides a robust framework for developing effective early warning systems for air pollution hazards.