A deep learning model integrating a wind direction-based dynamic graph network for ozone prediction.

Area of Science:

• Environmental Science
• Computer Science
• Data Science

Background:

• Ozone pollution poses a significant environmental challenge globally.
• Accurate ozone concentration forecasting is crucial for implementing effective mitigation policies.

Purpose of the Study:

• To develop a novel hybrid deep learning model for hourly ozone concentration prediction.
• To enhance the accuracy of air quality forecasting by incorporating dynamic spatial-temporal data.

Main Methods:

• Developed the wind direction-based dynamic spatio-temporal graph network (WDDSTG-Net).
• Utilized a dynamic directed graph structure based on hourly wind direction to model station relationships.
• Employed graph attention and sequence-to-sequence models for adaptive information aggregation and temporal dependency extraction.
• Integrated meteorological predictions to refine ozone forecasts.

Main Results:

• Achieved a mean absolute error of 6.69 μg/m³ for 1-h and 18.63 μg/m³ for 24-h predictions.
• Exceeded the performance of several classic forecasting models.
• Attained over 75% accuracy in Integrated Air Quality Index (IAQI) predictions across all stations.
• Demonstrated strong capability in predicting severe ozone pollution events with a 0.77 true positive rate for 24-h forecasts.

Conclusions:

• WDDSTG-Net highlights the importance of short-term wind fluctuations and transport dynamics in data-driven air quality modeling.
• The model offers a robust approach for predicting ozone concentration and potentially other airborne pollutants.
• Dynamic spatio-temporal graph networks represent a promising direction for advanced environmental monitoring and forecasting.