Forecasting PM_2.5 using hybrid graph convolution-based model considering dynamic wind-field to offer the benefit of...

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Air pollution, specifically PM2.5 concentration, is influenced by meteorological conditions and chemical factors.
• Data-driven spatio-temporal models show promise but struggle with causality due to reliance on statistical correlations.
• Integrating domain knowledge into data-driven models can enhance prediction accuracy and physical realism.

Purpose of the Study:

• To develop a physically realistic data-driven model for PM2.5 concentration prediction.
• To incorporate the influence of dynamic wind-fields into air pollution modeling.
• To improve the accuracy and interpretability of spatio-temporal PM2.5 predictions.

Main Methods:

• Proposed a hybrid deep learning framework: dynamic directed spatio-temporal graph convolution networks (DD-STGCN).
• Fused pollution diffusion distance with a deep learning model based on a wind-field surface.
• Utilized directed graph time-series and wind-field diffusion distance to model dynamic, anisotropic spatial dependencies between monitoring stations.

Main Results:

• The DD-STGCN model demonstrated superior prediction ability compared to LSTM, GC-LSTM, and STGCN.
• Achieved average improvements of 10.2% in MAPE, 9.7% in MAE, and 9.6% in RMSE for 12-hour predictions.
• Showcased better prediction distribution and spatial interpretability, especially during a haze period, compared to pure data-driven models.

Conclusions:

• Incorporating domain knowledge, specifically wind-field dynamics, significantly enhances PM2.5 prediction accuracy and physical realism.
• The DD-STGCN framework effectively models complex spatio-temporal dependencies in dynamic environments.
• The study highlights the benefit of integrating physical insights into data-driven approaches for environmental modeling.