Air quality prediction by integrating mechanism model and machine learning model

Area of Science:

• Environmental Science
• Computer Science
• Atmospheric Science

Background:

• Air Quality Prediction (AQP) is complex due to interacting meteorological, spatial, and temporal factors.
• Existing methods struggle to effectively model these intricate relationships.

Purpose of the Study:

• To propose a novel AQP method, Dynamic Multi-granularity Spatio-temporal Graph Neural Network (DM_STGNN), that integrates mechanism models and machine learning.
• To leverage the HYSPLIT model for dynamic spatio-temporal graph construction in AQP.
• To enhance AQP by considering meteorological, spatial, and temporal influences comprehensively.

Main Methods:

• Developed DM_STGNN with an encoder-decoder architecture for AQP.
• Utilized a multi-granularity graph structure with meteorological, time, and geographical features as node attributes.
• Employed the HYSPLIT model for dynamic edge construction and LSTM for time-series pollutant concentration learning.
• Incorporated an attention-based LSTM decoder and unsupervised pre-training for enhanced temporal dependency learning.

Main Results:

• The DM_STGNN model effectively captures fine-grained and long-term influences in AQP.
• Experimental results on a project-based dataset and the Yangtze River Delta city group dataset demonstrate superior performance.
• The model shows appealing performance over state-of-the-art AQP methods.

Conclusions:

• The proposed DM_STGNN model successfully integrates mechanism and machine learning approaches for AQP.
• The method effectively addresses the challenge of modeling spatio-temporal dependencies in air quality.
• DM_STGNN offers a promising advancement in accurate and comprehensive air quality prediction.