A new attention-based CNN_GRU model for spatial-temporal PM_2.5 prediction

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Accurate prediction of spatial-temporal PM2.5 distribution is hindered by missing data and model selection challenges.
• Effective data imputation requires considering inter-variable relationships while preserving data variability and uncertainty.

Purpose of the Study:

• To develop an innovative spatiotemporal hybrid model for enhanced PM2.5 concentration prediction in urban areas.
• To address data imputation challenges by leveraging machine learning for analyzing relationships between meteorological variables and other pollutants.

Main Methods:

• Employed machine learning techniques for missing data imputation, analyzing relationships between meteorological variables and pollutants.
• Introduced the AC_GRU model, integrating 1D CNN, GRU, and an attention-based network for spatiotemporal PM2.5 prediction.
• Utilized meteorological variables, nearby PM2.5 data, and other pollutant concentrations as model inputs.

Main Results:

• The AC_GRU model effectively learns spatiotemporal correlations within time-series data, improving PM2.5 prediction accuracy.
• The attention mechanism enhances prediction by dynamically weighting past input variables based on their relevance.
• Experimental results show the AC_GRU model outperforms existing state-of-the-art methods in PM2.5 forecasting.

Conclusions:

• The developed AC_GRU model offers a significant advancement in predicting urban PM2.5 concentrations.
• This model serves as a valuable tool for effective urban air quality management and safeguarding public health.