Air pollutant prediction model based on transfer learning two-stage attention mechanism

Area of Science:

• Environmental Science
• Data Science
• Machine Learning

Background:

• Atmospheric pollution poses significant risks to regional economies and human health, necessitating accurate prediction methods.
• Traditional prediction models struggle with new monitoring sites due to insufficient historical data.
• Transfer learning offers a promising approach to overcome data limitations in environmental monitoring.

Purpose of the Study:

• To develop an advanced prediction model for atmospheric pollutants that overcomes data scarcity at new sites.
• To enhance the accuracy and reliability of air quality forecasting.
• To improve the transferability of prediction models across different monitoring locations.

Main Methods:

• A two-stage attention mechanism model based on transfer learning (TL-AdaBiGRU) was proposed.
• The model segments pollutant sequences into periods using a temporal distribution characterization algorithm and a temporal attention mechanism.
• A multi-head external attention mechanism mines hidden layer features, with knowledge transferred from source to new sites.

Main Results:

• The TL-AdaBiGRU model effectively mines critical information from periodic segments of air pollutant sequences.
• The two-stage attention mechanism successfully captures complex nonlinear relationships within air pollutant data.
• The model demonstrated significant improvements over existing methods, reducing MAE by 14%, RMSE by 13%, and MAPE by 4%.

Conclusions:

• The proposed TL-AdaBiGRU model offers a robust solution for atmospheric pollution prediction, particularly in data-scarce environments.
• Transfer learning combined with a novel two-stage attention mechanism enhances prediction accuracy and model generalizability.
• The findings highlight the potential of advanced machine learning techniques for environmental monitoring and public health protection.