MTLPM: a long-term fine-grained PM2.5 prediction method based on spatio-temporal graph neural network.

Area of Science:

• Environmental Science
• Data Science
• Artificial Intelligence

Background:

• Particulate Matter 2.5 (PM2.5) is a key air quality indicator.
• Current PM2.5 prediction models often overlook spatial interdependencies and struggle with long-term forecasting.
• Existing methods typically focus on single monitoring stations, neglecting environmental factors like air circulation.

Purpose of the Study:

• To develop an advanced model for accurate long-term PM2.5 concentration prediction.
• To incorporate spatial dynamics and complex environmental factors into air quality forecasting.
• To improve upon existing methods by addressing limitations in spatial analysis and prediction duration.

Main Methods:

• Proposed MTLPM, a spatio-temporal graph neural network with an encoder-decoder architecture.
• Utilized a message passing mechanism with spatial and environmental data (temperature, humidity, wind) for real-time spatial information.
• Employed Multi-head ProbSparse Self-attention for extracting temporal features and long-term dependencies.
• Implemented a generative one-step decoder for simultaneous multi-station, long-term forecasting.

Main Results:

• MTLPM demonstrated superior performance compared to state-of-the-art methods.
• Achieved an average reduction of 1.6 in Mean Absolute Error (MAE).
• Obtained an average reduction of 0.02 in Symmetric Mean Absolute Percentage Error (SMAPE).

Conclusions:

• MTLPM effectively captures spatio-temporal patterns for improved PM2.5 prediction.
• The model offers enhanced accuracy in long-term air quality forecasting.
• The proposed approach provides a significant advancement in predicting PM2.5 concentrations across multiple stations.