Attention-based parallel networks (APNet) for PM_2.5 spatiotemporal prediction.

Area of Science:

• Environmental Science
• Data Science
• Artificial Intelligence

Background:

• Accurate forecasting of urban fine particulate matter (PM2.5) is crucial for air pollution management.
• Existing prediction methods struggle with the complex nonlinearity and spatiotemporal dependencies of PM2.5 concentrations.

Purpose of the Study:

• To develop an advanced model for precise 72-hour PM2.5 concentration forecasting.
• To simultaneously capture short-term and long-term temporal features and spatial dependencies.

Main Methods:

• Proposed an attention-based parallel network (APNet) integrating Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks.
• Utilized Maximum Information Coefficient (MIC) for spatiotemporal correlation analysis.
• Incorporated an attention mechanism and a Bi-LSTM module for enhanced feature extraction and interpretability.

Main Results:

• APNet demonstrated superior performance compared to existing state-of-the-art methods in PM2.5 prediction.
• Achieved high recall (0.790) and precision (0.848) for 72-hour forecasts.
• The model effectively extracts both short-term mutations and long-term periodic characteristics.

Conclusions:

• The proposed APNet model offers a feasible and effective solution for accurate PM2.5 forecasting.
• The methodology holds potential for predicting other multivariate time series data.
• APNet enhances air quality management through improved early warning systems.