A multiscale spatial-temporal-variable feature fusion network for predicting multiple air pollutants.

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Science

Background:

• Accurate urban air quality prediction is crucial for environmental management and public health.
• Existing models often focus on single pollutants, neglecting complex inter-pollutant relationships and limiting prediction accuracy.

Purpose of the Study:

• To develop a novel network for accurate, joint prediction of multiple air pollutants.
• To address the limitations of single-pollutant models by incorporating multi-scale spatial-temporal and inter-pollutant feature interactions.

Main Methods:

• Proposed a Multiscale Spatial-Temporal-Variable Feature Fusion Network (MSTVFFN).
• Utilized dedicated modules for extracting temporal (individual, local, global scales), spatial (global, local), and variable (inter-pollutant correlations) features.
• Employed a feature fusion module in the decoder to integrate multi-dimensional interactions.

Main Results:

• MSTVFFN demonstrated significant performance gains on Beijing, London, and Wuhan datasets for 12- and 24-hour joint predictions of four pollutants.
• Achieved 11%-33% reductions in Mean Absolute Error (MAE) and 3%-16% improvements in R-squared (R²) compared to state-of-the-art models.
• Showcased consistent performance across different pollutants and cities.

Conclusions:

• The MSTVFFN provides a robust framework for multi-pollutant air quality forecasting.
• This advancement offers a valuable tool for informed decision-making in environmental management and public health protection.
• Source code is publicly available for further research and application.