Air Quality Prediction Based on Confocal Time-Frequency Feature Fusion Residual Networks

In the realm of air quality pollutant forecasting, addressing the intricate spatiotemporal dynamics of pollutant dispersion across varying times and regions is a critical challenge. In this paper, we present an advanced model for analyzing spatiotemporal patterns of air quality index using residual networks with confocal fusion of time and frequency domain features (CTFF-AQP). The model is designed with a novel feature fusion approach to capture time-frequency correlation in the pollutant dispersion process and to improve the accuracy of prediction. Initially, a preprocessing module encodes air quality indicators and geospatial data into temporal and spatial features. Subsequently, two key modules are employed: a full-scale spatiotemporal feature extraction module and a regional attention-enhanced temporal convolution module. These modules are designed to extract spatiotemporal features from diverse data sources. The fusion module then effectively integrates these features, enabling CTFF-AQP to capture and synthesize spatiotemporal patterns and temporal correlations in air quality data across individual, global, and local scales. Experiments demonstrate CTFF-AQP's superior performance in both shortand long-term air pollutant prediction, achieving an <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{R}^{\mathbf{2}}$</tex> of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 9 4}$</tex> and an 11.3 % accuracy improvement over state-of-the-art models.