Environmental Assessment of PM2.5 Concentration Patterns Through TC-MixerInformer Modeling: Cross-Regional Analysis of Shanghai and London

Urban air quality prediction faces dual challenges: complex pollution processes and cross-regional diversity. Traditional methods and deep learning techniques often struggle with non-stationary time series data and fail to adapt to unique urban pollution patterns. Additionally, existing models face computational limitations when processing long sequences. Although the Informer model improves computational efficiency through ProbSparse self-attention mechanisms, its accuracy decreases with longer prediction horizons. This limitation stems from inadequate adaptability to non-stationary environmental changes and insufficient capture of temporal variations inherent in urban air quality data. Meanwhile, cities in developed and developing countries exhibit fundamentally different pollution mechanisms, challenging models’ cross-regional generalization capabilities. To tackle these two challenges, this study proposes TC-MixerInformer, combining Reversible Instance Normalization (RevIN) with a Temporal-Channel Mixer (TCMixer) module. Validation using Shanghai (Jing’an) and London (North Kensington) monitoring stations demonstrates excellent performance in both short-term (1-12 h) and long-term (24-168 h) predictions, with 8%-54% error reductions compared to baseline models. The model effectively handles Shanghai’s complex pollution patterns (2-378 μg/m³, mean 28.23 μg/m³) and London’s lower concentrations (0-121.56 μg/m³, mean 8.08 μg/m³). RevIN addresses time series non-stationarity while TCMixer enhances multi-scale feature expression, maintaining stable performance across different time scales. The model shows particular advantages in predicting extreme pollution events, especially capturing substantial peaks approaching 350+ μg/m³ in Shanghai. Our research provides a new technical approach for addressing scale diversity and temporal non-stationarity in urban air quality prediction.