CNN-LinATFormer: Enhancing PM2.5 Prediction Through Feature Assessment and Linear Attention Mechanism

Atmospheric fine particulate matter (PM2.5) poses a serious threat to public health, and its accurate prediction is crucial for environmental management and pollution control. However, existing prediction methods have difficulty in effectively capturing the complex nonlinear characteristics and multi-scale spatiotemporal dependencies of PM2.5 concentration changes. To address this challenge, this study proposes a CNN-LinATFormer hybrid deep learning architecture that combines the local feature extraction capabilities of CNN with the global dependency modeling advantages of the linear attention mechanism. The model innovatively introduces a feature evaluator to dynamically classify environmental features into three categories, and achieves targeted processing through three specially designed processing branches: CNN feature extraction, channel attention, and linear attention fusion. Based on the urban monitoring data of 9 environmental feature dimensions from 2020 to 2023, the experimental evaluation results show that CNN-LinATFormer outperforms the existing methods in all evaluation indicators, with an RMSE of 8.42μg/m³, which is 21.1% lower than the CNN-RF model with the closest performance; the ablation experiment confirms the effectiveness of each component, especKeywords-PM2.5 prediction; air quality forecasting; deep learning; convolutional neural network; linear attention mechanism; channel attention; feature assessment; hybrid model architecture; environmental monitoring; spatiotemporal modeling the channel attention mechanism; the case analysis reveals that the model performs well in the low concentration range (RMSE is 3.12μg/m³), but the high pollution range (>150μg/m³) still needs to be improved. This study provides a new technical path for air quality prediction, which is of great value to environmental monitoring and public health protection.