Transformer-based forecasting of national NO₂ concentrations with spatiotemporal and attributive dependencies

The prediction of nitrogen dioxide (NO₂) concentration is crucial for protecting human health and controlling environmental pollution. However, the complex temporal patterns and rapid time fluctuations pose significant challenges to accurate NO₂ forecasting. Some existing studies have introduced machine learning techniques like Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks to extract richer temporal features, while they are still struggling with effectively capturing long-term dependencies. Moreover, most studies focus primarily on individual monitoring stations, often overlooking the spatial correlations between stations, which limits the ability to make comprehensive predictions for larger regions. To address these issues, this study expands the scope to include all monitoring stations across China. By employing Transformer models, we aim to extract long-term dependencies at multiple temporal scales while incorporating spatial and attributive distances to facilitate information sharing among different monitoring stations. Our objective is to achieve holistic prediction of NO₂ concentrations nationwide and make analysis for the future trend. The findings of this research are expected to provide valuable theoretical support for proactive environmental pollution management and prevention.