Transformer-based approach for roadside air pollutant prediction using multi-site smart lamppost data

Roadside air quality significantly impacts public health and the urban ecosystem. This study develops a Transformer-based deep learning model to predict short-term roadside NO 2 and PM 2.5 concentrations using data from multi-sensor smart lampposts at three tested sites in Hong Kong. The model integrates traffic volume, speed, and meteorological factors through a self-attention mechanism that captures complex temporal dependencies. The smart lamppost data, collected directly at the roadside, provide high-fidelity measurements without the need for emission estimation algorithms. A year-long dataset spanning hourly, weekly, and seasonal variations is used for training and validation, with single-site, cross-site, and mixed-site testing strategies. The Transformer is compared against statistical, neural network, and tree-based machine learning baselines, achieving promising prediction accuracy. SHAP and sensitivity analyses reveal that NO 2 is more sensitive to traffic factors while PM 2.5 is more influenced by meteorological conditions.