DSTMA-BLSTM algorithm for roadside air pollutant time series prediction and sensitivity analysis

Road traffic pollution greatly affects urban air quality, making accurate prediction of roadside pollutant concentrations essential for effective environmental management. This study presents a novel DSTMA-BLSTM algorithm, which combines Dynamic Shared and Task-specific Multi-head Attention (DSTMA) with Bidirectional Long Short-Term Memory (BLSTM) networks, to forecast temporal changes in roadside pollutants and analyze their sensitivity. Using real monitoring data, the study identifies wind speed and the counts of gasoline and diesel vehicles as critical factors influencing roadside pollutant levels. The model achieved outstanding predictive performance for NO, NO₂, and CO₂, with R² values of 0.959, 0.944, and 0.949, respectively, demonstrating its exceptional ability to capture the dynamics of traffic-related pollutants. This work not only establishes the DSTMA-BLSTM model as a powerful tool for multi-pollutant forecasting but also proposes a fresh perspective for jointly predicting traffic and non-traffic-related pollutants in future research.