This study focuses on grid–hourly PM_2.5 modeling and identifying traffic impacts in the urban core of Hangzhou. A multi-source heterogeneous dataset is constructed by integrating air quality, meteorological, traffic, and spatial information, and an interpretable nonlinear prediction and scenario-evaluation framework is developed. By comparing Random Forest, XGBoost, and linear regression models, and combining ablation experiments with SHAP analysis, we quantify the contributions of traffic intensity and fleet composition to spatiotemporal PM_2.5 prediction. The results show that overall traffic intensity and heavy-duty fossil-fuel vehicle–related features significantly influence PM_2.5 variations, and these effects are more sensitive under low wind-speed conditions. Scenario simulations further indicate that traffic flow restrictions and fleet-structure optimization yield greater mitigation benefits in traffic-dense areas and during high-pollution periods. This study provides quantitative evidence and a modeling basis for refined urban traffic management and for evaluating the effectiveness of electrification policies.