A Temporal–Section–Lane–Class Network for Freeway Lane-Level Mixed Traffic Flow Prediction

In vehicle-infrastructure cooperative systems, mixed freeway traffic comprising human-driven vehicles (HVs) and connected and automated vehicles (CAVs) presents new challenges for fine-grained traffic state monitoring and control. Lane changes and queue propagation, together with differences in HV and CAV driving behavior, produce traffic states that vary across space and vehicle type. We formulate lane-level mixed-traffic flow prediction as a joint forecasting task over the section, lane, and vehicle-class dimensions and develop a Temporal–Section–Lane–Class Network (TSLCN). A shared gated causal temporal convolutional network captures common temporal dynamics, while class-specific adapters preserve class-dependent characteristics. A section–lane–class (SLC) interaction module models longitudinal associations across monitoring sections, lateral interactions between adjacent lanes, and local HV–CAV dependencies. These representations yield class-specific forecasts for every lane at each monitoring section. Experiments combined observed lane flow data with mixed traffic scenarios simulated in Simulation of Urban MObility (SUMO) using demand obtained from field data. At 30% CAV penetration, TSLCN achieved the lowest aggregate MAE, RMSE, and WAPE and the lowest mean error in 68 of 72 comparisons. Its MAE reductions over the strongest baseline were statistically significant at all forecast horizons. Ablation analysis showed that each principal component contributed to prediction accuracy. TSLCN also achieved the lowest overall WAPE at CAV penetration rates from 0% to 100% and the lowest aggregate errors on an alternative freeway, with only modest sensitivity to key SUMO parameters. Overall, TSLCN provides reliable forecasts by lane and vehicle class, with potential applications in fine-grained freeway management.