Maintaining Car-Following Prediction Under Naturally Occurring Sensing Loss in an Expressway Tunnel Cluster: A Route-Specific Projector-Transformer Case Study

To address the spatially concentrated loss of relative speed and spacing measurements in expressway tunnel clusters, this route-specific case study evaluates a Projector-Transformer Deep Learning (PTDL) pipeline that adds a linear reconstruction head to a Transformer following-speed predictor. Sixty trajectory segments came from two instrumented-vehicle trials on a 42.086 km route with 15 tunnels. Errors were summarized over tunnel cluster cells (Tce). Coupled variants share an encoder under a weighted dual-task loss. Decoupled variants train separate encoders sequentially, freezing reconstruction during forecasting. Under the reported development records, decoupling reduced missing-feature MSE by 40.7% (from 0.0246 to 0.0146), while complete-feature MSE was nearly identical (0.0075 versus 0.0074). With complete inputs, D-PTDL-S gave a lower ASE than the reported IDM, LSTM, and D-PTDL values over one and two Tce. In the naturally missing subset, its ASE was 45.0–52.6% lower than D-PTDL across one to five Tce. The 60 trajectory segments were partitioned segment-wise into disjoint training, validation and test subsets (80/10/10), with validation used solely for early stopping and the test subset read once after training. The baselines were neither information- nor calibration-matched, no repeated-seed inference was performed, and no ground truth is available within the naturally missing intervals. The results therefore establish route-specific feasibility rather than cross-route generalization. Safety, energy, emissions, and life-cycle outcomes were not measured. Sustainability relevance is limited to operational continuity under incomplete sensing.