Physics-guided machine learning for traffic-flow prediction and operating-regime attribution on a desert heavy-duty expressway

Sparse roadside infrastructure creates long monitoring gaps on desert expressways. This study develops a physics-guided machine-learning framework for downstream traffic-flow prediction and operating-regime attribution using sparse volume observations. For a 40.9-km section of the G30 Lianhuo Expressway in Xinjiang, a conservation-motivated boundary-flow imbalance feature was constructed from temporally aligned upstream and downstream flows. K-means was fitted to standardized training-period data, after which fixed preprocessing parameters and centroids were used to assign validation and test observations. A Tree-structured Parzen Estimator-optimized XGBoost model provided one-step-ahead predictions, and SHapley Additive exPlanations quantified model-based feature contributions. Under a chronological train-validation-test partition, the model achieved an R 2 of 0.9092 ± 0.0043, an RMSE of 129.43 ± 2.35 pcu/h, and an MAE of 98.83 ± 1.57 pcu/h. These were the lowest mean errors among the evaluated configurations, although the difference from the strongest baseline was modest. Ablation showed that the boundary-flow imbalance feature provided complementary predictive information. Attribution patterns varied across the data-driven regimes but are interpreted as model associations, not causal effects or evidence of physical traffic states. The framework supports interpretable short-term prediction where continuous speed and density observations are unavailable.