AutoML‐Enhanced Delay Forecasting With SHAP Interpretability in Highway Work Zones Under Diversion Constraints

Abstract

The increasing scale of highway reconstruction and expansion projects has intensified traffic management challenges in construction zones, particularly within sparse road networks constrained by limited diversion capacities and elevated freight truck ratios. This study proposes an integrated analytical framework that combines microscopic simulation, automated machine learning (AutoML), and explainable artificial intelligence. Traffic flow dynamics under high truck proportions (72%) were modeled using the VISSIM microsimulation, generating 1320 parameterized scenarios encompassing traffic volume, work zone length, speed limits, and vehicle composition. By leveraging the AutoGluon AutoML framework, we developed an ensemble delay prediction model using optimized feature engineering. SHapley Additive exPlanations (SHAP) interpretability analysis further decoded the multifactorial coupling mechanisms influencing traffic organization. The results demonstrate that while complex ensembles achieved the lowest error (RMSE = 1.49), the CatBoost_BAG_L1 model was identified as the optimal model for operational deployment, achieving identical accuracy with a more than 25‐fold improvement in computational speed. The SHAP‐based interpretation revealed traffic volume as the dominant delay contributor, exhibiting nonlinear dynamics with escalating marginal effects beyond 1400 vehicles/h. Increasing the speed limit to 80 km/h elevated delays by 0.58 units, while work zones exceeding 2 km in length induced length‐proportional delay amplification. This methodology advances intelligent decision‐making for dynamic lane control and truck scheduling optimization in diversion‐constrained environments.