An enhanced approach for weaving area capacity estimation combining high‐fidelity simulation and interpretable machine learning

Accurately assessing weaving area capacity is critical for optimizing traffic management. However, transportation agencies face a persistent challenge: Existing capacity models often inadequately characterize vehicle weaving behaviors and fail to quantify the interactions between influencing factors. This limitation hinders the development of precise traffic control strategies in weaving areas. Addressing this, we propose an integrated methodology combining enhanced microscopic simulation with interpretable machine learning. The proposed method is tested and validated on two datasets. The DIEGA calibration-improved method integrates density-based spatial clustering of applications with noise (DBSCAN) clustering, information entropy, and genetic algorithms (GA) to achieve superior modeling accuracy and 22.2% faster convergence than GA. Simulation experiments demonstrate that under a constant weaving flow of 1900 pcu/h, variations in ramp-to-freeway ( Q RF ${Q}_{\mathrm{RF}}$ )/freeway-to-ramp ( Q FR ${Q}_{\mathrm{FR}}$ ) ratios induce a capacity fluctuation of approximately 15% (ranging from 4277 to 4937 pcu/h) and indicate nonlinear coupling among Q RF ${Q}_{\mathrm{RF}}$ , Q FR ${Q}_{\mathrm{FR}}$ , weaving length ( L W ${L}_{\mathrm{W}}$ ), and capacity. The ML_RF capacity model outperforms the baseline model while providing sHapley additive exPlanations (SHAP)-based interpretability of factor interactions. The methodology's demonstrated capability to identify optimal capacity ranges (4635–4860 pcu/h at Q RF ${Q}_{\mathrm{RF}}$ / Q FR ${Q}_{\mathrm{FR}}$ ≈ 1 with L W ${L}_{\mathrm{W}}$ = 250–350 m) provides transportation agencies with a potential decision-support tool for both weaving area design optimization and operational traffic management.