Abstract: Traditional support parameter optimization for large-span excavated stations constructed using the arch-cover method relies on repetitive numerical simulations, resulting in low efficiency. To address this limitation, a data-driven intelligent optimization framework is proposed. First, a case study of a metro station constructed using the arch-cover method in upper-soft and lower-hard strata is conducted, and a specialized engineering parameter database is established based on Latin hypercube sampling and logical coupling rules for geological parameters, supported by cooperative automated computation between Python and FLAC3D. Subsequently, multiple machine learning algorithms are evaluated, and extreme gradient boosting (XGBoost) is selected as the displacement prediction model because of its superior predictive performance. The model hyperparameters are further optimized using a four-parameter adaptive growth optimizer (QAGO), which significantly improves prediction accuracy. Finally, to overcome the limitations of single metaheuristic algorithms, including susceptibility to local optimal and overly conservative solutions, a hybrid optimization strategy integrating the slime mold algorithm (SMA) and the honey badger algorithm (HBA) is developed, with support economy incorporated as an optimization objective. The results indicate that (1) the proposed database generation method efficiently produces datasets that conform to engineering logic, (2) the QAGO-optimized XGBoost model provides a reliable surrogate model for support parameter optimization, and (3) the SMA-HBA hybrid optimization framework rapidly identifies support schemes that satisfy safety constraints while achieving improved economic performance.

Key words: mined metro station, arch-cover method, support parameter optimization, displacement prediction, data-driven approach, machine learning