As the number of users in quantum key distribution (QKD) networks continues to grow, parameter selection (e.g., probability of X-bases, signal state strength, and decoy state levels) has become increasingly complex. Traditional methods such as local search algorithms (LSA) are commonly employed for parameter optimization but suffer from significant computational overhead, slow execution on low-power platforms, and resultant system delays. These limitations hinder the scalability and real-time requirements of large-scale QKD networks. To address these challenges, this paper proposes an innovative approach that combines Bayesian optimization with XGBoost model to optimize QKD system parameters. The proposed model is evaluated against random forest (RF) and back-propagation neural network (BPNN) models based on three critical metrics: prediction accuracy, model fitting, and training efficiency. Simulation results demonstrate that the proposed model significantly outperforms both RF and BPNN in terms of accuracy and computational efficiency. Notably, while achieving predictive performance comparable to LSA, the proposed method reduces computation time by three orders of magnitude. This research introduces a novel and efficient framework for QKD parameter optimization, providing new insights and directions for enhancing the performance, scalability, and real-time adaptability of QKD systems.