Abstract: To solve the problem of difficult deformation prediction of platform doors on in-service lines, this paper constructed two prediction models: a Bayesian theory-optimized Support Vector Regression (SVR) model and a Generalized Regression Neural Network (GRNN) model optimized via grid search and K-Fold Cross Validation. The paper adopted Bayesian optimization to determine the core parameters of the SVR model, thus avoiding local optima; for the GRNN model, grid search and K-Fold Cross Validation were applied to optimize its smoothing factor, which also prevents the factor from falling into local optima. Subsequently, the paper trained the models using deformation data of platform doors under dozens of actual operating conditions, before conducting prediction tests with the trained models. The prediction results show that the Bayesian-optimized SVR model outperforms the GRNN model optimized by grid search and K-Fold Cross Validation. Compared with the current optimal machine learning model for platform door deformation prediction, its average absolute error (MAE) is reduced by 0.0643, which can provide data support for structural health assessment and operation maintenance decision-making of platform doors.