Addressing nonlinearities and non-stationarity of load prediction in metro power systems, this study introduces a novel hybrid prediction model that combines PSO with Bayesian optimization, integrated into a VMD-LightGBM-LSTM framework. Firstly, to reduce VMD decomposition residuals and enhance the extraction of temporal periodicity, the PSO optimization algorithm, guided by a combined objective function, is employed to optimize the key parameters of VMD, yielding the optimal parameter configuration. Subsequently, the metro power load is decomposed into modal sequences at different frequencies. These modal sequences are then fed into LightGBM and LSTM prediction models. The prediction results of each model are weighted and combined, with Bayesian optimization applied to optimize the weighting coefficients to achieve the best prediction results. Finally, using real operational data from the Qingdao metro in Shandong, metro power load prediction simulations are performed to validate the proposed method’s effectiveness and superiority in multi-step short-term predictions over the next 12 h.