To address the challenges of ultra-short-term load forecasting in plateau traction power supply systems caused by low utilization of regenerative braking energy

severe voltage fluctuations

and the lack of real-time train operation commands

this paper proposes an ultra-short-term load forecasting method for intraday dispatch and real-time control of the source-grid-train-storage system. First

a second-level feature quantification framework is established

and a multi-stage dynamic feature selection model is proposed to eliminate redundant features and achieve lightweight modeling. Then

a hybrid forecasting model integrating complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN)

convolutional neural network (CNN)

bidirectional long short-term memory (BiLSTM)

and random forest (RF) is constructed. Bayesian optimization is introduced to adaptively optimize the model hyperparameters. Finally

case study results show that the proposed method can effectively improve the accuracy of ultra-short-term forecasting for plateau traction loads

exhibiting strong dynamic feature extraction capability

tracking performance

and stability. The method provides a reliable forward-looking decision-making basis for real-time energy management.