Ultra-short-term precise power prediction is crucial for achieving intelligent control of traction bidirectional converter devices in urban rail transit systems. Accurate power prediction can not only optimize energy distribution and improve system efficiency but also effectively reduce energy consumption and communication operation costs. Therefore, a power prediction method based on deep learning is proposed. This method collects information on the kilometer markers of trains, as well as the DC voltage and DC current information of traction substations, to perform ultra-short-term prediction of the active power at the station. This effectively improves the accuracy of the prediction, thereby achieving more intelligent and sustainable control. Tested with actual operating data from a metro system in China, the deep learning model developed in this paper demonstrates low error, low complexity, and prediction performance that meets engineering requirements when used for power prediction.