As the core component of the traction drive system of high-speed trains, the working condition of the traction motor is directly related to the operational safety of the train. During train operation, the temperature signal of the traction motor is constantly changing. Accurate prediction of the traction motor temperature using real-time data generated by sensors at the train end is beneficial for the early detection of abnormal motor conditions. The traditional approach is to apply prediction models from offline learning to the onboard side of the train for real-time temperature prediction, but its prediction accuracy degrades with the online data distribution changed. Therefore, this paper proposes an online deep learning model (ODL) for real-time traction motor temperature prediction. The ODL model uses DNN as its basic structure and hedging strategy to achieve online learning of the model parameters, which can dynamically adjust the model structure and parameters to adapt to streaming data with changing probability distributions. Finally, the study conducts multiple motor temperature prediction experiments using multi-sensor data from train operations. The results show that the ODL model outperforms the currently popular models in temperature prediction and can be effectively applied to real-time temperature prediction on the train end.