Virtual Coupling railway technology allows trains to conduct dynamic formation and reconciliation according to the actual passenger transport needs. The trains within the formation operate at small intervals close to the length of mechanical hooks, realizing local ultra-high density driving and collaborative control. In order to ensure that the train can also operate safely when the location and speed information is missing, this paper proposes to predict the train track and perceive the future train operation state in advance to realize the safe operation. Firstly, the existing urban rail train braking model is analyzed, and the highly accurate train braking model is identified by the improved Particle Swarm (PSO) algorithm; then after analyzing Long Short-Term Memory (LSTM) and Unscented Kalman Filter (UKF) respectively, the Virtual Coupling train track prediction algorithm based on LSTM and UKF is proposed. Finally, the actual operation data construction environment, simulation prediction and verification. The simulation experiment results show that: compared with the simple LSTM trajectory prediction, the formation train trajectory prediction model based on LSTM-UKF can further reduce the prediction error and instability, which is conducive to the safe operation of the virtual formation train.