In response to the problem of falls caused by incoordination between the walking speed of the elderly and the designated speed of the walking rehabilitation training robot during rehabilitation, this paper proposes a human-robot speed coordination anti-falling method, consisting of two parts: a falling prediction model and an anti-falling control method. First, the walking posture signals of the subject are collected by an inertial measurement unit (IMU), and a falling prediction model for the elderly is constructed using long short-term memory (LSTM) network and attention mechanism. Second, based on the falling prediction, a multi-dimensional phase space reconstruction (PSR) speed prediction model is designed for the anti-falling controller. Finally, the predicted speed of the subject is used as the target speed, and the PSR theory and model predictive control (MPC) technology are used to design an anti-falling controller for the walking rehabilitation training robot, achieving precise tracking of the subject’s walking speed and preventing falls caused by incoordination between the subject’s walking speed and the robot’s designated speed during rehabilitation training. Simulation and experimental results show that the prediction accuracy of the falling prediction model can reach 95.2%, and the lead time for falling prediction can reach 1.82 s. The human-robot speed coordination anti-falling method can effectively prevent falls caused by incoordination between the subject’s walking speed and the robot’s designated speed, enabling the subject to complete walking rehabilitation training safely.