When the non-invasive stimulation such as transcranial electric stimulation locks phases with the intrinsic neural electrical activity in the brain, the neural oscillatory activity can be regulated in a more effective manner. Due to the complex time-variation of EEG signal, the existing methods cannot meet the accuracy of phase estimation and real-time performance of the system at the same time. In this paper, a real-time phase estimation method for phase-locked stimulus system was proposed. In this method, the EEG signal was modeled by autoregressive (AR), then the AR model was used to predict the EEG signal and identify the phase feature points, and the phase to be stimulated was calculated by the phase feature points. The method was used to analyze the closed eye resting EEG of 20 subjects (aged 20~ 36, male 12, female 8) and it was found that the performance of the method is related to the updating time of the model coefficient, the prediction step and the narrow-band power of the EEG. It had better performance for the EEG with higher narrowband power. Under the optimal model parameters (the updating time of the model coefficient was 5 s and the predicted step length was 30), the average phase locking value (PLV) of the 20 subjects was 0. 968, and the average phase error was 13. 33. Compared with the average period method, this method has higher PLV value and lower phase error, which can be used in the development of closed-loop phase-locked electric stimulator.