With the development of wide area measurement system, it is possible for power system transient instability control based on real-time trajectory. Using real-time data to predict can take control measures as soon as possible, which can prevent out of step. Therefore, it is very important to study high-precision and long-time trajectory prediction. Proceeding from motion equations of generator, a rolling self-memory prediction method for power system is deduced. First, trigonometric function is used to fit unbalanced power; then, angular velocity is predicted by unbalanced power; at last, according to the self-memory prediction formula, power angle can be obtained. The simulation result of IEEE9 system shows that, the method mentioned in this paper can accurately predict at least 0.5s’ future trajectory. Compared with trigonometric function prediction and autoregressive prediction, this method can predict in a longer time with a higher accuracy.