Clock difference prediction is a key technology of time keeping work. Kalman algorithm as a kind of optimal prediction algorithm, has the characteristics of real-time, widely used in time keeping work. The classical Kalman algorithm needs to accurately know the random error of the model and measurement error, otherwise, the state estimate will bring error, which is characterized by atomic clocks noise and clock difference measure noise in the atomic time algorithm. Noise parameters of atomic clock is usually estimated through Allan variance, if estimate is not accurate, Kalman filter prediction error will appear. In this paper, the adaptive Kalman filtering prediction algorithm based on window of Sage was researched, the state model error was corrected real-time. The model error was reduced by adjusting the state prediction covariance matrix using adaptive factor, which improved the prediction accuracy. Finally, the effectiveness of the algorithm was verified by the measured data of two hydrogen atomic clocks and two cesium atomic clocks.