Time synchronization has been widely used in the distributed system. Clock calibration is the precondition of the time synchronization. Aimed at the problem that in the process of the invalidation of time signal's transmission in the calibration may exist in time synchronization, this paper introduces a clock error prediction into the synchronization system. Simultaneously, to improve precision of error prediction, the paper proposes a model based on GM-BP network. In the combined model, different GM (1,1) models based on observational data of clock errors are built. The models can predict the future clock errors. The paper employs the BP network to combine the prediction results in a nonlinear manner. The time signals transfer via troposcatter imitated by an attenuator to prove the precision of the combined model. The result shows that the accuracy of the combined model is superior to the single model, the error of the new model has a better stability and the precision increases by 53%~95%.