Based on the predicting problem of the non-equidistant GM(1,1) model, a method of optimizing initial condition is proposed. The initial value is constructed by the weighted average of each component of the non-equidistant first-order accumulative generation operator (1-AGO) sequence on the original sequence. According to the principle of information priority, the weighted coefficients are the respective components of unitization sequence of ordinal sequence which is corresponding to the non-equidistant 1-AGO sequence. The time parameter in initial condition is derived from a method of minimizing error summation of square between the raw sequence and the simulative sequence. Thus the non-equidistant GM(1,1) model is established. Finally, the validity and practicability of the non-equidistant optimal model proposed are verified by examples. And the result shows that the optimal model can significantly improve the simulation and prediction accuracy.