For the weighting method in combination forecasting with multiple single predictions, a non-negative time-variant weighting method is given by combining the sliding window model and an introduced modified variation coefficient. Traditional variation coefficient on the level of predicted data is transformed to the level of predicted residual error data, in which the affection of mean in high level is deleted to avoid traditional variation coefficient in low level. Besides, by introducing the sliding window model, current existed weighting methods and the proposed weighting based on modified variation coefficient are improved, and the transformation from time-invariant weights to time-variant weights is realized. The numerical study shows the validity of the developed modified variation coefficient, which also shows the efficiency of the sliding window model in improving the combined forecasting accuracy.