Abstract: The latest beginning date of late frost in springs in 1957-2013 about Yangcheng county of Shanxi province was as zero point, and then the relative dates that they were zero point minus those appearance dates could be obtained More recent，the abrupt date from zero point was the catastrophic threshold.Further, the interannual variation of the relative date and its intersection point with the threshold line were drawn Then, the distance about the second intersection point and the first point was worked out.And every other distance behind the second intersection point and the first point were obtained.Finally, the original date sequence of Model GM (1, 1) was obtained, and the traditional model GM (1, 1) was set up.This model could be improved through those methods, which had majorized sequence, the background value, pessimistic sequence of equidimensional filling vacancies method and so on.And different improved models had been tested by means of the relative error, variance ratio and small error probability.The optimal model tested could be applied to forecasting late frost.The results showed that, this model established through the three improvement way had the best prediction effect, .The results can be a referentially theoretical basis for defending late frost.

Key words: Improvement, Model GM(1,1), Late frost, Forecast