One of the major issues in large-scale wind power generation is dealing with wind power ramp-down events, but few common methods that properly describe the extreme value distribution are reported. Based on the field measurement, this paper studies statistical models of empirical quantile of wind power ramp-down magnitudes in different temporal scales. The ratio of the extreme ramp-down value to the standard deviation of wind power fluctuations is relatively large in shorter temporal scales. It's found that generalized Pareto distribution is suitable to identify the upper-tail probability of wind power ramp-down. The T year return level of wind power ramp-down magnitudes can be used as index of extreme ramp-down. The paper then establishes the generalized Pareto distribution model under different installed capacity and temporal scales. Analysis shows that the ratio of extreme wind power ramp-down magnitude to installed capacity gets smaller with a trend of saturation when the installed capacity increases and there is a non-linear slow growth in extreme wind power ramp-down magnitude when temporal scale increases.