The statistical parameters of 10-minute mean wind speeds based on the records of 29 meteorological stations in Taiwan were investigated. The third and fourth moments, skewness and kurtosis coefficients, were found highly related to the tail characteristics of extreme value distributions. It was pointed out that a Weibull or Frechet model showed better agreement with observed data at certain stations. Anon-Gaussian simulation technique was adopted for simulation of annual maximum wind speeds. From the variations of four moments, the non-identical and independent distributed feature was indicated and the significant scattering in the simulation was showed to imply the difficulty of prediction. It was then demonstrated that sometimes even a short period of meteorological records, say 10 years, could provide a fairly good prediction of extreme wind speeds once the higher statistical moments were estimated well.