In order to investigate the non-stationarity and prediction method of hydrological extremes under the climate change impact, statistical prediction models for the non-stationary generalized extreme value (GEV) distributions of hydrological extremes were developed. Annual mean North Hemisphere polar vortex intensity was selected as the predictor of distribution parameters. Observational hydrological data from the upper reaches of the Huaihe River basin (HRB) and 500 hPa circulation monitoring data during 1952-2010 and 17 CMIP5 GCM simulations for historical, RCP2.6, RCP4.5 and RCP8.5 greenhouse gas emission scenarios were used for the model fitting. Responses of hydrological extremes to climate change during 2006-2050 under the 3 future scenarios were analyzed by using the fitted models. The results show that hydrological extremes with less- and longer-than-10-year return periods from non-stationary GEV distributions may have decreasing and increasing trends, respectively. However, ensemble means of multi-model predictions for different return periods under the 3 future emission scenarios all show increasing trends. The higher the emission, or the longer the return period, the faster the increase of the trend. Compared with the averages of extremes, extremes of extremes (tails of GEV distributions) are more liable to change under these scenarios.