Considering the uncertainties of climate models and downscaling methods in the assessment of climate change impact will help us obtain more credible results in the simulation of rice water utilization response to future climate change. In this paper,two Global Climate Models named HadCM3 and CGCM3 under three climate scenarios (A1B,A2 and B2) were downscaled by Statistical Downscaling Model (SDSM) and Back-Propagation (BP) artificial neural network respectively, then the four downscaled results were merged by Bayesian model averaging (BMA) method. Rice growth duration,yield,water demand and water use efficiency in two future stages (2050s and 2080s) were simulated by ORYZA2000 rice model based on BMA merging results. The results show that (1) the BMA method is more competent to produce low bias in comparison with simple model averaging (SA) method;(2) in the two future stages, rice yield and growth duration would decline remarkably as the increasing of temperature and decreasing of solar radiation;and(3) water demand falls as the solar radiation,while in 2080s,the quickly increasing of temperature will bring the increase of water demand,but the value still higher than the historical reference stage, and the decrease of water demand cannot offset the negative effects on water use efficiency brings from yield decline.