Based on the 10-15 days extended range ensemble forecasts of CMC,ECMWF,UKMO and NCEP in the TIGGE dataset,the multi-model ensemble forecasts of surface air temperature in the Northern Hemisphere middle latitudes during the period from 1 July to 30 September 2008 have been conducted by using EMN(Ensemble Mean),BREM(Bias-Removed Ensemble Mean) and SUP(Multi-model Superensemble),respectively.Meanwhile,the Talagrand distribution,ensemble spread,and RMS error are utilized to evaluate the forecast skills of each single forecast system.As is known that the precision of daily forecasts in extended range is relatively low,people pay more attention to the prediction precision of weather process in extended range.The daily data of each single forecast system and the daily observed data are processed by the moving average method.The processed data are also conducted with multi-model ensemble means.Finally,the training period of SUP is tested.Results show that four single forecast system’s ensemble spread are low compared with RMS error.ECMWF is the best single system,NCEP follows and UKMO is the worst.The forecast skills of EMN,BREM and SUP are all higher than single system and SUP is the best ensemble method.Moving average methods further improve the forecast skill of surface air temperature and the longer of moving step is,the better of the prediction results are.For the training period of SUP about daily and 3-day moving average,75 days are the best training period for 10-12 d forecasts and 35 days are the best for 13-15 d forecasts.For 5-day and 7-day moving average,the best training period is 35 days at all forecasting time.