Based on the 1-7 days ensemble forecasts of the European Centre for Medium-Range Weather Forecasts (ECMWF),US National Centers for Environmental Prediction (NCEP),and the Japan Meteorological Agency (JMA),the UK Met Office (UKMO) as well as the Korea Meteorological Administration (KMA) in the TIGGE datasets,the multimodel ensemble forecasts of the surface air temperature in China and its adjacent area during the period from 1 January to 30 September 2015 were conducted by using long-term memory (LSTM) neural networks,neural networks (NN),bias-removed ensemble mean (BREM) and the superensemble (SUP) with sliding training period for the forecast period from September 5 to 30,2015.The results showed that the BREM forecast was no better than the ECMWF forecast due to the impact of low skill model forecasts among the five models.The forecast skill of SUP was better than that of all the single models.For 24-144h forecasts,the root mean square error (RMSE) of SUP was significantly smaller than that of ECMWF forecast.As the forecast lead-time increased,the RMSE increased as well.The forecast skill of NN was roughly equivalent to that of SUP.Overall,the LSTM approach showed the best forecast performance,especially when the forecast lead-time was longer than 72 h,the RMSE of the LSTM forecast was considerably smaller than that of ECMWF,BREM,NN,and SUP forecasts.The LSTM neural networks approach significantly reduced the forecast RMSE of the surface air temperature in the northwestern,northern,northeastern,southwestern,and southern China.However,the RMSE of the LSTM forecast was relatively larger in southern Xinjiang area compared with ECMWF forecast.