Abstract:Based on the ensemble forecasts of ECMWF, JMA, NCEP and UKMO in the TIGGE datasets in Northern Hemisphere middle latitudes during the period from 1 June until 31 August 2007, the multimodel superensemble forecasts of the surface temperature for the forecast period from 8 to 31 August 2007 have been conducted by using fixed training period and running training period, respectively. The root mean square error is utilized to evaluate the forecast errors of two kinds of superensemble forecasts, the best single model and the ensemble mean. Results show that SUP(the multimodel superensemble forecast with fixed training period) reduces the forecast error considerably. The forecast skill of the multimodel superensemble is higher than that of EMN(the ensemble mean) and the best single model among ECMWF, JMA, NCEP and UKMO models for the 24 h~144 h surface temperature forecast. However, for the 168h forecast the forecast skill of the superensemble is lower than that of EMN. The R-SUP(multimodel superensemble with running training period) further improves the forecast skill. It has higher forecast skill than EMN for the 24 h~168 h forecast. For the 168 h forecast, in particular, R-SUP improves the forecast skill and has higher forecast skill than EMN.