The moving mean bias correction method and the historical deviation correction method are used to correct the error of the extended-range 2 m temperature forecast of the ECMWF numerical model. Through the study of the moving training period, the optimal training period length of the moving mean bias correction method is 25-30 days for daily temperature forecast from 11 to 15 days. The verification of corrected temperature forecast in 2018 shows that the application of these two temperature deviation correction methods can correct the deviation that the temperature forecast in model is significantly lower than observation, and improve the forecast accuracy by 30% at least. The mean absolute error of the corrected temperature forecast is basically less than 2℃ from June to October, which has certain reference value and could provide good support for forecasters. There is no obvious difference between the two methods in the bias correction effect of the extended-range forecast within 15 days. With the extension of forecast lead time, the advantages of the historical deviation correction method show up gradually.