To obtain more accurate and faster grid temperature forecast products, the study used a high-frequency gridded observation fusion product and eight kinds of error correction methods to carry out a rolling correction forecast test for the 2 m temperature forecast filed of the European Centre for Medium-Range Weather Forecasts. The test was conducted on two forecast simulations from 1 January to 28 February 2017 and from 1 June to 30 July 2017, starting at 14:00 BT and 20:00 BT in Beijing, and was performed in 3-24 hours rolling prediction correction. The forecast results were tested and analyed using grid observation fusion data and site observation data. The results show that the eight methods have positive correction effects on the direct model output. The full-grid sliding error regression model correction and the full-grid sliding two-factor regression model have the best correction effect. Both schemes can make the average absolute error of the grid of the correction field below 2℃, and the grid accuracy of 3 h, 6 h and 9 h is above 0.9. The prediction results of the full-grid sliding error regression model are slightly better than the prediction results of the full-grid sliding two-factor regression model. This shows that the error field at the beginning of the forecasting time plays an important role in the correction as a predictor of the regression model.