In order to reduce the high deviation between the predicted value and the observed value of the traditional numerical prediction model (GRAPES_GFS) of 2m temperature and improve the prediction accuracy. Combining the grid data of GRAPES_GFS and the corresponding observation data, this paper proposes a 2m temperature forecast correction model based on Convolution Long Short Time Memory Network (ConvLSTM) and attention mechanism. The model mainly includes the following steps: First, the local shallow features of the input data are extracted by the local feature extraction module. Secondly, input the extracted feature map into the feature attention module, assign different weights to different channel dimensions and different spatial dimensions of the data, restrain the weight of meteorological elements with low correlation with 2m temperature, and achieve local enhancement of high-temperature areas in 2m temperature data. Finally, ConvLSTM network is used to capture the time dimension characteristics of data and output the forecast correction results. The experimental results show that, compared with the prediction results of GRAPES_GFS, each numerical evaluation index of the model is improved in the 2 m temperature prediction with an aging of 12-36 hours. The Pearson correlation coefficient increases from about 0.5 to about 0.88, the root mean square error decreases from 1.74-2.06 ℃ to 0.9-1.1 ℃, and the average absolute error decreases from 1.36-1.64℃ to 0.7-0.84℃; At the same time, compared with the mainstream revised models, the model also achieved better correction results.