In order to reduce the large deviations between predicted and observed values in the traditional numerical weather prediction model (GRAPES_GFS) for 2 m temperature and to improve forecast accuracy,this paper proposes a correction model based on a Convolution Long Short-Term Memory (ConvLSTM) network and attention mechanism,utilizing GRAPES_GFS grid data and corresponding observational data.The model consists of three main steps.First,shallow local features are extracted from the input data through a local feature extraction module.Second,the extracted feature maps are fed into a dual attention module,which assigns different weights to different channel and spatial dimensions of the data,suppresses the influence of meteorological elements weakly correlated with 2 m temperature,and enhances local features in high-temperature regions.Finally,a ConvLSTM network captures temporal dependencies in the data and outputs the corrected forecast.Experimental results show that,compared with the original GRAPES_GFS forecasts,the proposed model improves all evaluation metrics for 12- to 36-hour lead times in 2 m temperature prediction.The Pearson correlation coefficient increases from about 0.55 to approximately 0.87,the root mean squared error decreases from 1.74-2.06 ℃ to 0.90-1.10 ℃,and the mean absolute error decreases from 1.36-1.64 ℃ to 0.69-0.84 ℃.Moreover,the model outperforms other mainstream correction approaches.