Bias correction is crucial for optimizing the prediction accuracy of traditional numerical weather prediction models. In meteorology, temperature is a vital parameter that significantly influences weather phenomena, atmospheric circulation, climate patterns, and human activities. Therefore, it is essential to perform temperature bias correction on traditional numerical weather prediction products. This paper designs a novel deep learning network model, SWGU-ConvLSTM, for temperature bias correction, which adopts the U-Net and bidirectional adversarial network architectures. The model uses the ConvLSTM module to extract local information and the SwinTransform module to capture global features. The outputs of the ConvLSTM and SwinTransform modules are then fused using the IAFF module. Furthermore, the fused feature information is processed with U-shaped connections and skip connections to better combine shallow and deep features and capture multi-scale information. Finally, the fusion module is employed as both the generator and discriminator in a bidirectional adversarial training framework to enhance the model′s learning and prediction capabilities. In this study, the TIGGE numerical model data from the ECMWF platform is used as the input data to be corrected, and ERA5 reanalysis data is used as the target data. The model is applied to correct 6-hour temperature forecasts. Experimental results show that the proposed SWGU-ConvLSTM model outperforms other comparison models in terms of MSE, MAE, and SSIM. Compared to the simvp model, its MSE and MAE errors are reduced by 30% and 27%, respectively, improving the accuracy of temperature bias correction.