I designed a novel deep learning network model called SWGU-ConvLSTM for temperature bias correction, which incorporates a U-Net and bidirectional adversarial network architecture. The model utilizes the ConvLSTM module to extract local information and the SwinTransform module for global information. The IAFF module fuses the output features of both ConvLSTM and SwinTransform, employing U-shaped connections and skip connections to better integrate shallow and deep information, while capturing information at different scales. The model serves as both a generator and discriminator for bidirectional adversarial training, enhancing its learning and predictive capabilities. Using ECMWF"s publicly available TIGGE numerical model data as the corrected data and ERA5 reanalysis data as the label data, the model corrects 6-hour temperature forecasts. Experimental results indicate that the proposed SWGU-ConvLSTM model significantly outperforms other comparative models in metrics such as MSE, MAE, and SSIM.