Nowcasting is crucial for the prevention and mitigation of extreme weather events. At present, radar echo extrapolation is still the primary technique used in nowcasting due to the high spatio_temporal resolution of weather radar data, but this algorithm has some problems, such as limited lead time, low accuracy of long_distance prediction, fuzzy prediction images, false alarms and underestimation, etc.. To address these problems, a novel model, GA_MotionRNN, is proposed in this study, which integrates MotionRNN and video vision transformer (ViViT) architectures within a generative adversarial framework, incorporating a gradient_based structural similarity loss function, achieving the accurate extrapolation of 6 min radar echo in 2 h. In this study, the training set and test set are constructed by using the composite reflectivity mosaic data of Jiangxi Province and its surrounding areas from March to September in 2021 - 2023. The loss function is constructed by using the structural similarity index (SSIM) combined with Mean Square Error (MSE) and Binary Cross Entropy (BCE) to train the network to learn more real spatial and temporal distribution characteristics of radar echoes. The results show that the critical success index (CSI) and probability of detection (POD) of GA_MotionRNN are the highest at all the reflectivity thresholds of 20 dBz, 30 dBz, 40 dBz, and 50 dBz, reaching maximum scores of 0. 6814 and 0. 8164, respectively, and all are superior to those by the optical flow method and deep learning models. Moreover, GA_MotionRNN can still maintain high extrapolation accuracy in long_distance predictions. The case analysis shows that GA_MotionRNN can not only predict the position of the strong echo center better than other deep learning models, but can also retain more detailed information, which effectively solves the common problems of fuzzy output image and loss of detailed information in deep learning methods.