Based on four machine deep learning algorithms (PredRNN++, MIM, CrevNet and PhyDNet), radar data and precipitation data in Wuhan from 2012 to 2019, this study investigates the possibility of application of artificial intelligence (AI) technology in the nowcasting of Wuhan Region. The forecasting skills of radar echo nowcasting are examined in terms of mean square error (MSE), structural similarity index (SSIM), probability of detection (POD), false alarm rate (FAR) and critical success index (CSI), then compared with the semi-Lagrangian optical flow method. The results are summarized as follows. The MSE and FAR are the lowest and SSIM is the highest in the MIM algorithm. The POD and CSI of PredRNN++ are the highest. The POD, CSI and SSIM of machine learning are higher than semi-Lagrangian optical flow, while the FAR and MSE of machine learning are much lower, of which the SSIM, POD and CSI of machine algorithms are improved by 3.2%-24.7% than semi-Lagrangian optical flow, but the MSE and FAR are reduced by 13.1%-43.3%. Within 30 minutes, except the CrevNet algorithm, the skills of other algorithms are similar to that of semi-Lagrangian optical flow. 30 minutes later, the skills of both machine algorithm and semi-Lagrangian optical flow decline significantly with the increase of forecast lead time. However, the skill of machine algorithms declines much more slowly. Especially after 60 minutes, the skill of semi-Lagrangian optical flow descends more quickly indicating the advantage of machine learning algorithms for long-term prediction. In addition, the descending rates at different forecast lead times for different score indexes are different among the machine algorithms. The CIS of PredRNN++ is the highest in any intensity, MIM and PhyDNet performance is better than semi-Lagrangian optical flow for radar echo intensity exceeding 40 dBz, but CrevNet shows better skill for radar echo intensity exceeding 50 dBz. The POD and CSI of machine algorithms and semi-Lagrangian optical flow decline significantly with the increase of forecast intensity of radar echo, while the FAR increases quickly, but the increase of FAR rate of machine learning algorithm is much slower. To sum up, the analysis of four different echo patterns and different development trends shows that the machine learning algorithm has the ability not only to predict the change of radar echo intensity in a certain content, but also to predict the time node of the evolution tendency of intensity and acreage, which are basically consistent with the observation. These results suggest that the ability of machine deep learning to predict the movement of radar echo is better than that of semi-Lagrangian optical flow, indicating its possible wide prospect for operational application.