Typhoon cloud images are an important resource for weather forecasting, which can indicate the evolution of weather systems and help with accurate predictions. However, their practicality is hindered by challenges such as limited data availability and poor real-time updates. Therefore, accurate prediction of typhoon cloud maps is of great significance. The goal of typhoon cloud prediction is to predict subsequent frames in the image sequence. Essentially, this is a spatiotemporal sequence prediction problem. Due to the large observation area, predicting typhoon cloud maps is challenging. In this article, we introduce self attention mechanism into PredRNN-v2. Specifically, we incorporated a self attention mechanism that memorizes long-range dependent features in both spatial and temporal domains. On the basis of the self attention module, we aggregate the features of all input positions with memory features paired with similarity scores, and embed the self attention mechanism into PredRNN-v2 to construct a self attention PredRNN-v2 for spatiotemporal prediction. In the experiment, we applied SA-PredRNN to frame prediction of MovingMNIST and typhoon cloud images, and the results showed that SA-PredRNN achieved better performance on both datasets.