In response to the problems of imbalanced samples and low prediction accuracy, an enhanced predictive recurrent neural network EN_PredRNN is proposed. Firstly, the radar data is preprocessed and samples are selected to construct a highquality radar echo dataset; then, deep fusion of spatiotemporal long short-term memory units and dynamic convolution is used to design a dynamic convolution combined with spatio temporal long short term memory module DC_STLSTM, which adjusts convolution parameters in real-time to accurately capture the instantaneous changes in radar echoes. Then, stack 5 layers of DC_STLSTM to extract deeper features of radar echoes, and use gradient highways to alleviate gradient vanishing, improving the model′s generalization ability and prediction accuracy. The experimental results showed that EN_PredRNN performed the best, significantly improving the critical success index and reducing false alarm rates. Compared with PredRNN, it increased the critical success index by 19.3%, 17.3%, 16.5% and 14.0% at 25, 35, 45 and 65 dBZ, respectively, while reducing false alarm rates by 28.3%, 27.5%, 26.7% and 24.9%, effectively. This model effectively learned the spatiotemporal variation characteristics of radar data and accurately predicted the radar echo intensity and location.