In response to the issues of visual performance blurring and underestimation of high echo values in the modeling results of traditional short-term rainfall prediction models for historical radar data, we propose a short-term rainfall radar echo extrapolation model that integrates 3D convolution and dual-end attention mechanism 3DuA-Net. Using ST-LSTM space-time long short-term memory networks as the recurrent units, replacing ordinary convolution with 3D convolution enhances the model′s capability to capture short-term motion features from a global perspective. Additionally, an efficient dual self-attention module DuAtt is proposed to improve the model′s ability to preserve and integrate important local and global features in long-term radar image sequences. Experimentation conducted using publicly available Doppler radar datasets from the Shenzhen Meteorological Bureau shows that, at 10、20、 40 dBz thresholds, the model exhibits an average improvement of 7.74% in the CSI metric compared to the Conv-LSTM model, an average improvement of 5.54% in the HSS metric, a decrease of 3.8% in the MAE metric, and an improvement of 8.86% in the SSIM metric.