Severe convective weather poses a great threat to human production activities, and high-precision prediction results are of great value for accurate prevention and control of natural disasters. For example, the traditional short-term and imminent precipitation prediction methods mainly focus on the task of radar echo extrapolation. At present, the deep learning method to achieve radar echo extrapolation has become a research hotspot. These methods effectively improve the quality of prediction results from the perspective of data-driven and statistical. Although deep learning-based methods show excellent performance, they often fail to consider their high spatio-temporal resolution, resulting in high computing resource consumption and fuzzy extrapolation results. Radar echo extrapolation can be regarded as a spatio-temporal prediction problem. In this paper, this problem is modelled by a recurrent neural network (ST-ConvLSGRU), which uses LSTM to improve the GRU model by making gradient connections from the temporal and spatial dimensions to establish long-term dependencies. Also configure Convolutional Block Attention Modules (CBAM) and Depthwise Separable Convolution (DSC) to achieve lower model complexity when building multi-layer models. Experiment shows that ST-ConvLSGRU has better extrapolation prediction ability than other baseline methods.