Aiming at the existing problems of fuzzy distortion and low accuracy in predicting strong echoes in existing radar echo extrapolation models, this paper designs a radar echo extrapolation method based on self-attention and dense convolution improved convolutional long short-term memory network (ConvLSTM) by using the composite reflectivity mosaic image of Doppler radars data from Anhui Province from May to September 2016. Based on ConvLSTM, the model incorporates self-attention mechanism into each cell and encoder-decoder to enhance the ability of extracting features with long-term spatial dependence. Meanwhile, the model uses dense convolution instead of common convolution to improve the feature reuse ability. The experiment used the past 1h radar echo image to predict the future 2h radar echo image, and compared with the ConvLSTM before the improvement, proved that the proposed model can improve the accuracy of radar echo extrapolation.