A variable-rate dynamic point cloud compression network framework based on scene flow was proposed in order to address the problem of training multiple network models for existing dynamic point cloud compression neural networks. The raw dynamic point cloud was taken as input, and the scene flow network was utilized to estimate motion vectors. A channel gain module was introduced to evaluate and scale the latent vector channels while compressing motion vectors and residuals, achieving variable-rate control. A new joint training loss function was designed to end-to-end train the entire network framework by comprehensively considering the motion vector loss and rate-distortion loss. A human body dataset with motion vector labels was created based on the AMASS dataset for network training in order to solve the problem of lack of real motion information labels in dynamic point cloud datasets. The experimental results show that the compression bit rate of the method decreases by several orders of magnitude compared with existing deep learning-based dynamic point cloud compression methods. The method has a 5%~10% improvement compared with the reconstruction effect of static compression networks processing each frame separately.