The core of dynamic community detection lies in identifying groups of nodes (i.e., communities) that are closely connected internally and relatively independent externally by tracking the time-varying characteristics of nodes and edges. In this paper, we propose a community detection framework based on dynamic graph convolutional networks and model transfer (MT-SA-DGCN). The framework first constructs a node representation learning model based on self-attention and dynamic graph convolutional networks, which is based on a large model of dynamic graph representation learning (DGCN) based on graph convolutional networks. By incorporating the self-attention mechanism into DGCN, it is possible to pay more attention to the important nodes in independent time slices. To capture comprehensive structural information of dynamic graphs in independent time slices, mutual information between local and global representations of the graph is maximized. To capture the overall structural information of dynamic graphs in all time slices, the weight parameters of the graph convolutional network are updated using the long and short-term memory network. On this basis, an efficient model transfer strategy is proposed, which can realize the transfer of the above large model to small datasets, and design a hierarchical fine-tuning strategy to realize the adaptation to small datasets. Experimental results on real and synthetic datasets show that the framework achieves state-of-the-art results.