Traffic flow prediction plays a crucial role in alleviating traffic congestion. Many research methods did not fully explore dynamic hidden correlations in traffic data. To address this challenge, an encoder-decoder-based traffic prediction model was proposed by studying the dynamic spatio-temporal variation characteristics. In the model, both encoder and decoder mainly consisted of multi-head spatio-temporal attention mechanism modules, and a connection attention mechanism was added in between to analyze the spatio-temporal correlations of the road network. The model also used a dynamic embedding module consisting of a combination of both spatio-temporal embedding coding and adaptive graph convolution to analyze the dynamic and static information of nodes. Experiments on two real datasets demonstrated that the spatio-temporal model outperformed the baseline model for long- and short-term traffic prediction. Thus, the spatio-temporal encoder-decoder model can effectively handle complex spatio-temporal sequences and improve the traffic flow prediction accuracy.