Traffic flow prediction has a broad and promising application potential as a core technology in intelligent transportation systems. Due to the complexity of the spatio-temporal characteristics of traffic flow, it becomes a challenging task to effectively capture the complex and dynamic spatio-temporal dependencies in traffic data. A spatio-temporal encoded traffic prediction model, STEFM, is proposed as a means of capturing the complex dynamic dependencies of the transportation system. This is achieved by employing the coding layer of the transformer model. The temporal self-attention module is initially constructed in the coding layer to capture the temporal dependencies inherent to the transportation system. Then the geospatial self-attention module and the regional spatial self-attention module are designed. The capture of proximal dynamic spatial relationships and remote spatial dependencies is achieved through the utilisation of self-attention mechanisms and correlation coefficients. In the geospatial self-attention module, the detection of traffic events is performed with the objective of identifying and predicting anomalies on each roadway. The historical traffic data and anomalous event information are fused into the geospatial self-attention module to explicitly model the time delay of spatial event propagation. The experimental results on real road traffic datasets demonstrate that the STEFM model proposed in this paper outperforms the current mainstream modeling approaches in all aspects.