Short-term traffic flow prediction is a key technology for intelligent transportation systems to achieve efficient management and real-time regulation. To address the shortcomings of existing methods in spatio-temporal dependency representation and error accumulation, a dual-path spatio-temporal graph attention network model is proposed to achieve more accurate and robust short-term traffic flow prediction. Firstly, a dual-path graph attention mechanism for both static and dynamic conditions is designed to extract spatial features. The static attention branch builds stable spatial relationships based on the physical topology of the road network, while the dynamic attention branch adaptively captures sudden spatial correlations based on real-time traffic flow data. The two branches are weighted and fused through a learnable fusion parameter, adaptively balancing the influence between static topology and dynamic state, thereby extracting comprehensive spatial features. Secondly, a long short-term memory network encoder structure is adopted to hierarchically extract and flexibly map the features of time series. Finally, the decoder introduces a dynamic attenuation hybrid teaching-learning mechanism, integrating recursive and teaching decoding modes to predict traffic flow, in order to effectively prevent error accumulation. The PEMS04 and PEMS08 datasets were used for training, validation and testing. The results showed that the model achieved approximately 4.23%, 3.18%, 6.90% and 3.89%, 2.59%, 6.44% improvements over the optimal spatio-temporal comparison model CNN-LSTM in terms of MAE, RMSE, MAPE indicators respectively. It can be seen that the model is capable of effectively balancing the static topology and dynamic state of the road network, and significantly reduces error accumulation through the hybrid teaching-learning mechanism, demonstrating its superior performance and application potential in short-term traffic flow prediction.