The existing traffic flow prediction model has limited performance in dealing with abrupt phenomena and non-stationary features in the sequence,and fails to fully capture the spatial correlation features and ignores the dynamic correlation between nodes.To solve the above problems,this paper proposes a prediction model based on decoupling fusion framework combined with graph enhanced sparse convolutional attention network (DSGEAN).Firstly, the traffic flow sequence is decoupled into trend term and fluctuation term to better capture its inherent law.The causal convolution and attention mechanism are respectively applied to perceive local global time dependence,the multi-graph construction enhancement strategy and the sparse design of global graph attention are utilized to pay attention to local differences,the location coding is designed to simulate the dynamic association between nodes,and then the data-driven adaptive fusion is realized.The multi-supervision method is used to train the prediction results to avoid the accumulation of errors.The validity of the model was verified on two real data sets PEMS04 and PEMS08,and the results showed that the average absolute error of the prediction accuracy index was 18.18 and 13.29, respectively, which was lower than other advanced models.