In view of the non-stationary and random characteristics of the short-term traffic flow sequence, in order to improve the shortterm traffic flow prediction accuracy and model training speed, this paper proposes a combined prediction model based on adaptive variational modal decomposition (VMD) and bi-directional long-term memory network (BiLSTM) combined with attention mechanism. Firstly, the spatial and temporal traffic flow sequence is decomposed by the adaptive VMD method to a series of modal components with limited bandwidth, which can refine the traffic flow information, reduce non-stationarity, and improve the accuracy of modeling. Secondly, the spatio-temporal correlation in the short-time traffic flow sequence after decomposition is mined by BiLSTM combined with attention mechanism to reveal its spatio-temporal variation rules, which further improves the modeling accuracy. In addition, in order to accelerate the training convergence speed of the prediction network, the network weight optimization is carried out by the improved Adam algorithm. Finally, the predicted value of each modal component is superimposed as the predicted value of the final traffic flow prediction value. The experimental results show that the prediction performance of the model using modal decomposition is obviously better than that of the model without modal decomposition, and the RMSE of the self-adaptive VMD-Attention-BiLSTM prediction model is reduced by 47. 1% compared with that of the EEMD-attention-BiLSTM prediction model. The combined prediction model improves the prediction accuracy and can quickly predict the traffic flow time series.