End-to-end traffic flow prediction incorporating hidden variables and dynamic time series memory

In order to improve the accuracy and robustness of traffic flow prediction, MIEE-VAE-LSTM, a prediction model integrating deep learning techniques, is proposed to address the difficulty of modeling the complex nonlinear relationships and long-term temporal dependence of traffic data by traditional methods, It combines a bi-directional long and short-term memory network with a self-attention mechanism to extract key features from the bi-directional time dimension and identify key time steps by attention weights. time dimension to extract key features and identify key time steps by attention weights; a variational autoencoder is introduced for data noise reduction and potential feature extraction; and a dynamically weighted loss function is used to enhance the attention to high uncertainty time steps. Experiments on the real dataset PEMS, in which its MAE at the 3rd, 6th, 9th, and 12th future time steps was reduced by 42.23%, 24.18%, 10.34%, and 10.17%, respectively, and MSE was reduced by 66.46%, 40.74%, 15.76%, and 14.83%, respectively, and MAPE compared to the suboptimal model on the PEMS04 dataset was reduced by 40.03%, 26.62%, 16.35%, and 6.91%.Similar improvements were achieved on the PEMS08 dataset. The results show that the model effectively improves the prediction accuracy and robustness, and finally reveals the characteristics of traffic flow temporal dependencies to support traffic prediction model optimization.