Traffic-Flow Prediction of DLSTM-AE Model Based on RAdam Optimization

Abstract

In order to further fully extract the spatiotemporal and periodic characteristics of traffic flow, this paper adopts a combined model (DLSTM-AE) of autoencoder architecture and Deep Long Short-Term Memory neural networks and introduces rectified adaptive moment estimation (RAdam) algorithm for model training. First, the characteristics of traffic flow sequence information are collected using the deep long short-term memory network model, and the collected information is compressed into a fixed dimension representation vector with the help of the automatic encoder structure. Then, the vector is reconstructed by the decoder to realize further information mining. Finally, in the process of model training, the RAdam algorithm is used to optimize and update the momentum parameters in batches, to shorten the time to find the optimal solution. This can improve the timeliness and accuracy of model prediction. Simulated on real data sets of highway traffic flow and compared with other model methods, DLSTM-AE combined model not only has obvious advantages in prediction results, but also has good curve-fitting ability in traffic flow periodicity. The root mean square error (RMSE) value of DLSTM-AE combined model on the test set decreased by about 0.445 to 1.826, the average absolute error (MAE) value decreased by about 0.282 to 0.984, and the coefficient of determination (R2) value increased by about 0.005 to 0.023. In terms of periodicity, the prediction accuracy of working days corresponding to adjacent weeks is much higher than that of the control group. Simulation results show that the model can capture more potential spatiotemporal and periodic information in traffic flow sequence. It can better meet the needs of highway traffic-flow prediction.