A flexible deep learning-aware framework for travel time prediction considering traffic event

Abstract

Travel time is an important signal to measure the performance of transportation systems. However, urban traffic forecasting is still a challenging task due to the complicated nonlinear nature of urban traffic and the impact of abnormal traffic event. In this study, we propose a flexible deep learning-aware framework which is composed by integrating multiple components with a sequence-to-sequence (Seq2Seq) model as the main body. Specifically, we incorporate a newly designed hybrid adjacency matrix of graph convolution network and temporal attention mechanism to flexibly capture spatio-temporal dynamics accordingly. In addition, we design a responsive algorithm to obtain the latent representation of traffic event by applying stacked denoising autoencoder. We then conduct the baseline model comparison and ablation experiments to evaluate our model performance with real-word datasets. The results indicate that our method outperforms baselines and the fusion of traffic event data can improve prediction accuracy. Moreover, the case study and sensitivity analysis also have a reference value to the practical application of traffic prediction task.