Dilated LSTM Networks for Short-Term Traffic Forecasting using Network-Wide Vehicle Trajectory Data

Short-term traffic forecasting is anticipated as an always evolving research topic, boosted by the tremendous recent advances of Machine Learning and Deep Learning, as well as computational power of modern PCs. In this paper, the Dilated Recurrent Neural Networks are introduced in traffic forecasting. Their architecture promotes the deployment of long-term relations and prevents common issues of RNNs, such as exploding and vanishing gradients. The Dilated LSTM Network is exploited to perform traffic conditions forecasting using network-wide data. The data consist of GPS trajectories of ride-hailing company DiDi's vehicles from November of 2016. After preprocessing the data and organizing them into section's travel speed of five-minute time resolution timeseries for each one of the 498 road sections of the road network of Xi'an, China, we fed them to the Dilated LSTM Network. The model consists of four hidden layers, each of them implementing an LSTM Network with one, two and four-step dilation correspondingly. The model achieves 85% accuracy, which is improved over a classic LSTM structure, trained on the same data.