Graph convolutional LSTM model for traffic delay prediction with uncertainty

Abstract

Traffic flow in an urban environment exhibits a complex spatio-temporal in- teraction. The propogation of traffic flow through a transportation network depends on a number of factors, including the structure of the network and the time of day. Current analysis of this data by road controlling authorities is of- ten simplified and lacks a detailed understanding of how traffic moves through the network. A deep learning model which models both the spatial and tempo- ral interactions present in the data is able to capture complex patterns present in the data and allows for a more detailed understanding of traffic flow. A GC- LSTM model is explored for Hamilton City to predict traffic delay. It is found to have improved prediction accuracy over a standard LSTM by incorporating the spatial structure of the Hamilton road network. Additionally, Bayesian layers are integrated into the model to obtain a probability distribution over each prediction. By quantifying the uncertainty over each prediction, the de- cision making process based on the analysis can be carried out with a higher degree of confidence than a single point prediction from the model.