Traffic forecasting in the city of Madrid using Graph Neural Networks

Abstract

One of the greatest challenges facing modern societies is the management of road traffic and the problems that directly result from it: congestion, accidents and pollution. In order to mitigate and suppress these adverse effects, it is necessary to create increasingly accurate and realistic descriptions that can account for such a complex phenomenon as traffic. In particular, thanks to the ever-increasing volume of data currently available, in recent years there have been numerous works based on machine learning that have proposed predictive models by learning through real data sources. Specifically, the type of neural networks based on graphs is able to capture the spatial and temporal correlation that exists between the different elements of a road network understood as a graph. This work aims to evaluate a pair of predictive models based on graph neural networks, DCRNN and Graph WaveNet (GWN), applied to the case of the city of Madrid. For this purpose, a dataset has been compiled from open data corresponding to measurements taken during six months by a sensor network on the roads of the city of Madrid. The analysis of the obtained dataset shows a strong periodicity with two marked frequencies: a daily one, with ups and downs around the hours of entry and exit of the working day and a minimum of night activity; and a weekly one, with a clear separation between weekday and weekend values. The results of the evaluations reveal a null dependence on the spatial component; this may be due to the fact that the traffic roads of the M30 are one-way, with no two-way flow on these roads that can be picked up by the same sensor. In addition, the results of the GWN model show that the inclusion of an adaptive term to obtain the adjacency matrix at training time improves the prediction significantly. Although the inclusion of this term alone yields a clearly higher error than the other cases, the results show that it is possible to make a forecast with this type of model without prior knowledge of the network structure.