Identifying the spatial and temporal congestion patterns at the urban level for journey time forecasting

As part of the ERC MAGnUM project and in collaboration with Dittlab at the University of Delft, IFSTTAR has recently developed a novel but straightforward method for characterising traffic conditions at the urban scale.

The principal idea is to summarise all the data that are available at road section level within a single spatial and temporal map which identifies zones with uniform speeds. The different stages of congestion are thus described as a single volume that is located in both space and time and which can be characterised by a mean speed value. Congestion maps have been drawn up for each of the 35 days for which data were available on the network in Amsterdam (Netherlands) by applying advanced classification methods (clustering).

What is even more interesting is that the study shows that some daily congestion maps share similar characteristics. Thus, we were able to divide the 35 days for which we have data into just four groups of days. At this stage, a learning method was applied to identify a consensual congestion profile for each group. Our remarkable finding is that four speed maps each with only nine zones are sufficient to characterise the traffic conditions observed during the 35 studied days. It would thus seem that at the scale of the whole city and a whole day, the dynamic of traffic is more repetitive and easier to forecast than might be thought on the basis of a local analysis of traffic conditions.

By identifying consensual congestion maps we have also been able to develop a highly innovative method for estimating journey times. The idea is to identify, in real time, which consensual map is the closest to the observations which are currently available. This map is then used to predict future journey times for any journey within the study zone. By identifying the closest map every hour, it was possible to predict the journey times for 84% of the journeys with a margin of error of less than 25%.

A paper describing this study has recently been published in the very selective journal Scientific Reports, which is a multidisciplinary open source journal belonging to the Nature group.