Real-time city-level traffic prediction in the context of Stockholm City

Background: The ongoing POST (Prediktions- och Scenariobaserad Trafikledning) project and the previous project Mobile Millennium Stockholm (MMS) provided tools and frameworks for real-time estimation and prediction of travel times on the city-level. City-level prediction of the traffic state as well as the traffic demand is important for both traveler information applications, such as online navigation, and traffic management applications, such as scenario evaluation of incident management strategies. However, city-level prediction is very challenging and requires efficient processing of large amounts of data. Here we present the recent research about effects of the clustering on the prediction performance and computational cost. Partitioning of the road network based on spatial and temporal attributes can potentially result in clusters that provide more robust and accurate prediction with reasonable bias-variance tradeoff. Methods: The effects of the clustering on the prediction performance are studied on the three case studies, representing different travel time sources in Stockholm city. First represent 15 MCS radars as the sources of travel times. Second 420 segments on the major roads around Stockholm with travel times estimated from the MCS radars. Third, travel times of 11,340 links processed from GPS data of 1,500 taxis operating in Stockholm. With the computational experiments, we studied different clustering approaches based on the day classification, functional classes, spatial locations and temporal attributes, and how they can effect the prediction performance and computational cost. Results: reveal that partitioning can significantly improve the prediction accuracy and rapidly decrease the computational cost and time.QC 20230901