Using flows or speeds in traffic pattern clustering and prediction : does the data type matter?

• Data and knowledge of travel patterns play a key role in finding more cost-effective solutions and better utilization of existing resources to increase sustainability and decrease CO2 emissions, pollution, and noise. Understanding travel patterns and prediction of future traffic states is a central ingredient in Intelligent transport systems (ITS). Pre-clustering the data before applying the prediction models is a recommended practice. We consider in this work revealing day-to-day traffic regularities and grouping days into representative day-types based on their traffic similarities before training prediction models. Specifically for this presentation, we will present our recent work on day-type clusterings that concern the similarities and interchangeability of day-types recognized by flow and speed traffic measurements. We consider the speed and flow traffic measurements from the motorway control system in the highway system around Stockholm, Sweden. Different clustering methods are used and their performance is evaluated on short-term prediction models. The results reveal that day-types are similar across data types and clustering methods, and their similarity does not depend much on the number of clusters. As the baseline scenario, calendar-based day-types are used. The similarity is higher between flow and speed recognized day-types compare to calendar-based day-types. Considering short-term prediction performance, the data-driven day-types outperform calendar-based methods. However, for more sophisticated prediction models the difference becomes insignificant. The interchangeability of speeds and flows in traffic prediction is studied in a scenario where new days are classified into day-types based on speed observations. This could be particularly interesting for traffic management centers as speed observations may be collected in more affordable, sustainable, and scalable ways. However, results reveal that flow prediction is sensitive to whether the new day is classified to one of the clusters using speed instead of flow observations, and prediction performance is reduced by about 28%. This sensitivity can be overcome by using a more sophisticated prediction model. When classifying based on flow observations a more sophisticated model results in slight improvements in speed prediction.

Ämnesord

TEKNIK OCH TEKNOLOGIER  -- Samhällsbyggnadsteknik -- Transportteknik och logistik (hsv//swe)

ENGINEERING AND TECHNOLOGY  -- Civil Engineering -- Transport Systems and Logistics (hsv//eng)

Publikations- och innehållstyp

vet (ämneskategori)

kon (ämneskategori)