None

Abstract

In strategic transport models, road travel demand matrices are usually estimated using estimation methods that fuse prior or synthetic travel demand matrices with flow data observed on individual roads (‘links’) in the network. On the one hand, ever more data on flows, speeds and/or densities on link level is available, driven by technological advances (e.g. PnD’s, smartphones, IoT), trends in transport policy towards smarter usage instead of expansion of the network and the smart mobility concepts arising from them. On the other hand, the urgency of robust and sound estimation procedures is triggered by rising congestion levels on these networks that are at an all-time high.

In this paper we address the known difficulties when estimating travel demand using link flows observed on a network with high levels of congestion. Such a network incorporates at least several active bottlenecks, which influence flow values both upstream (queues will form) and downstream (flow is metered). This implies that, on such a network, observed link flow values may represent either 1) the unconstrained travel demand for that link, 2) a proportion of the capacity of a set of upstream links, 3) the capacity of the normative (in terms of capacity deficit) downstream link or 4) a combination of these quantities. Which quantity each observed link flow represents depends on the specific traffic conditions in the network. Note that in practice a very large portion of observed flows is affected by flow metering (2) whereas only a small portion is unaffected (1) or affected by queues (3 or 4).

Demand matrix estimation methods use a traffic assignment model to assess the relationship between travel demand and link flow in intercept information. If the assignment model that provides the intercept information does not strictly adhere to link capacity constraints, as with static traffic assignment models, flow metering effects of bottlenecks (2) are not taken into account and all traffic is considered unaffected (1), thereby forcing incorrect assumptions upon the estimation. Therefore, matrix estimation methods using these models should only be applied on observed flows values that are unaffected (1), rendering them mostly useless on networks with high congestion levels. Note that by nature these assignment models should actually not be applied on study areas with congestion altogether.

Current practice to use observed flows affected by congestion (2, 3 or 4) is to derive unconstrained link demand values from the observed flow values, for example using the ‘Tonenmethodiek’ (used in the Dutch LMS/NRM models), or similar techniques that shift observed flows to upstream unconstrained links. Then, instead of the actual observed flows, these post-processed link demand values are used during matrix estimation. As such, these methods exhibit poor tractability and robustness and do not integrate any information from the assignment model about the composition of routes on the observed links.

This paper describes and compares three novel demand matrix estimation methods for large scale strategic congested transport models that use assignment models that strictly adhere to link capacity constraints, allowing them to explicitly consider the conditions under which link flows are observed. It compares these methods to the current practice and gives practical insights from applications, thereby demonstrating that these methods allow for usage of (big) data sources such as floating car data, congestion patterns and (route) travel time observations. Using these novel approaches, the need to post-process synthetic link demands is taken away, thereby increasing tractability and robustness of the matrix estimation methods and allowing for use of observed congestion patterns as additional input. Furthermore, these methods more efficiently reveal inconsistencies between model link capacities and observed congestion patterns and inconsistencies between count values, allowing the modeler to correct the model network and other matrix estimation input.

Authors continue research on the topic, the next goal being to extent the methods to support estimation of OD demand covering multiple time period(s), which should eventually lead to a method that supports 24 hour estimation.