Dynamic Origin-Destination Matrix Estimation in Large-Scale Congested Networks (Schatting van dynamische herkomst-bestemmingsmatrices in grootschalige, congestiegevoelige netwerken)

Dynamic Network Loading (DNL) models and Dynamic Traffic Assignment (DTA)models are an important tool in the context of traffic planning and trafficmanagement. An essential input of these models is a dynamic origin-destinationmatrix. The dynamic origin-destination (OD) matrix specifies the number of travelersfrom each origin to each destination, and this for each time period. In general, theOD matrix is not directly observable. Therefore, estimation methods are used toderive this OD matrix. They make use of available information such as measuredflows on a road, measured travel times, etc. When reviewing the available literaturethat discusses these methods, we can identify an important gap in current state-ofthe-art methods, namely the impact of congestion on dynamic OD estimation. Thepresence of congestion has a non-linear effect on the relationship between link flowsand OD flows. This relationship is crucial in the dynamic OD estimation problem. Inthis thesis, this relationship is studies in two phases.In a first phase of this research, the effect of congestion spillback is analyzed by theuse of synthetic experiments. It is found that, when congestion spillback is present,the accuracy of the estimated OD matrix depends on the spillback mechanism that isused in the OD estimation. More importantly, the presence of congestion spillbackitself is identified as a source of inaccuracy. The ambiguity of traffic flowmeasurements is conceived as the source of this inaccuracy: a certain flowmeasurement can indicate both slow moving vehicles with a small headway (incongested conditions), and fast moving vehicles with a large headway (in free flowconditions). To overcome these problems, the use of additional information sourcesis suggested, combined with the use of specialized estimation methods. A novel ODestimation procedure based on density measurements is presented and tested in casestudy on a real-world network. It is shown that the manner in which this informationis used, has an impact on the performance of the estimation: when the speeds werecombined with the flow measurements into densities, better results were obtained.However, the performance of the developed methods on a real-world case study isnot yet satisfactory in terms of the match between estimated and measured link flowsand speeds.The second phase of the research focuses on the approximation of the relationshipbetween link flows and OD flows, and its impact on the OD estimation problem.When solving the OD estimation problem, this relationship is approximated. In stateof-the-art methods, this is usually done by assuming a constant mapping, whichimplicitly assumes separability of the link flows to the OD flows, which can lead toinaccurate results when dealing with congested networks. Another important sourceof error attributable to congestion dynamics is the presence of multiple local minimain the objective function. It is illustrated that these local minima are the result of anincorrect interpretation of the information from the detectors. This work proposessensitivity-based OD estimation (SBODE), which uses a linear instead of constantapproximation. It is shown to lead to superior OD estimated.Although the proposed SBODE method yields superior OD-estimates with respect toexisting methods, its computational complexity is also much larger due to the need tonumerically approximate the sensitivity of link flows to OD flows. Hence, in thefinal phase of this research, we focus on efficient algorithms allowing the applicationof the developed methodologies to real-world networks. A first technique is MarginalComputation (MaC), which is a computationally efficient method that performs aperturbation analysis (using kinematic wave theory principles) for deriving thesensitivity of the link flows to the OD flows. The calculation of this sensitivity is arequirement of the SBODE methodology. Next, a hierarchical approach fordecomposing the dynamic OD estimation problem in a number of smaller problemson subareas is presented. The main idea is to perform a more accurate dynamic ODestimation only on subareas with substantial congestion. Both approaches aresuccessfully tested, both on a study network and on a real network.status: publishe