Origin-destination (OD) demand modeling and estimation in urban road network is a foundamental component when a transportation planner or operator makes long term transportation planning and short term traffic management. Urban traffic route guidance, road space rationing, speed limit, or congestion control cannot be implemented without the accurate OD demands. OD demands can not only provide the decision of government investment construction projects and balance of supply-demand in urban road network, but also provide the data support of development scale of urban vehicles. Accurate OD demand estimation is helpful to catch on the traffic characteristics of road network, and hence alleviate traffic congestion when do some transportation planning and management. From the perspective of practical engineering application, the theories of full network and subnetwork OD demands estimation is studied. Specifically, the following main contributions are made in this thesis.
(1) Traffic flow based OD estimation: OD demand estimation under congested network
Two bi-level models to estimate OD demand under congested network are explored in terms of the observed link and route travel times, where one model has the known trajectories of observed route travel times and the other model has both known and unknown trajectories of observed route travel times. The proposed models leverage both the link and route traffic information to determine the network OD demand that minimizes the distances between the historical/observed and estimated traffic information (OD demand, link and route travel time) in the upper-level, and optimizes the stochastic user equilibrium (SUE) in the lower-level. Meanwhile, the observed information of travel time can capture the relationship between flow and cost (trip time) in congested network. The K-means (hard assignment) and Gaussian mixture model (GMM, soft assignment) clustering methods are presented to identify the trajectories of observed route travel times. An iterative solution algorithm, which contains the method of gradient descent, the method of successive average and Expectation-Maximization (EM) algorithm for solving the upper-level model, lower level model and GMM, respectively, is proposed to solve the built OD estimation models. Results from numerical experiments demonstrate the superiority of the travel time based model over the traditional flow based method in congested traffic network; using both the route and link information outperforms only using link information in the estimation of OD demand; and also suggest that the GMM is superior to the method of K-means especially in the case of observing some wrong data.
(2) Sample based OD estimation: scaling rate inference
The sample-based origin-destination (OD) demand estimation is to aggregate the survey data or mobile phone location data into the traffic network. Inevitably, the estimated OD demands should be scaled up to the population level counts. Hence, both the stochastic and deterministic scaling rate inferences are studied. A two-stage optimization model to determine the sensor location and stochastic scaling rate in an integrated manner, where the first stage which is sensor deployment model identifies the optimal sensor location strategy through minimizing the variability of the scaling rate inference under a budget constraint, and the second stage which is Bayesian-based scaling rate inference model leverages the date observed from the identified sensor locations and the prior information to determine the stochastic scaling rate. Deterministic scaling rate inference model which is a bi-level program is explored in terms of the observed link flows, which minimizes the distances between the observed and estimated link flows in the upper-level, and optimizes the SUE in the lower-level. A sequential identifying sensor location algorithm that avoids matrix inversions is proposed to solve the sensor deployment model. And the iterative solution algorithm is developed to solve the built Bayesian-based and deterministic scaling rate inference models. Results from numerical experiments demonstrate that the sensor deployment model can provide the most reliable scheme of sensor locations under certain budgets, further contribute to make a reliable estimation of stochastic scaling rate. The results also illustrate that both the endogenous information (i.e., prior information of scaling rate) and exogenous factors (i.e., observed link flows) can help make a better scaling rate inference.
(3) Subnetwork OD estimation using the method of full network route flow (obtained from incremental assignment method) aggregating
The user equilibrium (UE) based and SUE based incremental equibrium assignment methods are proposed to assign the OD demands and hence obtaion the UE and SUE solutions. In UE based incremental equibrium assignment method, the OD demand matrix is firstly divided into several matrixes, whose demands are divided from the whole OD demands in average. Each OD matrix is loaded on the shortest routes with the current link costs. When all the OD matrixes are loaded, the traffic flows on the route whose travel costs are not the shortest are picked up to aggregate the new OD matrix. Then divided the OD matrix and load each matix until the accuracy is satisfied. Unlike the UE based incremental equibrium assignment method, when loading the traffic flows in the SUE based incremental equibrium assignment method, the Logit formula is needed to compute the route choice probability. And also the ratio of current route flow to total demands needs to be derived. Hence, each OD matrix is loaded on the route with maximum difference between the route choice probability and the ratio. The redundant route flows that exceed the route choice probability are aggregated to form the new OD matrix. The subnetwork OD demands are obtained through aggregating all the route flows assigned from the UE based and SUE based incremental equibrium assignment methods. Results from numerical experiments demonstrate that the UE based and SUE based incremental equibrium assignment methods reserve all the route set in the iterative processes; and all these two methods can steadily converge to a high accuracy.
(4) Topology based subnetwork OD matrix estimation
A topology based subnetwork OD matrix estimation model under traffic demand constraints that explicitly considers both internal-external subnetwork connection and OD demand consistency between the subnetwork and full network is proposed. This new model uses maximum entropy of OD demands as the objective function, and uses total traffic generations (attractions) along with some fixed OD demands of subnetwork OD nodes as the constraints. The total traffic generations and attractions along with the fixed OD demands of subnetwork OD nodes are obtained through OD nodes transformation and subnetwork topology analysis. For solving the proposed model, a convex combination method is used to convert nonlinear topology based subnetwork OD matrix estimation model to classical linear transportation problem, and tabular method is used to solve the transportation problem. Results from experiments of the Sioux Falls network and Kunshan network demonstrate that the ratios of any two demands from different origin (destination) but the same destination (origin) are equivalent; the designed algorithm can rapidly make convergence to a high accuracy; and also the relative errors of traffic flows assigned from subnetwork and full network in most links are very small, hence the proposed model satisfies the application requirement.