Exploring sensitivity-based clustering of OD variables in dynamic demand calibration

The estimation of Origin - Destination (OD) matrix is a methodologically and computationally challenging, yet essential step in setting up a transportation planning model. In this process, demand (and supply) parameters need to be calibrated to match the simulation output with real observations. In this paper, we investigate how information extracted from a Jacobian matrix can be applied to improve efficiency and efficacy of Dynamic Origin-Destination estimation. Following [1], we formulate the OD estimation as a Least Square problem and solve it using a sensitivity-based solution algorithm (sensitivity-based OD estimation (SBODE), using Levenberg-Marquardt (LM)) algorithm. Information from the Jacobian matrix is used to define clusters of OD-demand variables. Opposed to the original SBODE where all elements can influence the optimization direction, in this paper the optimization direction of different clusters is computed separately. We also introduce in the LM algorithm a modified damping factor, which improves efficacy and efficiency of the algorithm by providing a step size tailored to each cluster of OD-demand variables. Several experiments have been conducted on a severely congested corridor network, confirming the potential of the clustering method. The clustering breaks the OD estimation problem into smaller sub-problems, herewith reducing the risk of being trapped in local optima.