Duration: 21:51

132 views

0

We introduce a new perspective on the calibration of microscopic traffic simulations for use-cases where the focus is on reproducing observed travel times of individual vehicles rather than traffic counts. This contribution therefore deals with the calibration of network-wide traffic states to match trajectories of observed and simulated ego-vehicles spatio-temporally. The procedure is split into a global and local traffic state optimization: First, the daytime-dependent relative mean speed on all active edges is aligned network-wide by altering demand-, routing- and congestion-influencing measures in the simulation. Second, all edges passed by ego-vehicles are calibrated dynamically by inserting or removing artificial cars in neighboring roads to achieve the desired traffic states locally. The main goal is to delay the ego-vehicles in a similar manner as in observations, or in travel-time calibrated mesoscopic transport simulations -- for our specific use case. This paper focuses on the global optimization step: We identify suitable measures to influence travel times effectively and assess the measures' impact on global network states for a medium-sized network extract of downtown Berlin with over 1300 simulations.
We also evaluate the measures' robustness to stochastic noise to ensure random-seed independent results, and determine a quasi-optimum parameter set by means of a genetic algorithm. In summary, our global calibration technique aligns travel times during day-time hours very well, but requires further improvements to produce more satisfactory results at night.