When calibrating microscopic or macroscopic traffic flow models to trajectory or stationary detector data, one often obtains surprising results such as unexpected parameter estimates that may also vary widely between comparable situations. Furthermore, one often observes insignificant goodness-of-fit differences between models, even when comparing very simple (and unrealistic) models with highly refined models having many parameters. As a result, rankings of different models according to their descriptive power are somewhat arbitrary. In this contribution, I discuss several confounding factors underlying these somewhat frustrating results, namely inter-driver variation, intra-driver variation, data selection, and the choice of the calibration method/objective function. Interestingly, it turned out that data preparation (e.g., changing the sampling rate, smoothing out noise, or making trajectories internally and platoon consistent) had comparatively little effect on the estimation results.
Copyright. All Rights Reserved.