Traffic modeling practice has witnessed a plethora of studies applying different simulation tools to a wide range of applications. These tools span the gamut of modeling resolutions, from the detailed micro-simulation to less computationally expensive macroscopic simulation. In the middle lies a set of mesoscopic (meso) tools that use a variety of techniques to trade off the accuracy of micro-simulation to achieve running times similar to their macroscopic counterparts. Running time has historically been the primary motivation for the adoption of macro and meso models. However, the level of detail of micro-simulation has generally been desirable for many applications such as emissions modeling, control delay estimation and the evaluation of lane-based phenomena such as weaving.

Powerful computing hardware combined with software optimization allow today’s modelers the luxury of re-evaluating the above trade-off. A survey of the traffic simulation literature indicates a general lack of studies that objectively compare micro and meso approaches on a common dataset. A comparison of tests performed across tools is also confounded by the differences in algorithms (e.g. shortest path calculations and route choice models) and their implementations. We address some of these limitations in this paper, by executing micro and meso simulations on the same network and demand data. Critically, we also use a common software platform that can run meso and micro simulations on the same dataset. This approach minimizes the impacts arising from software implementation choices, and allows a comparison of the underlying core model components instead.

We present numerical results for a large-scale network in Phoenix, Arizona, derived from the regional model for the Maricopa Association of Governments (MAG). The network spans about 500 square miles and has a demand of about 2 million trips in each peak period. A detailed, lane-level simulation database was created from the planning center line network and aerial imagery, and traffic control data were coded for more than 1,800 intersections. Microscopic and mesoscopic simulation models were calibrated to match traffic count and Inrix speed data. The results from the two calibrated models are compared and conclusions are drawn about any reduction in accuracy in the meso model.