Beauty or the Beast: Stability of Large Scale Microsimulation Models
S Ahuja, T van Vuren, Mott MacDonald Ltd, UK
In this paper we present our research, experiences, problems and issues associated with building and managing large scale microsimulation models
In this paper we present our research, experiences, problems and issues associated with building and managing large scale microsimulation models. We present our findings into the stability of large scale models and the issues concerned with the convergence, runtimes and calibration of such models. We conclude with guidelines on building large scale microsimulation models and highlight areas of further research.
With the increasing popularity of microsimulation models and faster computing technology becoming cheaper, there has been a trend to build larger and more complex microsimulation applications. Traditionally microsimulation models were applied for optimising traffic flow on a few intersections or a small section of the highway network. More recently, there has been a demand for simulation models to cover large parts of the urban and highway network, to segment road vehicles into vehicle and driver classes, to include other road users such as pedestrians and to model responses such as parking choice and land use effects.
Our experience shows that there are three fundamental problems that arise from the increasing size of microsimulators:
- run time problems
- convergence problems
- calibration problems
To some extent these are all related. In the paper we suggest the following solutions to these problems:
- linking microsimulation with conventional assignment models, with the routes being calculated by the assignment model and being fixed in the microsimulation application. We will present comparative statistics for link flows and delays, and run time implications
- as a pre-processor calculating a small number of routes in the microsimulation model directly without simulation at junctions, and using these as fixed routes in full simulation. We will present the impact on run times, and measures of fit between flows in the pre-process and after full simulation
- collecting calibration data over an extended period and explicitly allowing for variability in the calibration statistics. We will present illustrative examples
This paper is based on the development of three very large and complex microsimulation modes for urban town centres in the Birmingham City region, United Kingdom. Each of the models has around 200 zones and 600 zone connectors. They each cover a geographic area of over 40 square kilometres and have more than 250 kilometres of road length modelled. They include over 500 junctions of which 10% are signalised intersections that operate on real time optimisers such as SCOOT, MOVA or VA signal control. In addition all models include over 45 pedestrian intersections modelled with realistic pedestrian flow and signal control data. The travel demand in all models is disaggregated into car, LGV, HGV and bus trips, further disaggregated by departure time and trip purpose. Over a typical peak hour, the models simulate over 40,000 vehicles. Two of the models have an additional sophisticated layer of car parking choice model. In these simulations, the vehicles can choose between various car parks based on the attractiveness and capacity of a particular car park and simulate the effect of car parking guidance systems.
The primary objective of building such complex and sophisticated microsimulation models has been to quantify and test the impact of redevelopment and land use changes due to large regeneration projects. The requirements of public consultation make the use of microsimulation tools attractive; however, it is the modeller?s responsibility to ensure that the visualisation is supported by robust models.
This paper is subject to approval of the clients for the project on which it is based.
Association for European Transport