Mesoscopic Simulation of Multi-Modal Urban Traffic

Urbanisation is one of the most dominant global megatrends. As many metropolitan transport networks are already under pressure, the prospect of having to fit even more people into such often car dominated urban systems is disturbing. Not only does it result in countless hours wasted in traffic, but also increasing local pollution and green house gas emissions.

One way to relieve motorised traffic in urban areas is to provide appropriate alternatives to car traffic. Public transport systems where passengers can reach their destinations reliably and dedicated bicycle infrastructure allowing people to travel smoothly by bicycle can for many be seen as sustainable and effective alternatives to car traffic in urban areas.

However, even in cities with a high share of cyclists, the models for evaluating future projects and initiatives concerning bicycle traffic are of much lower quality than their counterparts for car traffic, for instance regarding their ability to estimate travel times. As such projects are regarded as socioeconomic investments with the return being received in travel time savings, there is a risk that the lack of appropriate models is setting back bicycle-friendly development.

In order to be able to evaluate transport policies with a higher level of detail, in the past decade transport models have shifted towards agent-based models. Agent-based models follow every individual of the population for the entire day, and makes it possible to evaluate very detailed effects that macroscopic models cannot.

Despite agent-based models being capable of simulating very large areas in great detail, the aforementioned capability gap between models for car traffic and other modes of transport persists, at least when disregarding dedicated microscopic models limited to small geographical areas. This PhD thesis aims at reducing this gap by modelling detailed individual behaviour in large-scale agent-based transport simulation models.

The trade-off between detail level and computational performance is a vital focus point, as the developed models are intended for application on a metropolitan scale. Even while doing so, the thesis contributes diversely to the literature through five papers on agent-based modelling divided into three parts: i) Output variability in agent-based simulation of transport systems (Paper 1), ii) Agent-based passenger delay modelling with real-time information (Paper 2), and iii) Agent-based simulation of bicycle traffic and interaction with cars (Papers 3-5).

Part I deals with one of the downsides of most agent-based simulation models – the strong dependence on pseudo-random numbers. Whereas most aggregated models can get away with stating choices by their choice distribution, the individual constituents of agent-based models require explicit choices to be made, either by resolving to completely by their random seed, changing the seed will also change modelled choices, eventually alter the final model output.

As such, the output of stochastic agent-based models can be considered as samples from – extremely complex – random distributions. The first paper of the thesis investigates the variability of such outputs by conducting an experiment in MATSim with a large-scale scenario of Santiago de Chile using 100 different runs – each using a different random seed.

The corresponding link flows are analysed, and the between-seed variation is generally found to be small with tolerable coefficients of variation. However, for some links the relative error is occasionally serious when compared to other random seeds, indicating that the results based on a single seed may be deceiving.

Finally, it is found that the variation between the last and penultimate iteration is almost entirely dominated by the variation across the different seeds, suggesting the need for using multiple random seeds when analysis results. Part II of the thesis is concerned with agent-based passenger delay modelling under consideration of real-time information.

Public transport systems are often uncertain, and automated vehicle location (AVL) data of their vehicles can be used to measure vehicle punctuality. Although being extremely relevant, evaluating how such delays influence passenger travel times is rarely done, presumably because doing so is inherently tricky as passengers may adapt to changes along the way, especially when real-time information is present.

The objective of the second paper of the thesis is to formulate a simulation model that based on recorded vehicle delays from AVL data can determine the corresponding passenger delays at a large scale under different levels of real-time information prevalence. The model lets its passengers search for new alternatives every two and a half minutes while they travel through the system based on the at any time accessible real-time information.The model is applied to the public transport system of Metropolitan Copenhagen and 812,359 daily trips are modelled for 65 days where real-life AVL data from the rail and bus network was collected.

The study shows how a particularly irregular railway line causes many passengers to pursue alternative routes, and that information of better alternatives often occur at stations with many high-classed options. In line with existing literature it is found that despite some trips arriving earlier than expected, on average the passenger delays are far larger than the vehicle delays that cause them.

The study further discovers that passenger delays can be reduced considerably by using real-time information obtained at the beginning of the trip and even further when acquiring information en-route. Part III consists of three papers that successively develop a novel, tailor-made methodology for simulating bicycle traffic and interaction with car traffic at a metropolitan scale.

The first paper of the part – the third paper of the thesis – develops a computationally fast methodology for realistically simulating congestion in separated bicycle traffic. As the heterogeneity of cyclists is profound, taking overtaking into account is essential for deterministic, rule-based setups or by using pseudo-random numbers to draw such choices stochastically.

However, as sequences of pseudo-random numbers are entirely determined being able to simulate bicycle traffic appropriately. The research uses video tracking data of low intensity bicycle traffic to investigate how desired speeds and preferred headway distances vary across cyclists, and estimates appropriate probability distributions for these cyclist characteristics used as input for the developed model.

The model is based on simple assumptions, and allows overtaking by having cyclists explicitly choosing lanes when entering a link. A bottleneck network is used for testing the methodology under a wide spectrum of traffic intensities. As designed for, the model is shown to be more likely to pose excess travel time on cyclists with high desired speeds, whereas high traffic intensities are required to consistently delay cyclists with preferences for slower speeds.

Still, the derived fundamental diagrams of the small-scale experiments are validated with observed video tracking data of 3,763 cyclists from a morning peak hour at Queen Louise’s Bridge in Central Copenhagen. The research carried out in the paper exposes that recognising cyclist heterogeneity is essential for realistic simulation of bicycle traffic.

The fourth paper is a natural extension of the previous paper. Before this paper, no dedicated bicycle traffic assignment models with feedback between route choice and travel times were present in the literature. The paper alters this by using the methodology from the third paper to simulate bicycle traffic and integrates it in a traffic assignment model with meaningful feedback between supply and demand.

The method is implemented in MATSim and applied to a large-scale case study of Metropolitan Copenhagen with 1,082,958 bicycle trips. Although the excess travel time of cyclists is low compared to other modes, through scenarios using better and worse bicycle infrastructure the study shows that this can to a large degree be contributed to the high level of bicycle infrastructure in Copenhagen.

Still, it is shown that flows differ significantly between the initial and final iterations, demonstrating that feedback from the network is needed when modelling bicycle traffic in cities with a high share of cyclists – even more so if the infrastructure is insufficient. Whereas the two preceding papers focus on simulated bicycle traffic on links, the fifth and final paper addresses how to model the intersections of the network.

In cities with a high level of separated bicycle traffic, intersections are of particular interest as practically all interactions between bicycles and motorised traffic occur here. The paper formulates a joint car and bicycle traffic assignment model capable of modelling right-of-way at every network node, ensuring that conflicting moves do not take place simultaneously.

The agent-based model is implemented in MATSim and applied to the same large-scale case study of Metropolitan Copenhagen as the fourth paper, but with the addition of 3,210,685 car trips, 299,416 truck trips, and a car network increasing the network size to 572,935 links and 144,060 nodes. Simulations are run with the right-of-way node model and the original node model of MATSim for setups with car/truck, bicycle, and all three modes, respectively.

Without increasing computation times notably, the study shows that omitting to include yielding due to right-of-way at intersections underestimates travel times and causes too much traffic to be led through the city centre. It is furthermore shown that inter-modal conflicts between bicycle traffic and motorised traffic at intersections delay traffic more than the intra-modal conflicts, emphasising the need for joint modelling of multi-modality.

In summary, this PhD thesis has contributed to the literature on large-scale agent-based simulation of urban transport systems, spanning across output variability, public transport passenger delay modelling, and dynamic traffic assignment of bicycle traffic. This includes recommendations for future practices when dealing with agent-based modelling, but also in terms of developing methodologies for new types of models for detailed largescale modelling of passenger delays and congested bicycle traffic, allowing estimating effects that were previously ignored when planning and evaluating metropolitan transport systems.

Although there are still additional steps to be taken within each of the topics covered by the five papers of the PhD thesis, their contributions constitute considerable improvements to, not only the understanding, but also the capabilities of agent-based transport simulation models for multi-modal urban traffic.