Towards Developing an Integrated Microscopic Traffic Simulation Model for Large Road Networks

Abstract

Traffic jams have become one of the key topics in the 21st century and cause immense losses in productivity, increase CO2 emissions and affect driver stress, travel time predictability and increased wear and tear on vehicles. To test possible countermeasures, optimise existing infrastructure or develop new Intelligent Transport Systems (ITS), traffic has to be modelled. The inherent complexity which is a result of inter- and intra-driver heterogeneity, macroscopic feedback loops, local interactions, multi-modal transport and many more is tackled with simulations. By representing the (longitudinal) dynamics of individual vehicles (“microsimulation”), high model fidelity can be achieved. By aggregating the resulting trajectories, macroscopic phenomena emerge and can be incorporated to answer a wide range of traffic-related questions. Since human drivers vary in their perception of stimuli, preferences and reaction, randomness is inevitably and circumvents the construction of perfectly accurate models. Which microscopic features are needed and how they should be mathematically represented runs like a central thread through this thesis. First, it develops a systematic classification scheme to identify modelling strategies and evaluate advantages and shortcomings of (partly) discrete microsimulations. Because real-world trajectories are continuous in time and space, representing them discretely leads to artefacts which induces an upper precision boundary for all models operating on this level of discretisation. Such model-independent errors will be measured based on empirical, naturalistic and synthetically generated trajectories. It is also evaluated to which extent driver heterogeneity and randomness may be compensated with discrete components to simplify modelling and increase computational efficiency. Based on the gained insights, a new, integrated microscopic model is developed. The second major theme in this thesis are gridded, discrete road topologies (chequerboards, Manhattan layouts) populated with vehicles “hopping” from one location to the next. An extensive literature review summarise existing approaches and it is discussed how these standardised road networks and extremely simplified dynamics are well-suited as testbed for ITS. Based on identified research gaps a Timed Automata-based particle hopping model is developed. The two main tools to achieve the outlined objectives are literature reviews, data analysis and computer simulations. To construct the classification scheme, existing genealogies and typologies for traffic models and dynamical systems in other scientific disciplines were reviewed. Another approach was taken for conducting the literature review of chequerboard models: starting from the prototypical BML model, the citing sources are surveyed in reverse order, the behaviour of relevant models is synthesized and contradictions and gaps critically analysed. To quantify the model-independent error, datasets are sampled, quantised and discretised over a wide range of step sizes. Making use of naturalistic driving data, synthetically generated trajectories and high-quality experimental observations, the highest achievable errors for (partly) discrete microsimulations is measured. The dataset recorded under experimental conditions with a group of homogeneous drivers is reused to identify to quantify driver heterogeneity and find the maximum achievable correlations between car-following stimuli and reactions. In summary, this thesis shows that randomness and heterogeneity in human drivers is significant and exceeds the model-independent discretisation errors for a wide range of quantisation parameters. This justifies representing trajectory features discretely to increase computational efficiency and improve the modelling process. Based on this proposition two new formulations for simplistic and advanced microsimulations are developed. The former is able to reproduce all macroscopic free-flow to congestion phase transitions while the latter integrates lateral and longitudinal dynamics based on statistical microscopic properties. Both approaches fill gaps identified by a systematic literature review.