Dealing with uncertainty in detailed calibration of traffic simulation models for safety assessment

• Calibrations of driving behaviour models using aggregated data are likely to produce biased detailed outputs.
• Uncertainty in traffic micro-simulation applications can be assessed using multi-step global sensitivity analysis.
• Multi-step global sensitivity analysis can significantly reduce the calibration experiment dimensions.
• Detailed calibration of complex driving behaviour models can result in a good replication of detailed traffic variables.

With the increasing level of detail of traffic simulation models, the need for a consistent understanding of simulators’ performance and the adequate calibration and validation procedures to control uncertainty is crucial, particularly in applications focusing on complex driving behaviour and detailed outputs, such as road safety analysis.In this work the calibration of traffic microscopic simulation models for safety analysis is analyzed considering four different key uncertainty sources: the input data, the calibration methodology, the model structure and its parameters, and the output data. The use of a multi-step sensitivity analysis (SA) framework is proposed and applied to the simulation of an urban motorway scenario, using a complex traffic simulation model with more than one hundred parameters. A three-level analysis is presented: (1) different advanced SA and calibration methods are described, compared and integrated in a multi-step global SA framework; (2) the proposed method is tested using both vehicle trajectory and aggregated traffic data to assess the impact of model parameters uncertainty and different types of input data on relevant outputs; and (3) accident and non-accident scenario-specific calibrations are performed to test the capacity of the simulator in replicating changes in detailed traffic and safety related measurements. Different techniques are adopted in each phase of the global SA and calibration method, attending to the problem complexity, the dimensionality of the experiment, and minimizing the necessary number of model evaluations.The proposed method successfully identified the role played by all parameters and by the model stochasticity on different safety outputs. The final model calibration, carried out by explicitly considering the presence of uncertainty at different levels, confirmed the potential of advanced microscopic traffic models to adequately replicate detailed traffic and safety measurements, shedding light on different aspects of the interaction between road safety and traffic dynamics.