Traffic simulation optimization considering driving styles

Abstract

Parameter calibration is essential for ensuring the accuracy of microscopic traffic simulations. The expected speed is a critical parameter that characterizes behaviors of vehicles in most simulation models, which is influenced by road traffic conditions and the driving characteristics of different drivers. Most existing parameter calibration methods typically concentrate on micro-level parameters such as time headway and lane change motivation, while overlooking the calibration of vehicle expected speeds in consideration of driver behavior habits. This study combines data from highway electronic toll collection (ETC), gantries, and 100-m mileage average speed data, and proposes a method for calibrating vehicle expected speed that considers driving style clustering. The Gaussian mixture model (GMM) algorithm is used to develop driver models with three distinct driving styles: aggressive, moderate, and conservative. To ensure driving diversity and enhance parameter calibration efficiency, we rebuild vehicle driving models and representative parameters based on the classification results. Moreover, the Bayesian optimization algorithm is modified in conjunction with a microscopic traffic simulation model to perform automatic calibration of expected speeds. Experiments conducted on the Shanghai–Hangzhou–Ningbo highway demonstrate that the proposed method significantly reduces the mean absolute percentage error (MAPE) from 20.2% (using default parameters) to 3.1%. Additionally, in the model robustness test, the MAPE reaches 5.01%, indicating a certain level of stability and scalability. This method proposes a tailored calibration method accounting for the heterogeneous driving behaviors of micro-traffic simulation models, achieving satisfactory calibration results for simulation models in highway scenarios.