Improved multi-lane traffic flow simulation based on weigh-in-motion data

Vehicle load is the main variable load of highway bridges, which significantly affects the security and duration of bridges. Highly precise vehicle load models are the basis for studying vehicular load responses and providing reliability assessments of highway bridges, whose core is the spatio-temporal distribution of vehicular load on a bridge deck. Vehicle type, axle weight, axle distance and position are the main parameters of spatio-temporal simulation of traffic flow. Weigh-in-motion (WIM) systems can obtain these original vehicle parameters efficiently. However, there are two main problems about how to accurately simulate the above traffic flow parameters based on WIM data: (1) A large number of WIM systems have low accuracy for vehicle classification and an efficient classification method for vehicles from massive WIM data has not been proposed; (2) The most popular MC simulation method ignored the statistical dependence of parameters (vehicle sequence, axle weight, axle distance, etc.), which results in the deviation between traffic flow simulation samples and measured data. In this study, to obtain traffic flow simulation results closer to measured data, a multi-lane traffic flow simulation method via fusion of efficient vehicle classification and the statistical dependence of parameter was proposed. First, the overall framework of the traffic simulation was introduced. Subsequently, the vehicle classification process and traffic flow simulation were introduced respectively. For the vehicle classification, the abnormal vehicles were removed based on the box diagram method, and then the vehicle classification was completed based on the improved K-means++ clustering algorithm. The vehicle sequence and the axle weight/distance samples were obtained by MCMC simulation method and Copula function theory respectively. Finally, a case study was presented using WIM measured data of an 8-lane expressways. The results showed that the optimal number of clusters determined based on the K-means++ clustering algorithm was reasonable. The vehicle simulation samples considering the statistical dependence of parameter were closer to the measured data than previous methods and the normal copula function was slightly better than the t-copula function.