A data-driven two-lane traffic flow model based on cellular automata

Abstract

In this paper, a data-driven two-lane traffic flow model based on cellular automata is proposed. Long Short-Term Memory (LSTM) and Support Vector Machine (SVM) are used to learn the characteristics of car following behavior and lane changing behavior, respectively, from real operation data of vehicles. Under optimal network parameters, the mean absolute errors of the LSTM network for training and testing data are only 0.001 and 0.006, respectively; while the prediction accuracy of the SVM classifier for both data reaches higher than 0.99. Moreover, forward rules and lane changing rules which are more consistent with actual situation are designed. The simulation results show that: (1) the new model can reflect the first-order phase transition from free flow to synchronized flow; (2) the frequency of unsuccessful lane changing is near zero in low-density traffic areas, but increases sharply in high-density regions; and (3) the lane changing duration and unsuccessful lane changing frequency display similar trends as traffic densities increase.