Identification of Bunching State of Bus Lines Based on a LightGBM Model

Actual headways of adjacent buses of a same line can be significantly shortened, compared with the departing intervals, due to the influences of road situations and other factors, resulting in adjacent buses arriving at the same bus station in a relatively short period of time. This is called "bus bunching" in the transit industry. Identification of the bunching state of bus lines(i.e., bunching or non-bunching)is a key step to improve the operation of the urban public transit system. A LightGBM model with its parameters optimized by a Bayesian algorithm is proposed and applied to identify the bunching state. First, 20 factors related to the following five aspects including bus stops, operation, passengers, time, and weather, which potentially influence the bus bunching state, are selected. Spearman correlation test and variance inflation factor are used to diagnose their multi-collinearity. Then, a binary Logit model is developed to identify the significant impact factors, based on which the LightGBM model is developed to identify the bus bunching state. The super parameters of the LightGBM model(which are used to determine model attributes and training process)are optimized by a Bayesian optimization and a random search optimization, respectively. Finally, bus operation data from the City of Xi'an, China is used to verify the proposed model. The efficiency of the above two parameter optimization methods(i.e., Bayesian and random search)are compared, and the identification accuracy of the proposed LightGBM model is compared with XGBoost, Random Forest(RF), Decision Tree(DT)and AdaBoost models. Study results show that: first, the following factors, including number of passengers, number of signal lights, number of business districts within a short range, driving length on the main road within a short-range and traffic congestion index have a significant impact on the bus bunching state; second, the accuracy of the LightGBM model with its parameters optimized with the Bayesian method is 1.31%higher than that model with its parameters optimized by the random search method; third, the accuracy of the proposed Bayesian optimized LightGBM model for identifying the two bus bunching states(i.e., bunching or non-bunching)reaches 82.89%, which is found to be better than the above competing models.