XGBoost-Based Travel Time Prediction between Bus Stations and Analysis of Influencing Factors

Abstract

Real-time and accurate travel time information between bus stations is critical for passengers to make suitable travel plans to reduce waiting time at the stops. By mining and analyzing bus operational data, it can be obtained that factors such as the variation of vehicle speed in adjacent sections and the proportion of bus lanes between stations have affected the travel time between bus stations. Therefore, considering the temporal feature, spatial feature, and weather feature as the prediction model’s input, travel time between bus stations prediction model based on eXtreme Gradient Boosting (XGBoost) was trained and established. The 28-day bus operation data of a certain bus line in Guangzhou was used for training and verification, and they were compared with the prediction models based on K -Nearest Neighbors (KNN), BP neural network, and Light Gradient Boosting Machine (LightGBM). In comparison with other models, the lowest MAPE of 11.96% was found for the XGBoost prediction model, which is 9.30% lower than other models on average. The sensitivity analysis of the proposed prediction model was further conducted: temporally, the accuracy of the prediction model was best during the flat peak hours; spatially, the MAPE of the model gradually decreased as the number of line units increased, and when the number of line units exceeded 18, the accuracy of the prediction model stabilized and was lower than 7%. The results confirm that the XGBoost model outperforms the KNN, BP, and LightGBM in terms of fitting, accuracy, and stability.