Abstract: In order to overcome the shortcomings of the existing prediction methods, such as short prediction steps and insufficient utilization of spatial-temporal characteristics of road networks, and to predict the freeway travel time accurately, five commonly used prediction models, namely RF(Random Forests),XGBoost(Extreme Gradient Boosting),LSTM(Long Short-Term Memory),KNN(K-Nearest Neighbor), and SVR(Support Vector Regression), were taken to carry out multi-steps prediction of freeway travel time based on the origin and destination data set. A fusion model based on Bayesian linear regression method was proposed. The Long-gang to Bu-long section of Shui-guan Expressway in Guangdong province was taken as a case study. We predicted the travel time of every 15 minutes in the next 2 hours. The results show that the prediction performance of RF model and XGBoost model is stable under multi-steps; the LSTM model has superior prediction performance in the case of short prediction steps; the fusion method integrates the advantages of various prediction methods and has higher accuracy and robustness. The experiments also demonstrate that it has the best prediction performance.

Key words: travel time prediction, machine learning, spatial-temporal characteristics, freeway, Bayesian regression