Travel Time Prediction Using Machine Learning Algorithms: Focusing on k-NN, LSTM, and Transformer

Background Travel time prediction is fundamentally important in advanced traveler information systems. Particularly, in the Dedicated Short-Range Communications (DSRC) traffic information system targeted by this study, travel times are obtained after vehicles terminate trips on the route, indicating that a time-lag phenomenon is inevitable. Therefore, travel time prediction is even more emphasized in DSRC systems. With advances in artificial intelligence technologies, many new machine learning algorithms have recently been introduced. Methods In this study, three machine learning algorithms—k-Nearest Neighbor (k-NN), Long Short-Term Memory (LSTM), and Transformer—were applied to predict travel times collected on a DSRC-equipped national highway. Travel time outliers caused by entry and exit maneuvers at intersections were filtered using a robust outlier treatment algorithm. Results From applying the prediction algorithms, errors of 12.4%, 11.7%, and 10.3% were revealed for k-NN, LSTM, and Transformer, respectively. To identify the statistical significance of the differences in prediction performance, paired t-tests were performed. Consequently, all the differences between the algorithms were proven to be statistically significant. The differences in performance were found to be generally more significant under congested conditions compared to uncongested conditions. Conclusion The enhanced reliability of real-time travel time information can increase the effectiveness of DSRC traffic information systems by encouraging drivers to divert from congested routes or adjust their trip schedules to less congested periods.