Developing a time-series speed prediction model using Transformer networks for freeway interchange areas

Abstract

This study investigates the lane-level speed distribution in the freeway interchange area and develops a time-series speed prediction model using Transformer networks. The full-sample real-time speed data from interchange areas under heavy traffic was extracted by the YOLO v3 detection algorithm and Deep-Sort tracking algorithm based on Unmanned Aerial Vehicle (UAV) videos. A short-term prediction model of lane-level driving speed was constructed using the Time-Series Transformer (TST) framework. Results showed that the greatest magnitude of speed change occurred during the process of switching from the passenger lane to the truck lane. The inner lane consistently demonstrated higher mean speeds compared to the outer lane. The TST model proposed in this study could achieve an accuracy of 98.35% in predicting lane-level driving speed. These findings suggest the need to consider the speed transition between passenger and truck lanes in freeway engineering design, marking setting, and optimization of safety facilities.