The inability to accurately predict the following behavior of vehicles for long-time in common traffic scenarios is a shortcoming of existing following models. In this research, we propose a novel car-following model based data-driven methodology that has ability to predict the long-time car-following behavior with the high-resolution traffic data which is collected by the NGSIM (Next Generation Simulation). The proposed model first predicts the velocity, space headway, and velocity difference of the following vehicle 25-time steps individually using the Seq2seq (Sequence to Sequence) architecture with attention mechanism. Then the key features of each time step of the predicted data are extracted using CNN (Convolution Neural Network). Three different levels of traffic flow data are conducted with NGSIM data to validate the performance of the proposed model. The results indicate that the proposed model is better in terms of prediction accuracy, and it can be used for congested or uncongested traffic flow generation.
A Novel Long-time Prediction Car-following Model Based Mixed Seq2seq Architecture
2022-10-08
824597 byte
Conference paper
Electronic Resource
English
A Real-Time Seq2Seq Beamforming Prediction Model for C-V2X Links
Springer Verlag | 2022
|Traffic accident real-time detection and early warning method based on Seq2Seq auto-encoder model
European Patent Office | 2022
|European Patent Office | 2024
|