A TCN-Transformer Model of Acceleration Prediction for Preceding Vehicle in CACC Systems

This study establishes an acceleration prediction model of the preceding vehicle tailored for cooperative adaptive cruise control (CACC) systems that address the common challenges posed by V2V communication delays or no communication from the preceding vehicle. The designed model integrates a temporal convolutional network (TCN) with the Transformer architecture block to create a sophisticated mechanism for temporal data analysis. This integration leverages the strengths of TCN in capturing spatial features and the multi-attention mechanism of Transformers to enhance the recognition of long-term dependencies in the time series of data. The synergy between these techniques not only effectively narrows the identification gap between short-term and long-term dependencies but also helps to process spatial information for accurate time-series prediction, significantly improving prediction accuracy. Simulations using the NGSIM dataset confirm the model's prediction performance over the traditional model. Further, the potential of the method to solve the communication delay problem is verified, highlighting the effectiveness of TCN and Transformer multi-attention mechanism in improving the CACC capability.