What question did this study set out to answer?

The aim is to improve trajectory prediction for vehicle platoons in the presence of traffic oscillations using a new predictive model.

May 31, 2026Communications in Transportation Research0 citationsOpen Access

Vehicle platoon trajectory prediction under traffic oscillation: A causal physics-informed deep learning approach

JLJipu Li STShan Tian DNDong Ngoduy

Key Points

The aim is to improve trajectory prediction for vehicle platoons in the presence of traffic oscillations using a new predictive model.
Developed the Causal Physical Information Deep Learning Model (CPIDLM) with a causal graph attention mechanism.
Implemented a physics-informed enhancement architecture to integrate physical model insights into deep learning.
Created a dynamic adaptive weighting module for balancing physical laws and data-driven patterns.
CPIDLM achieved a 16.7% reduction in gap prediction error compared to existing models, indicating superior accuracy.
The model significantly outperformed current state-of-the-art methods in dynamic environments.

Abstract

Abstract Autonomous driving based on single-vehicle perception still cannot avoid the negative impacts caused by traffic oscillations when traveling in mixed-vehicle platoons. In this scenario, real-time and accurate prediction of the future trajectories of the vehicle platoon ahead of the autonomous vehicle is key to addressing this issue. However, in mixed platoons dominated by heterogeneous and uncertain human-driven vehicles (HDVs), this remains a significant challenge. Existing model-driven, data-driven, and hybrid approaches often suffer from poor interpretability, weak physical constraints, and insufficient accuracy in dynamic environments. To overcome these limitations, this study proposes a Causal Physical Information Deep Learning Model (CPIDLM) for high-fidelity platoon trajectory prediction. First, CPIDLM constructs a novel causal graph attention mechanism that explicitly captures behavioral heterogeneity and causal interactions among vehicles. Second, a physics-informed enhancement architecture is developed to embed prior knowledge from physical models into the deep learning network. Additionally, a dynamic adaptive weighting module is designed to achieve a dynamic balance between contributions from physical laws and data-driven patterns. Extensive validation based on real-world trajectory datasets shows that, compared to existing models, CPIDLM reduces gap prediction error by 16.7%, significantly outperforming current state-of-the-art methods in accuracy. This study establishes a powerful new paradigm for vehicle platoon trajectory prediction under traffic oscillation scenarios.

Ask AI

Helpful

Bookmark

View Full Paper