Hybrid data-model driven trajectory prediction on highways: Integrating anticipatory interaction awareness and personalized driving preferences

Accurate trajectory prediction of human-driven vehicles (HDVs) in mixed traffic environments is critical for enabling safe and efficient interactions between connected autonomous vehicles (CAVs) and human-driven vehicles on highways. The coexistence of HDVs and CAVs introduces complex dynamics: HDVs exhibit heterogeneous driving preferences influenced by driver behavior patterns, while CAVs’ planned trajectories create anticipatory interactions that reshape HDV motion patterns. Existing methods often overlook the dual challenges of personalized driving preference modelling and anticipatory interaction awareness, particularly in scenarios where CAV trajectories are dynamically integrated into HDV prediction frameworks. To address these challenges, we propose a hybrid data–model driven framework that integrates physics-based behavioral calibration with data-driven interaction modeling. At the core of this framework is the Kepler Optimization-based Temporal Attention Fusion Transformer Network (KO-TAFTN), which enables unified modeling of dynamic historical interactions, anticipatory interactions, and static driving preferences. A driving preference extraction module first derives individualized behavioral traits using Kepler-based physical modeling. These preferences are encoded into context vectors via a static encoder and static enhancement layers, and then incorporated into the network. To improve interpretability and robustness, a variable selection module is applied to evaluate the relevance of input features. A dynamic encoder and a temporal attention fusion module jointly capture and fuse historical and anticipatory interactions by modeling temporal dependencies. Finally, a multimodal trajectory prediction module generates diverse candidate trajectories that reflect potential future motion patterns of HDVs. Experiments demonstrate that the proposed framework consistently outperforms benchmark methods in mixed traffic environments, particularly under complex and congested scenarios. Visualization results further validate the effectiveness of integrating human driving preferences and anticipatory interaction cues. These findings underscore the framework’s potential to improve interaction safety and trajectory accuracy during the transitional evolution of mixed traffic systems.