Multi-Expert Trajectory Prediction for Highway Weaving Sections Using Conflict Potential Energy and GAN

To address the challenge of lane-changing prediction for autonomous vehicles (AVs) within highway weaving sections characterized by dynamic interactions in human-machine mixed traffic flows, this paper proposes a multi-expert collaborative prediction model based on the Conflict Potential Field and Generative Adversarial Networks, termed G-MoE-WGAN. This model quantifies the dynamic game intensity among interacting vehicles via a Conflict Potential Field (CPF). The proposed model constructs a Mixture of Experts (MoE) system to facilitate intention decision-making and trajectory generation, employing adversarial training to optimize the physical plausibility and interaction adaptability of the prediction results. Furthermore, a Physics-Informed Neural Network (PINN) is introduced for the reconstruction and smoothing of raw naturalistic driving trajectories to validate the model. Experimental results demonstrate that the G-MoE-WGAN achieves a lane-changing intention classification accuracy of 94.35%, representing an improvement of 3.89% to 16.55% compared to baseline models. Within a 3-second prediction horizon, the Final Displacement Error (FDE) and Average Displacement Error (ADE) of the proposed model are significantly reduced by 8.61%–40.55% and 3.21%–40.27%, respectively. In the 5-second long-term prediction, benefiting from the dynamic expert fusion mechanism, the ADE and FDE metrics exhibit a reduction in error of 3.19%–39.29% relative to comparative methods.The study indicates that the proposed conflict potential representation and the multi-expert adversarial training mechanism effectively capture the spatiotemporal heterogeneity of high-dynamic interactions in weaving sections significantly enhancing the robustness and interpretability of lane-changing predictions in complex scenarios.