Prediction of Vehicle Lane‐Changing Trajectories in Highway Merging Areas Based on Physics‐Enhanced Residual Learning

Abstract

To improve lane‐changing efficiency and reduce safety risks for ramp vehicles in highway merging areas, this paper presents a method for predicting vehicle trajectories in these types of scenarios, and it is based on physics‐enhanced residual learning. Focusing on ramp lanes and adjacent mainline lanes, the model considers the influence of both the current and target lanes on the vehicle’s velocity during lane‐changing maneuvers. A hybrid prediction model is constructed by integrating a physics‐based model with a data‐driven approach. Specifically, an improved speed prediction model based on the Gipps general collision avoidance algorithm is introduced to calculate vehicle speed variations during lane‐changing maneuvers, and its parameters are calibrated using a genetic algorithm. The next trajectory point of the vehicle is predicted, and the corresponding residual is computed using the calibrated physical model. A long short‐term memory network is constructed to learn and predict the residuals. The final trajectory prediction is obtained by combining the physical model’s output with the predicted residuals. The experimental results based on real‐world traffic data show that the approach introduced in this study outperforms traditional neural network models significantly in terms of both accuracy and stability. The model achieves a higher determination coefficient and notably reduces both overall and longitudinal prediction errors. Additionally, ablation studies confirm that incorporating a Gipps‐based residual learning mechanism into the data‐driven model significantly enhances prediction performance, thereby validating the effectiveness of integrating physical information with residual learning. The proposed trajectory prediction model offers a novel and effective solution for improving trajectory prediction accuracy for ramp vehicles in highway merging areas.