Abstract
Lane-change risk prediction models are critical for enhancing driving safety and reducing traffic accidents. However, existing studies have not sufficiently considered the effects of traffic context and driving styles on lane change risk, limiting their adaptability in complex traffic environments. This study proposes a lane-change risk prediction model that integrates traffic context and driving styles. The traffic context model identifies different lane changing types using Attention-LSTM by analyzing traffic levels and vehicle types; The LSTM-Stacked Denoising Autoencoder (LSTM-SDAE) identifies and further characterizes distinct driving styles. The Light Gradient Boosting Machine (LGBM) algorithm predicts lane change risk using the Lane Change Risk Index derived from driving styles (LCRI_DS), and applies Shapley additive explanations (SHAP) to analyze feature importance across different LC types. The results show that incorporating traffic context improves the model's prediction accuracy from 88.79 to 95.05%. The LGBM algorithm, combined with traffic context and driving style features, outperforms other algorithms. Feature importance analysis reveals that the risk of left lane changes is primarily influenced by the vehicle's lateral velocity, the relative longitudinal velocity to the preceding vehicle in the target lane, and the driving style, whereas the risk of right lane changes is more strongly associated with the relative longitudinal velocities to following vehicles in both the current and target lanes, along with the driving style. This work provides valuable insights for improving driving safety and developing decision-support systems in complex traffic environments.