Lane-changing risk significantly increases on icy and snowy surfaces, resulting in frequent accidents. However, current lane-changing risk prediction models fail to incorporate driving intention recognition, which limits their accuracy and practical applicability. Therefore, this study proposes a dynamic lane-changing risk prediction framework for icy and snowy surfaces that incorporates driving intention recognition. Driving simulation experiments were designed to replicate icy and snowy driving conditions, and the resulting simulation data were used for model training and testing. First, a Bidirectional Multi-layer Long Short-term Memory (Multi-BiLSTM) network was employed to recognize the drivers’ lane-changing intentions. Second, lane-changing risk was analyzed from both temporal and spatial dimensions, quantified using fault tree analysis and the Lane-changing Risk Index, and categorized using a k-means clustering algorithm. Finally, the Light Gradient Boosting Machine (LightGBM) algorithm was applied to predict lane-changing risks. Results indicate that the average duration of lane-changing intentions on icy and snowy surfaces was 6.12 s, a 36.3% increase compared to normal road surfaces. The Multi-BiLSTM model achieved recognition accuracies of 98.22%, 97.74%, and 96.31% for left lane-changing, right lane-changing, and lane-keeping, respectively. The LightGBM model achieved an overall accuracy of 97.49% in predicting lane-changing risk, outperforming other machine learning algorithms. These findings provide theoretical support for developing risk warning systems and control strategies for intelligent vehicles on icy and snowy surfaces.
Zhao et al. (2026) studied this question.