An integrated multi-dimensional framework for real-time traffic risk prediction and its simulation-based verification in path planning

An integrated real-time traffic risk prediction framework is developed to enhance prediction reliability and effectiveness by addressing model input, modeling, and output comprehensively. The framework was applied to a case study using real vehicle trajectory data from the HighD dataset to characterize traffic state variables and categorize associated risks into four classes: no risk (NR), low risk (LR), medium risk (MR), and high risk (HR). Connected vehicle (CV) features were extracted under the assumption of 20% market penetration rate (MPR) and uniform distribution of CVs. Interaction terms among these variables were also incorporated. The optimal prediction model was selected from a pool of task-specific models using a surrogate model (SM), with variable pre-screening via the smoothly clipped absolute deviation (SCAD) method for models lacking inherent variable selection. Cost-sensitive learning and dynamic thresholds were integrated to account for varying prediction error costs and to enhance multi-class performance, with their corresponding parameters being optimized using a genetic algorithm (GA). Simulations confirmed the framework's practical applicability in improving traffic safety and efficiency. The results demonstrate that the optimal model varies by lane and segment, underscoring the need for the proposed approach in selecting the most suitable prediction model. The identified risk prediction models meet real-time prediction requirements in terms of both precision and computational speed. Furthermore, path replanning based on these predictions can improve both safety and operational efficiency. Findings of this study are expected to contribute to enhancing real-time risk prediction and advancing proactive safety management in the future.