Prediction of Traffic Accident Injury Degree Based on the XC-RL-MLP model

Xiao Shengjie; Ma Sheqiang

doi:10.3969/j.issn.1004-3403.2025.06.011

Abstract：With the rapid increase in the number of motor vehicles,people are paying increasing attention to traffic safety issues.In order to further explore the factors affecting the traffic accident injury degree in Shenzhen,this pa-per takes traffic accident injury degree as the study object and constructs an XC-RL-MLP model for predictive mod-eling and analysis.To verify the validity of the model,the traffic accident data from Shenzhen between 2016 and 2020 were taken as the example.Firstly,a dataset containing 1975 accident records was obtained through prepro-cessing,covering 7 key elements such as pedestrians and vehicles,and oversampling technology was used to bal-ance the dataset to ensure the reliability of model training.Secondly,the XC-RL-MLP model was systematically trained and evaluated by methods such as cross-validation to ensure itsgood generalization ability and stability.Fi-nally,the SHAP value was used for visual analysis of model characteristics,revealling the importance and influence mechanism of each characteristic in predicting accident injury degree.The results showed that the model reached an accuracy rate of 90.38%,higher than that of other Stacking models and 0.51%better than XGBoost,the optimal sin-gle model,and its F1 score also superpassed that of other single models.The five characteristics with the greatest in-fluence on the accident injury degree are use of safety protection devices,motor vehicle state,transportation mode,on-site status and state of vehicles'turn signals.This study can provide certain reference for traffic managers.