基于FHSPSO-RF的交通事故严重程度预测

王淑椒，胡学龙，唐磊

安徽理工大学矿业工程学院，安徽 淮南  232001

摘要：为提高交通事故严重程度的预测准确性，构建一种基于觅食生境选择粒子群优化（foraging habitat selection particle swarm optimization，FHSPSO）算法优化随机森林（random forest，RF）算法关键超参数的FHSPSO-RF模型，以美国华盛顿州西雅图市2022年1月至2023年2月的交通事故数据为基础，选取12项特征指标，采用合成少数类过采样技术，增加重伤和死亡事故的样本数，改善事故类别分布均衡性；将FHSPSO-RF模型与支持向量机（support vector machine，SVM）、K近邻（K-nearest neighbors，KNN）和逻辑回归（logistic regression，LR）等模型进行性能对比，并通过SHAP（Shapley additive explanations）分析法，解析各特征对交通事故严重程度的影响机制。结果表明：过采样后重伤、死亡事故的召回率显著增大，FHSPSO-RF模型的整体性能更均衡；FHSPSO-RF模型的准确率、精确率、召回率、F1分数均高于其他3个模型，对交通事故严重程度的预测效果最好；在所有交通事故类型中，受伤人数和车辆数量均为最关键的驱动因素，对交通事故严重程度具有显著正向影响，行人数量、高冲击碰撞类型（如正面撞击）与复杂道路环境（如交叉口、匝道）是重伤及死亡事故的关键风险组合；财产损失事故与是否碰撞路边停放车辆密切相关。FHSPSORF模型在交通事故严重程度预测中表现出良好的性能与可解释性，可为交通事故风险预测与防控决策提供依据。

关键词：交通事故；严重程度预测；FHSPSO；RF

Traffic accident severity prediction based on FHSPSO-RF

WANG Shujiao, HU Xuelong, TANG Lei

School of Mining Engineering, Anhui University of Science & Technology, Huainan 232001, China

Abstract: To enhance the accuracy of traffic accident severity prediction, an FHSPSO-RF model is developed, where foraging habitat selection particle swarm optimization (FHSPSO) is used to optimize key hyperparameters of a random forest (RF). Using traffic accident data from Seattle, Washington, from January 2022 to February 2023, twelve features are selected. The synthetic minority oversampling technique is applied to increase the samples of severe-injury and fatal crashes and improve class balance. The FHSPSO-RF model is compared with support vector machine (SVM), K-nearest neighbors (KNN), and logistic regression (LR). Shapley additive explanations (SHAP) are used to interpret how each feature influences severity. Results indicate that after oversampling, the recall of severe-injury and fatal crashes increases significantly, and the FHSPSO-RF model achieves more balanced overall performance. The model attains higher accuracy, precision, recall, and F1 score than the three benchmarks, yielding the best severity prediction. Across all crash types, the numbers of injuries and vehicles are the most influential drivers with significant positive effects on severity. The number of pedestrians, high-impact collision types such as head-on crashes, and complex road environments such as intersections and ramps form key risk combinations for severe-injury and fatal crashes. Property-damage-only crashes are closely related to whether a parked roadside vehicle is struck. The FHSPSO-RF model demonstrates strong predictive performance and interpretability, providing support for crash risk prediction and prevention-control decision making.

Keywords: traffic accident; severity prediction; FHSPSO; RF