为准确、快速地预测高速公路事故多发路段，获得事故时空数据的特征样本，明确事故的时空演化规律及关联机制，根据时空热点分析结果鉴别事故多发点的位置和时空演变模式，构建了GA-XGBoost事故多发点预测组合模型。首先，依据样本数据分别构建年尺度、日尺度下的时空立方体，并进行热点分析，根据分析结果得到样本高速的事故多发点位置及其时空演化模式；经过比较分析和相关性检验，选取事故发生时间、里程、事件类型、处理时长、影响车道数、是否处于汇入口附近、是否节假日7项特征预测事故是否处于事故多发点。然后，分别使用CNN-LSTM、CNN-LSTM-ATT、随机森林、XGBoost模型4种算法对事故多发点进行预测，结果显示：相比其他3种，XGBoost模型的预测准确率最高。接着，采用遗传算法对XGBoost模型进行优化，构建了GA-XGBoost组合模型，使预测准确率提高0.06，F1分数提高0.07，精确率提高0.08。这表明，相比既有算法，GA-XGBoost模型能够较准确地预测出路段是否处于事故多发点，明确事故多发点事故的时空特征。最后，通过SHAP分析对预测结果进行解释，发现处于汇入口附近、事件类型为侧翻、故障、处于国庆假期和影响车道数为2的样本处于事故多发点的概率相比不处于汇入口附近和其他事故类型更大。据此，在交通安全和应急管理中可采取预防性措施，提升交通管理的效率和应急响应能力，营造安全、高效的交通环境。

In order to accurately and quickly predict the accident-prone sections of highways, obtain the characteristic samples of accident spatio-temporal data, clarify the spatio-temporal evolution patterns and correlation mechanisms of accidents, and identify the location and the spatio-temporal evolution patterns of accident-prone points based on spatio-temporal hotspot analysis results, this paper constructed GA-XGBoost accident-prone point prediction model. Firstly, based on the sample data, the spatio-temporal cube was constructed under annual scale and daily scale respectively, and hotspot analysis was carried out. According to the spatio-temporal hotspot analysis results, the locations of accident-prone point of the sample highway and their spatio-temporal evolution pattern were obtained. After comparative analysis and relevance test, seven characteristics were selected to predict whether the accident was located in the accident-prone location, including accident occurrence time, mileage, event type, processing time, number of affected lanes, whether it was in the vicinity of the confluence, and whether it was a holiday. Then, four algorithms, including CNN-LSTM, CNN-LSTM-ATT, Random Forest, and XGBoost model, were used to predict the accident-prone points respectively, and the results showed that the XGBoost model had the highest prediction accuracy compared to the other three algorithms. Subsequently, the XGBoost model was optimized with GA(Genetic Algorithm), and a GA-XGBoost combination model was constructed, which improved the prediction accuracy by 0.06 and F1 score by 0.07, and the precision by 0.08. This indicated that compared to existing algorithms, the GA-XGBoost model could more accurately predict whether a road section was located in an accident-prone area, and clarify the spatio-temporal feature of accidents in accident-prone areas.Finally, the prediction results were interpreted by SHAP value analysis, and it was found that the samples located near the confluence, with incident types of rollover, breakdown, during National Day holidays and with 2 affected lanes were more likely to be at an accident-prone point compared to those not located near the confluence and with other accident types. Based on this, preventive measures could be taken in traffic safety and emergency management to improve the efficiency and emergency response capabilities of traffic management, so as to creat a safe and efficient traffic environment.