随着城市化推进，交通事故数量逐年上升，危害人类的生命安全并造成严重的经济损失。事故的发生受诸多因素影响，本研究利用多源时空数据来探究交通模式的内在机理以进行事故预测。

（1）针对当前公开交通事故数据集来源单一等问题，构建基于多源时空数据的交通事故数据集。以天气数据、流量信息和交通事故等数据为基础，通过数据获取、数据处理及数据表征等步骤，获取道路级别的萨克门托数据集和网格区域级别的纽约和芝加哥数据集。

（2）在交通事故预测场景中，针对当前道路级别预测方法存在忽视道路属性结构及路段扩散模式等问题，提出一种基于扩散性时空关联感知的交通事故预测模型DSTCPnet。首先，异质性时空感知模块通过表征路段属性及POI信息构建图结构，并结合MS-Block模块探索不同地点的空间异质性，随后T-Block模块学习不同时间步的时间周期性；其次，扩散性时空感知模块通过GAT-D模块，利用路段有向图以捕捉不同地点的空间扩散性，随后GRU-T模块引入神经微分方程，学习车流量的时间波动性；最后，融合异质性时空感知模块和扩散性时空感知模块以生成事故预测结果。在萨克门托数据集的验证实验中，与基线模型相比，DSTCPnet模型在Recall和RMSE指标上分别提高了4.45%和5.10%。

（3）在交通事故风险等级预测场景中，针对当前研究忽视数据因果关联及未能完全获取数据的高阶关系等问题，提出基于演化性时空因果感知的事故风险等级预测模型ARLPnet。首先，在相似性时空感知模块中，利用Spatial Block模块学习不同相似区域的空间关联性，随后使用Temporal Block模块学习时间周期性，以理解交通流量在不同时间的时空分布特征；其次，在演化性因果感知模块中，利用DA-Block模块自适应学习区域网格数据之间的空间关联性，通过网格数据构建超图，结合Hy-Block模块获取数据的高阶关联关系；接着CP-Block模块采用因果卷积捕获时间序列数据中的因果关联模式）；最后，融合相似性时空感知模块和演化性因果感知模块以预测事故风险等级。在纽约和芝加哥数据集的实验结果显示，与基线模型相比，ARLPnet模型在Recall和RMSE指标上分别提高了9.06%和6.56%。

With the advancement of urbanization, the number of traffic accidents has increased year by year, endangering human life and causing serious economic losses. The occurrence of accidents is affected by many factors. This study uses multi-source spatiotemporal data to explore the internal mechanism of traffic patterns for accident prediction.

(1) In view of the problem that the current public traffic accident dataset has a single source, a traffic accident dataset based on multi-source information fusion is constructed. Based on weather data, traffic information, traffic accidents and other data, the road-level Sacramento dataset and the grid-level New York and Chicago datasets are obtained through data acquisition, data processing and data characterization.

(2) In the traffic accident prediction scenario, in view of the problems that the current road-level prediction method ignores the road attribute structure and the road section diffusion pattern, a traffic accident prediction model DSTCPnet based on diffuse spatiotemporal correlation perception is proposed. First, the heterogeneous spatiotemporal perception module constructs a graph structure by characterizing the road segment attributes and POI information, and combines with the MS-Block module to explore (the spatial heterogeneity of different locations), and then the T-Block module learns the temporal periodicity of different time steps; secondly, the diffuse spatiotemporal perception module uses the road segment directed graph through the GAT-D module to capture the spatial diffusion of different locations, and then the GRU-T module introduces neural differential equations to learn the temporal volatility of traffic flow; finally, the heterogeneous spatiotemporal perception module and the diffuse spatiotemporal perception module are integrated to generate accident prediction results. In the verification experiment of the Sacramento dataset, compared with the baseline model, the DSTCPnet model improves the Recall and RMSE indicators by 4.45% and 5.10% respectively.

(3) In the scenario of traffic accident risk level prediction, in view of the problems that current research ignores data causal association and fails to fully obtain the high-order relationship of data, an accident risk level prediction model ARLPnet based on evolutionary spatiotemporal causal perception is proposed. First, in the similarity spatiotemporal perception module, the Spatial Block module is used to learn the spatial correlation of different similar regions, and then the Temporal Block module is used to learn the temporal periodicity to understand the spatiotemporal distribution characteristics of traffic flow at different times; secondly, in the evolutionary causal perception module, the DA-Block module is used to adaptively learn the spatial correlation between regional grid data, and a hypergraph is constructed through grid data, and the high-order correlation relationship of data is obtained by combining the Hy-Block module; then the CP-Block module uses causal convolution to capture the causal correlation pattern in time series data); finally, the similarity spatiotemporal perception module and the evolutionary causal perception module are fused to predict the accident risk level. The experimental results on the New York and Chicago datasets show that compared with the baseline model, the ARLPnet model improves the Recall and RMSE indicators by 9.06% and 6.56%, respectively.