在智能交通服务中，准确估计行程时间是确定出行时间、路径规划、车辆资源配置调度、交通管理决策的重要参考。轨迹数据具有成本低、时覆盖率高、易获取等优点，蕴含了行进速度、交通路况、出行需求、驾驶经验等丰富信息。针对如何充分地挖掘路网和轨迹数据并有效利用先验信息、捕捉城市交通系统的时空特性、以及顾及众多行程时间影响因素等问题，获得更高精度的行程时间估计，本研究设计了一种集成轨迹和路网的时空表征模型，支持路线估时。该模型运用图注意力网络捕获路网的空间结构信息，并将其融入到路段表征中，随后采用序列自注意力网络提取路径内部各路段间的时间与空间相互依赖关系，最后拼接长短时记忆网络提取的轨迹特征一起估计行程时间。此外，模型训练时进行多任务学习，同时估计全局路径和局部路段的行程时间，减小局部路段时间误差以优化整体性能。在武汉市大规模轨迹数据上的实验表明，该模型在各项性能指标上超越了其他基线模型，且模型中各模块对估计准确性均有所贡献。

In intelligent transport services, accurate estimation of travel time is an important reference for determining travel time, path planning, vehicle resource allocation scheduling, and traffic management decision-making. Trajectory data has the advantages of low cost, high time coverage, easy access, containing rich information such as travelling speed, traffic conditions, travelling demand, driving experience, etc. This study designs a model with a multilevel architecture to address the issues of how to adequately mine road network and trajectory data and effectively use the a priori information, capture the spatio-temporal characteristics of the urban transport system, and take into account the many factors affecting the travel time to obtain a more accurate travel time estimation. The model employs a graph attention network to capture spatial structural information of the road network and incorporate it into the road segment representation. Subsequently, self-attention networks are used to extract the temporal and spatial interdependencies between road segments within the path. The obtained dynamic features are finally concatenated with the trajectory features extracted by the long and short-term memory network to estimate the travel time. In addition, multi-task learning is performed during model training to simultaneously estimate the travel times of global paths and local road segments, and reduce the local road segment time error to optimise the overall performance. Experiments on large-scale trajectory data in Wuhan city show that the model outperforms other baseline models in all performance metrics, and each module in the model contributes to the estimation accuracy.