Resilience offers a broad social-technical framework to deal with breakdown, response and recovery of transportation networks adapting to various disruptions. Although current research works model and simulate transportation resilience from different perspectives, the real-world resilience of urban road network is still unclear. In this paper, a novel end to end deep learning framework is proposed to estimate and predict the spatiotemporal patterns of transportation resilience under extreme weather events. Diffusion Graph Convolutional Recurrent Neural Network and a dynamic-capturing algorithm of transportation resilience jointly form the backbone of this framework. The presented framework can capture the spatiotemporal dependencies of urban road network and evaluate transportation resilience based on real-world big data, including on-demand ride services data provided by DiDi Chuxing and grid meteorological data. Results show that aggregate data of related precipitation events could be used for transportation resilience modeling under extreme weather events when facing sample imbalance problem due to limited historical disaster data. In terms of observed transportation resilience, transportation network demonstrates different characteristics between sparse network and dense network, as well as general precipitation events and extreme weather events. The response time is double or triple of the recovery time, and an elastic limit exists in the recovery process of network resilience. In terms of resilience prediction, the proposed model outperforms competitors by incorporating topological information and has better predictions of the system performance degradation than other resilience indices. The above results could assist researchers and policy makers clearly understand the real-world resilience of urban road networks in both theory and practice, and take effective responses under emergent disruptive events.

抵御能力提供了广泛的社会技术框架，以应对适应各种破坏的交通网络的故障，响应和恢复。尽管当前的研究工作从不同的角度对交通运输弹性进行建模和模拟，但城市道路网络的真实弹性仍不清楚。在本文中，提出了一种新颖的端到端深度学习框架来估计和预测极端天气事件下运输弹性的时空模式。扩散图卷积递归神经网络和运输弹性动态捕获算法共同构成了该框架的骨干。提出的框架可以捕获城市道路网络的时空依赖关系，并根据现实世界的大数据评估交通运输的弹性，包括DiDi Chuxing提供的按需乘车服务数据和网格气象数据。结果表明，由于历史灾害数据有限，当面临样本不平衡问题时，有关降水事件的汇总数据可用于极端天气事件下的运输复原力建模。就观测到的运输弹性而言，运输网络表现出稀疏网络和密集网络之间的不同特征，以及一般的降水事件和极端天气事件。响应时间是恢复时间的两倍或三倍，并且网络弹性的恢复过程中存在弹性极限。在弹性预测方面，与其他弹性指标相比，该模型通过合并拓扑信息胜过竞争对手，并且对系统性能下降的预测更好。以上结果可以帮助研究人员和政策制定者在理论和实践上清楚地了解城市道路网络的真实应变能力，并在突发的破坏性事件下采取有效的应对措施。