Assessing Road Network Vulnerability and Resilience by Simulation-based Graphs and Machine Learning

Abstract: Understanding transportation network vulnerability and resilience is crucial for urban mobility and economic vitality. This is particularly important as natural hazards become more frequent due to environmental challenges. While graph representations are commonly used for such assessments, the choice of graph weights or metrics often lacks clear justification, which limits their reliability. This study presents a comprehensive approach to addressing this gap through a multi-step process.

Initially, we introduce a novel methodology integrating outputs from numerous disruption simulations in a national transport model with graph theory analysis. This involves a comparison of weight-metric combinations, reflecting four weighting schemes and nine graph metrics, against traffic simulation results. Our findings revealed that a weighting scheme incorporating travel time and volume-to-capacity ratio consistently outperformed others. When combined with this scheme, Total Node Strength, Average Shortest Path, and Average Closeness Centrality metrics showed the strongest correlation in the given sequence with simulation outcomes. Additionally, the analysis demonstrated the advantage of weighted graphs over unweighted ones in capturing network performance.

However, while effective, this dynamic weighting approach proved time-consuming, especially for large networks, as it required manual disruption modeling and extraction of simulation outputs for each scenario. To address this challenge, we examined the use of a static weighting approach. This method involved conducting a single traffic simulation of the base (undisrupted) scenario to create a weighted base graph, from which disruption scenarios were generated by removing links. This approach aimed to substantially reduce computational requirements by eliminating the need for repeated simulations. This static weighting approach yielded weak correlations between graph-based and simulation-based performance outputs. This outcome highlighted the critical importance of accounting for traffic rerouting in each disruption scenario, emphasizing the need to establish new edge weights based on redistributed traffic volumes and travel times for each event.

To streamline this process and avoid the impractical option of rerunning simulation models for each scenario, we propose developing a Machine Learning (ML) model. This model accounts for various aspects of disruptive events and establishes new traffic distributions efficiently. The ML model was trained and validated using dynamic analysis disruption scenarios data, considering variables such as initial volume, travel time, capacity, edge betweenness, and disruption-specific factors like distance to impacted links.

Implementing an ML approach would enable the analysis of random disruption scenarios, producing good estimates for network-wide traffic distributions. This presents a feasible and efficient option for conducting comprehensive network assessments across numerous disruption scenarios without relying on time-intensive simulation software. Moreover, it creates an easily integrable code for establishing weighted transportation networks/graphs, potentially facilitating integration with other critical infrastructure networks (e.g., water, power) to reflect their interdependencies and enable holistic resilience and vulnerability evaluations.

Future research should focus on validating the findings across diverse network types, sizes, and disruptive events to generalize the applicability of the proposed methods. The combination of dynamic weighting, graph theory, and ML presents a promising avenue for advancing the analysis of transportation network resilience in the face of environmental challenges.