Learning-Based Travel Prediction in Urban Road Network Resilience Optimization

Urban mobility is a key part of routine operations in cities, but is increasingly at risk due to floods. To mitigate these risks, urban planning and disaster management agencies must anticipate the potential impacts of road network connectivity losses on travel flows between different locations in an urban area. However, detailed travel flow data are not widely available, necessitating the use of models for estimating origindestination travel volumes based on geospatial and socioeconomic features. This paper examines several travel demand prediction models in the context of their suitability for informing road network resilience planning. We first evaluate the capacity of these models to capture census-tract-level urban travel demand patterns. We then use the predicted travel flows as input to a budget-constrained optimization scheme for identifying which roads to reinforce against flooding in order to preserve connectivity. We find that the road upgrade prioritization performs equally well when provided the ground truth travel demand data and moderately accurate predictions from the learning-based models. Thus, models with moderately lower prediction accuracy but with other computational and practical advantages may be favored for specific decision processes.