Design of strategic evacuations with conventional and self-driving vehicles




Lee, Jooyong (Ph. D. in civil engineering)

Journal Title

Journal ISSN

Volume Title



This dissertation delivers efficient strategies for evacuating Houston and Galveston neighborhoods in the Texas Gulf Coast to various safe destinations. Effective network assignment and vehicle-operation strategies are pursued to maximize evacuation efficiency and improve regional resilience to various threats. Behavioral characteristics of evacuees and modeling techniques for traffic flows are investigated via large-scale modeling tools for complex regions. Intermodal evacuation (bringing carless and other vulnerable residents to evacuation buses) is also analyzed. In Chapter 3, this dissertation presents an efficient evacuation plan that minimizes all coastal residents’ total travel time and late-arrival costs (at 8 regional exit points or shelters), while reflecting dynamic road closures that mimic the 2017 Hurricane Harvey flooding events. With different penetration rates of privately-owned self-driving or “autonomous” vehicles (AVs), each household’s efficient departure time choices are estimated upstream, in an aggregated model using genetic algorithms. That model relies on a bi-level algorithm, where the upper level searches for departure time schedules, and the lower level evaluates such choices via an agent-based transportation simulation across the region’s detailed roadway network. Higher AV penetration rate results in more favorable evacuation conditions, thanks to greater evacuation order compliance and smaller inter-vehicle headways. Chapter 4 models coastal Houstonites’ evacuation decisions in a disaggregate fashion, reflecting each evacuee’s home location, surrounding traffic conditions, while relying on cumulative prospect theory (CPT) for households’ departure time and destination choices. Recognizing both panicked and patient evacuees’ preference functions, results illustrate how more patient evacuees deliver lower traffic congestion, arrive at more distant inland destinations, and stagger their departures to prioritize the most distant (from shelters) residents first. Similar departure-time and destination choice models based on multinomial logit (MNL) specifications were then used to compare evacuation conditions to those emerging under a CPT framework. Comparison of CPT and MNL model results highlight how the CPT assumptions can better reflect evacuee decision-making under uncertainty, which can lead to an analyst preference for CPT over MNL. Finally, in Chapter 5, shared autonomous vehicles (SAVs) are used to provide a first-mile connection to evacuation buses for those residing in zero-car or insufficient vehicles owned households. Various SAV fleet size scenarios, along with SAV seating, dynamic ride-sharing (DRS) strategies, and evacuee behavioral assumptions are simulated to evaluate this setup’s evacuation efficiency and performance. In terms of cost-effectiveness, one 5-seat SAV per 14 (carless) residents is recommended. Ride-sharing coordination between the SAVs and larger (37-seat) evacuation buses is a good option if evacuees are willing to wait for the SAVs to arrive in a DRS-coordinated way. Carless evacuees’ unwillingness to share rides in SAVs (for first-mile services to evacuation buses) had a more negative impact on total evacuation costs than those who make decisions in a more panicked way.


LCSH Subject Headings