A BEHAVIORALLY-INTEGRATED INDIVIDUAL-LEVEL STATE-TRANSITION MODEL THAT CAN PREDICT RAPID CHANGES IN EVACUATION DEMAND DAYS EARLIER.

Timely and accurate forecast of evacuation demand is key for emergency responders to plan and organize effective evacuation efforts during a disaster. The state-of-the-art in evacuation demand forecasting includes behavior-based models and dynamic flow-based models. Both lines of work have critical limitations: behavioral models are static, meaning that they cannot adjust model predictions by utilizing field observation in real time as the disaster is unfolding; and the flow-based models often have relatively short prediction windows ranging from minutes to hours. Consequently, both types of models fall short of being able to predict sudden changes (e.g., a surge or abrupt drop) of evacuation demand in advance. This paper develops a behaviorally-integrated individual-level state-transition model for online predictions of evacuation demand. On a daily basis, the model takes in observed evacuation data and updates its forecasted evacuation demand for the future. An individual-level survival model formulation is innovatively devised for the state-transition model to account for history-dependent transition probabilities and allow individual heterogeneity. A Bayesian updating approach is employed to assimilate observed evacuation data in real time. To enable a longer-term perspective on how evacuation demand may evolve over time so that rapid surges or drops in demand can be predicted days in advance, the model integrates insights from existing behavioral curves (either from past disasters or simply expert opinions). Using a likelihood-based approach, the state-transition model integrates the future trends of evacuation demand informed by the behavioral curve when updating its forecasts. Theoretical proof is also provided showing that the likelihood function guarantees a unique global solution to the state-transition model. The proposed model is tested in six scenarios using mobile app-based data for Hurricane Harvey that hit the US in 2017. The results demonstrate overall robustness of the proposed model: in all six scenarios, the model is able to predict accurately the occurrence of the rapid surges or drops in evacuation demand at least two days ahead. The current study contributes to the field of evacuation modeling by integrating the two lines of work (behavior-based and flow-based models) using mobile app-based data.