Validation of a Data-Driven Evacuee Distribution Prediction Method Integrating Graph Neural Networks and Data Assimilation

Abstract

Immediately after a major earthquake, it is important to understand how people stay and evacuate in real time to prevent secondary damage. In a previous paper [Takasaki 2021], we proposed a data-driven evacuee distribution prediction model that integrates a graph neural network (GNN) model called Gretel [Cordonnier 2019] and a particle filter, a data assimilation method. In this paper, based on the proposed method, we construct a model that can predict the distribution of evacuees considering the global trend of evacuation behavior by incorporating the status of road blockage due to fire spread and collapse of buildings as feature values that change over time. Additionally, we compared the model with the results of an agent-based simulation used to generate a pseudo-evacuee distribution. Furthermore, we verify the model's usefulness regarding computation time and prediction accuracy.