Development of a Data-driven Evacuee Distribution Prediction Model by Integrating Graph Neural Networks and Data Assimilation

Abstract

In the disaster response immediately after a large earthquake, it is important to grasp the real-time situation of people's stay and evacuation. However, it is a challenge to shorten the computation time and improve the prediction accuracy. This paper aims to build a new data-driven evacuee distribution prediction model, which integrates graph neural network (GNN) and data assimilation, to efficiently and accurately predict the distribution of the number of evacuees per road link and time unit: (1)We train a GNN model to obtain a likelihood distribution corresponding to the probability of movement from a node to its connecting links; (2) We apply a particle filter, one of the data assimilation methods, to the probabilistic predictive evacuee distribution obtained from the likelihood distribution mentioned above. The sequential calculation by the model is expected to improve the speed and accuracy of evacuee distribution prediction.