Countrywide Origin-Destination Matrix Prediction and Its Application for COVID-19

ABSTRACT

Modeling and predicting human mobility are of great significance to various application scenarios such as intelligent transportation system, crowd management, and disaster response. In particular, in a severe pandemic situation like COVID-19, human movements among different regions are taken as the most important point for understanding and forecasting the epidemic spread in a country. Thus, in this study, we collect big human GPS trajectory data covering the total 47 prefectures of Japan and model the daily human movements between each pair of prefectures with time-series Origin-Destination (OD) matrix. Then, given the historical observations from past days, we predict the countrywide OD matrices for the future one or more weeks by proposing a novel deep learning model called Origin-Destination Convolutional Recurrent Network (ODCRN). It integrates the recurrent and 2-dimensional graph convolutional components to deal with the highly complex spatiotemporal dependencies in sequential OD matrices. Experiment results over the entire COVID-19 period demonstrate the superiority of our proposed methodology over existing OD prediction models. Last, we apply the predicted countrywide OD matrices to the SEIR model, one of the most classic and widely used epidemic simulation model, to forecast the COVID-19 infection numbers for the entire Japan. The simulation results also demonstrate the high reliability and applicability of our countrywide OD prediction model for a pandemic scenario like COVID-19. © 2021, Springer Nature Switzerland AG.