Real-time Estimation Method of time-varying Activity Parameters for Traffic Demand Prediction under disaster restoration period

Abstract

This research developed a fast and robust parameter estimation algorithm for activity simulators. The algorithm enables the real-time parameter estimation which is necessary especially in unstable situations like a recovery-period from a disaster when activity parameters change considerably. The developed algorithm is based on surrogate-based optimization, which can save much time for optimization with the help of machine learning. We proposed and validated three expansions of the existing algorithm. The study applied our proposed algorithm to estimate the parameters of a multi-agent activity simulator using the data collected during the 2016 Kumamoto Earthquake. The numerical example showed that the proposed algorithm could estimate well-calibrated parameters within ten minutes after the observed data arrived, while the conventional GA method takes nearly one day to estimate parameters. Additionally, the parameters updated every hour improved the accuracy of the prediction of the populations in 1 km square zones compared with the prediction by less updated parameters.