Transfer Learning for Transportation Demand Resilience Pattern Prediction Using Floating Car Data

Abstract

Understanding the response of a transportation system to disruptive events is significant for evaluating the resilience of the system. However, data collection during such events is always challenging, and the data volume is insufficient for building a robust model. Transfer learning provides an effective solution to this problem. In this study, we propose a floating car data (FCD) driven transfer learning framework for predicting the resilience of target transportation systems to similar disruptive events to those that have ever occurred in the source systems. The core of the framework is an unsupervised pattern extractor that combines the k-Shape clustering and Bayes inference methods for extracting resilience patterns from the FCD collected in the source systems during the disruption period. The extracted patterns can then be used to assist in the prediction of the resilience of the target systems. We examine the effectiveness of the proposed framework by conducting a case study under the context of the COVID-19 pandemic, in which the source domain cities include Antwerp and Bangkok, and the target domain city is Barcelona. Results show that the extracted resilience patterns can improve the prediction performance of transfer learning neural networks with less pre-event information and limited data volume.