Collaborative generative adversarial networks for fusing household travel survey and smart card data to generate heterogeneous activity schedules in urban digital twins

Abstract

Conventional activity-based models (ABMs) relying on household travel survey (HTS) data suffer from low spatiotemporal heterogeneity and outdated information due to low sampling rates and infrequent data collection of HTS. Transit smart card (SC) data can aid with continuously collected spatiotemporally heterogeneous mobility patterns of transit users. However, its integration with HTS in activity schedule generation is challenging due to differences in spatial resolutions, missing activity purpose, and lack of non-transit trips. To tackle these issues, we propose a novel data fusion method based on a deep generative model: collaborative generative adversarial networks (CollaGAN), which leverage the complementary strengths of HTS and SC data. CollaGAN generates activity schedules by harmonizing the differences in spatiotemporal heterogeneity and information between the two datasets in the latent space. The novel architecture of CollaGAN involves a discriminator for each HTS and SC data to simultaneously preserve the comprehensive information in the small-scale HTS data and the heterogeneous patterns in the large-scale SC data. We also devise novel boundary-based and domain-specific regularization to maintain the feasibility of the generated activity schedules. Using HTS and SC data in Seoul, a semi-synthetic simulation study quantitatively demonstrates multi-fold enhancements in the heterogeneity of the activity schedules from CollaGAN compared to those from single-source models, with a pronounced increase in heterogeneity of spatial attributes. A case study qualitatively shows that mobility patterns overlooked by HTS are captured through fused joint probability distributions, generating heterogeneous mobility patterns of non-transit users that existing data fusion methods fail to capture. By simulating heterogeneous activity schedules, the model provides more precise and policy-relevant insights into urban mobility, ultimately enhancing the overall accuracy of the urban digital twin by preventing error propagation to other interconnected energy and land use systems.