Applicability of Long Short-Term Memory Traffic Volume Imputation Model to Drive Connected Corridor Simulation

Abstract

For effective implementation of connected corridor applications, it is imperative to study the characteristics of the high-resolution connected corridor data streams leveraged in smart city applications. In a previous effort, a smart city application – real-time corridor data-driven traffic simulation model, i.e. Digital Twin – is developed. Investigation of the corridor volume data streams revealed the presence of data gaps. To address the data gaps, deep Long Short-Term Memory Recurrent Neural Network univariate and multivariate volume imputation models are developed. In this paper, the impact of the developed model imputations on the digital twin generated travel times are investigated. Simulation runs are conducted for two scenarios – typical and atypical (holiday) traffic, for three volume input cases: base data (original volumes), univariate model imputations, and multivariate model imputations. Results indicated that: 1) the travel times generated using multivariate imputations are closer to the travel times generated using base data, 2) the impact of imputations on travel times is more focused on congested routes, and 3) the impact on travel time is minimal despite input volume overestimation on routes that have capacity to accommodate higher volumes. These finding demonstrate the need to prioritize data streams based on the given application and underlying corridor conditions.