Affiliations: [a] Department of Civil and Environmental Engineering, University of Maryland, MD, USA | [b] Department of Mathematics, University of Maryland, MD, USA | [c] Joint Program in Survey Methodology, University of Maryland, MD, USA
Correspondence:
[*]
Corresponding author: Kartik Kaushik, Department of Civil and Environmental Engineering, University of Maryland, MD, USA. E-mail: kartikk@umd.edu.
Abstract: Short term traffic forecasting cannot be more important in the current world of cash strapped economies, placing ever increasing importance on managing existing facilities as opposed to building new infrastructure. The advent of autonomous vehicles further stresses the need for robust and fast frameworks to forecast traffic over the horizon of a typical trip length so that the best routing decision might be made. There is an extensive amount of research on this topic already. However, most of the techniques in literature do not scale well with data or the size of the network in terms of model complexity, computational power and time. Proposed in this paper is a flexible synthetic time series framework that aims to solve the complexity and scalability problems with most models in literature. The synthetic time series framework takes advantage of the repeatability of the traffic patterns such that real-time predictions can be quickly made. It is flexible enough to work with most models in literature, and extendable quite easily with additional parameters to make predictions more robust. Presented in this work are the Autoregressive Integrated Moving Average (ARIMA) models within the synthetic time series framework. It is shown that reasonably accurate predictions can be made by using just the basic structure of ARIMA without any auxiliary variables accounting for the upstream/downstream conditions, incidents or weather. With a robust model fitted within the synthetic framework, prediction errors can be further reduced, while ensuring scalable computation power. Predictions are performed online, where incoming data is fed to the fitted model as the independent variable, and predictions are obtained as the dependent variable.
Keywords: Traffic forecasting, ARIMA, network, probe data
