An operator-theoretic approach for traffic prediction
Author(s)
Ling, Esther P.
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
In this thesis, we develop a case for data-driven modeling of traffic flow at signalized intersections using an operator-theoretic framework. Traffic at signalized arterials is known to be problematic to model, due to significant short-term fluctuations and recurring unstable conditions. Moreover, the collection of large volumes of traffic data from traffic sensors begs the use of creating data-informed models. We highlight the Koopman Operator as a unifying framework and elaborate on existing data-driven algorithms that have been connected to it. We divide these algorithms into two broad classes: dynamic mode decomposition and the Arnoldi method. This thesis focuses mainly on dynamic mode decomposition. First, we demonstrate the performance of dynamic mode decomposition in learning dynamics of the highly oscillatory signalized traffic data. We show the rank-deficiency limitation, and how to overcome it using time-shifted observations. Additionally, we show that modeling signal phases as exogeneous control input performs better compared to solely using time-shifted observations. Second, we apply dynamic mode decomposition to the inverse problem of signal and phase timing recovery from vehicular flows. Third, we develop an algorithm for real-time instability detection that is able to distinguish between queue growth arising from a traffic accident and regular peak hours. Finally, we propose a method for simulating a modified signal phase for queue-breakdown mitigation, using dynamic mode decomposition with control.
Sponsor
Date
2018-04-25
Extent
Resource Type
Text
Resource Subtype
Thesis