Traffic State Prediction and Control Using a Koopman Operator Framework with Polynomial Observables

Abstract

In this paper, we develop a Koopman operator framework for modeling and control of traffic flow governed by the Lighthill--Whitham--Richards (LWR) conservation law. We show that the nonlinear dynamics induced by the Greenshields flow--density relation naturally motivate a dictionary of polynomial observables, while the controlled dynamics become bilinear in the lifted Koopman coordinates. A reduced-order Koopman model is identified from spatio-temporal traffic-density snapshots and used as the prediction model in a model predictive control (MPC) framework for ramp metering. The resulting Koopman-MPC is compared with two predictive control baselines that use, respectively, a fixed linearized LWR model and the nonlinear finite-volume LWR model. In a one-hour freeway experiment, the Koopman-based controller maintains the traffic density below the critical Greenshields value and closely reproduces the behavior of the nonlinear PDE-based MPC, with a maximum-density difference below 1%. The Koopman predictor also reduces the mean online model-evaluation cost by approximately one order of magnitude relative to the nonlinear PDE predictor. In contrast, the linearized LWR predictor is computationally cheaper but exhibits substantially larger prediction errors and yields a more conservative metering policy. These results demonstrate that the analytical structure of the nonlinear traffic conservation law can be exploited to construct an interpretable data-driven surrogate for efficient traffic prediction and control