Abstract
This paper develops a boundary observer for the estimation of congested freeway traffic states based on the Aw-Rascle-Zhang (ARZ) partial differential equations (PDE) model. Traffic state estimation is critical for freeway management but is often limited by sparse sensor data. We propose a boundary observer design that estimates aggregated traffic states in a freeway segment using only boundary measurements of flow and velocity. By linearizing the $2 \times 2$ coupled nonlinear hyperbolic PDEs of the ARZ model and applying PDE backstepping methods, we construct an observer that guarantees exponential stability and finite-time convergence of the estimation error. Simulations validate the design, demonstrating robust performance even without knowledge of initial conditions.
Introduction
Effective freeway management strategies, such as ramp metering and variable speed limits, rely on accurate real-time knowledge of traffic density and velocity across entire road segments. However, practical constraints often limit data collection to specific boundary points. Traditional first-order models, like the Lighthill-Whitham-Richards (LWR) model, fail to capture the complex "stop-and-go" oscillations characteristic of congested traffic. This paper addresses these limitations by utilizing the second-order Aw-Rascle-Zhang (ARZ) model and a novel boundary observer design.
The Aw-Rascle-Zhang Model
The ARZ model represents macroscopic traffic dynamics through two coupled nonlinear hyperbolic PDEs for traffic density ($\rho$) and velocity ($v$). In congested regimes, disturbances travel both downstream with the vehicle flow and upstream as shockwaves. This "hetero-directional" nature makes state estimation particularly challenging. The model captures the relaxation of traffic toward an equilibrium state, allowing for the representation of non-equilibrium phenomena like stop-and-go waves.
Boundary Observer Design via PDE Backstepping
The core of our approach is a boundary observer designed using the PDE backstepping method. The process involves:
- Linearization: The nonlinear ARZ equations are linearized around steady-state traffic flow and velocity.
- Riemann Transformation: The system is transformed into Riemann coordinates to decouple the downstream and upstream wave dynamics.
- Backstepping and Output Injection: We design output injection gains based on the error between boundary measurements and model predictions. A Volterra transformation maps the error system into a stable target system, ensuring that estimation errors decay to zero in finite time.
- Nonlinear Implementation: These gains are applied to a copy of the original nonlinear ARZ model, creating a robust observer capable of tracking real-world traffic fluctuations.
Simulation and Results
The observer's performance was evaluated through numerical simulations of a 500-meter freeway segment. The simulation utilized a two-stage Lax-Wendroff scheme to solve the hyperbolic PDEs with second-order accuracy.
As shown in the profiles above, the ARZ model successfully generates the complex spatial-temporal patterns of congested traffic. The observer is tasked with reconstructing these hidden internal states using only the data available at the segment boundaries ($x=0$ and $x=L$).
Performance and Convergence
A key advantage of this design is its guarantee of finite-time convergence. In simulations, the observer demonstrated remarkable speed and accuracy.
The estimation error for both density and velocity dropped to less than 2% of the setpoint value within approximately 75 seconds (1.25 minutes). Furthermore, the observer proved robust to unknown initial conditions, successfully "pulling" the estimates toward the true state regardless of the starting guess.
Conclusion
This work presents the first boundary observer for the nonlinear ARZ traffic model. By leveraging PDE backstepping, we provide a mathematically rigorous solution for estimating congested traffic states from limited boundary data. Future research will focus on validating this observer using empirical traffic field data to assess its performance in the presence of sensor noise and unmodeled driver behaviors.
Summary updated