Abstract
Modeling traffic flow requires balancing physical realism with computational feasibility. Traditional local models often fail to capture driver anticipation and the smoothing effects of speed adjustments. This paper presents a high-order Local Discontinuous Galerkin (LDG) scheme for a 1D non-local diffusive traffic flow model. By utilizing a perceived density framework, the method ensures physical consistency—specifically avoiding negative velocities—while leveraging high-order approximations to overcome the $O(n^2)$ computational complexity inherent in non-local simulations.
Introduction
Traffic flow modeling has historically relied on the Lighthill-Whitham-Richards (LWR) model, a first-order conservation law that assumes a driver's speed is determined solely by the local density at their exact position. However, real-world driving involves anticipation—drivers look ahead to adjust their speed—and diffusion, as speed changes are not instantaneous.
While incorporating non-local and diffusive terms improves model fidelity, it introduces significant numerical challenges. Standard diffusive models frequently produce non-physical artifacts, such as negative velocities. Furthermore, calculating non-local interactions requires evaluating convolution integrals, which scales quadratically ($O(n^2)$) with the number of grid points. This study addresses these issues by proposing a robust, high-order numerical framework designed for both accuracy and efficiency.
The Non-local Diffusive Model
The core of the proposed approach is a model where velocity is a function of "perceived density" ($\hat{\rho}$). Unlike standard density, perceived density accounts for the average density within a specific look-ahead window and the gradient of that density.
A critical theoretical contribution of this work is the proof (Theorem 2.1) that this perceived density $\hat{\rho}$ remains strictly within the bounds of $[0, 1]$. This mathematical guarantee ensures that the model never predicts negative velocities, a common failure point in previous diffusive traffic models that lacked such constraints.
Numerical Methodology: Local Discontinuous Galerkin (LDG)
To solve the governing equations, the authors employ a Local Discontinuous Galerkin (LDG) method. The LDG approach is particularly suited for this problem due to its flexibility in handling high-order approximations and its ability to manage the diffusive (second-order) terms by splitting them into a system of first-order equations.
Key Technical Components:
- High-Order Basis Functions: The solution within each grid cell is represented using Legendre polynomials defined on Chebyshev nodes. This allows for high-order spatial accuracy.
- Non-local Integration: The convolution terms, which represent the driver's look-ahead behavior, are computed using piecewise integrals across multiple intervals. This ensures the integration remains accurate even when the look-ahead window spans several grid cells.
- Temporal Discretization and Stability: The system is evolved in time using a 3rd-order Runge-Kutta method. To prevent artificial oscillations near sharp traffic shocks (such as the sudden onset of a traffic jam), a Generalized Slope Limiter (GSL) is applied.
Results and Performance
The method was evaluated using standard synthetic benchmarks, including rarefaction waves, forward/backward shockwaves, and smooth sine waves for convergence testing.
Accuracy and Convergence
The numerical experiments confirmed the theoretical convergence rates. For instance, using $p=3$ polynomials (fourth-order accuracy) resulted in a convergence rate of approximately 4.1. This implies that doubling the grid resolution reduces the error by a factor of 16, demonstrating the high precision of the LDG scheme.
Overcoming the $O(n^2)$ Bottleneck
The most significant finding regarding performance is the efficiency of high-order methods in the context of non-local interactions. Because non-local terms require $O(n^2)$ operations, the computational cost grows rapidly as the grid is refined. However, the authors demonstrate that a high-order simulation (e.g., $p=3$) requires significantly fewer grid points to achieve a target accuracy than a low-order simulation ($p=1$).
In practice, the high-order method was both faster and more accurate than its low-order counterparts. This is vital because, as the number of grid partitions ($n$) increases, the non-local convolution begins to dominate the CPU time—accounting for over 80% of the total computation at $n=320$.
Conclusion
This research provides a robust framework for simulating non-local diffusive traffic flow. By combining a physically consistent perceived density model with a high-order LDG scheme, the authors successfully mitigate the computational burden of non-local convolutions while maintaining high accuracy and physical reliability.
Future work will need to address the scalability of these methods to 2D environments, where the complexity could reach $O(n^4)$. Potential avenues for optimization include the use of GPU acceleration or Fast Multipole Methods (FMM) to further reduce the burden of the convolution calculations.
Summary updated