A Physics-Informed Deep Learning Paradigm for Traffic State Estimation and Fundamental Diagram Discovery

Rongye Shi, Zhaobin Mo, Kuang Huang, Xuan Di, Qiang Du

TL;DR

This paper introduces PIDL+FDL, a novel hybrid deep learning approach that combines physics-informed neural networks with a learned fundamental diagram (FD) to improve traffic state estimation (TSE). By integrating a learned FD, the method simultaneously estimates traffic states, identifies model parameters, and discovers the underlying FD relation, outperforming existing methods in accuracy and data efficiency.

What Problem Does This Paper Address?

Traffic state estimation (TSE) is crucial for understanding and managing traffic flow. Traditional approaches fall into two camps: model-driven methods that rely on physical traffic models but can be sensitive to model inaccuracies and limited data, and data-driven methods (like machine learning) that require vast amounts of data to learn complex patterns without explicit physical understanding. Physics-Informed Deep Learning (PIDL) emerged as a hybrid solution, but existing PIDL methods can struggle with complex traffic models and often require explicit FD equations. This paper addresses the challenge of developing a more robust and flexible PIDL framework that can learn the fundamental diagram (FD) – the relationship between traffic density and flow/speed – directly from data, without needing pre-defined FD equations, thus improving TSE accuracy and data efficiency.

Methods

The core innovation is the Physics-Informed Deep Learning with a Fundamental Diagram Learner (PIDL+FDL). This framework builds upon the PIDL concept by integrating a Fundamental Diagram Learner (FDL), implemented as a neural network, into the model-driven component.

Here's a breakdown:

  • Physics-Informed Neural Network (PINN): This component enforces the underlying traffic flow physics (e.g., conservation laws) by penalizing deviations from these physical laws in the loss function.
  • Fundamental Diagram Learner (FDL): This is a novel addition. Instead of hard-coding an FD equation, a separate neural network learns the FD relation (e.g., density-flow mapping) directly from the traffic state variables estimated by the main neural network. This FDL is trained implicitly through the physical regularization provided by the PINN.
  • Simultaneous Learning: The PIDL+FDL framework trains the traffic state estimation network, the FDL, and any unknown model parameters (like diffusion coefficients or relaxation times) concurrently.
  • Regularization: The paper also introduces a regularization term to reshape learned FD curves, particularly to address potential abnormal shapes in data-sparse regions.

Datasets

  • Simulated Data: The authors generated ground truth data for two traffic flow models:
  • Greenshields-based LWR (Linear-Wong-Richards) model: A first-order traffic flow model.
  • Greenshields-based ARZ (Aw-Rascle-Zhang) model: A second-order traffic flow model.
  • These simulations were used to test the method's ability to estimate traffic states, model parameters, and the FD.
  • Real-World Data:
    • NGSIM (Next Generation SIMulation) dataset: This dataset, collected from US Highway 101, provides real-world vehicle trajectories. The authors processed this data to obtain spatio-temporal fields of traffic density and velocity, using loop detector data for training and evaluation.

Key Insights and Ideas

  • Learning the FD is Key: The paper strongly suggests that learning the fundamental diagram directly, rather than relying on pre-defined equations, significantly enhances the flexibility and accuracy of physics-informed traffic models.
  • Hybridization Pays Off: The PIDL+FDL approach demonstrates that combining physics-based constraints with data-driven learning (via the FDL) is more effective than purely model-driven or data-driven methods, especially with limited data.
  • Model Complexity vs. Trainability: The study highlights a crucial trade-off: while more sophisticated traffic models (like ARZ) are theoretically richer, they can be harder to train within a PIDL framework. The simpler LWR model, when combined with the FDL, surprisingly outperformed the ARZ-PIDL in some real-world scenarios due to better trainability.
  • FD Discovery from Indirect Evidence: The FDL learns the FD relation solely from the physical consistency enforced by the PINN and the observed traffic states, without direct measurements of flow or velocity being used to train the FD itself. This is a powerful way to infer hidden relationships.

Key Results

  • Improved Accuracy: PIDL+FDL consistently achieved lower estimation errors (e.g., L² error for density) compared to various baselines (EKF, LSTM, pure NN, PIDL without FDL) on both simulated and real-world NGSIM data. For instance, on the LWR model, the L² estimation error dropped from 0.6021 with 2 loops to 0.01287 with 4 loops.
  • Data Efficiency: The method performed well even with a small number of loop detectors (e.g., 3-4 loops), demonstrating its ability to extract meaningful information from limited observations.
  • Successful FD Reconstruction: The learned FD curves from PIDL+FDL closely matched the actual density-flow data points from the NGSIM dataset, especially when the reshaping regularization was applied.
  • Model Parameter Identification: The framework successfully identified unknown model parameters (e.g., diffusion coefficient $\epsilon$ for LWR, relaxation time $\tau$ for ARZ) with high accuracy, often converging to values very close to the true parameters. For the LWR model, $\epsilon^*$ was estimated to be 0.00495 when the true value was 0.005.

Limitations and Future Work

  • Hyperparameter Tuning: The authors acknowledge that hyperparameter tuning remains a challenge for deep learning methods, including PIDL+FDL. The model requires careful tuning for different spatial/temporal resolutions and data characteristics, making it complex for real-world deployment.
  • Sensitivity to Noisy Data: PIDL methods can be sensitive to noisy data. Further investigation is needed on how to robustly handle high noise and corruption in real traffic data.
  • Expanding Surrogate Components: The paper suggests exploring the integration of ML surrogates for other unobserved traffic quantities (beyond FD) within the traffic flow models to further enhance performance.

Summary updated

Abstract

Traffic state estimation (TSE) bifurcates into two categories, model-driven and data-driven (e.g., machine learning, ML), while each suffers from either deficient physics or small data. To mitigate these limitations, recent studies introduced a hybrid paradigm, physics-informed deep learning (PIDL), which contains both model-driven and data-driven components. This paper contributes an improved version, called physics-informed deep learning with a fundamental diagram learner (PIDL+FDL), which integrates ML terms into the model-driven component to learn a functional form of a fundamental diagram (FD), i.e., a mapping from traffic density to flow or velocity. The proposed PIDL+FDL has the advantages of performing the TSE learning, model parameter identification, and FD estimation simultaneously. We demonstrate the use of PIDL+FDL to solve popular first-order and second-order traffic flow models and reconstruct the FD relation as well as model parameters that are outside the FD terms. We then evaluate the PIDL+FDL-based TSE using the Next Generation SIMulation (NGSIM) dataset. The experimental results show the superiority of the PIDL+FDL in terms of improved estimation accuracy and data efficiency over advanced baseline TSE methods, and additionally, the capacity to properly learn the unknown underlying FD relation.
BibTeX
@article{shi2022physics,
  author = {Rongye Shi and Zhaobin Mo and Kuang Huang and Xuan Di and Qiang Du},
  title = {{A Physics-Informed Deep Learning Paradigm for Traffic State Estimation and Fundamental Diagram Discovery}},
  journal = {IEEE Trans. Intell. Transp. Syst.},
  year = {2022},
  url = {https://openreview.net/forum?id=byxRMVxIZEL},
}
Figure