Reducing Detailed Vehicle Energy Dynamics to Physics-Like Models

Nour Khoudari, Sulaiman Almatrudi, Rabie Ramadan, Joy Carpio, Mengsha Yao, Kenneth Butts, Alexandre M. Bayen, Jonathan W. Lee, Benjamin SeiboldarXiv 2023 · DOI · arXiv

Abstract

Quantifying vehicle energy consumption at scale requires a balance between simulation fidelity and computational efficiency. High-fidelity models, while accurate, are often computationally expensive and contain discontinuities that hinder optimization. This paper presents a systematic model reduction pipeline that transforms complex vehicle simulations into smooth, "physics-like" polynomial formulas. By utilizing a Virtual Chassis Dynamometer (VCD) to extract stateless behavior, the researchers developed models that maintain industry-standard accuracy while enabling real-time traffic flow optimization and large-scale energy estimation.

Introduction

In the fields of traffic modeling and automated vehicle control, the ability to accurately predict fuel consumption is critical. However, researchers face a significant technical hurdle: high-fidelity simulation tools like Autonomie provide detailed, validated data but operate as "black boxes." These models are computationally intensive and often incorporate discrete logic, such as gear shifting, which introduces discontinuities or "jumps" in data.

Such discontinuities pose a major challenge for control engineers. Optimization algorithms typically require smooth, differentiable functions; mathematical jumps can cause these algorithms to become trapped in local minima, preventing the discovery of energy-efficient driving strategies. Furthermore, the computational overhead of high-fidelity models makes them unsuitable for simulating the energy impact of thousands of vehicles in a city-wide traffic model. This research bridges the gap by creating simplified mathematical expressions that remain faithful to real-world vehicle dynamics.

Methodology

The authors propose a two-stage reduction pipeline designed to compress high-fidelity data into a usable, optimization-friendly format.

1. The Semi-Principled Model

The first stage involves a "Virtual Chassis Dynamometer" (VCD). This tool probes the original Autonomie models by running them through a structured grid of speeds ($v$), accelerations ($a$), and road grades ($\theta$). This process extracts the underlying "stateless" behavior of the vehicle, effectively removing the "memory" or hysteresis associated with complex simulations (e.g., specific gear-shift timing). The result is a model where fuel consumption depends solely on the current state $(v, a, \theta)$.

2. The Simplified Physics-Like Model

In the second stage, the extracted data is compressed into smooth polynomial functions. Unlike "black-box" machine learning approaches, these functions are structured based on physical laws. The models include specific terms for:

  • Aerodynamic Drag: Proportional to $v^3$.
  • Rolling Friction: Proportional to $v$.
  • Potential Energy: Accounting for road grade.

Crucially, the researchers enforce "convexity-like" properties during this stage. By smoothing out the jumps caused by gear changes, they ensure the model is "optimization-friendly," allowing computers to calculate efficient paths without mathematical interference from discrete shift points.

Experimental Setup and Datasets

The study utilized Autonomie Rev 16SP7 as the ground truth for model generation. The researchers curated a portfolio of six representative vehicle classes to ensure the pipeline's versatility:

  • Light-Duty: Compact Sedan, Midsize Sedan, Midsize SUV, and Midsize Pickup.
  • Medium/Heavy-Duty: Class 4 Pickup-and-Delivery (PND) and Class 8 Tractor-Trailer.

Validation was conducted against standard regulatory drive cycles to test the models under realistic conditions. These included the EPA’s UDDS, HWFET, and US06 cycles, the UNECE’s WLTC, and the HHDDT variants for heavy-duty trucks.

Key Insights

Smoothing for Fleet Representation

A pivotal finding of this research is that removing gear-shift discontinuities actually improves the representation of vehicle fleets. While an individual vehicle experiences a sharp change in fuel consumption during a shift, an aggregate fleet of vehicles will shift at slightly different times. A smooth, continuous model effectively represents the mean behavior of a traffic stream more accurately than a jagged, single-vehicle model.

The Feasibility Boundary

The pipeline explicitly models the "feasible region"—the physical limits of a vehicle's powertrain. By defining the maximum possible acceleration and torque for a given speed, the model prevents optimization algorithms from suggesting maneuvers that are physically impossible, such as a heavy-duty truck accelerating at the rate of a passenger car.

Results and Discussion

The reduced models demonstrated high fidelity to the original simulations. On standard flat-road drive cycles, the simplified models achieved a percentage relative error of less than 4% across all vehicle classes.

Beyond accuracy, the primary advantage of these models is computational speed. Because the models are expressed as explicit mathematical formulas rather than iterative Simulink-based simulations, they are orders of magnitude faster to evaluate. While the models performed well on uphills, the researchers noted increased error margins on steep downhills due to the complexity of "fuel-cut" dynamics (where the engine stops consuming fuel during coasting). However, because absolute fuel consumption is extremely low in these scenarios, the impact on total energy estimation remains negligible.

Conclusion and Future Work

This research provides a robust framework for reducing complex vehicle dynamics into physics-informed, differentiable models. Future work will focus on several key areas:

  • Electric Vehicles (EVs): Extending the pipeline to Hybrid and Electric Vehicles, which requires new structures for regenerative braking and battery chemistry.
  • Manual Transmissions: Adapting the logic for driver-controlled shifting behaviors.
  • Real-World Validation: Comparing these "models of models" against physical chassis dynamometer experiments to further verify their real-world applicability.

Summary updated

BibTeX
@misc{khoudari2023reducing,
  author = {Nour Khoudari and Sulaiman Almatrudi and Rabie Ramadan and Joy Carpio and Mengsha Yao and Kenneth Butts and Alexandre M. Bayen and Jonathan W. Lee and Benjamin Seibold},
  title = {{Reducing Detailed Vehicle Energy Dynamics to Physics-Like Models}},
  year = {2023},
  eprint = {2310.06297},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.06297},
}
Figure