Mitigating Metropolitan Carbon Emissions with Dynamic Eco-driving at Scale

Vindula Jayawardana, Baptiste Freydt, Ao Qu, Cameron Hickert, Edgar Sanchez, Catherine Tang, Mark Taylor, Blaine Leonard, Cathy Wu

TL;DR

Researchers have conducted the first metropolitan-scale assessment of "dynamic eco-driving," simulating one million traffic scenarios across three major US cities. They found that programming semi-autonomous vehicles to smooth out stop-and-go traffic at intersections can reduce city-wide intersection CO2 emissions by 11-22% without sacrificing safety or traffic throughput.

What Problem Does This Paper Address?

Decarbonizing the transportation sector is a monumental challenge. A significant portion of emissions—roughly 14.6% of US land transport CO2—stems from the unproductive idling and frequent acceleration characteristic of signalized intersections. While "eco-driving" (optimizing vehicle speed profiles to save fuel) has been studied for decades, research has been stuck in a "small-scale" trap, focusing on single intersections or simplified routes.

The authors tackle the "curse of dimensionality" inherent in large-scale traffic analysis. Evaluating a city requires accounting for at least 33 influential factors, including intersection geometry, weather, vehicle age, and varying adoption rates of automation. This paper moves beyond the technical question of how to eco-drive to the socio-technical question of whether it can move the needle on climate change at scale.

Methods

The authors developed a comprehensive pipeline to model and control traffic across 6,011 intersections in Atlanta, Los Angeles, and San Francisco.

  • Scenario Modeling: They used a data-driven approach to prune a massive scenario space into one million representative simulations using the SUMO microscopic traffic simulator.
  • Multi-Task Deep Reinforcement Learning (MT-DRL): Instead of traditional model-based methods that struggle with complex vehicle dynamics, they trained generalizable policies using MT-DRL and Proximal Policy Optimization (PPO).
  • Network Decomposition: To make optimization tractable, they optimized intersections individually but applied strict constraints on throughput, queue length, and waiting times to ensure that local improvements didn't cause network-level overflows.
  • Surrogate Emission Modeling: Since the industry-standard MOVES model is too slow for millions of scenarios, they trained fast neural network surrogate models to estimate instantaneous emissions based on velocity, acceleration, and environmental factors.
  • Zero-Shot Transfer: Policies were trained on a synthetic distribution of scenarios and then evaluated across the million real-world-inspired scenarios.
A comparison of eco-driving control methodologies, highlighting the advantages of model-free RL in handling scale and complex assumptions.

Datasets

The study fused several high-fidelity data sources:

  • OpenStreetMap (OSM): For real-world intersection topologies, lane counts, and speed limits.
  • Annual Average Daily Traffic (AADT): To calibrate realistic vehicle inflows.
  • CitySim: Drone-based trajectory data used to calibrate the Intelligent Driver Model (IDM) for realistic human driving behavior.
  • EPA MOVES Database: For national-level distributions of vehicle age, fuel type, and engine characteristics.
  • US Geological Survey: For accurate road grade (slope) data.

Key Insights and Ideas

  • Non-linear Benefit Scaling: One of the most striking findings is that just 10% adoption of eco-driving technology captures 25-50% of the total potential emission benefits. This suggests that we don't need a fully autonomous fleet to see massive climate gains.
  • The 70/20 Rule: Approximately 70% of the total city-wide benefits are concentrated in just 20% of intersections.
  • Dynamic Hotspots: The specific intersections that provide the most benefit change significantly as adoption levels rise. An intersection that is a "hotspot" at 10% adoption might not be in the top tier at 100% adoption, complicating long-term infrastructure planning.
  • Implicit Cooperation: Controlled vehicles act as "pace cars," smoothing the flow for the human-driven vehicles behind them. This "Lagrangian control" means a few smart cars can improve the efficiency of the entire fleet.
Figure 1: A schematic illustration of eco-driving at signalized intersections. Controlled vehicles are driven in a way that reduces their carbon emissions. Through car following dynamics and implicitly forming local cooperative teams, controlled vehicle behaviors can affect human-driven vehicle behaviors, leading to a reduction in carbon emissions in human-driven vehicles as well.

Key Results

  • CO2 Reduction: City-wide intersection emissions dropped by 11-22%. The authors estimate this is equivalent to the entire national carbon footprint of Israel or Nigeria.
  • Throughput: Contrary to fears that eco-driving might slow down traffic, intersection throughput actually increased by up to 5% due to smoother flow.
  • Safety: Using surrogate safety measures (like Time-to-Collision), the researchers found that eco-driving behaviors are as safe as, or safer than, conventional human driving.
  • Electrification Synergy: The benefits remain significant even when projected alongside vehicle electrification and hybridization through 2050, as the US power grid is not yet carbon-neutral.

Limitations and Future Work

  • Conservative Estimates: The network decomposition strategy was intentionally conservative to ensure stability; the actual benefits might be even higher if network-wide optimization becomes tractable.
  • Human Factors: The study assumes human drivers will follow SAVs naturally. Real-world human responses to "counterintuitive" SAV behaviors (like gliding slowly toward a red light) require further study.
  • Infrastructure Requirements: Realizing these gains requires widespread deployment of V2I (Vehicle-to-Infrastructure) communication to provide vehicles with Signal Phase and Timing (SPaT) data.
  • Simplified Turns: The current model assumes all turns are protected and focuses on light-duty trucks, leaving room to explore the impact of heavy-duty freight and unprotected left turns.

Summary updated

Abstract

The sheer scale and diversity of transportation make it a formidable sector to decarbonize. Here, we consider an emerging opportunity to reduce carbon emissions: the growing adoption of semi-autonomous vehicles, which can be programmed to mitigate stop-and-go traffic through intelligent speed commands and, thus, reduce emissions. But would such dynamic eco-driving move the needle on climate change? A comprehensive impact analysis has been out of reach due to the vast array of traffic scenarios and the complexity of vehicle emissions. We address this challenge with large-scale scenario modeling efforts and by using multi-task deep reinforcement learning with a carefully designed network decomposition strategy. We perform an in-depth prospective impact assessment of dynamic eco-driving at 6,011 signalized intersections across three major US metropolitan cities, simulating a million traffic scenarios. Overall, we find that vehicle trajectories optimized for emissions can cut city-wide intersection carbon emissions by 11-22%, without harming throughput or safety, and with reasonable assumptions, equivalent to the national emissions of Israel and Nigeria, respectively. We find that 10% eco-driving adoption yields 25%-50% of the total reduction, and nearly 70% of the benefits come from 20% of intersections, suggesting near-term implementation pathways. However, the composition of this high-impact subset of intersections varies considerably across different adoption levels, with minimal overlap, calling for careful strategic planning for eco-driving deployments. Moreover, the impact of eco-driving, when considered jointly with projections of vehicle electrification and hybrid vehicle adoption remains significant. More broadly, this work paves the way for large-scale analysis of traffic externalities, such as time, safety, and air quality, and the potential impact of solution strategies.
BibTeX
@misc{jayawardana2024mitigating,
  author = {Vindula Jayawardana and Baptiste Freydt and Ao Qu and Cameron Hickert and Edgar Sanchez and Catherine Tang and Mark Taylor and Blaine Leonard and Cathy Wu},
  title = {{Mitigating Metropolitan Carbon Emissions with Dynamic Eco-driving at Scale}},
  year = {2024},
  eprint = {2408.05609},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.05609},
}
Figure