Carbon-Aware EV Charging

Kai-Wen Cheng, Yuexin Bian, Yuanyuan Shi, Yize Chen

Introduction

As electric vehicles (EVs) become a cornerstone of global decarbonization efforts, a critical realization has emerged: an EV is only as clean as the electricity used to charge it. In many regions, the carbon intensity of the power grid fluctuates significantly throughout the day based on the mix of renewable and fossil fuel generation. This paper proposes a "carbon-aware" charging framework designed to shift EV charging loads to periods of low carbon intensity, effectively aligning vehicle energy demand with the availability of clean energy.

The Problem: The Carbon Intensity Mismatch

In regions like California, the grid's carbon intensity ($C(t)$) is highly dynamic. During the "solar peak" at midday, the grid is relatively clean due to an abundance of photovoltaic generation. Conversely, in the evening and early morning, the grid relies more heavily on natural gas to meet demand, leading to higher emissions per kilowatt-hour.

Existing smart charging strategies typically prioritize cost-optimization (leveraging Time-of-Use rates) or grid-stability (preventing local transformer overloads). However, these objectives do not always align with carbon minimization. The authors identify a significant "mismatch" where EV charging often peaks during the dirtiest hours of the grid cycle, highlighting the need for an explicit carbon-minimization objective in charging infrastructure management.

Methodology

The researchers developed a constrained optimization framework that treats EV charging as a flexible load. The core components of the system include:

1. Carbon Intensity Modeling and Forecasting

Using energy mix data from the California Independent System Operator (CAISO), the researchers calculated real-time carbon intensity. To facilitate real-time (online) decision-making, they implemented a forecasting module using a linear regression model. This model utilizes features such as time of day, month, and moving averages of historical intensity to predict the next day's carbon profile.

2. Optimization Algorithm

The objective function minimizes the product of carbon intensity and charging power over a given time horizon. A critical innovation in this model is the introduction of a balanced factor ($\lambda$). This parameter allows the system to navigate the trade-off between two competing goals:

  • Carbon Minimization: Reducing the total emissions associated with the charging session.
  • Energy Delivery: Ensuring the vehicle receives the total energy requested before its departure.

3. Operational Constraints

The model respects physical and operational limits, including battery State-of-Charge (SoC) requirements, individual vehicle charging rate limits, and the total power capacity of the charging station's transformer to prevent infrastructure damage.

Data and Experimental Setup

The study validated the framework using two primary real-world datasets from 2021:

  • EV Charging Data: Granular session data (arrival/departure times and energy requests) from the Lawrence Berkeley National Laboratory (LBNL) campus.
  • Grid Data: Electricity generation fuel mix and load forecasts provided by CAISO.

Analysis of the LBNL data revealed significant flexibility: most vehicles arrived before noon and stayed for several hours. While they arrived during the "clean" part of the day, they often remained plugged in long after their required energy could have been delivered, providing a "slack" period that the algorithm could exploit to find the absolute cleanest charging windows.

Key Insights and Results

The Carbon "Duck Curve"

Similar to the well-known California "duck curve" for net load, the researchers identified a seasonal carbon curve. Emissions are lowest in the spring (due to high solar and hydroelectric output) and highest in the winter. The algorithm naturally adapts to these seasonal shifts to maximize savings.

Performance Metrics

  • Zero-Sacrifice Gains: Using a conservative balanced factor, the algorithm achieved a 3.81% reduction in carbon emissions with 0% loss in energy delivery compared to standard "Earliest Deadline First" (EDF) scheduling.
  • Aggressive Decarbonization: By prioritizing carbon reduction over full energy delivery, the system achieved up to a 26% reduction in emissions, though this necessitated a 12.61% reduction in total energy delivered.
  • Online Efficiency: The real-time "online" algorithm performed remarkably close to the "offline" version (which assumes perfect foresight), proving that simple linear forecasting is sufficient for practical, real-world deployment.

Economic and Social Impact

The authors estimate that if this carbon-aware scheduling were applied across California, it could reduce the "Social Cost of Carbon" by approximately $5.47 million per year.

Limitations and Future Directions

While the results are promising, the study notes several areas for future research. The current model focuses on a single charging station; scaling this to thousands of stations simultaneously could create "rebound" peaks in demand that might stress the grid. Additionally, the model assumes users provide accurate departure times, which may not always happen in practice. Future work will explore adaptive methods for choosing the balanced factor ($\lambda$) and incorporating battery health considerations to ensure that frequent power ramping does not negatively impact vehicle longevity.

Conclusion

Carbon-aware EV charging represents a significant opportunity to enhance the environmental benefits of electric mobility. By leveraging the inherent flexibility in EV charging sessions and real-time grid data, charging providers can significantly reduce the carbon footprint of transportation without requiring major changes to user behavior or infrastructure.

Summary updated

BibTeX
@inproceedings{cheng2022carbon,
  author = {Kai-Wen Cheng and Yuexin Bian and Yuanyuan Shi and Yize Chen},
  title = {{Carbon-Aware EV Charging}},
  booktitle = {SmartGridComm},
  year = {2022},
  url = {https://openreview.net/forum?id=TlZgpEhcvIy},
}
Figure