A Machine Learning Approach to Boost the Vehicle-2-Grid Scheduling

Gabriele Agliardi, Giorgio Cortiana, Anton Dekusar, Kumar Jang Bahadur Ghosh, Naeimeh Mohseni, Corey O'Meara, Víctor Valls, Kavitha Yogaraj, Sergiy Zhuk

As the world shifts toward renewable energy, the "intermittency problem"—the fact that the sun doesn't always shine and the wind doesn't always blow—has become a major hurdle. One promising solution is using Electric Vehicles (EVs) as a giant, distributed battery, a concept known as Vehicle-to-Grid (V2G). However, managing thousands of cars with different battery levels and owner needs is a computational nightmare.

A new paper from researchers at E.ON and IBM Research proposes a clever workaround: using machine learning to "learn" optimal scheduling policies, allowing for near-instantaneous grid balancing that traditional solvers simply can't match.

TL;DR

The researchers developed a two-step approach that uses Approximate Dynamic Programming (ADP) to generate optimal EV charging schedules offline, then trains a Machine Learning model to predict those schedules in real-time. This method matches the performance of industry-standard solvers for large fleets but operates significantly faster, making it feasible for the high-frequency demands of wholesale energy markets.

What Problem Does This Paper Address?

To make V2G work, an aggregator (like a utility company) needs to decide every few minutes how much energy to buy from or sell to the grid using a fleet of EVs. This is usually modeled as a Mixed-Integer Linear Program (MILP).

While standard solvers like CPLEX are great at finding the "perfect" answer for small groups, they struggle as complexity increases. The authors identify "Level 3" complexity—where you must limit the number of charging cycles to protect battery health—as a breaking point. For a fleet of 200+ EVs, a traditional solver might take over 10 minutes to find a solution. In a market where prices change every 15 minutes, that’s simply too slow to be useful.

Methods

The authors propose a "Learn then Infer" pipeline:

  1. Data Generation (The Teacher): They use Approximate Dynamic Programming (ADP) on historical data to find "good" scheduling policies. To keep things manageable, they simplify the fleet's possible actions into three aggregate states: Charge (C), Idle (I), and Discharge (D).
  2. The Learning Model (The Student): They trained a Support Vector Classifier (SVC)—and even experimented with a Quantum Support Vector Classifier (QSVC)—to look at the current state (time, energy prices, battery levels) and predict which of the three aggregate actions (C, I, or D) is best.
  3. Heuristic Execution: Once the ML model picks an aggregate action (e.g., "Charge"), a simple heuristic determines which specific cars to charge (e.g., those with the lowest battery levels).

Datasets

The study utilized:

  • EPEXSPOT Market Data: Real-world energy price traces from 2015–2019 were used to simulate the wholesale market environment.
  • Simulated EV Fleets: The authors generated scenarios for fleets ranging from 50 to 300 EVs, incorporating realistic constraints like 100 kWh batteries and ±10 kW charging rates.

Key Insights and Ideas

  • The Complexity Wall: The most interesting finding is how specific constraints affect computation. While varying battery sizes (Level 1) or group charging limits (Level 2) are manageable, adding precedence constraints (Level 3, like battery cycle caps) makes the search space explode for traditional solvers.
  • Aggregate Actions are Enough: By reducing the complex decision of "what should every individual car do?" to a 3-class classification problem (Charge/Idle/Discharge), the researchers made the problem solvable by ML without a significant loss in economic performance.
  • Scalability is Flat: Unlike traditional solvers, where runtime grows exponentially or linearly with fleet size, the ML model's inference time is nearly "flat." It takes roughly the same amount of time to decide for 300 cars as it does for 50.

Key Results

  • Speed: The SVC model provides answers in a fraction of a second, whereas the ADP "teacher" takes hundreds of seconds as the fleet grows.
  • Economic Performance: For fleets larger than 230 vehicles, the SVC actually outperformed CPLEX (when CPLEX was limited to a 10-minute runtime). This is because CPLEX "timed out" before finding a high-quality solution, while the ML model provided a solid approximation instantly.
  • Accuracy: The classical SVC achieved a 92% classification accuracy in predicting the optimal aggregate action.
  • Quantum Potential: The QSVC (Quantum SVC) was able to scale similarly to the classical version, though its accuracy was slightly lower (87%), likely due to the limitations of current quantum simulators.

Limitations and Future Work

The current model uses a relatively simple heuristic to pick which specific cars to charge once the aggregate action is chosen. The authors suggest that more sophisticated "ranking" models could improve performance further.

Additionally, the study assumes EVs are always plugged in between 6 PM and 6 AM. In the real world, cars arrive and leave at random times. Future iterations will need to account for this "stochastic" connectivity to be truly field-ready. Finally, while the Quantum SVC showed promise in scalability, its lower accuracy suggests that quantum advantage in this specific scheduling task is still a horizon goal.

Summary updated

Abstract

Electric Vehicles (EVs) are emerging as battery energy storage systems (BESSs) of increasing importance for different power grid services. However, the unique characteristics of EVs makes them more difficult to operate than dedicated BESSs. In this work, we apply a data-driven learning approach to leverage EVs as a BESS to provide capacity-related services to the grid. The approach uses machine learning to predict how to charge and discharge EVs while satisfying their operational constraints. As a paradigm application, we use flexible energy commercialization in the wholesale markets, but the approach can be applied to a broader range of capacity-related grid services. We evaluate the proposed approach numerically and show that when the number of EVs is large, we can obtain comparable objective values to CPLEX and approximate dynamic programming, but with shorter run times. These reduced run times are important because they allow us to (re)optimize frequently to adapt to the time-varying system conditions.
BibTeX
@inproceedings{agliardi2024machine,
  author = {Gabriele Agliardi and Giorgio Cortiana and Anton Dekusar and Kumar Jang Bahadur Ghosh and Naeimeh Mohseni and Corey O'Meara and Víctor Valls and Kavitha Yogaraj and Sergiy Zhuk},
  title = {{A Machine Learning Approach to Boost the Vehicle-2-Grid Scheduling}},
  booktitle = {iSPEC},
  year = {2024},
  doi = {10.1109/ispec59716.2024.10892547},
}

Loading related content...

Figure