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

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

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.

Introduction

As the global energy landscape shifts toward renewables, the intermittency of solar and wind power presents a significant challenge. Vehicle-to-Grid (V2G) technology offers a solution by utilizing Electric Vehicle (EV) fleets as distributed battery storage. However, scheduling thousands of vehicles with varying battery states and user requirements is a computational nightmare.

Aggregators must make rapid decisions—often every few minutes—on energy transactions in wholesale markets. This is typically modeled as a Mixed-Integer Linear Program (MILP). While standard solvers like CPLEX are effective for small fleets, they struggle with "Level 3" complexity, which includes constraints like battery cycle caps to preserve health. For fleets exceeding 200 EVs, traditional solvers can take over 10 minutes to find a solution, exceeding the 15-minute market price update window.

Methodology: The "Learn then Infer" Pipeline

To overcome these computational barriers, researchers from E.ON and IBM Research developed a two-step framework that shifts the heavy computation offline, allowing for near-instantaneous real-time scheduling.

The two-step pipeline: offline learning via ADP and SVM training, followed by real-time inference.
  1. Data Generation (Learn): The system uses Approximate Dynamic Programming (ADP) on historical EPEXSPOT market data (2015–2019) to identify optimal scheduling policies. To simplify the decision space, the fleet's actions are aggregated into three states: Charge (C), Idle (I), and Discharge (D).
  2. Model Training: A Support Vector Classifier (SVC) is trained to predict the optimal aggregate action based on real-time features like energy prices and current battery levels. The study also explored a Quantum SVC (QSVC) to test future scalability.
  3. Inference: In the live environment, the trained model predicts the aggregate action near-instantly. A simple heuristic then distributes this action among individual vehicles (e.g., charging cars with the lowest state-of-charge first).

Results and Discussion

The primary advantage of the ML approach is its scalability and performance under time constraints. While traditional solvers like CPLEX provide optimal answers for small groups, their performance degrades as fleet size increases due to the "complexity wall" of battery health constraints.

Objective value comparison: For fleets larger than 230 vehicles, the SVC outperforms CPLEX when the latter is limited to a 10-minute runtime.

Key findings include:

  • Economic Performance: For fleets larger than 230 vehicles, the SVC actually outperformed CPLEX (when CPLEX was limited to a 10-minute runtime). CPLEX "timed out" before finding a high-quality solution, while the ML model provided a solid approximation instantly.
  • Computational Speed: The SVC provides scheduling actions in milliseconds, and its inference time remains nearly "flat" regardless of fleet size, unlike the exponential growth seen in traditional solvers.
  • Predictive Accuracy: The classical SVC achieved 92% accuracy in predicting optimal aggregate actions. The QSVC achieved 87% accuracy, demonstrating potential for quantum-enhanced scheduling as hardware matures.

Conclusion

This machine learning approach enables high-frequency V2G scheduling that is both scalable and economically efficient. By reducing complex individual decisions to a 3-class classification problem, the model maintains high performance without the computational overhead of MILP solvers. Future work will focus on more sophisticated heuristics for individual vehicle selection and accounting for stochastic connectivity to make the system field-ready for real-world grid balancing.

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 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},
}
Figure